feat(gateway): add pull-mode relay outbound with outbox list and ack (#2742) (#2844)

Adds an optional pull mode to the local gateway relay, for clients that cannot run a push listener (#2742, by @w4rc0n). A relay profile with OPENRIG_RELAY_MODE=pull is ready with an inbound token only. GET /api/gateway/relay/outbox lists complete frozen envelopes the transport staged for that connector (authenticated, direct loopback, limit 1-200, a read changes nothing); POST /api/gateway/relay/outbox/ack accepts any listed deliveryId and records the same acceptance receipts, guards and drain as a push 2xx, completing a partial receipt set on retry. Duplicate and unknown acks are answered from the delivery seen stores. A pull notice stays quiet only while the outbox would list it, and reads never-posted at 120 s otherwise or after 24 hours if never acknowledged. Push replay completes a partial receipt set only from this delivery's own receipts. Relay profiles load once per delivery-outcome batch. No broker, leases, read receipts or polling requirement.

.claude-plugin/marketplace.json

@@ -0,0 +1,15 @@
+{
+  "name": "openrig-app",
+  "description": "The OpenRig plugin: work with your OpenRig team from this app.",
+  "owner": {
+    "name": "Mike Schwarz",
+    "url": "https://github.com/mvschwarz"
+  },
+  "plugins": [
+    {
+      "name": "openrig",
+      "source": "./packages/daemon/assets/plugins/openrig",
+      "description": "Work with your OpenRig team from your coding agent: hand it work, follow that work, answer its questions and read what it returns."
+    }
+  ]
+}

.cursor-plugin/marketplace.json

@@ -0,0 +1,15 @@
+{
+  "name": "openrig-app",
+  "description": "The OpenRig plugin: work with your OpenRig team from this app.",
+  "owner": {
+    "name": "Mike Schwarz",
+    "url": "https://github.com/mvschwarz"
+  },
+  "plugins": [
+    {
+      "name": "openrig",
+      "source": "./packages/daemon/assets/plugins/openrig",
+      "description": "Work with your OpenRig team from your coding agent: hand it work, follow that work, answer its questions and read what it returns."
+    }
+  ]
+}

.github/workflows/tests.yml

@@ -79,7 +79,7 @@ jobs:
 
   linux-seat-environment:
     if: ${{ github.event_name != 'push' }}
-    # This integration test reads /proc/<pid>/environ and is skipped on macOS.
+    # Cover /proc seat environments and Linux canonical-input limits that macOS cannot exercise.
     runs-on: ubuntu-latest
     timeout-minutes: 20
     steps:
@@ -108,7 +108,7 @@ jobs:
           ip link set lo up
           exec runuser -u "$1" -- env -i HOME="$2/home" PATH="$3" TMPDIR="$2" TMUX_TMPDIR="$2" \
             CI=true TERM=xterm-256color \
-            npm run test -w packages/daemon -- test/seat-env-parent-session.integration.test.ts --maxWorkers=1 --reporter=verbose
+            npm run test -w packages/daemon -- test/seat-env-parent-session.integration.test.ts test/pi-shell-launch-native.test.ts --maxWorkers=1 --reporter=verbose
           SUITE
 
   package-tests:
@@ -116,7 +116,7 @@ jobs:
     # macOS, the platform OpenRig runs on: several daemon and cli tests read macOS ps columns,
     # file-inode behaviour and release tags.
     runs-on: macos-latest
-    timeout-minutes: 45
+    timeout-minutes: 60
     strategy:
       fail-fast: false
       matrix:

CHANGELOG.md

@@ -8,14 +8,733 @@ deprecations, and behavioral changes. Breaking changes are called out explicitly
 
 ---
 
+## [0.7.10]
+
+Published 2026-11-08 as `@openrig/[email protected]`, built from the commit tagged
+`v0.7.10` (`a226b1b6`).
+
+OpenRig 0.7.10 is a patch for people running their own agent teams. After a
+managed compaction, a step held because OpenRig can't read the seat's input box
+now goes in after two idle polls instead of waiting with no end. Wakes and sends
+get through a seat pane left in more than one tmux mode, and `rig exec` hands a
+queue row's task to Claude Code or Codex as one finite run that outlives the
+command (not yet run with a real adapter). Installing through the rigs skill tries a user-writable npm prefix
+before a project-local install, and agent spec advisories now show at launch.
+There are fixes across refocus, routing, status and speed under load, many
+contributed by people using OpenRig. The OpenShell worker now does real work and
+exports results, but its cleanup still fails against OpenShell 0.1.6, so don't
+rely on cleanup or recreate yet: see [the OpenShell worker](#the-openshell-worker-in-0710).
+
+### Before you upgrade
+
+- **Restart the daemon after installing:**
+  ```bash
+  npm install -g @openrig/cli@0.7.10
+  rig daemon stop && rig daemon start
+  ```
+  This release adds no migrations.
+- **The bundled openrig-core plugin advances from 0.1.25 to 0.1.26.** Its Claude
+  hooks add a `PermissionRequest` hook that reports permission requests to seat
+  status, and its refocus hook and the refocusing skill's trace script changed.
+  Managed refresh preserves replaced files under
+  `$OPENRIG_HOME/backups/managed-refresh/<id>/`; reapply local edits you still
+  need from that copy. ([#2615](https://github.com/mvschwarz/openrig/pull/2615), [#2562](https://github.com/mvschwarz/openrig/pull/2562), [#2582](https://github.com/mvschwarz/openrig/pull/2582), [#2587](https://github.com/mvschwarz/openrig/pull/2587))
+- **Check your agent specs:** `rig agent validate <path>` now prints each
+  advisory. A profile's `uses.<category>` written as a single value now selects
+  that resource instead of being dropped, so a member using that profile fails to
+  resolve if the resource pool doesn't hold it (see Behaviour that changed).
+  ([#2585](https://github.com/mvschwarz/openrig/pull/2585), [#2606](https://github.com/mvschwarz/openrig/pull/2606))
+- **If you run an OpenShell environment, rebuild its guest image** with this
+  release's `guest.py` (`daemon/assets/openshell/README.md` describes the build).
+  The CA repair below is in that file, which the image carries. ([#2602](https://github.com/mvschwarz/openrig/pull/2602))
+
+### Finite runs: `rig exec`
+
+- **`rig exec` runs one turn of a queue row's task on Claude Code or Codex**
+  through its maintained ACP adapter, in a connector that outlives the command.
+  `rig exec harnesses` shows the versions installed,
+  `rig exec run <row> --harness claude` (or `codex`) starts it (`--plan` starts
+  nothing, `--wait` waits), and `rig exec status` and `rig exec cancel` follow
+  it. ([#2597](https://github.com/mvschwarz/openrig/pull/2597))
+- **One run per row, and the row stays open.** Running it again returns the
+  row's run instead of sending the task twice; `--again` starts another turn once
+  it has ended. A run whose connector died without recording an outcome reads
+  `interrupted`, outcome unknown. The outcome is the agent's report, and
+  accepting the work stays with you. `--wait` exits 0 for `completed`, 2 for any
+  other outcome and 3 when you stop waiting; a Ctrl-C there stops the wait, not
+  the run. ([#2597](https://github.com/mvschwarz/openrig/pull/2597), [#2601](https://github.com/mvschwarz/openrig/pull/2601))
+- **An ACP run's outcome note,** `rig acp run`'s included, carries the usage and
+  cost the agent reported, each with its source, or unknown. ([#2597](https://github.com/mvschwarz/openrig/pull/2597))
+- **`docs/reference/acp-agents.md`** covers both adapters
+  (`@agentclientprotocol/claude-agent-acp` 0.101.0 and
+  `@agentclientprotocol/codex-acp` 2.5.0): installing them, which native
+  versions run, and how their permissions, refusals and cancels differ. ([#2596](https://github.com/mvschwarz/openrig/pull/2596))
+
+### The OpenShell worker in 0.7.10
+
+**The OpenShell worker hasn't run end to end on 0.7.10. Don't rely on this path
+yet.**
+
+0.7.10 changes the two things our release testing on 0.7.9 found. Tested against
+OpenShell 0.1.6, the first fix holds and the second doesn't:
+- **The guest's CA settings reach its Claude.** The guest launcher built
+  Claude's environment from scratch with only the provider settings, so a CA the
+  OpenShell exec supplied never arrived, and API calls failed with
+  `SELF_SIGNED_CERT_IN_CHAIN`. It now also passes `NODE_EXTRA_CA_CERTS`,
+  `SSL_CERT_FILE`, `REQUESTS_CA_BUNDLE` and `CURL_CA_BUNDLE`, each only when the
+  exec supplies it. Nothing else is copied, and nothing turns TLS checks off.
+  With OpenShell 0.1.6 the guest now authenticates, does useful work, makes a
+  verified export and completes its task.
+  Rebuild the guest image to get it (above). ([#2602](https://github.com/mvschwarz/openrig/pull/2602))
+- **Deleting an owned environment reads Podman's inventory.** The cleanup step
+  asked for routes Podman serves only under a version prefix, misread the "Not
+  Found" reply, and never sent the delete. It now uses routes Podman serves
+  without one, and a volume counts as gone only when one complete volume listing
+  leaves it out. Any other answer stops as `environment_cleanup_unconfirmed`,
+  with the HTTP status. The maintained cleanup still fails against OpenShell
+  0.1.6, which doesn't mount a workspace volume (see Known limits). ([#2603](https://github.com/mvschwarz/openrig/pull/2603))
+
+An exact restore hasn't been shown: recreate and restore can't run while the
+cleanup fails.
+
+### Compaction, delivery and terminals
+
+- **A held post-compaction step now goes in.** Each step after a managed
+  compaction (the turn boundary, the restore request, the audit) went only into
+  an input box that read empty, so a box OpenRig couldn't read held the restore
+  request on every poll, and the seat sat idle without its restore. Now, after
+  two idle polls in a row that can't read the box, the next poll sends the step,
+  and `rig compact <seat> --state` and `rig seat status` say it went in over a
+  box OpenRig couldn't read, never that the box was empty. Some cases still hold
+  (see Known limits). ([#2576](https://github.com/mvschwarz/openrig/pull/2576), refs [#2575](https://github.com/mvschwarz/openrig/issues/2575))
+- **Wakes and sends leave every tmux mode.** A pane scrolled into copy mode, with
+  run-shell output, a chooser, customize or clock mode on top, kept a mode after
+  one cancel, and the send was refused until someone left it by hand. Delivery
+  now runs tmux's `copy-mode -q`, which leaves them all, with a single cancel
+  where that fails (tmux before 3.1a), and a refusal names why leaving failed.
+  ([#2594](https://github.com/mvschwarz/openrig/pull/2594))
+- **A scrolled pane with a stale record no longer blocks delivery.** tmux
+  reuses pane IDs after a server restart, so a stale binding, from another rig
+  for example, can name the same pane as a live seat. Leaving the mode was
+  addressed by that raw pane ID, which then matched two seats, so it was
+  refused before anything was written and the send failed with nothing sent.
+  Leaving the mode now goes through the seat's address, as the text and Enter
+  writes already did. The stale binding itself is unchanged.
+  ([#2644](https://github.com/mvschwarz/openrig/pull/2644))
+- **A writable rig, pod or seat view sizes the pane it shows,** instead of
+  keeping an older attached client's size. Read-only views are unchanged. ([#2614](https://github.com/mvschwarz/openrig/pull/2614))
+- **Herdr views kept across an upgrade keep working,** on Linux Herdr 0.9.5 /
+  protocol 22. A viewer or follow view kept from 0.7.5 through 0.7.9 now counts
+  as the same attachment when only OpenRig's tile script changed, and moves to
+  the current command, so the viewer stays confirmed and the follow view goes on
+  adding and closing tiles. A kept viewer that's already running keeps its
+  earlier sizing for as long as it runs, its own re-attaches included; writable
+  viewers opened after the upgrade size their pane. A different session, tmux,
+  server or read-only setting still holds the view, as before. Checked against
+  the scripts those releases used, not yet with a live Herdr. It follows
+  @natefaerber's approach in
+  [#2467](https://github.com/mvschwarz/openrig/pull/2467). ([#2634](https://github.com/mvschwarz/openrig/pull/2634))
+
+### Coordination and status
+
+- **Routing and receipts:**
+  - a send to `<seat>@<rig>@<this host's ID>` goes to this host even when a
+    remote host is selected;
+  - `rig broadcast` prints `UNCONFIRMED`, not `FAILED`, for a recipient its text
+    or Enter may have reached, and points to `rig capture` before any resend;
+  - `rig seat stop`, `clean` and in-place restart, when they refuse a seat
+    handed over before 0.7.9, say why and no longer suggest `rig unclaim`;
+  - `rig roster show` and `find` read a seat's activity as unknown
+    ("Conflicting activity observations") when two routes to it disagree;
+  - `rig seat set-runtime` finds a pod seat's authored recap.
+  ([#2600](https://github.com/mvschwarz/openrig/pull/2600), [#2572](https://github.com/mvschwarz/openrig/pull/2572), [#2574](https://github.com/mvschwarz/openrig/pull/2574), [#2593](https://github.com/mvschwarz/openrig/pull/2593))
+- **Refocus:** unresolved typed work, and a `rig queue whoami` that timed out,
+  print a named trace gap with project-level orientation, labelled as not
+  evidence of a current mission, instead of counting as a refocus failure.
+  ([#2582](https://github.com/mvschwarz/openrig/pull/2582), [#2587](https://github.com/mvschwarz/openrig/pull/2587))
+- **Updates to people:** the stuck sweep no longer files an informational update
+  to a registered person as an unclaimed obligation, unless its quiet-hours hold
+  was orphaned. ([#2586](https://github.com/mvschwarz/openrig/pull/2586))
+- **Slack:** `rig slack status` shows relay attempts to another host still
+  awaiting a response. It's a count for the current connector at that moment, so
+  zero doesn't prove nothing is in flight. ([#2569](https://github.com/mvschwarz/openrig/pull/2569))
+- **Permission requests in seat status:** `rig seat status` shows a Claude
+  seat's latest permission request as openrig-core's hook observed it, with
+  Claude's own suggested rules labelled as suggestions. Nothing is applied,
+  approved or stored. ([#2562](https://github.com/mvschwarz/openrig/pull/2562))
+
+### Install, specs and errors
+
+- **The rigs skill's install step,** when npm's global prefix isn't writable,
+  now has the agent point npm's prefix at a folder the person owns whose `bin` is
+  on their login PATH, or use a version manager's Node, saying first what that
+  changes, and install inside the project only if neither works. ([#2611](https://github.com/mvschwarz/openrig/pull/2611))
+- **Plain `rig` works in managed Codex seats however OpenRig was installed,**
+  with a Codex that supports `allow_login_shell` (checked in Codex's source at
+  0.162.1 and 0.166); not yet run with a live Codex seat. OpenRig now passes Codex `-c allow_login_shell=false`, so tool commands run in
+  a non-login shell and inherit the environment their pane's shell set up,
+  OpenRig's command folder included. To defer to Codex's own setting instead,
+  run `rig config set launch.codex_login_shell true`; it applies at each seat's
+  next launch, and a restart in place keeps the running choice. ([#2613](https://github.com/mvschwarz/openrig/pull/2613))
+- **Agent spec advisories show where you'll see them:** `rig up`, a fresh restore
+  and a first-start retry report each as a launch warning naming the member,
+  spec and path, and `rig agent validate` prints them whether or not the spec is
+  valid. They never stop a launch. ([#2606](https://github.com/mvschwarz/openrig/pull/2606))
+- **An unknown key under a profile's `uses`,** such as a typo,
+  gets an advisory instead of being ignored silently. ([#2618](https://github.com/mvschwarz/openrig/pull/2618))
+- **Clearer errors:**
+  - a rejected terminal-token request says "Terminal auth failed" and names
+    `OPENRIG_TERMINAL_BEARER_TOKEN`, or `OPENRIG_AUTH_BEARER_TOKEN` when
+    terminal auth falls back to it;
+  - the TUI shows the daemon's own message when it refuses a launch;
+  - `rig doctor` gives up on its cmux status and `--spec` topology reads after 2
+    seconds each, so a listener that never answers can't hold it.
+  ([#2577](https://github.com/mvschwarz/openrig/pull/2577), [#2556](https://github.com/mvschwarz/openrig/pull/2556), [#2564](https://github.com/mvschwarz/openrig/pull/2564))
+- **Docs:** a Codex config fragment can't install a named Codex profile, with
+  how to keep a profile with full access or a `codex_permissions` envelope; and
+  a Claude seat that selects openrig-core loads its skills with the plugin.
+  ([#2604](https://github.com/mvschwarz/openrig/pull/2604), [#2584](https://github.com/mvschwarz/openrig/pull/2584))
+
+### Speed under load
+
+- **The execution view** (`GET /api/views/execution`, behind the TUI's execution
+  view and `rig view show execution`) reads the review registry once per request
+  instead of once per slice, and yields to the event loop while it does. Its
+  other synchronous steps are unchanged. ([#2578](https://github.com/mvschwarz/openrig/pull/2578))
+- **A queue listing** looks recovery rows up through one index per listing
+  instead of a scan for each listed row. It hasn't been measured on a live
+  daemon. ([#2588](https://github.com/mvschwarz/openrig/pull/2588))
+- **The event notification drain** reads pages of at most 500 rows instead of
+  the whole remaining tail. It bounds rows per read, not bytes, and still runs to
+  the end synchronously; no latency change on a live install is claimed.
+  ([#2610](https://github.com/mvschwarz/openrig/pull/2610))
+
+### For runner authors and integrators
+
+- **The runner contract says what has run with a real runner.**
+  `docs/reference/harness-runner-contract.md` and the dsh example say what has
+  run natively, at what scope, and what hasn't; any other runner, provider or
+  path stays unqualified until shown. The 0.7.8 Known limits line below that said
+  no harness had run end to end is struck and corrected in place.
+  ([#2592](https://github.com/mvschwarz/openrig/pull/2592), [#2595](https://github.com/mvschwarz/openrig/pull/2595))
+- **Telemetry:** `docs/reference/telemetry.md` says OpenRig's guarantee ends at
+  its finite reads, versioned by `schemaVersion` 1, and links the separately
+  maintained openrig-telemetry exporter, which OpenRig doesn't test or
+  guarantee. ([#2589](https://github.com/mvschwarz/openrig/pull/2589))
+- **Also:** the real-tmux transport test waits for the shell's prompt before
+  typing. ([#2559](https://github.com/mvschwarz/openrig/pull/2559))
+
+### Behaviour that changed
+
+- **Managed Codex seats run tool commands in a non-login shell** (above).
+  Inherited setup isn't re-run for each command, so setup that lives only in
+  login or rc files (such as `~/.bashrc` lines your login file doesn't source, or
+  a tmux `default-command` that doesn't start a login shell), shell functions and
+  aliases may be missing from tool commands. A model request for `login: true`
+  gets Codex's own refusal, "login shell is disabled by config". With
+  `launch.codex_login_shell` on, OpenRig omits its override rather than forcing
+  login shells, so a login file that sets `PATH` outright can then hide `rig`.
+  ([#2613](https://github.com/mvschwarz/openrig/pull/2613))
+- **A held post-compaction step goes in after two idle polls that can't read the
+  input box** (above). A draft OpenRig can't recognise in that box may receive
+  the restore text. ([#2576](https://github.com/mvschwarz/openrig/pull/2576))
+- **A send or wake leaves every tmux mode in a seat's pane,** including a
+  chooser, customize or clock mode a person has open, where before it was
+  refused. ([#2594](https://github.com/mvschwarz/openrig/pull/2594))
+- **A send to `<seat>@<rig>@<this host's ID>`** goes to this host, not the
+  selected remote host. ([#2600](https://github.com/mvschwarz/openrig/pull/2600))
+- **A profile's `uses.<category>` written as a single value** (for example
+  `uses.skills: review`) used to validate and then be dropped. It now counts as a
+  one-item list, with an advisory suggesting list syntax: a member using that
+  profile gets the resource, or fails to resolve with "not found in resource
+  pool" when the pool doesn't hold it. Other malformed `uses` shapes are
+  advisories, not new errors. ([#2585](https://github.com/mvschwarz/openrig/pull/2585))
+- **`rig up` can print new warnings,** one per agent spec advisory, and
+  `rig agent validate` prints them. ([#2606](https://github.com/mvschwarz/openrig/pull/2606))
+- **`rig broadcast`** prints `UNCONFIRMED` for a recipient it may have reached;
+  its exit code and `N/M delivered` count are unchanged. ([#2572](https://github.com/mvschwarz/openrig/pull/2572))
+- **`rig roster show` and `find`** read conflicting activity as unknown.
+  ([#2574](https://github.com/mvschwarz/openrig/pull/2574))
+- **The refocus hook** no longer counts unresolved typed work or a timed-out
+  `rig queue whoami` as a failure, and a timed-out refocus clears the pending
+  post-compaction retry. ([#2582](https://github.com/mvschwarz/openrig/pull/2582), [#2587](https://github.com/mvschwarz/openrig/pull/2587))
+- **The stuck sweep** no longer flags informational updates to registered people
+  as unclaimed obligations. ([#2586](https://github.com/mvschwarz/openrig/pull/2586))
+- **A rejected terminal-token request** gets "Terminal auth failed" instead of
+  "Mission Control auth failed"; its status code and fields are unchanged.
+  ([#2577](https://github.com/mvschwarz/openrig/pull/2577))
+
+### Known limits
+
+- **OpenShell:** maintained cleanup (`delete-owned`) fails against OpenShell
+  0.1.6, which doesn't mount a workspace volume, so recreate and restore can't
+  run (above). A cleanup error that's thrown still omits its stage and whether
+  its delete was sent, and the inventory is a snapshot, not an atomic read.
+  0.7.8's requirements, cost note
+  and cleanup limits still apply, including "Never grant the guest access to the
+  daemon's main API port (normally 7433)." A same-generation reconciled launch
+  opens no viewer, so run `rig environment connect` yourself; non-text startup
+  actions aren't supported on environment seats. ([#2603](https://github.com/mvschwarz/openrig/pull/2603))
+- **`rig exec`:** no real adapter has run under it yet; one run per row,
+  detached ownership and reconciling were checked against a stand-in agent.
+  Concurrent runs, long or background tasks, recovery across a reboot, MCP
+  servers and Windows aren't covered. A cancel is acknowledged, not settled, and
+  says nothing about further spend. Cancel an exec run with `rig exec cancel`:
+  `rig acp cancel` given its ID while it's still requested can signal the wrong
+  process. A Ctrl-C just before the connector is launched isn't caught by
+  `--wait`.
+  ([#2597](https://github.com/mvschwarz/openrig/pull/2597), [#2601](https://github.com/mvschwarz/openrig/pull/2601))
+- **After a managed compaction, a step can still hold:** when the input box
+  reads as typed, with or without a suggestion; on a seat whose draft-aware
+  delivery or typing guard holds OpenRig's input; and after a `/compact` whose
+  result was lost. To take over by hand, run `rig compact <seat> --cancel`, then
+  send the seat a restore request naming its newest restore map, its restore
+  marker and its packet; a step already being pasted can still land. Which
+  reading the reported seat's box gave isn't established, so the report stays
+  open. ([#2575](https://github.com/mvschwarz/openrig/issues/2575))
+- **Draft-aware delivery to Claude:** on 0.7.8 with Claude Code 2.1.319,
+  deliveries to idle Claude agents, queue wakes included, were reported held at
+  an empty prompt when a narrow pane showed
+  `✔ Update installed · Restart to update` on its own line. 0.7.10 still doesn't
+  recognize that bare banner, though it recognizes `/rc` and the full versioned
+  banner. A message the draft-aware delivery path holds isn't replayed
+  automatically by that path. ([#2605](https://github.com/mvschwarz/openrig/issues/2605))
+- **Grant-service assistants aren't woken:** "an item stays queued until its
+  receiver claims it."
+- **Open updates for a person** read a 1,001-row window, so a page can come back
+  short, marked `truncated`.
+- **Slack:** a private channel needs the documented route, which we don't test,
+  and direct messages aren't supported. The channel and people reads, and a
+  click on a re-parked ask, haven't been run against a live Slack workspace.
+  Redaction of the bot token from a transport error can still leave a malformed
+  token fragment shorter than six characters, so review an error before sharing
+  it. ([#2503](https://github.com/mvschwarz/openrig/pull/2503))
+- **SSH host status** reads `unknown` behind a ProxyJump or ProxyCommand; that
+  isn't proof the connection is broken. ([#2488](https://github.com/mvschwarz/openrig/pull/2488))
+- **Shared asks:** "a click's option written before its answer row … can stand
+  beside or inside the closer's answer without the first clicker's authorship."
+- **Existing seats** last launched fresh before 0.7.9 pick up Codex launch
+  instructions and the single skill load at their next fresh launch. A
+  `CLAUDE.md` or `AGENTS.md` that already grew isn't trimmed.
+- **Permit records** last as long as stream retention keeps them. A permit a
+  runner makes on its own, when it can't reach OpenRig, is in its pane only. For
+  an ask opened before 0.7.9, a record may omit the option's kind. ([#2361](https://github.com/mvschwarz/openrig/pull/2361))
+- **Permission requests in seat status** show only the latest one, held in
+  memory for five minutes and lost on a daemon restart. Claude's suggested rules
+  can contain sensitive values, and `rig seat status` and both seat-status routes
+  can return them. ([#2562](https://github.com/mvschwarz/openrig/pull/2562))
+- **Agent spec advisories** don't show for a native resume that continues a
+  saved conversation, since it doesn't reload the spec. ([#2606](https://github.com/mvschwarz/openrig/pull/2606))
+- **Seats handed over before 0.7.9** keep their adopted origin until a later
+  launched handover or managed launch records a new session, so until then
+  `rig seat stop`, `restart` and `clean` refuse them; the refusal now says so.
+  ([#2440](https://github.com/mvschwarz/openrig/pull/2440), [#2572](https://github.com/mvschwarz/openrig/pull/2572))
+- **Codex hooks after an upgrade:** Codex sessions kept running across the
+  upgrade to 0.7.9 showed "Hook failed — plugin contents changed; hook approval
+  required" until relaunched; in that upgrade test, one seat restarted with
+  OpenRig's managed restart showed no review afterwards. 0.7.10 changes
+  openrig-core's files again.
+- **The runner contract** has run natively only at the scope its status section
+  states (`docs/reference/harness-runner-contract.md`); any other runner,
+  provider or path is unqualified until shown.
+- **Fixed or changed from 0.7.9's limits:** the OpenShell worker's CA settings
+  (fixed, and shown working with OpenShell 0.1.6) and owned-delete inventory
+  (changed, but the cleanup still fails against 0.1.6; see above), a
+  restore request held after a managed compaction, `rig broadcast`'s `FAILED`
+  for an uncertain recipient, the refusal for a seat handed over before 0.7.9,
+  where to read what has run under the runner contract, and Codex seats'
+  `rig: command not found` (now a non-login shell by default; see Behaviour that
+  changed).
+- **The 0.7.8 limits that 0.7.9 didn't fix or change, and the 0.7.7 and 0.7.6
+  limits, below still apply.**
+
+### Credits
+
+Thanks to @dajiaohuang, @Aummadour, @billbenetti, @smota, @Nikhila-Dhevi-R and
+@Vaishnavi220506 for merged contributions, and to @AlephCasara, whose work on a
+fork pointed us to the unpaged event notification drain
+([#2610](https://github.com/mvschwarz/openrig/pull/2610)), and to @natefaerber, whose
+approach in [#2467](https://github.com/mvschwarz/openrig/pull/2467) the Herdr view fix follows. Thanks to
+@billbenetti, @KM199, @ziadeh, @dajiaohuang, @jajabong, @george661 and
+@aanestesia for reports carried into this release, and to @smota for the
+OpenTelemetry discussion behind [#2589](https://github.com/mvschwarz/openrig/pull/2589).
+
+## [0.7.9]
+
+Published 2026-11-06 as `@openrig/[email protected]`, built from the commit tagged
+`v0.7.9` (`ddb9d25b`).
+
+OpenRig 0.7.9 follows through on 0.7.8. A person's typed words on a runner
+seat's permission ask now reach the agent, each permit OpenRig grants a runner
+by default is recorded, and a runner can keep a long, quiet turn reading
+working. New stream observations go to the queue worker by default. Fresh Pi
+seats, and fresh Codex seats on Codex 0.164.0 or later, can keep OpenRig's
+guidance out of your repository's `AGENTS.md`. There are fixes across delivery,
+asks, handover, install and recovery, many reported or contributed by people
+using OpenRig. The OpenShell worker that wasn't usable in 0.7.8 still isn't
+usable in 0.7.9; read [The OpenShell worker](#the-openshell-worker-in-079)
+before relying on it.
+
+### Before you upgrade
+
+- **Restart the daemon after installing:**
+  ```bash
+  npm install -g @openrig/cli@0.7.9
+  rig daemon stop && rig daemon start
+  ```
+  Four migrations (146 to 149) record a Pi seat's guidance file, what OpenRig
+  last wrote in each managed guidance block, an index for reading a rig's
+  outbox, and the stream intake position.
+  ([#2432](https://github.com/mvschwarz/openrig/pull/2432), [#2461](https://github.com/mvschwarz/openrig/pull/2461), [#2496](https://github.com/mvschwarz/openrig/pull/2496), [#2321](https://github.com/mvschwarz/openrig/pull/2321))
+- **Stream intake is on by default.** Each `rig stream emit` now opens an
+  intake row for the kernel queue worker (`queue.worker`), or adds a note to the
+  one still unclaimed. The row wakes the queue worker like any queue row, and
+  earlier stream history isn't replayed. **To keep 0.7.8's behaviour,** with no
+  intake rows, run this after installing: `rig config set stream.intake off`.
+  ([#2321](https://github.com/mvschwarz/openrig/pull/2321))
+- **The bundled openrig-core plugin advances from 0.1.24 to 0.1.25.** Its
+  seat-continuity-and-handover skill changed. Managed refresh preserves replaced
+  files under `$OPENRIG_HOME/backups/managed-refresh/<id>/`; reapply local edits
+  you still need from that copy. ([#2420](https://github.com/mvschwarz/openrig/pull/2420), [#2554](https://github.com/mvschwarz/openrig/pull/2554))
+- **If you run an OpenShell environment, rebuild its guest image** with this
+  release's helper (`daemon/assets/openshell/README.md`). An older helper
+  rejects the new startup step, and the error names the rebuild. ([#2418](https://github.com/mvschwarz/openrig/pull/2418))
+
+### Runner seats
+
+- **A person's typed words reach the agent, and the ask says what they do.** A
+  reply of just yes or no counts as the agent's allow or reject. "Any other
+  words choose nothing": OpenRig's default permits the request when the agent
+  offered an allow option, and the words come back whole in the read result as
+  `said`. Section 7 of `docs/reference/harness-runner-contract.md` lets a runner
+  queue them for the agent "as an ordinary prompt" after the current turn:
+  "They're never the answer." A runner that drops them still keeps the contract.
+  ([#2363](https://github.com/mvschwarz/openrig/pull/2363), [#2561](https://github.com/mvschwarz/openrig/pull/2561))
+- **Each permit OpenRig grants by default is recorded,** once, as a stream item
+  from the seat: when no one can be asked, and when an ask is closed without a
+  choice (a row passed to an agent seat included). The record names the tool
+  call, why no person answered (for a closed ask, who closed which row and how),
+  the option the agent got with its kind, and the conversation and request or
+  the ask. Read the records with `rig stream list --tag runner-permission-permit`,
+  or `--source <seat>`. A record has no destination, and one that can't be
+  written doesn't change the permit. Stream intake reads past it, so it opens no
+  intake row and wakes no one. ([#2361](https://github.com/mvschwarz/openrig/pull/2361), [#2563](https://github.com/mvschwarz/openrig/pull/2563))
+- **A working beat keeps a long, quiet turn reading working.** A runner may post
+  `active` / `liveness` while a prompt is outstanding and its agent is alive
+  (section 4); one that sends no beats behaves as before. The Pi runner and the
+  dsh example post it. ([#2370](https://github.com/mvschwarz/openrig/pull/2370), [#2336](https://github.com/mvschwarz/openrig/pull/2336), [#2399](https://github.com/mvschwarz/openrig/pull/2399), [#2388](https://github.com/mvschwarz/openrig/pull/2388))
+- **The conformance check** covers the beat, and no longer falsely fails a
+  correct runner on a busy machine (cases 1, 2 and 4b).
+  ([#2368](https://github.com/mvschwarz/openrig/pull/2368), [#2375](https://github.com/mvschwarz/openrig/pull/2375), [#2388](https://github.com/mvschwarz/openrig/pull/2388), [#2530](https://github.com/mvschwarz/openrig/pull/2530))
+- **Section 8 names ACP's `session/resume`,** and the dsh example reads
+  `DSH_SEAT_BIN` and `DSH_SEAT_HOME`, which a member's `env:` can set, and
+  reports dsh's real version. ([#2364](https://github.com/mvschwarz/openrig/pull/2364), [#2414](https://github.com/mvschwarz/openrig/pull/2414))
+
+### The OpenShell worker in 0.7.9
+
+**The OpenShell worker still doesn't work in 0.7.9. Don't rely on this path.**
+
+0.7.9 repairs the layer that stopped it in 0.7.8: in our release testing the
+guest's agent now starts and reaches its first input. Startup reaches the guest
+through a hash-checked first-input file, and nothing is typed through the host
+viewer; the case where the viewer exits wasn't run in our release testing.
+Readiness checks the guest's own execution, and a failed owned delete reports
+the stage that failed.
+([#2418](https://github.com/mvschwarz/openrig/pull/2418), [#2404](https://github.com/mvschwarz/openrig/pull/2404))
+
+Our release testing on 0.7.9 then found two more problems:
+- **The worker can't reach its model.** The guest's Claude starts, but every API
+  call fails with `SELF_SIGNED_CERT_IN_CHAIN`. The guest launcher builds a clean
+  environment and leaves out the CA settings the OpenShell exec supplies
+  (`NODE_EXTRA_CA_CERTS`, `SSL_CERT_FILE`, `REQUESTS_CA_BUNDLE` and
+  `CURL_CA_BUNDLE`).
+- **Deleting an owned environment doesn't run.** Its inventory step misreads a
+  "Not Found" reply, so the delete is never sent and the owned resources stay.
+  In our test they were removed with the vendor's own cleanup.
+
+Attached resources were kept, as intended, and the worker stopped cleanly. With
+the four CA settings passed through by hand, a worker did useful work; a
+verified export, an exact restore and a finished task weren't shown. Fixes for
+both problems are on `main` ([#2602](https://github.com/mvschwarz/openrig/pull/2602),
+[#2603](https://github.com/mvschwarz/openrig/pull/2603)). They aren't in 0.7.9
+and haven't been run end to end.
+
+### Assistants connected through a grant service (experimental)
+
+**Experimental and unsupported. Off by default. Use at your own risk.** This path
+needs a grant service that you supply and run; ordinary use of OpenRig doesn't
+need one. With `rig mcp serve --http --grant-service <url>`, a connected
+assistant can now claim, update and park the rows addressed to its own seat, and
+see its team with `rig_team`, within the scope its existing tools have. Serving
+MCP over HTTP behind your own tunnel "is like opening a port in your firewall."
+See `docs/reference/hosted-members.md`. ([#2384](https://github.com/mvschwarz/openrig/pull/2384))
+
+### For client authors
+
+- **New reads:** `GET /api/gateway/participants` (people and the default
+  person), `GET /api/queue/open-human-updates` (a person's open updates, without
+  Slack), `GET /api/queue/outbox/list?rig=<rig>` (a rig's messages in one
+  request), and `GET /api/gateway/slack/channels` and `/users` (a page of public
+  channels or of people, without the client holding the bot token).
+  ([#2495](https://github.com/mvschwarz/openrig/pull/2495), [#2497](https://github.com/mvschwarz/openrig/pull/2497), [#2496](https://github.com/mvschwarz/openrig/pull/2496), [#2501](https://github.com/mvschwarz/openrig/pull/2501))
+- **Settle exactly the decision a person reviewed:** `GET /api/queue/:qitemId`
+  returns `decisionRevision`, and answer, resolve, approve and deny take an
+  optional `expectedRevision`. A stale one gets 409 `decision_revision_changed`
+  with nothing written; a caller that sends none acts as before. ([#2448](https://github.com/mvschwarz/openrig/pull/2448))
+- **A client reference:** `docs/reference/daemon-http-clients.md`, with a recipe
+  for people, open updates and a rig's outbox. ([#2490](https://github.com/mvschwarz/openrig/pull/2490), [#2524](https://github.com/mvschwarz/openrig/pull/2524))
+
+### Seats, delivery and the TUI
+
+- **Delivery:**
+  - a Claude prompt suggestion isn't taken for a person's draft;
+  - a long line Claude wraps short of the box is still submitted;
+  - a guarded wake recovers its own paste instead of holding later deliveries;
+  - a wake isn't typed into Claude's open Theme picker, or refused because a
+    reply quotes a permission question;
+  - a seat that starts, or takes over a seat, with held messages is told about
+    them; each watchdog job holds one reminder, and handoffs pass a full held
+    quota;
+  - a reminder that couldn't reach its seat, for example after a reboot, stays
+    owed and retries;
+  - a stalled tmux client is ended after 30 seconds.
+  ([#2435](https://github.com/mvschwarz/openrig/pull/2435), [#2436](https://github.com/mvschwarz/openrig/pull/2436), [#2447](https://github.com/mvschwarz/openrig/pull/2447), [#2514](https://github.com/mvschwarz/openrig/pull/2514), [#2520](https://github.com/mvschwarz/openrig/pull/2520), [#2437](https://github.com/mvschwarz/openrig/pull/2437), [#2442](https://github.com/mvschwarz/openrig/pull/2442), [#2438](https://github.com/mvschwarz/openrig/pull/2438), [#2478](https://github.com/mvschwarz/openrig/pull/2478), [#2452](https://github.com/mvschwarz/openrig/pull/2452))
+- **Status and receipts:**
+  - an idle Claude with newer footer rows reads idle, and a busy Codex with a
+    quiet window reads working;
+  - window samples older than their bound no longer decide a seat's activity, so
+    a seat whose pane can't be read reads unknown, not its last verdict;
+  - `rig send --verify` and `rig walk` report unconfirmed when they can't see the
+    result, and uncertain sends and broadcast recipients read `indeterminate`;
+  - a broadcast that reached nobody doesn't say "Delivered";
+  - replies to `human@kernel`, `human@host` and people forwarded from another
+    host get a usable `rig queue create` hint.
+  ([#2470](https://github.com/mvschwarz/openrig/pull/2470), [#2500](https://github.com/mvschwarz/openrig/pull/2500), [#2558](https://github.com/mvschwarz/openrig/pull/2558), [#2509](https://github.com/mvschwarz/openrig/pull/2509), [#2510](https://github.com/mvschwarz/openrig/pull/2510), [#2505](https://github.com/mvschwarz/openrig/pull/2505), [#2507](https://github.com/mvschwarz/openrig/pull/2507), [#2512](https://github.com/mvschwarz/openrig/pull/2512), [#2359](https://github.com/mvschwarz/openrig/pull/2359), [#2493](https://github.com/mvschwarz/openrig/pull/2493), [#2504](https://github.com/mvschwarz/openrig/pull/2504))
+- **Guidance and handover:**
+  - an edit to a culture or startup file reaches the next start, and a block you
+    edited by hand is kept and named;
+  - a shared `CLAUDE.md` or `AGENTS.md` stops growing by a line each launch;
+  - a Claude seat loads openrig-core's skills once;
+  - a fresh or rebuild handover successor gets its guidance back for every
+    runtime and a list of its startup files, role first, and is marked when a
+    required file is missing;
+  - launched successors can be stopped, restarted and cleaned, and the rebuild
+    receipt reports its lineage record and the old process truthfully;
+  - a Claude launched through the native installer's symlink can be rebuilt, a
+    slow classic resume under a non-exec launch shell no longer lands in
+    attention, and a `local:` Claude seat gets its activity hooks back on a
+    contained resume; restores use the seat's launch directory.
+  ([#2441](https://github.com/mvschwarz/openrig/pull/2441), [#2443](https://github.com/mvschwarz/openrig/pull/2443), [#2432](https://github.com/mvschwarz/openrig/pull/2432), [#2461](https://github.com/mvschwarz/openrig/pull/2461), [#2376](https://github.com/mvschwarz/openrig/pull/2376), [#2433](https://github.com/mvschwarz/openrig/pull/2433), [#2457](https://github.com/mvschwarz/openrig/pull/2457), [#2387](https://github.com/mvschwarz/openrig/pull/2387), [#2402](https://github.com/mvschwarz/openrig/pull/2402), [#2474](https://github.com/mvschwarz/openrig/pull/2474), [#2458](https://github.com/mvschwarz/openrig/pull/2458), [#2466](https://github.com/mvschwarz/openrig/pull/2466), [#2440](https://github.com/mvschwarz/openrig/pull/2440), [#2420](https://github.com/mvschwarz/openrig/pull/2420), [#2371](https://github.com/mvschwarz/openrig/pull/2371), [#2518](https://github.com/mvschwarz/openrig/pull/2518), [#2479](https://github.com/mvschwarz/openrig/pull/2479), [#2373](https://github.com/mvschwarz/openrig/pull/2373))
+- **Asks and answers:**
+  - in a shared workspace, any registered person settles another's ask as
+    themselves, and the first answer wins everywhere;
+  - a person may have no connector binding, and is then asked on Steering and in
+    the Feed only;
+  - a click on a re-parked Slack ask records a new answer and wakes its asker;
+  - a Slack ask's post shows it was handled once its row settles;
+  - For You no longer offers Approve on an ask with questions.
+  ([#2386](https://github.com/mvschwarz/openrig/pull/2386), [#2419](https://github.com/mvschwarz/openrig/pull/2419), [#2421](https://github.com/mvschwarz/openrig/pull/2421), [#2472](https://github.com/mvschwarz/openrig/pull/2472), [#2415](https://github.com/mvschwarz/openrig/pull/2415))
+- **Compaction:** managed compaction reads a small Claude pane at a readable
+  size; a manual `rig compact` that times out says whether the restore map was
+  ready; a Pi seat shows a failed or cancelled compaction as such.
+  ([#2445](https://github.com/mvschwarz/openrig/pull/2445), [#2400](https://github.com/mvschwarz/openrig/pull/2400), [#2537](https://github.com/mvschwarz/openrig/pull/2537))
+- **Install, hosts and setup:**
+  - the openrig-upgrade skill backs up without the sqlite3 CLI when Node has
+    `node:sqlite`;
+  - preflight warns when a Codex seat's login shell can't find `rig`;
+  - the experiment config can be written where Node has no `process.getuid`;
+  - bundle unpack restores archived owner bits;
+  - the Pi Node check is bounded, and `rig doctor` no longer hangs on a port held
+    by a listener that never replies;
+  - a scoped launch on tmux 3.4 keeps its default command;
+  - the kernel's template survives a daemon restart;
+  - `rig host ls` reads your ssh config with `ssh -G`, which runs any
+    `Match exec` command in it, and `rig host pair` names the bearer token the
+    target needs;
+  - the install docs describe the installer's step-4 outcome and the first
+    kernel's provider choice.
+  ([#2394](https://github.com/mvschwarz/openrig/pull/2394), [#2522](https://github.com/mvschwarz/openrig/pull/2522), [#2449](https://github.com/mvschwarz/openrig/pull/2449), [#2475](https://github.com/mvschwarz/openrig/pull/2475), [#2451](https://github.com/mvschwarz/openrig/pull/2451), [#2549](https://github.com/mvschwarz/openrig/pull/2549), [#2557](https://github.com/mvschwarz/openrig/pull/2557), [#2471](https://github.com/mvschwarz/openrig/pull/2471), [#2488](https://github.com/mvschwarz/openrig/pull/2488), [#2491](https://github.com/mvschwarz/openrig/pull/2491), [#2551](https://github.com/mvschwarz/openrig/pull/2551), [#2560](https://github.com/mvschwarz/openrig/pull/2560))
+- **The CLI:**
+  - unknown `--state`, `--kind` and `--priority` values warn and still run;
+  - `rig workspace validate` prints a readable refusal, and `rig seat set-runtime`
+    says what to do for a seat with no pane;
+  - a cut `rig queue show` body says how to get the rest;
+  - `rig queue update --closure-required-at` moves a claimed row's deadline;
+  - `rig seat set-typing-guard` applies from an ordinary terminal, recorded as
+    unattributed;
+  - `rig scope` tells shipped, superseded and closed slices apart, flags a slice
+    name two missions share, and skips fenced examples;
+  - `rig context add <url>` installs into a new nested name;
+  - a retained stuck-sweep recovery names its current reason, and a refused
+    context-pressure setting says what it compared.
+  ([#2357](https://github.com/mvschwarz/openrig/pull/2357), [#2365](https://github.com/mvschwarz/openrig/pull/2365), [#2358](https://github.com/mvschwarz/openrig/pull/2358), [#2367](https://github.com/mvschwarz/openrig/pull/2367), [#2383](https://github.com/mvschwarz/openrig/pull/2383), [#2422](https://github.com/mvschwarz/openrig/pull/2422), [#2460](https://github.com/mvschwarz/openrig/pull/2460), [#2351](https://github.com/mvschwarz/openrig/pull/2351), [#2434](https://github.com/mvschwarz/openrig/pull/2434), [#2492](https://github.com/mvschwarz/openrig/pull/2492), [#2395](https://github.com/mvschwarz/openrig/pull/2395), [#2546](https://github.com/mvschwarz/openrig/pull/2546), [#2552](https://github.com/mvschwarz/openrig/pull/2552), [#2553](https://github.com/mvschwarz/openrig/pull/2553))
+- **Launch options:** `specs.builtins_visible: false` leaves built-in specs out
+  of the inventory (`rig specs ls --include-builtins` still lists them).
+  `plugin_channels` passes a Claude member's own plugin channels to
+  `--channels`; beside `claude_mcp: none` it warns that no channel could
+  register, and the member still launches. We haven't run it with a real channel
+  plugin. ([#2379](https://github.com/mvschwarz/openrig/pull/2379), [#2439](https://github.com/mvschwarz/openrig/pull/2439))
+- **The TUI and web UI:**
+  - body text reads on a light terminal, and the selected row shows in
+    16-colour mode;
+  - JUST FINISHED and RECENTLY FINISHED show what closed last, in the configured
+    time zone, even on a busy fleet;
+  - a fresh seat opened in tiles is sized to its tile, and grid fillers open your
+    login shell;
+  - in the web UI, the Topology filter no longer hangs, and the kernel card and
+    maintenance notice display correctly.
+  ([#2499](https://github.com/mvschwarz/openrig/pull/2499), [#2506](https://github.com/mvschwarz/openrig/pull/2506), [#2392](https://github.com/mvschwarz/openrig/pull/2392), [#2393](https://github.com/mvschwarz/openrig/pull/2393), [#2459](https://github.com/mvschwarz/openrig/pull/2459), [#2464](https://github.com/mvschwarz/openrig/pull/2464), [#2465](https://github.com/mvschwarz/openrig/pull/2465), [#2411](https://github.com/mvschwarz/openrig/pull/2411), [#2413](https://github.com/mvschwarz/openrig/pull/2413), [#2412](https://github.com/mvschwarz/openrig/pull/2412))
+- **Also:**
+  - transcript search scans a long line once instead of once per chunk, with the
+    same results;
+  - a Slack transport error has the bot token removed (see Known limits);
+  - the hook-review check recognises a kernel Codex pane that runs under a shell;
+  - the reference docs cover a Slack private-channel route, missed-message
+    recovery, checked stops and slice status words, the spec scan, bundle hashes
+    and the Pi selector path.
+  ([#2397](https://github.com/mvschwarz/openrig/pull/2397), [#2503](https://github.com/mvschwarz/openrig/pull/2503), [#2416](https://github.com/mvschwarz/openrig/pull/2416), [#2494](https://github.com/mvschwarz/openrig/pull/2494), [#2401](https://github.com/mvschwarz/openrig/pull/2401), [#2489](https://github.com/mvschwarz/openrig/pull/2489), [#2462](https://github.com/mvschwarz/openrig/pull/2462), [#2543](https://github.com/mvschwarz/openrig/pull/2543), [#2477](https://github.com/mvschwarz/openrig/pull/2477), [#2369](https://github.com/mvschwarz/openrig/pull/2369), [#2403](https://github.com/mvschwarz/openrig/pull/2403))
+
+### Behaviour that changed
+
+- **Stream intake is on by default** (above). `off` keeps 0.7.8's behaviour.
+- **Fresh Pi seats, and fresh Codex seats on 0.164.0 or later, get OpenRig's
+  guidance outside the repository `AGENTS.md`:** Pi in its own agent folder
+  (unless that file or folder is a link), Codex as launch developer
+  instructions. Codex still uses the repository file for a resume, fork,
+  restore or handover successor, an older or unknown version, or a config layer
+  that sets `developer_instructions` or can't be read. If only your shell
+  startup sets `CODEX_HOME`, OpenRig reads `~/.codex`, so a
+  `developer_instructions` in your home is replaced for that session; declare
+  the same absolute path as `CODEX_HOME` in the member's or rig's `env:` to have
+  it checked.
+- **`rig down`** removes OpenRig's own additions from a working folder once no
+  live seat uses them.
+- **After upgrading, the Slack connector edits asks that ended earlier** to show
+  they were handled, at Slack's edit rate: about an hour for 450 asks.
+  `rig slack status` shows pending edits.
+- **A later answer to a settled ask** gets 409 `qitem_already_terminal` and
+  records nothing.
+- **`rig send --verify`** reports unconfirmed and exits 1, not delivered, when it
+  can't see what its Enter did.
+- **`rig export` with no `-o`** writes `rig-export.yaml` beside an existing
+  `rig.yaml` instead of replacing it. ([#2356](https://github.com/mvschwarz/openrig/pull/2356))
+- **`rig workspace validate`** exits 1 for malformed optional frontmatter.
+  ([#2396](https://github.com/mvschwarz/openrig/pull/2396))
+- **An active rig name that matches more than one rig** makes `rig env down`,
+  `status` and `logs` list the matching rig IDs and stop, and
+  `rig doctor --spec` report conformance as unknown, instead of using the first
+  match. ([#2548](https://github.com/mvschwarz/openrig/pull/2548), [#2550](https://github.com/mvschwarz/openrig/pull/2550))
+- **New default kernels** start in `<workspace.root>/kernel`. Existing kernels
+  stay where they are; `docs/reference/agent-startup-guide.md` describes the
+  optional move.
+
+### Known limits
+
+- **OpenShell:** the worker doesn't work in 0.7.9 (see
+  [The OpenShell worker](#the-openshell-worker-in-079)). 0.7.8's requirements,
+  cost note and cleanup limits still apply, including "Never grant the guest
+  access to the daemon's main API port (normally 7433)." A same-generation
+  reconciled launch opens no viewer, so run `rig environment connect` yourself;
+  non-text startup actions aren't supported on environment seats.
+- **Grant-service assistants aren't woken:** "an item stays queued until its
+  receiver claims it."
+- **Open updates for a person** read a 1,001-row window, so a page can come back
+  short, marked `truncated`.
+- **Slack:** a private channel needs the documented route, which we don't test,
+  and direct messages aren't supported. The channel and people reads, and a
+  click on a re-parked ask, haven't been run against a live Slack workspace in
+  our release testing. Redaction of the bot token from a transport error can
+  still leave a malformed token fragment shorter than six characters, so review
+  an error before sharing it. ([#2503](https://github.com/mvschwarz/openrig/pull/2503))
+- **SSH host status** reads `unknown` behind a ProxyJump or ProxyCommand; that
+  isn't proof the connection is broken. ([#2488](https://github.com/mvschwarz/openrig/pull/2488))
+- **Shared asks:** "a click's option written before its answer row … can stand
+  beside or inside the closer's answer without the first clicker's authorship."
+- **Existing seats** pick up Codex launch instructions and the single skill load
+  at their next fresh launch. A `CLAUDE.md` or `AGENTS.md` that already grew
+  isn't trimmed.
+- **Permit records** last as long as stream retention keeps them. A permit a
+  runner makes on its own, when it can't reach OpenRig, is in its pane only. For
+  an ask opened before 0.7.9, a record may omit the option's kind. ([#2361](https://github.com/mvschwarz/openrig/pull/2361))
+- **Draft-aware delivery to Claude:** on 0.7.8 with Claude Code 2.1.319,
+  deliveries to idle Claude agents, queue wakes included, were reported held at
+  an empty prompt when a narrow pane showed
+  `✔ Update installed · Restart to update` on its own line. 0.7.9 still doesn't
+  recognize that bare banner, though it recognizes `/rc` and the full versioned
+  banner. A message the draft-aware delivery path holds isn't replayed
+  automatically by that path.
+  ([#2605](https://github.com/mvschwarz/openrig/issues/2605))
+- **After a managed compaction, the restore request can stay held.** OpenRig
+  delivers each post-compaction step only when it can read the seat's input box
+  as empty. When it can't, it holds the restore request and keeps retrying on
+  its polls, so the seat can sit idle at an empty prompt with its restore not
+  yet requested. `rig compact <seat> --state` and `rig seat status` show the
+  restore as owed, and from the second failed attempt the seat's owner gets a
+  queue item naming it. To restore by hand, send the seat a restore request
+  naming its newest restore map, its restore marker and its packet, then settle
+  the owed attempt with `rig compact <seat> --cancel`; a poll can still race
+  that step. Reported on 0.7.8, and 0.7.9 behaves the same way. ([#2575](https://github.com/mvschwarz/openrig/issues/2575))
+- **Seats handed over before 0.7.9** keep their adopted origin until a later
+  launched handover or managed launch records a new session, so until then
+  `rig seat stop`, `restart` and `clean` refuse them as before. ([#2440](https://github.com/mvschwarz/openrig/pull/2440))
+- **`rig broadcast`:** an uncertain recipient's outbox receipt now reads
+  `indeterminate`, but the broadcast's own output for that recipient still says
+  FAILED. Check that seat with `rig capture` before resending. ([#2507](https://github.com/mvschwarz/openrig/pull/2507))
+- **Codex hooks after an upgrade:** Codex sessions kept running across the
+  upgrade show "Hook failed — plugin contents changed; hook approval required"
+  until relaunched. In our upgrade test from 0.7.8, one seat restarted with
+  OpenRig's managed restart showed no review afterwards.
+- **Codex seats can still report `rig: command not found`** when `rig` isn't on
+  the login shell's PATH: with Node outside it, or when npm can't install
+  globally and OpenRig is installed inside the project instead. Preflight now
+  warns at launch. In our testing the agents recovered by calling `rig` by its
+  full path; the getting-started guide describes a `~/.profile` fix, which we
+  didn't run.
+- **What we ran end to end under the runner contract:** this release's dsh
+  example runner, natively, with dsh 0.2.1-rc.2 on DeepSeek's
+  `deepseek/deepseek-v4-flash` through dsh's pi-ai OpenRouter chat-completions
+  profile, against 0.7.8 Core: two useful queue tasks, queued and multiline
+  input, a quiet tool interval of about 26 seconds that still read working,
+  allow-once and reject-once with their file effects, cancel, a return to the
+  same conversation with recall, and shutdown. The permission answers came from
+  an automated test operator, and the run's preset added `resume: true`, which
+  the shipped preset doesn't declare. dsh's own DeepSeek Messages route through
+  OpenRouter failed with HTTP 400, and the cause isn't established. Not run: a
+  missing-conversation refusal, cross-host restore, OpenShell, and other
+  adapters, models or providers. 0.7.8's release tests had already run its dsh
+  example end to end, so the 0.7.8 CHANGELOG's "No harness has run end to end
+  under the contract yet" was out of date.
+- **Fixed or changed from 0.7.8's limits:** the dsh example's `env:` names, the
+  conformance check's false failures, a person's typed words, the README's
+  operator address, a rebuild behind a symlink, the rebuild receipt, a click on
+  a re-parked ask, and recording a permit made when no one can be asked: a
+  person's answer is recorded on the ask, and a default permit as the stream
+  item above.
+- **The other 0.7.8 limits, and the 0.7.7 and 0.7.6 limits, below still apply.**
+
+### Credits
+
+Thanks to @porterchild, @Mohsintl, @Nikhi00718, @oodadoudou, @AdityaxDeore,
+@w4rc0n, @genius8267, @Aschenbath, @DeryFerd, @mohitdixit02, @Vaishnavi220506,
+@Nikhila-Dhevi-R, @natefaerber, @dajiaohuang, @Aummadour, @KM199 and @Totopo27
+for merged contributions, and to @todie and @seanmcn, whose fork work is carried
+in [#2439](https://github.com/mvschwarz/openrig/pull/2439) and [#2400](https://github.com/mvschwarz/openrig/pull/2400). Thanks to @ziadeh, @billbenetti, @eisenbergjerry,
+@ch-davidriott, @puniserv, @oneandonlydean, @sabnanikl-dev, @gifflen,
+@pvl-agent, @EZotoff, @adam-nectar, @bpamiri, @tgrundtvig, @aanestesia,
+@AdamWhitehurst and @eisenrigbot for reports carried into this release. Thanks
+also to @z4cc, @hitianqi, @jimallen, @m3ac-AllbrittenJ and @lkraider for reports
+and field feedback.
+
 ## [0.7.8]
 
-**Not yet published.**
+Published 2026-11-05 as `@openrig/[email protected]`, built from the commit tagged
+`v0.7.8` (`12df562c`).
 
 OpenRig 0.7.8 lets a person answer a runner seat's permission requests, with a
-conformance check and an example runner to build against. It adds a maintained
-way to run a Claude worker in an NVIDIA OpenShell sandbox, and includes 0.7.7's
-fixes plus repairs across seats, tmux, Slack and the TUI.
+conformance check and an example runner to build against. It includes 0.7.7's
+fixes plus repairs across seats, tmux, Slack and the TUI. It also includes a
+reference for running a Claude worker in an NVIDIA OpenShell sandbox, but that
+worker isn't usable in this release: see [the OpenShell reference](#the-openshell-reference-isnt-usable-in-078).
 
 ### Before you upgrade
 
@@ -61,20 +780,15 @@ fixes plus repairs across seats, tmux, Slack and the TUI.
 - **A seat activity grant** lets a runner outside the host's hook setup report
   activity for one seat and launch, without the install-wide hook token. ([#2295](https://github.com/mvschwarz/openrig/pull/2295))
 
-### Claude workers in OpenShell sandboxes
+### The OpenShell reference isn't usable in 0.7.8
 
-- **`rig environment setup` installs a maintained reference** that binds a seat
-  to one NVIDIA OpenShell sandbox (`rig environment create` or `attach`, then
-  `rig up`).
-  - **Context and results:** the seat's guidance, skills, plugins and startup
-    files go into the sandbox, and verified JSON results come back through
-    queue work.
-  - **Stop and delete:** stopping the agent, stopping compute and deleting the
-    sandbox are separate actions.
-  - **Restore:** a recreated sandbox resumes the same conversation.
-
-  The steps are in `daemon/assets/openshell/README.md`.
-  ([#2296](https://github.com/mvschwarz/openrig/pull/2296), [#2300](https://github.com/mvschwarz/openrig/pull/2300))
+0.7.8 includes `rig environment` (`setup`, `create`, `attach`) and a reference
+for binding a seat to an NVIDIA OpenShell sandbox (`daemon/assets/openshell/README.md`).
+**In our release testing the worker never became usable:** after the sandbox was
+created and the native process started, `rig environment connect` failed when its
+check timed out after 5 seconds, so no team work ran. Don't rely on this path in
+0.7.8.
+([#2296](https://github.com/mvschwarz/openrig/pull/2296), [#2300](https://github.com/mvschwarz/openrig/pull/2300), [#2378](https://github.com/mvschwarz/openrig/pull/2378))
 
 ### Seats, delivery and the TUI
 
@@ -136,8 +850,13 @@ fixes plus repairs across seats, tmux, Slack and the TUI.
 
 ### Known limits
 
-- **No harness has run end to end under the contract yet.** The dsh example and
-  the conformance check were checked against stand-ins.
+- ~~**No harness has run end to end under the contract yet.** The dsh example and
+  the conformance check were checked against stand-ins.~~ *Corrected after
+  release:* 0.7.8's release testing ran a real dsh 0.2.1-rc.2 through the
+  shipped dsh example runner for a person's allow-once and reject-once, and
+  through an outside runner, dsh-seat 0.3.0, for launch, queue work, multiline
+  input, activity, cancel and exact resume. The conformance check itself runs
+  against a stand-in agent.
 - **The answer owner is a registered person.** "A colleague's answer is
   followed on this host. If the ask is handed to another host, the request
   waits, shown as needs input, until the agent's turn ends."
@@ -150,14 +869,13 @@ fixes plus repairs across seats, tmux, Slack and the TUI.
 - **The dsh example keeps one hold per seat:** "if dsh asks twice in one turn,
   settling one ask releases the hold while the other may still wait on a
   person."
-- **The OpenShell reference requires:**
-  - NVIDIA OpenShell 0.1.6;
-  - a Linux image;
-  - an API-backed provider using `ANTHROPIC_API_KEY`. Subscription or OAuth
-    credentials aren't supported.
-
+- **The worker isn't usable** (above). The reference also requires NVIDIA
+  OpenShell 0.1.6, a Linux image, and an API-backed provider using
+  `ANTHROPIC_API_KEY`; subscription or OAuth credentials aren't supported.
   "Provider registration and image construction can incur costs." The worker's
   activity reads unknown, from pane sampling, never idle.
+- **The OpenShell README example uses `operator@kernel`;** the actual kernel
+  operator is `operator-agent@kernel`.
 - **"Never grant the guest access to the daemon's main API port (normally
   7433)."** The guest-only listener is what confines a guest's reach. Its grant
   only authorizes requests for its own seat and launch.

README.md

@@ -75,7 +75,7 @@ curl -fsSL https://raw.githubusercontent.com/mvschwarz/openrig/v0.7.5/scripts/in
 curl -fsSL https://raw.githubusercontent.com/mvschwarz/openrig/v0.7.5/scripts/install.sh | sh
 ```
 
-If a provider isn't signed in yet, the last step reports `FAILED [4/4]`, and `rig setup` lists the sign-in under "Some steps need attention". If the only remaining failures are provider sign-ins, the install steps finished; sign in to each selected provider as below and continue.
+If the last step isn't ready, the installer reports `INSTALLED, SETUP NOT COMPLETE` and points to the details from `rig setup` above. Follow that output and any reported Fix, then rerun `rig setup`; the CLI does not need reinstalling. If the only remaining failures are provider sign-ins, sign in to each selected provider as below and continue.
 
 **Or step by step:**
 

ROADMAP.md

@@ -6,9 +6,9 @@ This isn't a schedule or a promise: we release often, and an item ships when a g
 To see what's being worked on, by release and theme, open the [OpenRig roadmap board](https://github.com/users/mvschwarz/projects/1).
 Issues labelled `good first issue` are also listed on the [contribute page](https://github.com/mvschwarz/openrig/contribute).
 
-## Released in 0.7.6
+## Released in 0.7.10
 
-[0.7.6 is published](https://github.com/mvschwarz/openrig/releases/tag/v0.7.6): follow a slice's work in one terminal workspace, plus steadier daemon start and delivery. See its release notes for what shipped, known issues, upgrade instructions and contributor credits.
+[0.7.10 is published](https://github.com/mvschwarz/openrig/releases/tag/v0.7.10): it is a patch for people running their own agent teams. A step held after a managed compaction now goes in instead of waiting forever. See its release notes for what shipped, known issues, upgrade instructions and contributor credits.
 
 ## How to help
 

docs/as-built/architecture/adapters-and-runtimes.md

@@ -398,6 +398,21 @@ adapter never closes a session on it and reports `rex_outcome_unknown`.
 Herdr and Rex build their layout trees from one grid helper
 (`buildTileGrid`, `tile-layout.ts`).
 
+`HerdrAppProvider` (`herdr-app-provider.ts`) opens a view in a herdr-app
+window (one provider per app home, cached in `startup.ts`). In the daemon it is
+used only when named (`resolveHerdrAppProvider` in `terminal-service.ts`). Its
+transport (`herdr-app-transport.ts`) sends one JSON line per
+connection to `<app home>/herdr-app.sock` and reads the window's own "did not
+answer" as `unknown`. A view is a child of OpenRig's stack in a derived
+`layouts/rig.json` (`deriveRigLayout`, `herdr-app-layout.ts`), written
+atomically, with viewers started by `hc run --bg --no-parent` and journalled
+first (`herdr-app-state.ts`). Removal rewrites the file before a viewer ends,
+and ends it only after the `rig-strip` mod's acknowledgement
+(`<app home>/openrig/ack.json`) shows a layout without it. An `hc events`
+watcher (`startup.ts`) turns a viewer that ended elsewhere into an ended cell.
+Its four mods ship in `packages/daemon/assets/herdr-app/mods/` and are
+installed by `herdr-app-install.ts`.
+
 ### Tmux server selection
 
 Before reconciliation, `startup.ts` calls `selectTmuxServer` in

docs/as-built/architecture/architecture-rules-and-event-system.md

@@ -214,16 +214,26 @@ inside a caller-managed transaction, `eventBus.persistWithinTransaction(...)`
 (`event-bus.ts:70`) with subscribers notified after commit — across domain
 services (`stream-store.ts`, `workflow-runtime.ts`, `restore-orchestrator.ts`,
 `node-launcher.ts`, etc.) and route handlers. The event log is append-only and
-SQLite-backed.
+SQLite-backed, and it holds stored events only. `view.changed`, which the view
+event bridge (`view-event-bridge.ts`) produces for each queue, inbox, proof or project-classification change,
+is transient: `eventBus.publishTransient(...)` delivers it to the live
+subscribers without a row, without a `seq` of its own and without moving the
+notify watermark. It carries `transient: true` and, as its `seq`, the `seq` of
+the stored event that caused it.
+
+Upgrade note: `view.changed` rows stored by older versions stay in the log, so
+`/api/events` replays and the telemetry export still return them until an events retention
+(a separate change) removes them.
 
 SSE delivery surfaces include the following. The daemon has **11**
 `streamSSE(` call sites across **10** files in `packages/daemon/src/routes/`
 (`git grep -o 'streamSSE(' -- packages/daemon/src/routes | wc -l`;
 `git grep -l 'streamSSE(' -- packages/daemon/src/routes | wc -l`).
 
-- `GET /api/events` — global stream of all events (`server.ts:730`
-  `app.route("/api/events", eventsRoute)`); `?rigId=` keeps one rig's. Each
-  event's SSE `id` is its `seq`. Before going live, a connection replays
+- `GET /api/events` — global stream of the stored events (`server.ts:730`
+  `app.route("/api/events", eventsRoute)`); `?rigId=` keeps one rig's.
+  Transient events (`view.changed`) are not stored, so they appear here
+  neither live nor in a replay. Each event's SSE `id` is its `seq`. Before going live, a connection replays
   stored events, read and written 500 at a time (`routes/events.ts`):
   - with a `Last-Event-ID` header (an `EventSource` sends it when it
     reconnects), every event after that `seq`, whatever `replay` says;
@@ -234,6 +244,10 @@ SSE delivery surfaces include the following. The daemon has **11**
   Either way the replay ends at the newest event stored as the connection
   subscribes; the events stored meanwhile were buffered and follow it, each
   once and in order, before the stream goes live.
+- `GET /api/views/sse` (alias `/api/views/watch`) and
+  `GET /api/views/:viewName/sse` — `view.changed`, live only (`routes/views.ts`).
+  An event's `id` there is the `seq` of the stored event that caused it, so it
+  repeats across the views one change touches; it is not a replay cursor.
 - `GET /api/stream/watch` — new stream items (`routes/stream.ts:194`).
 - `GET /api/queue/watch` — queue/inbox coordination events
   (`routes/queue.ts:1054`, with a `/sse` alias at `:1055`).

docs/as-built/architecture/coordination-primitive.md

@@ -237,7 +237,9 @@ The sweep reports six finding kinds. Its custody check covers closures that
 name a successor qitem or a seat (`closure_reason = handed_off_to`) in the last
 24 hours, plus older closures with open findings. A seat target is satisfied by
 a local successor carrying this row's lineage, and any target by a
-`custody-verified: <target> …` note on the closed row. A new custody finding for
+`custody-verified:` note on the closed row: per note, its first word that names one
+of the row's targets exactly (or a comma list of them) is what it verifies, and later
+mentions don't count; the finding prints each target's note. A new custody finding for
 a human destination goes to the agent that closed the item, using the closing
 transition from active or archived history. Later notes and notification
 receipts do not change that owner. If the closing actor is missing, human, system
@@ -461,6 +463,23 @@ in the current update window. Current-update coverage is independent of the
 receipt window, so an unavailable current reader cannot retain closed delivery
 history beyond a successful history read. Both reads leave queue state intact.
 
+Clients that need only current FYIs can use `GET /api/queue/open-human-updates`.
+It returns `{items, limit, truncated}` with open human-addressed updates, ordered
+critical then urgent then routine, without requiring a delivery receipt. `limit`
+defaults to 20 and accepts integers from 1 to 100. Optional `view` follows the
+same person perspective as `/api/attention`; a perspective read scans at most
+1000 rows and signals `truncated` when the window may omit matching items.
+
+A finished item's result file is its `evidenceRef`, read through the daemon's
+shared roots with `rig_artifact_read` (MCP) or `rig queue artifact <qitemId>
+--json` (CLI). For a code result, the assignee writes a JSON file of kind
+`openrig.code-result/v1`, with `repo` (an absolute path or a URL), `base` and
+`revision` (full commit ids, SHA-1 or SHA-256), and optionally `paths`, `branch`
+and `fetch: {remote, ref}`. Both reads then add `codeResult`, the exact pair a
+client can diff. A manifest whose base or revision isn't a full commit id gets
+`codeResultIssue` instead. `branch` and `fetch` are hints for finding the commits,
+never the revision. Other files read as before.
+
 - `/api/stream` (`server.ts:783`) — `POST /emit` (`routes/stream.ts:56`),
   `GET /list` (`:91`, including `sourceSession`, `hintDestination`, `hintTag`,
   `since`, `until` and `direction` filters), `GET /watch` + `/sse` SSE

docs/as-built/architecture/lifecycle-snapshot-restore.md

@@ -160,8 +160,9 @@ Restore behavior, each point checked in `restore-orchestrator.ts`:
   `:780`; if `status === "restoring"` the node is skipped with a warning and
   reported as `fresh`, `:783–785`; `degraded` only adds a warning).
 - Reuses the seat's durably bound tmux session name when the prior binding
-  has a usable one (`:840`), and otherwise falls back to the derived name
-  with a warning (`:869`).
+  has a usable one, and otherwise falls back to the derived name. A missing
+  binding after the latest session's recorded intentional exit is quiet;
+  unclean/legacy missing bindings and incomplete or invalid binding rows still warn.
 - Replays restore-safe startup using persisted startup context;
   prefilters missing optional artifacts into warnings; **hard-fails a node if
   a required startup file is missing** (`:1161` — status `"failed"`, error

docs/as-built/architecture/transport-and-transcripts.md

@@ -169,9 +169,19 @@ Routes: `packages/daemon/src/routes/{transport,transcripts,ask,chat,whoami}.ts`
      posture, one for each mode OpenRig can launch: bypass
      (`⏵⏵ bypass permissions on (shift+tab to cycle) · ← for agents`), accept
      edits (OpenRig's default), auto, don't ask, plan, and default
-     (`⏸ manual mode on · ← for agents`); no statusLine row between the
-     closing rule and the footer (OpenRig's statusLine collector prints
-     nothing); a blank, effort or tmux-notice row above the box; and above
+     (`⏸ manual mode on · ← for agents`); between the closing rule and the
+     footer, either no row (OpenRig's statusLine collector prints nothing) or
+     the rows a person's own statusLine prints, with the indicator Claude
+     draws at the right of the first one (`● high · /effort`), trusted only by
+     position, as the composer reader trusts them (#2703, #2798): the closing
+     rule is the first rule above the footer, the frame above it must check
+     out, and no row between them may be blank or look like a menu or dialog
+     (`❯`, `›` or `»` anywhere in the row, a row that starts with a numbered
+     choice such as `1. ` or `2) `, with or without a border or gutter before
+     it, a box corner, or a row framed by side borders); so a statusLine that
+     uses `›`, `»` or `❯` as a separator, prints a blank row or starts a row
+     with a number and a dot reads `unknown`; a blank, effort or tmux-notice
+     row above the box; and above
      that, in the status slot, the completion row of the seat's last turn
      (`✻ <verb> for <duration> · done <time>` at column 0, matched by the
      Claude Code 2.1.296 bundle's grammar: one of its eight past-tense verbs,
@@ -204,13 +214,15 @@ Routes: `packages/daemon/src/routes/{transport,transcripts,ask,chat,whoami}.ts`
      status row, `esc to interrupt`, `paste again to expand`, a queued label,
      Codex's pending-input headers or `↳` rows, a spinner). Any other frame,
      including every running, streaming, steer-preview, paste-hint or unknown
-     layout, another seat's name, a statusLine that prints a row, a background
-     shell in the footer and a fullscreen renderer with a statusLine (which
-     reserves a row for it), reads `seat: not-idle` or
-     `unknown`, and the send reads unconfirmed. This holds when a running turn
-     hides its signal (Codex hides its `• Working` row while it streams;
-     Claude Code shows only `paste again to expand` in its footer for 8 s
-     after a paste, and no `esc to interrupt` under a statusLine). So a send to
+     layout, another seat's name, a second rule drawn under the box, a blank,
+     prompt-, menu- or dialog-shaped row between the closing rule and the
+     footer (so a fullscreen renderer that reserves a blank row for the
+     statusLine reads as before), and a background shell in the footer, reads
+     `seat: not-idle` or `unknown`, and the send reads unconfirmed. This
+     holds when a running turn hides its signal (Codex hides its `• Working`
+     row while it streams; Claude Code shows only `paste again to expand` in
+     its footer for 8 s after a paste, and no `esc to interrupt` under a
+     statusLine). So a send to
      a busy or unknown seat always reads unconfirmed.
 
      **Versions.** The layouts were recorded from Claude Code 2.1.296 and
@@ -272,7 +284,13 @@ Routes: `packages/daemon/src/routes/{transport,transcripts,ask,chat,whoami}.ts`
      characters never widen it. Any other break, and a break inside a word of
      the text, reads `ambiguous` and the send stays unconfirmed. A collapsed paste label, Claude's
      queued-messages label or `Try "…"` suggestion, or another draft is never
-     the message. The default `--verify` matcher above is unchanged.
+     the message. The default `--verify` matcher above is unchanged. Claude
+     Code's box ends at the last rule on the screen; the rows under it up to
+     the footer are a person's statusLine, read by position with the same
+     rule as the seat check above (#2798), so a box with statusLine rows
+     under it reads as one without them, and a blank, prompt-, menu- or
+     dialog-shaped row there leaves it `unrecognized`. The startup Enter
+     guard reads the same region.
    - **The box is empty after Enter on the same pane process.** The capture
      after Enter must show the box empty, with the pane process
      (`#{pane_pid}`) and the seat's binding unchanged.

docs/as-built/architecture/workflow-runtime.md

@@ -84,6 +84,15 @@ Diagnostic-only rows, moved sources, and versions retained for unfinished
 work have separate paths in `readThrough`, `writeDiagnostic`, and
 `isPinnedByUnfinishedWork`.
 
+Only instantiate (and `rig workflow run`, which posts to the same route) and
+the workflows-folder scan register specs through `readThrough`.
+`WorkflowRuntime.validate` (`rig workflow validate`, `POST /validate`) reads
+the file with `WorkflowSpecCache.readFromDisk`: the same file and parse
+checks, in memory, with no `workflow_specs` write. A spec that was only
+validated from outside the workflows folder is not listed by
+`rig workflow specs` and is not found by its bare name; its path still
+instantiates (#2726).
+
 ## Durable state
 
 The migration files record the storage evolution; they are not a count of

docs/as-built/cli-reference.md

@@ -9,23 +9,23 @@ applies-when: |
   the source that implements a command. Use command help for invocation details.
 siblings: [README.md, codemap.md]
 prerequisite-reads: [../reference/help.md]
-last-verified-against-source: 34fdcb80543903b054a982cc48bb302ae6871682
-last-updated: 2026-10-26
+last-verified-against-source: b29a70f14fc11797d46cbd9f12d4e61a523c3d06
+last-updated: 2026-11-09
 ---
 
 # CLI Reference — Registered Commands and Options
 
-Verified against source commit `34fdcb80543903b054a982cc48bb302ae6871682`.
+Verified against source commit `e1dedbf531b6db9961b4c77c0a74564e58274a50`.
 The inventory below comes from the actual Commander tree returned by
 [`createProgram()`](../../packages/cli/src/index.ts), not a grep of command
 strings or an installed CLI from a different commit.
 
-There are **90 top-level registrations**, **410 registered command objects**
+There are **92 top-level registrations**, **440 registered command objects**
 below `rig` (including groups and the hidden `restore apply` command), and
-**1,304 explicitly registered option objects**, including the root version
+**1,460 explicitly registered option objects**, including the root version
 option. Aliases do not add command objects; short/long spellings of one
 option do not add option objects. Commander-generated help is additional.
-These are counts for that source snapshot, not a claim about a deployed release. The provider-switch and set-runtime rows include the later account-home options; those additions are not included in the snapshot counts. The `environment` section and the `host edit`, `host remove` and `terminal attach` rows were added later from source at `522d635c`, and aren't in the snapshot counts either. With `environment`, main registers 91 top-level commands.
+These are counts for that source snapshot, not a claim about a deployed release. Every registered family is inside the snapshot: the `exec` and `environment` families, the `roster` and `telemetry` rows, and the `host edit`, `host remove` and `terminal attach` rows that earlier snapshots listed as later additions are all counted here, as are the herdr-app provider's `terminal close` row and its `--app-home` options.
 
 ## Reading this reference
 
@@ -188,7 +188,7 @@ the sections that were not read are `null`.
 |---|---|---|
 | `rig gateway` | — | — |
 | `rig gateway human` | — | — |
-| `rig gateway human add <entityId>` | — | `--display-name <name>` **required**<br>`--binding <kind:connectorRef:secretsRef:role[:handle=<id>]>` **required** (repeatable)<br>`--delivery-class <A\|B\|C\|D>` **required**<br>`--away`<br>`--replace`<br>`--reason <reason>`<br>`--actor <actor>`<br>`--organization <id>`<br>`--group <id>` (repeatable)<br>`--policy <text>`<br>`--policy-file <path>`<br>`--default` |
+| `rig gateway human add <entityId>` | — | `--display-name <name>` **required**<br>`--binding <kind:connectorRef:secretsRef:role[:handle=<id>]>` (repeatable; without one, the person is reached in the Feed and Steering only)<br>`--delivery-class <A\|B\|C\|D>` **required**<br>`--away`<br>`--replace`<br>`--reason <reason>`<br>`--actor <actor>`<br>`--organization <id>`<br>`--group <id>` (repeatable)<br>`--policy <text>`<br>`--policy-file <path>`<br>`--default` |
 | `rig gateway human list` | — | `--json` |
 | `rig gateway human quiet` | — | — |
 | `rig gateway human quiet set <entityId>` | — | `--daily <window>`<br>`--until <time>`<br>`--tz <zone>`<br>`--actor <actor>`<br>`--json` |
@@ -226,6 +226,7 @@ the sections that were not read are `null`.
   `groups`, `warnings` and `participantFile`. `default` with no argument prints the current default and its source
   (`chosen` or `only-person`), or `null` with the reason. See `docs/reference/gateway-participants.md`.
 - `human set <entityId> slack-subscription <value>` — sets the person's Slack subscription: `every-ask` (the default) posts every ask to them to Slack; `urgent-only` posts only asks created with `--alert`, and other asks wait on Steering and the TUI; `unset` removes the authored subscription so the default applies. `human show` prints the value with its origin, `authored` or `default` (`slack-subscription: every-ask (default)`); `--json` carries it as `record.slackSubscription` `{ value, source }`.
+- `human set <entityId> binding.<n> <spec|unset>` — replaces `binding.<n>` with a full `kind:connectorRef:secretsRef:role[:handle=<id>]` spec, or removes that slot when `<value>` is `unset`. Removing a binding keeps the person, their preferences, their other bindings and their open questions, and prints `remainingBindings` (with their shifted `binding.<n>` slot indexes; the primary's slot shifts down too when a lower-index secondary is removed) plus `note`. No replacement `primary` is chosen silently: unsetting a `primary` while secondary bindings remain is refused, teaching how to replace the primary slot in place, promote an existing secondary binding (first save that binding's full spec from `rig gateway human show <entityId> --json`, then unset that secondary slot, then find the primary's current slot in `remainingBindings` or `rig gateway human show <entityId> --json` and set it to the saved spec with role `primary` — promotion replaces the former primary binding, and its connector reference and inbound handle are not retained as a secondary), or unset the secondary slots first using their current slot indexes from `remainingBindings` or `rig gateway human show <entityId>` and finish at `binding.0` when removing every binding. Removing the last binding makes the current delivery episode inactive so new missing-post findings stop (an already-open recovery item for an undelivered notice may remain, and the current delivery-outcome display may disappear while historical receipt notes remain); it does not answer, close or cancel open questions, cancel or reroute already-prepared deliveries, or retract anything a bridge has already accepted.
 
 ### parked
 
@@ -356,6 +357,17 @@ the sections that were not read are `null`.
 | `rig host list` | `ls` | `--json` |
 | `rig host doctor <id>` | — | `--posture <profile>`<br>`--public-addr <ip>`<br>`--json` |
 
+`rig host pair` needs the target daemon to run with `OPENRIG_AUTH_BEARER_TOKEN` set,
+even when it is bound to a tailnet address and so starts without one: pairing hands
+over that token. Without it, the target's `POST /api/hosts/pair-request` answers 409
+`pair_target_no_bearer`; it answers 409 `pair_human_required` when no single
+registered person can approve (none, or several with no default; name one with
+`--human`) (`packages/daemon/src/routes/hosts.ts:167-186`).
+
+The request stays open on the target for as long as `--timeout` waits (600 seconds by default), and the person's
+approval says how long and until when, in UTC. A target older than this CLI ignores the timeout and keeps the request
+10 minutes.
+
 Registration makes a host visible in topology; remote FEED attention remains opt-in.
 The human-readable success output of `host add` and `host pair` now includes the FEED
 next step alongside the existing `host doctor` suggestion. For an HTTP entry named
@@ -379,7 +391,7 @@ do not establish connectivity to a live peer.
 |---|---|---|
 | `rig ps` | — | `--json`<br>`--no-cleanup`<br>`--resources`<br>`--nodes`<br>`--full`<br>`--verbose`<br>`--limit <n>`<br>`--fields <list>`<br>`--summary`<br>`--filter <key=value>`<br>`--active`<br>`--running`<br>`-A, --all-rigs`<br>`--rig <name>`<br>`--session <name>`<br>`--include-archived`<br>`--host <id>`<br>`--all-hosts`<br>`--hosts <ids>` |
 
-- `--nodes --json` — each seat carries the daemon's served `needsAttention` flag; the CLI doesn't re-derive it, and it is absent when an older daemon doesn't serve it. `arbitratedNeedsInput` and `needsInputEvidence` (`{ rung, observedAt }` or null) come with `--full`, or by name through `--fields`.
+- `--nodes --json` — each seat carries the daemon's served `needsAttention` flag; the CLI doesn't re-derive it, and it is absent when an older daemon doesn't serve it. `arbitratedNeedsInput` and `needsInputEvidence` (`{ rung, observedAt }` or null) come with `--full`, or by name through `--fields`. `startedWithoutRequired` lists the required startup files a handover's successor started without, and the compact read keeps it.
 
 ### roster
 
@@ -395,7 +407,7 @@ do not establish connectivity to a live peer.
 | Invocation | Aliases | Declared options |
 |---|---|---|
 | `rig mcp` | — | — |
-| `rig mcp serve` | — | `--port <port>`<br>`--http`<br>`--bind <host:port>`<br>`--grant-service <url>`<br>`--grant-service-header-file <path>`<br>`--oauth-resource <url>`<br>`--oauth-issuer <url>` |
+| `rig mcp serve` | — | `--port <port>`<br>`--http`<br>`--bind <host:port>`<br>`--grant-service <url>`<br>`--grant-service-header-file <path>`<br>`--oauth-resource <url>`<br>`--oauth-issuer <url>`<br>`--front-door`<br>`--task-destination <seat>` |
 | `rig mcp token` | — | — |
 | `rig mcp token add <name>` | — | `--seat <session>` **required**<br>`--rig <name...>`<br>`--root <name...>`<br>`--workspace`<br>`--owner` |
 | `rig mcp token list` | — | `--json` |
@@ -407,9 +419,21 @@ do not establish connectivity to a live peer.
 |---|---|---|
 | `rig acp` | — | — |
 | `rig acp run <qitemId> <agent...>` | — | `--cwd <dir>`<br>`--permissions-to <session>`<br>`--load-session <id>`<br>`--poll-seconds <n>` |
+| `rig acp serve` | — | `--to <member@rig>`<br>`--as <address>`<br>`--poll-seconds <n>` |
 | `rig acp cancel <runId>` | — | — |
 | `rig acp stop <runId> [stopCommand...]` | — | — |
 
+### exec
+
+| Invocation | Aliases | Declared options |
+|---|---|---|
+| `rig exec` | — | — |
+| `rig exec harnesses` | — | `--json` |
+| `rig exec run <qitemId>` | — | `--harness <name>` **required**<br>`--cwd <dir>`<br>`--permissions-to <session>`<br>`--load-session <id>`<br>`--wait`<br>`--plan`<br>`--again`<br>`--poll-seconds <n>` |
+| `rig exec status <runIdOrQitemId>` | — | `--json` |
+| `rig exec cancel <runId>` | — | — |
+| `rig exec connector <runId>` (hidden) | — | — |
+
 ### agent
 
 | Invocation | Aliases | Declared options |
@@ -455,6 +479,11 @@ do not establish connectivity to a live peer.
 | `rig stream read` | — | `--limit <n>`<br>`--reader <session>`<br>`--json` |
 | `rig stream act <streamItemId>` | — | `--work <qitemId>`<br>`--decline <reason>`<br>`--duplicate <id>`<br>`--reader <session>`<br>`--json` |
 
+Stream intake: the `stream.intake` setting (default `queue.worker`, or a canonical `<seat>@<rig>`) receives each emitted item.
+It opens an intake row for that seat, which wakes it, or adds a note to that seat's intake row still pending; a `critical`
+or interrupting item makes the row urgent. `rig config set stream.intake off` keeps emits passive. A runner's
+default-permit record (tag `runner-permission-permit`) opens no row.
+
 ### queue
 
 | Invocation | Aliases | Declared options |
@@ -463,16 +492,17 @@ do not establish connectivity to a live peer.
 | `rig queue create` | — | `--source <session>`<br>`--destination <session>` **required**<br>`--body <text>`<br>`--body-file <path>`<br>`--body-context <ref>`<br>`--mission <id>`<br>`--slice <id>`<br>`--gate <role>`<br>`--priority <priority>`<br>`--tier <tier>`<br>`--tags <tags>`<br>`--expires-at <iso>`<br>`--id <qitemId>`<br>`--target-repo <name>`<br>`--summary <text>`<br>`--human-intent <intent>`<br>`--alert`<br>`--human-detail-file <path>`<br>`--reply-to <qitemId>`<br>`--human-questions-file <path>`<br>`--evidence-ref <path>`<br>`--host <id>`<br>`--no-nudge`<br>`--verify`<br>`--full`<br>`--json` |
 | `rig queue claim <qitemId>` | — | `--destination <session>`<br>`--full`<br>`--json` |
 | `rig queue unclaim <qitemId>` | — | `--destination <session>`<br>`--reason <text>`<br>`--full`<br>`--json` |
-| `rig queue update <qitemId>` | — | `--actor <session>`<br>`--state <state>`<br>`--reopen`<br>`--closure-reason <reason>`<br>`--closure-target <target>`<br>`--resolution <text>`<br>`--blocked-on <blocker>`<br>`--wake-watchdog <jobId>`<br>`--wake-after <duration>`<br>`--summary <text>`<br>`--evidence-ref <path>`<br>`--note <text>`<br>`--full`<br>`--json` |
-| `rig queue block <qitemId>` | — | `--on <blocker>` **required**<br>`--actor <session>`<br>`--summary <text>`<br>`--evidence-ref <path>`<br>`--note <text>`<br>`--continuation <text>`<br>`--wake-watchdog <jobId>`<br>`--wake-after <duration>`<br>`--full`<br>`--json` |
+| `rig queue update <qitemId>` | — | `--actor <session>`<br>`--state <state>`<br>`--reopen`<br>`--closure-reason <reason>`<br>`--closure-target <target>`<br>`--closure-required-at <iso>`<br>`--resolution <text>`<br>`--blocked-on <blocker>`<br>`--wake-watchdog <jobId>`<br>`--wake-after <duration>`<br>`--summary <text>`<br>`--evidence-ref <path>`<br>`--note <text>`<br>`--host <id>`<br>`--full`<br>`--json` |
+| `rig queue block <qitemId>` | — | `--on <blocker>` **required**<br>`--actor <session>`<br>`--summary <text>`<br>`--evidence-ref <path>`<br>`--note <text>`<br>`--continuation <text>`<br>`--wake-watchdog <jobId>`<br>`--wake-after <duration>`<br>`--host <id>`<br>`--full`<br>`--json` |
 | `rig queue resolve <qitemId>` | — | `--decision <text>` **required**<br>`--actor <session>`<br>`--bearer <token>`<br>`--no-notify`<br>`--json` |
 | `rig queue handoff <qitemId>` | — | `--from <session>`<br>`--to <session>` **required**<br>`--body <text>`<br>`--body-file <path>`<br>`--note <text>`<br>`--priority <priority>`<br>`--tier <tier>`<br>`--tags <tags>`<br>`--gate <role>`<br>`--target-repo <name>`<br>`--summary <text>`<br>`--evidence-ref <path>`<br>`--alert`<br>`--host <id>`<br>`--no-nudge`<br>`--full`<br>`--json` |
 | `rig queue handoff-and-complete <qitemId>` | — | `--from <session>`<br>`--to <session>` **required**<br>`--body <text>`<br>`--body-file <path>`<br>`--note <text>`<br>`--priority <priority>`<br>`--tier <tier>`<br>`--tags <tags>`<br>`--gate <role>`<br>`--target-repo <name>`<br>`--summary <text>`<br>`--evidence-ref <path>`<br>`--alert`<br>`--host <id>`<br>`--no-nudge`<br>`--full`<br>`--json` |
 | `rig queue whoami` | — | `--session <session>`<br>`--recent-limit <n>`<br>`--work-candidates`<br>`--json` |
 | `rig queue fallback <qitemId>` | — | `--destination <session>` **required**<br>`--reason <text>`<br>`--json` |
-| `rig queue show <qitemId>` | — | `--full`<br>`--json` |
-| `rig queue transitions <qitemId>` | — | `--json`<br>`--full` |
-| `rig queue list` | — | `-a, --all`<br>`-A, --all-rigs`<br>`--full`<br>`--owned`<br>`--mine`<br>`-o <format>`<br>`--destination <session>`<br>`--source <session>`<br>`--state <state>`<br>`--target-repo <name>`<br>`--limit <n>`<br>`--json` |
+| `rig queue show <qitemId>` | — | `--full`<br>`--json`<br>`--host <id>` |
+| `rig queue artifact <qitemId>` | — | `--root <name>`<br>`--json` |
+| `rig queue transitions <qitemId>` | — | `--json`<br>`--full`<br>`--host <id>` |
+| `rig queue list` | — | `-a, --all`<br>`-A, --all-rigs`<br>`--full`<br>`--owned`<br>`--mine`<br>`-o <format>`<br>`--destination <session>`<br>`--source <session>`<br>`--state <state>`<br>`--target-repo <name>`<br>`--limit <n>`<br>`--json`<br>`--host <id>` |
 | `rig queue overdue` | — | `--rig <name>`<br>`-A, --all-rigs`<br>`--full`<br>`--limit <n>`<br>`--json` |
 | `rig queue undelivered` | — | `--rig <name>`<br>`-A, --all-rigs`<br>`--full`<br>`--limit <n>`<br>`--json` |
 | `rig queue inbox-drop <destinationSession>` | — | `--sender <session>`<br>`--body <text>`<br>`--body-file <path>`<br>`--tags <tags>`<br>`--urgency <urgency>`<br>`--audit <pointer>`<br>`--id <inboxId>`<br>`--json` |
@@ -481,6 +511,7 @@ do not establish connectivity to a live peer.
 | `rig queue inbox-pending <destinationSession>` | — | `--json` |
 | `rig queue outbox-record` | — | `--sender <session>`<br>`--destination <session>` **required**<br>`--body <text>`<br>`--body-file <path>`<br>`--tags <tags>`<br>`--urgency <urgency>`<br>`--audit <pointer>`<br>`--id <outboxId>`<br>`--json` |
 | `rig queue outbox-list <senderSession>` | — | `--limit <n>`<br>`--json` |
+| `rig queue erase` | — | `--members <file>`<br>`--rig <id>`<br>`--ids <file>`<br>`--action <action>` (`delete`, `scrub`)<br>`--routine`<br>`--apply`<br>`--actor <session>`<br>`--json` |
 
 ### slack
 
@@ -540,13 +571,16 @@ do not establish connectivity to a live peer.
 | Invocation | Aliases | Declared options |
 |---|---|---|
 | `rig terminal` | — | — |
-| `rig terminal open <view>` | — | `--provider <name>`<br>`--session <name>`<br>`--window`<br>`--expected-plan <id>`<br>`--json` |
+| `rig terminal open <view>` | — | `--provider <name>`<br>`--session <name>`<br>`--window`<br>`--app-home <dir>`<br>`--expected-plan <id>`<br>`--follow`<br>`--no-follow`<br>`--max-tiles <count>`<br>`--pin <seats...>`<br>`--unpin <seats...>`<br>`--suppress <seats...>`<br>`--unsuppress <seats...>`<br>`--suppression-ms <ms>`<br>`--grace-ms <ms>`<br>`--focus-retention-ms <ms>`<br>`--json` |
 | `rig terminal views` | — | `--json` |
-| `rig terminal status` | — | `--provider <name>`<br>`--session <name>`<br>`--json` |
+| `rig terminal close <view>` | — | `--app-home <dir>`<br>`--json` |
+| `rig terminal status` | — | `--provider <name>`<br>`--session <name>`<br>`--app-home <dir>`<br>`--json` |
 | `rig terminal attach <session>` | — | `--read-only` |
 
 See [named terminal sessions](../reference/terminal-sessions.md) to select a
 Herdr session per request, including session-bound previews and window reuse.
+herdr-app views (`--provider herdr-app`, plain `rig terminal open <view>` inside an answering herdr-app terminal versus `--window` for a desktop window, `close`, `--app-home`) are in
+[herdr-app](../reference/herdr-app.md).
 
 ### watchdog
 
@@ -576,9 +610,10 @@ Herdr session per request, including session-bound previews and window reuse.
 | `rig workflow specs` | — | `--json` |
 | `rig workflow guidance <instanceId>` | — | `--packet <id>`<br>`--component <id>`<br>`--full`<br>`--json` |
 | `rig workflow show <instanceId>` | — | `--json` |
-| `rig workflow trace <instanceId>` | — | `--json` |
+| `rig workflow trace <instanceId>` | — | `--full`<br>`--json` |
 | `rig workflow continue <instanceId>` | — | `--json` |
 | `rig workflow run <specPath>` | — | `--root-objective <text>` **required**<br>`--created-by <session>` **required**<br>`--entry-owner <session>`<br>`--rig <name>`<br>`--project <id>`<br>`--mission <id>`<br>`--slice <id>`<br>`--tags <tags>`<br>`--json` |
+| `rig workflow unregister [spec]` | — | `--prune-missing`<br>`--json` |
 | `rig workflow watch <instanceId>` | — | `--json` |
 | `rig workflow route <instanceId>` | — | `--to <session>` **required**<br>`--actor-session <session>` (only when the caller's session can't be derived)<br>`--packet <qitem-id>`<br>`--reason <text>`<br>`--json` |
 | `rig workflow resume <instanceId>` | — | `--actor-session <session>` (only when the caller's session can't be derived)<br>`--occurrence <failed-qitem-id>`<br>`--decision <text>`<br>`--json` |
@@ -644,6 +679,9 @@ Only Claude and Codex tenures can be woken.
 | Invocation | Aliases | Declared options |
 |---|---|---|
 | `rig context` | — | — |
+| `rig context reconstruction` | — | — |
+| `rig context reconstruction inspect` | — | `--store <absolute-path>`<br>`--store-map <file>`<br>`--project <id>`<br>`--mission <id>`<br>`--slice <id>`<br>`--cwd <path>`<br>`--item <id>`<br>`--category <name>`<br>`--tag <name>`<br>`--check-sources`<br>`--max-bytes <bytes>`<br>`--json` |
+| `rig context reconstruction view` | — | `--store <absolute-path>`<br>`--store-map <file>`<br>`--project <id>`<br>`--mission <id>`<br>`--slice <id>`<br>`--cwd <path>`<br>`--item <id>`<br>`--category <name>`<br>`--tag <name>`<br>`--check-sources`<br>`--max-bytes <bytes>`<br>`--json` |
 | `rig context work-install` | — | `--project <id>`<br>`--mission <id>`<br>`--slice <id>`<br>`--deliver`<br>`--runtime <runtime>`<br>`--cwd <path>`<br>`--topology <ids>`<br>`--apply-skills`<br>`--json` |
 | `rig context trace` | — | `--rig <rig>` **required**<br>`--pod <pod>`<br>`--seat <seat>`<br>`--name <file>` **required**<br>`--json` |
 | `rig context compose` | — | `--out <ref>` **required**<br>`--from <files...>` **required** |

docs/as-built/test-layers.md

@@ -8,7 +8,7 @@ applies-when: |
   read it before claiming a stub-agent scenario covers a behaviour.
 siblings: [arteries.md, README.md, codemap.md]
 last-verified-against-source: e8f0ab340db773392ec8be75b072d1c0f3068a50
-last-updated: 2026-10-08
+last-updated: 2026-11-09
 ---
 
 # Test layers: what to run before you push, and what each layer proves
@@ -448,9 +448,12 @@ not as current status.
 
 ## Help wanted: command families without a behavioural scenario
 
-> This list may be stale. It was taken from the `rig --help` tree of 0.6.3: 85 visible
-> top-level families. At `e8f0ab34`, `packages/cli/src/index.ts` registers 87 top-level
-> commands; `roster` and `telemetry` were added since and have no row yet. Before picking one up, check `packages/test-system/scenarios/`, the cases in
+> This list may be stale. The tables below cover the families that have a proposed first
+> check: at `b29a70f14`, the Commander tree returned by `createProgram()`
+> (`packages/cli/src/index.ts`) registers 92 top-level families, 87 of which have a row
+> below (7 of them marked partial); `roster` and `telemetry` now have rows too. These 5
+> families still have no row: `acp`, `community-harness`, `edge`, `environment`, `exec`.
+> Before picking one up, check `packages/test-system/scenarios/`, the cases in
 > `scripts/run-pr-scenarios.sh`, and the current `rig --help`.
 > One scenario doesn't cover a family's every option, sequence or platform. A `--help`
 > check only proves the command is discoverable. Existing unit tests are not counted here.
@@ -522,6 +525,8 @@ Before you start, two honest constraints:
 | reads | `status` | Two rigs in different states. Status matches their exact identities and transitions. |
 | reads | `ps` | Two rigs with overlapping seat names. Rig and node filters return exactly the intended members. |
 | reads | `usage` | Deterministic stub telemetry. top and series respect seat and time bounds. Missing data isn't reported as zero. |
+| reads | `roster` | A scratch workspace with two rigs and a rosters folder: `rig roster find` returns the expected specialists, and an empty folder reports none without an error. |
+| reads | `telemetry` | Deterministic retained-metadata fixtures: one bounded page returns exactly the recorded rows, and a history gap is named rather than skipped silently. |
 | reads | `health` | Known fresh, stale and cleared findings. Scope, explain and diagnosis values match the fixture. |
 | continuity | `snapshot` | Snapshot two seats. Rig, membership and snapshot bytes survive a daemon restart. |
 | continuity | `restore` | Restore after a daemon and tmux reset. Seat identity and nonce answers survive. |

docs/as-built/ui/shell-and-routing.md

@@ -181,7 +181,11 @@ UI event consumption is centralized through one shared `/api/events`
 connection (`lib/topology-events.ts`) so related surfaces avoid duplicate
 SSE connections. Current consumers: For You feed hydration, topology
 activity rings + hot-potato movement, the activity feed/system surfaces,
-and rig event surfaces. `useGlobalEvents` reports whether that connection
+and rig event surfaces. That connection carries stored events only:
+`view.changed` is delivered on the live view streams (`/api/views/sse`) and
+is not stored, so it no longer appears in the activity feed, as a For You
+observation card or in the telemetry export (entries stored by older
+versions still replay until an events retention, a separate change, removes them). `useGlobalEvents` reports whether that connection
 is up and refetches review, slice and mission data on reconnect and on
 `proof.*` events. Reduced-motion preference is honoured for pulse/packet
 animation. (Daemon SSE surfaces: `/api/events`, `/api/stream/watch`,

docs/reference/acp-agents.md

@@ -13,7 +13,8 @@ The row's notes get, in order:
 - **progress,** about once a minute (tool calls so far, output so far); the full transcript is in the run's log file, named on the row;
 - **each permission request and its answer** (below);
 - **the outcome:** `completed`, `stopped` (for example `max_tokens`), `cancelled`, `failed` or `unknown`, with the
-  agent's output;
+  agent's output and the tokens and cost it reported, each labelled with where it came from (unknown when it reported
+  none);
 - **what happened to the agent's processes** after the run ended.
 
 The outcome is the agent's report, not a judgment of the work. The row stays open and yours: read the result, then
@@ -62,6 +63,78 @@ shared one running; stop your own with its own command.
 - `failed`: the agent returned an error. The row keeps its original code, message and data.
 - `cancelled`, `stopped`: as above.
 
+## Claude Code and Codex, through their ACP adapters
+
+Both have maintained ACP adapters. What follows was checked on 2026-11-06 with OpenRig 0.7.8: one small file task per
+case, with `claude-haiku-5-5` and with `gpt-6.1-sol` at low effort. Other models weren't checked.
+
+Install the adapters once; running them through `npx` adds about 10 seconds a turn:
+
+```sh
+npm install --prefix <dir> @agentclientprotocol/[email protected] @agentclientprotocol/[email protected]
+export PATH="<dir>/node_modules/.bin:$PATH"
+
+rig acp run <row> --cwd <work folder> -- env CLAUDE_CODE_EXECUTABLE=<your claude> ANTHROPIC_MODEL=<model> claude-agent-acp
+rig acp run <row> --cwd <work folder> -- codex-acp
+```
+
+**Which versions run.** The binding note on the row names the adapter and its version.
+- `claude-agent-acp` 0.101.0 runs Claude Code through the Claude Agent SDK 0.3.318. With `CLAUDE_CODE_EXECUTABLE` set,
+  it runs that `claude`, with that `claude`'s own login; the check ran Claude Code 2.1.310 this way. Without it, an npm
+  install runs the Claude binary from the SDK's platform package (Claude Code 2.1.318, by the SDK's package metadata),
+  which wasn't checked. Installing with `--omit=optional` leaves that binary out, and then, without
+  `CLAUDE_CODE_EXECUTABLE`, the adapter refuses to start; it never falls back to a `claude` on PATH. The adapter also
+  loads your Claude configuration, skills included.
+- `codex-acp` 2.5.0 depends on `@openai/codex` `^0.165.0`, and 0.165.0 was the only release in that range when this
+  was checked. A later install can resolve a newer 0.165 release; `npm ls @openai/codex` in `<dir>` shows yours. It uses
+  the machine's existing Codex login.
+- `@zed-industries/claude-code-acp` and `@zed-industries/codex-acp` are deprecated on npm in favour of these. The zed
+  `codex-acp` 0.16.0's Codex was too old for the model in the check.
+
+**Permissions.** Each agent keeps its own permission mode, and OpenRig adds no restriction. In the check, neither
+asked anything for file writes or shell commands: Claude started in `auto` mode and Codex in `workspace-write`. To
+have the agent ask, so that each request becomes a permission row:
+- **Claude:** give it a dedicated `CLAUDE_CONFIG_DIR` whose settings set `permissions.defaultMode` to `"default"`.
+- **Codex:** set `INITIAL_AGENT_MODE=read-only` in its environment. Approval and sandbox keys in `CODEX_CONFIG` didn't
+  make it ask.
+
+These are the settings the check used to make the agents ask. Whether to use them is the caller's choice; they're not a
+recommended default. Send permission rows to a seat that can be woken: a row to a `terminal` seat sets off the stuck
+sweep.
+
+**Answers differ between the two.**
+- **Claude** offers `allow-once`, `allow-with-updates` (allow always) and `reject`. `reject` refuses that one action,
+  and the turn carries on.
+- **Codex** offers `allow_once`, `accept_execpolicy_amendment` (allow always) and `cancel`. Its only refusal, `cancel`,
+  ends the whole turn, and the outcome is `cancelled`.
+
+**Continuing, and a session that doesn't exist.**
+- **Continuing:** `--load-session` continued an earlier session on a second row with both adapters.
+- **A made-up session id:** both refused it and started nothing, but they report it differently. Claude returns
+  `-32002 Resource not found`, and Codex returns `-32603 Internal error` with "no rollout found" in its data. The row
+  shows `failed` with the agent's own error. Neither code by itself means a missing session, so read the message.
+
+**A completed turn isn't finished work.** In the check, Claude was asked to run a command and wait for it. It moved the
+command into a background task, said it would report when it finished, and ended the turn `completed`. The task died
+with the adapter, and its file never appeared. Check the work itself before you close the row. Beyond this one case,
+long and background work weren't checked.
+
+**Cancel.** `rig acp cancel` was acknowledged in 20-30 ms for both, sent while the agent waited on a permission
+answer; a cancel mid-generation or mid-tool wasn't timed. Straight after, Claude's run reported `incomplete`
+with one process still running, which was gone a few seconds later; Codex's reported `settled`. An acknowledgement
+isn't a finished shutdown, and it says nothing about further spend.
+
+**Usage and cost.** Claude's adapter reports token usage and a cost in USD; Codex's reports tokens only. Both are in
+the run's log file, not on the row.
+
+**Not checked:**
+- long or multi-tool tasks, and background work;
+- concurrent runs;
+- `rig acp stop` with a real stop command;
+- an adapter that crashes mid-turn;
+- continuing after a daemon restart;
+- MCP servers.
+
 ## Notes for specific agents
 
 - **OpenClaw:** `openclaw acp` is a bridge to a running OpenClaw Gateway, so start the Gateway first, then run, for
@@ -70,8 +143,109 @@ shared one running; stop your own with its own command.
 - **Continuing a session:** `--load-session <id>` continues an earlier session when the agent advertises loading. If it
   doesn't, or the load fails, the run reports that and starts nothing new.
 
+## Running it in the background: `rig exec`
+
+`rig exec` runs the same single turn on a named harness, `claude` or `codex`, through that harness's ACP adapter, in a
+connector that outlives the command:
+
+```sh
+rig exec harnesses                               # the adapter and native versions installed here, and how each differs
+rig exec run <row> --harness claude --plan       # what would run, with versions; starts nothing
+rig exec run <row> --harness claude [--wait]     # start it; --wait waits here for the outcome
+rig exec status <runId|row>                      # state, outcome, usage and cost, and what still runs
+rig exec cancel <runId>
+```
+
+- **The connector owns the run.** Closing the terminal, or leaving `--wait` with Ctrl-C, doesn't stop it. Its record
+  in `$OPENRIG_HOME/acp/runs/` holds its process and its state: `requested`, `started`, `finished` or `interrupted`.
+  The row gets the same notes as with `rig acp run`.
+- **One run per row.** Running it again returns the row's run, still going or ended, instead of sending the task
+  twice. `--again` starts another turn once that run has ended.
+- **Interrupted.** A run whose connector died without recording an outcome reads `interrupted`, outcome unknown, with
+  one note on the row, the next time `status`, `run` or `--wait` looks at it. Whether any of the agent's processes
+  remain is said there too.
+- **Missing.** With `--load-session`, a session reads `missing` only on the adapter's own evidence: for Claude,
+  `-32002` naming that session; for Codex, its "no rollout found for thread id" error for that session. Any other error
+  reads `failed`, with the agent's code and message, and no new session is started either way.
+- **The outcome is the agent's report.** `finished` with `completed` means the turn ended; accepting the work stays
+  with you, on the row. `--wait` exits 0 for `completed`, 2 for any other outcome, and 3 when you stop waiting.
+  Without `--wait`, a run that has already ended exits the same way, and one still going exits 0. Either way the code
+  follows the row's run, including one another caller started.
+- **Permissions:** the agent keeps its own mode, and `rig exec` adds no policy; your environment passes through.
+  `rig exec harnesses` lists what each harness reads.
+
+`rig acp run` is unchanged: it stays in your terminal for the length of the turn.
+
 ## Limits
 
 - One run is one turn for one row.
 - The client offers the agent no file-system or terminal methods; the agent works in its own environment.
-- Spend is whatever the agent's provider charges; OpenRig doesn't measure it.
+- Spend is whatever the agent's provider charges. OpenRig doesn't measure it; it records only what the agent reports.
+
+## OpenRig as an agent in ACP apps
+
+The other direction: `rig acp serve`. An app that speaks ACP and can launch a custom agent command can offer your OpenRig team as its agent. The app starts
+`rig acp serve` and talks to it on stdin and stdout:
+
+```sh
+rig acp serve --to <member@rig>
+```
+
+Without `--to`, the receiver is this machine's `acp.to` setting: `rig config set acp.to <member@rig>`, set once. An
+app that starts the agent from the ACP registry passes no arguments, so that's how it reaches your team.
+
+- **A prompt** becomes a queue row from you to `--to`, the seat that takes your team's new work. The row's body is your
+  prompt, plus a line saying which app sent it and that its closing resolution is your answer.
+- **Progress:** each row the work reaches shows in the app as a step (`Sent to …`, then `Handed to …` at each handoff),
+  with the row's notes as they're written, each with the seat that wrote it. The step's title also shows its latest
+  note, for apps that show steps collapsed.
+- **The reply** is the closing resolution, with who closed the row and how (`done`, `done (canceled)`, `denied`, …). The
+  turn ending says only that the turn ended: read the reply for how the work ended.
+- **A follow-up** in the same session goes to `--to` as a new row whose body names the earlier rows and their states,
+  and the app first shows any earlier row that's still open.
+- **A question the team asks you** while a turn is open comes to the app with the ask's own choices, when its row is
+  tagged for that chat (`acp-session:<id>`; the task's row tells the team how). An app that takes forms (it
+  advertises ACP elicitation's form mode, as T3 Code does) gets a one-choice form with each choice by its label; any
+  other app gets a permission request with one option per choice, as Zed shows them. An untagged
+  question from a seat the work reached is named in the chat with where to answer it, because it could be another
+  chat's. Your choice is recorded as your answer, as a click in Slack would be, and the seat that asked gets it.
+  Dismissing it, or cancelling the turn, records nothing: the question stays open in the TUI and Slack, and no further
+  question opens after a cancel. A question that ends unanswered here (dismissed, cancelled, closed with the work, or
+  one the app couldn't show) is named in the chat as still open, as well as on its card. A question that wants an
+  answer in words is shown with where to answer it.
+- **Sessions are kept in the queue:** the app can list your earlier sessions (`session/list`) and reopen one
+  (`session/load`). Reopening replays each prompt, its work, the questions the team asked you during it with your
+  answers, and the reply, as they stand now. Nothing is asked again on reopening: a question still open comes up in
+  your next prompt's turn, and a new prompt continues the session.
+- **Cancel** stops `rig acp serve` waiting. The team's work carries on, and the reply says so.
+- **Work handed to a seat on another machine** can't be followed from here; the reply says where it went.
+
+**Who you are:** `--as <address>`, else `OPENRIG_SESSION_NAME`, else this machine's default person: the only
+registered person, or the one chosen with `rig gateway default <id>` when there are several (the address is
+`<id>@external`). The app's sign-in step, "OpenRig on this machine", checks
+that the daemon answers and that you resolve to an address; until then, signing in and new sessions fail with the
+reason and the fix. The agent talks to the same daemon as `rig` on this computer: the local one, so start it first
+(`rig daemon start`), unless `OPENRIG_URL` names another.
+The team's questions reach you only as a registered person: an `<id>@external` address needs
+`rig gateway human add <id> --display-name "<your name>" --delivery-class <A|B|C|D>` first. Until then, the app says
+so at your first prompt, and your work is still sent.
+
+**In Zed**, add the agent to `settings.json`:
+
+```json
+"agent_servers": {
+  "OpenRig": { "type": "custom", "command": "rig", "args": ["acp", "serve", "--to", "lead@my-team"] }
+}
+```
+
+Other apps take the same command and arguments wherever they configure a custom ACP agent.
+
+**Limits:**
+- One receiver per agent command (or the `acp.to` setting); the app has no team picker.
+- Questions show only while a turn is open; one asked while no turn is open waits in your usual places (the TUI, Slack).
+- An app that shows a permission request's options only as fixed buttons, such as Approve and Decline, and doesn't
+  take forms, can't show the choices: answer those questions in the TUI or Slack.
+- Prompts are text. Links and embedded text files are passed on as written; images and audio aren't accepted.
+- Progress is read every 2 seconds while a turn is open (`--poll-seconds`).
+- Sessions are found among the 2,000 newest rows you've sent, and a reopened session's questions among the 2,000 newest
+  sent to you, so a caller with a lot of traffic, such as a busy seat, can't list or fully reopen its oldest sessions.

docs/reference/agent-spec.md

@@ -229,9 +229,10 @@ This gives access to the shared pool: skills such as `systematic-debugging`, `ve
 `development-team` and `test-driven-development`; the `openrig-core` plugin, which carries `openrig-user` and OpenRig's
 other skills; and runtime resources such as `claude-default-settings`, `claude-default-mcp`, `codex-default-config` and
 `claude-activity-hooks`. Agents select the ones they need via profile `uses`, for example
-`plugins: [shared:openrig-core]`. That selection projects the plugin's skills into the seat. For a Claude seat it also loads
-the plugin's hooks, through a `CLAUDE_CODE_PLUGIN_DIRS` the daemon sets for a seat that selected `openrig-core` (see
-`refocus-channel.md`); a seat that didn't select it gets neither.
+`plugins: [shared:openrig-core]`. For a Claude seat selecting the vendored `openrig-core`, the daemon loads the plugin
+itself through `CLAUDE_CODE_PLUGIN_DIRS`, including its skills and hooks, without projecting a second copy of its skills
+(see `refocus-channel.md`). Other selected plugin skills follow the projection rules below. A seat that didn't select
+`openrig-core` does not get its hooks through this selection.
 
 ---
 
@@ -287,7 +288,11 @@ Each array contains resource IDs. These can be:
 - **Qualified** (`shared:systematic-debugging`) — resolves against a specific imported spec's resources
 
 The `uses` categories are: `skills`, `guidance`, `subagents`, `plugins`, `runtime_resources`. A `uses.hooks` key, even an
-empty list, is refused: hooks ship inside plugins.
+empty list, is refused: hooks ship inside plugins. Any other key selects nothing and gets a validation advisory, shown at
+launch and by `rig agent validate`, so a typo such as `skils` doesn't go unnoticed. A single value written in place of a
+category's list (`skills: vault-user`) is read as a one-item list, with an advisory suggesting list syntax, and resolves
+exactly like that list: when the profile is selected, a resource that isn't available fails that member's resolution with
+"not found in resource pool".
 
 **Skills resolve more widely than other resources.** For Claude Code and Codex seats, a skill ID that no declared
 resource provides is looked up at launch in the project's runtime skill folder in the seat's cwd (`<cwd>/.claude/skills` or
@@ -370,10 +375,10 @@ version.
 
 | Category | Fields | Description |
 |----------|--------|-------------|
-| `skills` | `id`, `path` | Skill directories containing a SKILL.md. Projected by the runtime adapter. Claude and Codex seats get them in the seat's own storage, `<OpenRig home>/state/seats/<rig id>/<logical seat>/<runtime>/`, outside the project: Claude to `plugin/skills/<id>` in the seat's generated plugin, where they show as `openrig:<id>`, and Codex to `skills/<id>` beside the seat's skill catalog. A Claude or Codex seat without that storage gets `<cwd>/.claude/skills/<id>` or `<cwd>/.agents/skills/<id>`. Pi and OMP project to the seat's own agent directory under OpenRig's state. |
+| `skills` | `id`, `path`, `on_restore`* | Skill directories containing a SKILL.md (see "Skill Resources"). Projected by the runtime adapter. Claude and Codex seats get them in the seat's own storage, `<OpenRig home>/state/seats/<rig id>/<logical seat>/<runtime>/`, outside the project: Claude to `plugin/skills/<id>` in the seat's generated plugin, where they show as `openrig:<id>`, and Codex to `skills/<id>` beside the seat's skill catalog. A Claude or Codex seat without that storage gets `<cwd>/.claude/skills/<id>` or `<cwd>/.agents/skills/<id>`. Pi and OMP project to the seat's own agent directory under OpenRig's state. |
 | `guidance` | `id`, `path`, `target`*, `merge`* | Guidance files, merged as a managed block (see "Guidance Resources"). |
 | `subagents` | `id`, `path` | Subagent definitions. Claude copies them to `<cwd>/.claude/agents/`, Codex to `<cwd>/.agents/`; Pi and OMP don't project them. |
-| `plugins` | `id`, `source`, `plugin_type`* | Local plugin directories; selected with `uses.plugins`. `source.kind` must be `local`. `plugin_type` is `claude`, `codex` or `auto` (default): Claude projects a plugin to `<cwd>/.claude/plugins/<id>` when it is `claude`, or `auto` with a `.claude-plugin/plugin.json`; Codex projects to `<cwd>/.codex/plugins/<id>` when it is `codex`, or `auto` with a `.codex-plugin/plugin.json`. Pi and OMP don't project plugins. Neither Claude Code nor Codex reads skills from that plugin folder, so when a seat is instantiated a selected plugin's skills are also projected into the seat's skill folder under their plain names: the seat's own storage (see `skills`), or `.claude/skills/<name>` or `.agents/skills/<name>` for a seat without it. A folder already there with the same name, ignoring letter case, that OpenRig doesn't own is left alone; unless it already holds the same bytes, launch warns `plugin_skill_kept`. A skill the seat already selects by that name (from its profile, the System World or the project) keeps that source, an edited OpenRig-owned copy is kept with `plugin_skill_kept`, a retained plugin copy whose remembered source can no longer be projected is also kept with `plugin_skill_kept` (a seat selecting that plugin refreshes it at its next launch), and if the plugin skills can't be projected the seat launches without them and warns `plugin_skills_not_projected`. Hooks ship inside plugins. |
+| `plugins` | `id`, `source`, `plugin_type`* | Local plugin directories; selected with `uses.plugins`. `source.kind` must be `local`. `plugin_type` is `claude`, `codex` or `auto` (default): Claude projects a plugin to `<cwd>/.claude/plugins/<id>` when it is `claude`, or `auto` with a `.claude-plugin/plugin.json`; Codex projects to `<cwd>/.codex/plugins/<id>` when it is `codex`, or `auto` with a `.codex-plugin/plugin.json`. Pi and OMP don't project plugins. Claude Code and Codex do not read skills from that projected plugin folder, so selected plugin skills are also projected into the seat's skill folder under their plain names: the seat's own storage (see `skills`), or `.claude/skills/<name>` or `.agents/skills/<name>` for a seat without it. The exception is the vendored `openrig-core` selected by a Claude seat: Claude loads that plugin directly, skills included, so its skills are not projected a second time. A folder already there with the same name, ignoring letter case, that OpenRig doesn't own is left alone; unless it already holds the same bytes, launch warns `plugin_skill_kept`. A skill the seat already selects by that name (from its profile, the System World or the project) keeps that source, an edited OpenRig-owned copy is kept with `plugin_skill_kept`, a retained plugin copy whose remembered source can no longer be projected is also kept with `plugin_skill_kept` (a seat selecting that plugin refreshes it at its next launch), and if the plugin skills can't be projected the seat launches without them and warns `plugin_skills_not_projected`. Hooks ship inside plugins. |
 | `runtime_resources` | `id`, `path`, `runtime`, `type` | Runtime-specific resources. `runtime` and `type` are required. An entry whose `runtime` isn't the seat's runtime is skipped. Pi and OMP project none. |
 
 *Fields marked with `*` are optional.
@@ -419,6 +424,16 @@ Where a fragment's table is one the user already declares, the user's table
 wins: the managed table is dropped, their values are never merged, rewritten or
 overwritten, and the rest of the fragment still applies.
 
+**A fragment can't install a named Codex profile.** Codex versions with named
+profiles read each one from its own file, `<CODEX_HOME>/<name>.config.toml`,
+with top-level keys, and won't load a profile while `config.toml` still holds a
+legacy `[profiles.<name>]` table. A fragment only adds tables to `config.toml`,
+so neither form works there. Create the profile file yourself and select it with
+the member's `codex_config_profile`; see "Codex: select sandbox and approvals
+together" in [getting-started.md](getting-started.md#codex-select-sandbox-and-approvals-together).
+Before a launch or restore, OpenRig checks that a selected profile loads, and for
+a legacy table it names the file to move the settings into.
+
 ### Resource Path Rules
 
 - All resource paths must be safe relative paths (no `..` traversal, no absolute paths), except a plugin's
@@ -427,6 +442,55 @@ overwritten, and the rest of the fragment still applies.
 - Resource IDs must be unique within their category
 - Resource IDs are the identifiers used in `uses` references
 
+### Skill Resources
+
+```yaml
+skills:
+  - id: house-style
+    path: skills/house-style
+    on_restore: follow_source   # optional — keep_saved by default
+```
+
+| Field | Type | Required | Default | Description |
+|-------|------|----------|---------|-------------|
+| `id` | string | yes | — | Resource identifier, used in `uses.skills`. |
+| `path` | string | yes | — | Relative path to the skill's folder, which holds its `SKILL.md`. |
+| `on_restore` | string | no | `keep_saved` | What a restore (`rig up <rig>`) does with a Claude or Codex seat's copy of this skill in the seat's own storage: `keep_saved` or `follow_source`. |
+
+An unknown `on_restore` value produces an advisory and uses `keep_saved`; it does not invalidate the spec.
+
+A restore keeps the skills a seat was launched with. With `keep_saved`, it puts back the saved copy and never takes
+today's bytes from the source. With `follow_source`, before the seat's session starts, the restore reads the seat's agent
+spec as it is now, using the saved launch recipe and rig folder, and:
+
+- replaces the seat's copy with the source folder as it is now, and records it, so `rig down` removes it and the next
+  restore puts it back as usual;
+- copies in a `follow_source` skill the profile now selects that the seat didn't have;
+- removes a skill that followed its source and that the profile no longer selects.
+
+A copy edited in the seat (changed bytes or recorded file permissions, or an added file), or a folder with the same id that OpenRig didn't write, is kept as
+it is and named in the restore's warnings. If the agent spec can't be read or resolved, the seat keeps its saved skills
+and the restore says why; a source folder without a `SKILL.md` keeps that skill's saved copy. A skill changed to
+`keep_saved` stops following at the next restore, and its copy stays. Older seat records without file permissions
+retain their byte-only ownership checks until a new projection or refresh records permissions.
+
+Source file symlinks are dereferenced, including links to files outside the skill folder, just as during Claude and
+Codex launch projection. The seat receives an ordinary file containing the target's bytes. A broken link or a link to
+a directory makes the source unreadable and keeps the saved copy; links in the seat's destination remain protected.
+
+Only a restore does this; a running seat isn't changed, and neither is a Codex seat restarted in place. A Claude seat
+loads its seat plugin when its session starts, so it sees the new copies. A Codex seat with following skills receives
+its rebuilt catalog after its resumed session is ready. The catalog lists the skills actually present, replaces its
+earlier OpenRig skill list, and asks it to read current skill files before using them. This includes newly selected
+skills and the removal of the last following skill. An edited or linked catalog is kept and not sent, with a warning;
+a delivery failure is reported instead of claiming a complete restore. Neither runtime erases skill instructions
+already read into the retained conversation.
+
+Following covers the whole selected skill folder, including scripts and references. It does not install external
+tools, change MCP configuration, or update bundled plugins. The saved launch recipe supplies the rig directory,
+member and profile; today's agent spec supplies the skill selection. Today's rig YAML and other agent settings are
+not applied by this refresh.
+
 ### Guidance Resources
 
 ```yaml
@@ -440,7 +504,7 @@ guidance:
 |-------|------|----------|---------|-------------|
 | `id` | string | yes | — | Resource identifier. A `rig-role` guidance resource is never merged; deliver a role with a `send_text` startup file. |
 | `path` | string | yes | — | Relative path to the guidance file. |
-| `target` | string | no | — | Parsed but not used: the runtime decides the file. Claude merges into the rig's `managed_blocks` file (`CLAUDE.md` by default, or `CLAUDE.local.md`); Codex, Pi and OMP merge into `<cwd>/AGENTS.md`. |
+| `target` | string | no | — | Parsed but not used: the runtime decides the file. Claude merges into the rig's `managed_blocks` file (`CLAUDE.md` by default, or `CLAUDE.local.md`); Codex, Pi and OMP merge into `<cwd>/AGENTS.md`, except that a fresh Pi seat normally uses its own agent folder's `AGENTS.md` and a fresh Codex 0.164.0+ seat normally gets launch developer instructions; the conditions are in [Codex guidance at launch](agent-startup-guide.md#codex-guidance-at-launch) and [Runtime Differences](agent-startup-guide.md#runtime-differences). |
 | `merge` | string | no | `managed_block` | Only `managed_block` is applied. Any other value, such as `append`, merges nothing. |
 
 ---
@@ -482,6 +546,7 @@ restore behavior, and the distinction between readiness and verified proof.
 9. All resource paths must be safe relative paths (a plugin's `source.path` excepted).
 10. Resource IDs must be unique within their category, and each `resources.<category>` must be an array.
 11. `runtime_resources` entries require `runtime` and `type` fields.
+11a. An unknown skill `on_restore` value produces an advisory naming the known values and falls back to `keep_saved`.
 12a. `resources.hooks` and `profiles.<name>.uses.hooks` are refused: "removed in plugin-primitive (Phase 3a)". Declare
    a plugin instead.
 12b. A plugin needs a unique non-empty `id`, a `source` object with `kind: local` and a non-empty `path`, and a

docs/reference/agent-startup-guide.md

@@ -63,6 +63,33 @@ delivery has been attempted, it cannot be replayed by `continue`, including
 after a daemon restart: inspect `rig seat status <seat>` and the native pane
 before deciding how to recover an uncertain delivery.
 
+## After a restored seat's prompt
+
+A restore that stops at a prompt, such as Claude's resume selection, records the
+seat's restore outcome as `attention_required`. `rig seat continue` doesn't change
+that record: it only delivers pending startup context. While the record stands, it
+alone keeps the seat's lifecycle at `attention_required`, even with startup `ready`:
+`rig ps --nodes` shows LIFECYCLE `att`, `needsAttention` is true, and the seat counts
+in the rig's `attentionCount` and the TUI's attention totals.
+
+After a full restore of a Claude or Codex seat, answering the prompt is enough. When
+the seat's activity next turns to running or idle, or when the whole restore completes
+after the seat already reads running or idle, the daemon checks the live seat: its tmux
+session is alive, and the pane runs the agent resuming the conversation the seat
+records. When the check passes, the restore outcome becomes `operator_recovered`, so
+the restore no longer makes the seat need attention. When it fails, the restore outcome
+stays and the seat still needs attention, which is true when the pane runs another
+conversation. A check that started before a newer restore of the seat doesn't settle
+the newer outcome.
+
+`rig seat clear-attention <seat>` runs the same check on demand. When it fails, the
+command says which check failed, and `--reason` doesn't bypass it. If
+the seat records the wrong conversation and you know the right one, set it with
+`rig seat set-resume-token <seat> --token-stdin --reason <text>`, then run
+`clear-attention` again. After a restore of only some seats, or a community
+harness's assisted startup, `--reason "<what you saw>"` can acknowledge the
+outcome instead.
+
 ## Two Categories of Startup
 
 Everything an agent receives at boot time falls into one of two categories:
@@ -367,7 +394,15 @@ Each startup file and action specifies when it applies:
 - `restore` — delivered when a seat is restored without its earlier conversation: a restore whose resume falls back to a fresh conversation, or a retried start
 - Default: `[fresh_start, restore]` (both)
 
-**An exact resume delivers neither.** When a restore reopens the seat's saved conversation (`rig up <rig> --existing` with a resume token), OpenRig types no startup actions and delivers no startup files, `restore` ones included: the conversation already has them, and typing them again would add turns to it. The one exception: a managed guidance block missing from the guidance file is put back before the harness starts.
+**An exact resume delivers neither.** When a restore reopens the seat's saved conversation (`rig up <rig> --existing` with a resume token), OpenRig types no startup actions and delivers no startup files, `restore` ones included: the conversation already has them, and typing them again would add turns to it. The one exception: the guidance file's managed blocks are brought up to date before the harness starts. A missing block is put back, and a stale one is refreshed (see below).
+
+**A seat handover does the same put-back.** Before a fresh, fork or rebuild successor starts, the handover puts the seat's missing managed guidance blocks back in its guidance file from the saved startup context. That's the rig's Claude file (`CLAUDE.md`, or `CLAUDE.local.md` when the rig uses it) for Claude, and `AGENTS.md` for Codex, Pi and omp. So a seat rebuilt into a new working directory keeps its guidance.
+
+**A block you edited by hand is never overwritten.** OpenRig records what it last wrote in each managed block, so a launch, an exact resume, a handover or a runtime switch can refresh a block whose source file changed since. The record follows the file itself, so two links to one file share it. A block you've edited since OpenRig wrote it is left as you wrote it. The start's result names it (`guidance_block_edited`), and the seat is told in one line once it's ready, or in a handover's packet. A block OpenRig has no record of writing (any block written before these records were kept) that differs from its source works like this:
+- a launch replaces it, as launches always did, and names it (`guidance_block_refreshed_unrecorded`);
+- an exact resume or a handover leaves it, and names it as possibly stale (`guidance_block_may_be_stale`).
+
+An exact resume doesn't type anything into the resumed conversation. So there, the restore result names the block (`guidanceNotes`), but the seat isn't told.
 
 Use this to avoid re-sending context that a fresh conversation doesn't need twice. For example, a one-time project briefing might only apply on `fresh_start`, while identity grounding should apply on both, so a seat that comes back fresh gets it again.
 
@@ -386,6 +421,7 @@ running seat only through the action in the last column.
 | Reasoning effort | `rig seat set-effort` | `rig seat effort <seat>`: requested, saved, sent at launch, and what the session recorded, which is not verified until it takes a turn | Its next managed continuation; nothing restarts |
 | Permissions | `rig seat set-permissions` (`--mode`, `--rules`, or for Codex `--codex-permissions` or `--codex-profile`), shown by `rig seat status` | What it launched with; a status read doesn't prove the running process enforces the selection | Its next managed launch or resume ([per-seat permission mode](getting-started.md#per-seat-permission-mode)) |
 | Working directory | `rig seat set-cwd`, shown by `rig seat status` as **Working directory for future launches** | The directory it started in | `rig seat launch --fresh --stop` (a blank conversation) or `rig handover --source rebuild` (durable context); [details](#change-an-existing-seats-working-directory) |
+| Launch wrapper (`launch_wrapper`) | Member `launch_wrapper` in `rig.yaml` (`pi`, `codex` and `claude-code` only; `launch_wrapper: []` clears, omitting retains on existing-seat apply; surfaced as `launchWrapper` in `rig ps --nodes --json` when set; no live `rig seat` edit verb) | The literal wrapper argv that prefixes the resolved Codex entry or Claude executable (preserving TTY stdio) or the resolved Pi child executable inside OpenRig's host runner (preserving RPC pipes and runner lifecycle) when that session launched; a configured prefix is not proof of isolation, tooling, memory or governance | Supported spec create/recreate/apply flow, then the seat's next managed launch, resume, restore or handover (fails closed when wrapper support is unavailable; never falls back to bare child launch; [details](rig-spec.md#launch-wrapper-runtime-pi-and-codex)) |
 | Startup files and skills | The spec's startup files, skills and actions | What its conversation read when it started; editing a file afterwards doesn't reach it | A fresh launch or a handover delivers them again; for the running seat, send or walk the change (`rig walk <seat> --through <files>`) |
 | Tools (MCP servers, plugins) | Selected fragments, merged into the harness's own config (`.mcp.json` for Claude, `config.toml` for Codex) | Unknown to OpenRig: registered isn't active, Claude asks before it uses a new server, and no OpenRig command reports which tools a running harness loaded | Check in the harness itself (Claude Code's `/mcp`); a running session may need a restart to pick up newly written config ([runtime config disclosure](#runtime-config-disclosure)) |
 
@@ -560,6 +596,10 @@ The guidance still goes to the repository `AGENTS.md`, as before, when:
 A seat's role still arrives as startup text, and the seat-skills line, after the person's yes, still goes to
 `AGENTS.md`.
 
+A launch that carries developer instructions writes no new blocks to the repository `AGENTS.md`. But blocks an earlier
+handover or resume left there are refreshed from their sources, by the same rule as every other managed block. That
+matters because the launch text asks Codex to follow `AGENTS.md` where the two differ.
+
 **If your shell sets `CODEX_HOME`.** When neither the seat nor the daemon declares a `CODEX_HOME`, a seat's pane keeps
 its own, and OpenRig reads `~/.codex`. A `CODEX_HOME` that only your shell startup sets (in `~/.zshrc`, for example),
 with the daemon started without it, isn't seen. A `developer_instructions` set in that home is then replaced by the
@@ -723,6 +763,29 @@ When authoring startup content, note which instructions are runtime-specific. Fo
 
 ---
 
+## Files OpenRig writes into a project
+
+These are the files OpenRig writes into a seat's working directory, and what `rig down` does with each. Paths are
+relative to that directory.
+
+| Path | When OpenRig writes it | What `rig down` does |
+| --- | --- | --- |
+| Managed blocks in `CLAUDE.md` (or the rig's `managed_blocks` file, such as `CLAUDE.local.md`) or `AGENTS.md` | Each launch merges the seat's guidance blocks. A fresh Codex launch that takes developer instructions puts them in its launch text instead: in `AGENTS.md` it adds only the seat-skills line, after your yes, and refreshes blocks already there. A fresh Pi seat writes its own agent-directory `AGENTS.md` outside the project, except when that folder or file is a link: then its guidance stays in the repository `AGENTS.md`. | Removes OpenRig's blocks, current or legacy, and keeps everything else. When nothing else is left, it deletes the file if the file began with an OpenRig block, or held only blocks in an older format. A block-only file isn't proof that OpenRig created it. The seat-skills line is kept if you edited it. |
+| `.openrig/context-collector.cjs` | Claude seats, at launch. | Removes it while its bytes are still the shipped collector's; otherwise keeps it and says so. |
+| `.claude/settings.local.json` | Claude seats, at launch: OpenRig's status line and its activity hooks are merged in, separately. Your own status line is never replaced, but the activity hooks are still merged beside it. A file that doesn't parse gets neither. | Removes OpenRig's status line and hooks and keeps the file, even if `{}` is all that's left. |
+| `.claude/settings.local.json` fragments and `.mcp.json` | When a profile or spec declares `claude_settings_fragment` or `claude_mcp_fragment`. | Leaves the merged content in place. |
+| `.claude/agents/`, `.claude/plugins/<id>/`, `.claude/extensions/<id>/`; `.agents/<id>.yaml`, `.codex/plugins/<id>/`, `.agents/extensions/<id>/` | When the seat's projection includes subagents, plugins or runtime resources (Claude, then Codex). | Leaves them in place. |
+| `.claude/skills/<id>/`, `.agents/skills/<id>/` | Only for a seat without seat storage. New Claude and Codex seats keep their skills under `$OPENRIG_HOME/state/seats/`, outside the project. | Leaves them in place. |
+| The managed compaction preparation folder described in [Claude compaction preparation directory](#claude-compaction-preparation-directory), with its `.gitignore` of `*` and each attempt's `RESTORE-MAP.md` | Managed Claude compaction, unless a preparation root is configured elsewhere. | Leaves it in place. |
+| `.rigged-checkpoint.md` | A restore that didn't resume a seat's native conversation (by choice or not) and had a checkpoint to hand it. | Leaves it in place. |
+
+`rig down` leaves a file alone while another seat still uses it: a running seat of the same rig, or any seat of
+another rig that isn't archived. That is decided file by file. A guidance file is kept while another seat's
+guidance targets the same file. The Claude settings and collector are kept while another Claude seat works in that
+directory. So another rig's Codex seat keeps `AGENTS.md` without keeping `CLAUDE.md` or the Claude settings. It
+removes `.openrig/` and `.claude/` only when they are empty.
+`rig down --delete` does nothing more to project files. OpenRig doesn't edit your project's own `.gitignore`.
+
 ## Patterns and Anti-Patterns
 
 ### Good Patterns

docs/reference/agent-state-taxonomy.md

@@ -23,7 +23,10 @@ Only the activity axis is served today. The session and resumability values are
 `activity-taxonomy.ts` but no surface reads them: `rig ps --nodes --full` shows the raw session
 STATUS, both node views show a LIFECYCLE of `running`, `detached`, `recoverable` or `attention_required` (`run`, `det`, `rec`,
 `att`), and node inventory serves `occupantLifecycle` (`active`, `retiring`, `retired`,
-`context_walled`, `compacted`, `crashed`, `unknown`).
+`context_walled`, `compacted`, `crashed`, `unknown`). Node inventory's `sessionStatus` is the
+recorded session status, with two observed corrections: a running session whose pane is verified
+to have exited (identity reason `pane_exited`) reads `exited`, with `startupStatus:
+"attention_required"`, and a running session confirmed missing reads `detached`.
 
 The axes never blend. Reachability is strict; revival optimism is a different field
 (omnigent's strict-liveness split). `unknown` is a first-class honest value — the oracle
@@ -59,6 +62,21 @@ the existing live-hook fallback in effect; lifecycle, startup, held and error si
 can independently raise attention. A false attention flag means no known trigger, not
 proof that an unobserved input box is clear.
 
+A handover whose successor started without a required file also raises attention. That's
+a projected guidance file or a guidance startup file marked `required` still missing from
+the guidance file, or, for a fresh or rebuild successor, a required `send_text` startup
+file that's missing. The seat serves `startedWithoutRequired` with each file's path, the
+table's reason column names each file, and the handover's own output says it's degraded.
+The mark belongs to that successor's session, so the seat's next launch or handover
+clears it. Time alone never does. Repairing the file in place doesn't clear it either,
+because no route delivers guidance into a running conversation.
+
+A fresh or rebuild successor's first turn opens with the seat's `send_text` startup files,
+the ones a fresh start delivers, as absolute paths with the role first. A pointer asks the
+successor to read the file; it isn't the pasted text a start sends. A fork successor
+carries its parent's conversation and gets no list. `send_text` startup actions, such as
+an `after_ready` preload, aren't run for a successor.
+
 The compact human table's ATTN column renders `yes`, `no`, or `?` when the flag is
 unavailable. Compact JSON retains `needsAttention`; use `--full` or
 `--fields needsAttention,arbitratedNeedsInput,needsInputEvidence` with `--json` for the
@@ -116,7 +134,10 @@ time-bounded evidence. Rungs, top down, with each rung's retirement condition:
   PER-HARNESS when that harness's hook rung earns trust through a measured agreement
   window in production — never by fixture pass alone (symmetric admission: a fixture
   pass admits a rung to TRIAL, where it is measured but not consulted; production agreement
-  promotes it). Undeclared and partial-coverage rungs are identity-only, which is also where
+  promotes it). For Codex, this fallback also accepts a fresh working status adjacent to
+  the composer, or an advancing counter across two fresh captures when output separates
+  them. This positive evidence expires after five seconds; unrecognized, frozen, old or
+  pre-occupant frames add nothing. It uses the existing pane capture. Undeclared and partial-coverage rungs are identity-only, which is also where
   a degraded rung goes, with one declared exception: for a pi or omp seat or a community
   harness, sampling is positive-only: a changing pane reads working, and a quiet one never reads idle on its
   own, since the pane can be quiet mid-turn.
@@ -185,6 +206,13 @@ LIFECYCLE); see the
 - **Restore outcome:** `resumed`, `fresh-primed`, `awaiting-decision`, `attention_required`,
   `failed` (plus `rebuilt`, `operator_recovered` and the legacy `fresh`). A seat joined with
   `rig attach` as an external CLI reads `left_attached`: restore leaves it as it is.
+  An `attention_required` outcome, or `failed` with a running session, makes the seat's
+  lifecycle `attention_required`, whatever its startup status, so it raises `needsAttention`.
+  After a full restore, the daemon reconciles it to `operator_recovered` when the seat's
+  activity next turns to running or idle (or reads so when the restore completes) and the
+  live seat passes its check, and `rig seat clear-attention <seat>` runs that check on
+  demand; `rig seat continue` doesn't change it
+  ([startup guide](agent-startup-guide.md#after-a-restored-seats-prompt)).
 
 ## Rejected oracles — dated receipts (2026-08-26), so nobody re-derives them expensively
 

docs/reference/author-contracts.md

@@ -16,7 +16,7 @@ This draft was published with **OpenRig 0.7.6** (tag [`v0.7.6`](https://github.c
 | Community launcher presets, RigSpec members and AgentSpec role/startup declarations | Implemented interfaces | They compose configuration. A launcher must implement its declared effects; successful parsing does not prove delivery or native behavior. |
 | `runner: true` with `launch_result: true` | Released in 0.7.6 | The declaration is retained at launch, and Core uses the runner identity/activity path. It does not supply a runner executable. |
 | `passes_model: true` | Released in 0.7.6 | Passes the declared model through `OPENRIG_MODEL`. It does not observe or verify the effective native model. Cursor and Hermes examples leave it off pending native qualification. |
-| The dsh/ACP input, cancellation and answered permission recipe (a person answers, and the default is permit) | OpenRig's side released in 0.7.8; runner side not qualified | OpenRig asks a person for a runner's permission requests and returns the answer. No end-to-end dsh integration is established, and the native verdict is pending. Native reopen capabilities and the tools-off case remain implementer questions. |
+| The dsh/ACP input, cancellation and answered permission recipe (a person answers, and the default is permit) | OpenRig's side released in 0.7.8; runner side shown on two runners, not qualified in general | OpenRig asks a person for a runner's permission requests and returns the answer. In 0.7.8's release testing on a Claude Haiku model, dsh-seat 0.3.0 (dsh 0.2.1-rc.2 over ACP) showed input, activity, cancel, exact resume and shutdown, and the dsh example runner showed a person's allow-once and reject-once; see the [runner contract](harness-runner-contract.md)'s status for what wasn't shown. Native reopen capabilities and the tools-off case remain implementer questions. |
 | Exact native resume via an author launcher | Existing interface; integration-specific evidence | A requested conversation must reopen exactly or fail honestly. One provider/harness result does not establish portability to another. |
 | Environment preparation, projection, artifact export and owned-resource cleanup | OpenShell reference driver released in 0.7.8; generic seam proposed | The platform chapter makes the responsibilities reusable. Core ships one maintained reference driver, OpenShell 0.1.6, as `rig environment`. It doesn't provide a generic platform driver or `DeploySpec` loader. |
 | `HarnessSpec`, `DeploySpec`, generic lifecycle fields and wake-on-demand declarations | Proposed vocabulary or unsupported configuration | They are not new accepted YAML keys. Use existing composition and an explicit author implementation. |

docs/reference/checked-seat-stop.md

@@ -8,4 +8,15 @@ Read `GET /api/seat/stop-precondition/:seatRef` immediately before an operator-c
 
 Pass `expected` unchanged in the body of the existing `POST /api/seat/stop/:seatRef`, alongside `reason` and optional `operator`. The contract marker lets clients detect support. `expected` is optional; a request without it retains the existing stop behavior. A malformed explicit value returns HTTP 400. A null session ID can describe an empty seat, but stopping it returns `no_session`.
 
-Checked stop requires the delivery guard. It refuses a changed node, latest session, binding or occupant before a terminal effect with HTTP 409 `selection_changed`. If the selection changes after client detach, the response reports `partialEffect: true` and no kill is sent. A kill failure or a failure to save stop state after a kill reports `stop_uncertain`; inspect the seat before retrying. Success retains the usual `seat`, `sessionName` and `sessionId` fields.
+Checked stop requires the delivery guard. It refuses a changed node, latest session, binding or occupant before a terminal effect with HTTP 409 `selection_changed`. A kill failure or a failure to save stop state after a kill reports `stop_uncertain`; inspect the seat before retrying. Success retains the usual `seat`, `sessionName` and `sessionId` fields.
+
+A `selection_changed` or `stop_uncertain` response from a checked stop says how far the stop got in `stopPhase`. A refusal before the stop starts carries neither `stopPhase` nor `partialEffect`, and nothing was sent. The stop detaches the seat's clients, checks the selection again, then kills the session:
+
+| `stopPhase` | What was sent to the seat |
+| --- | --- |
+| `none` | Nothing. |
+| `detach_attempted` | The client detach, which may not have finished. No kill. |
+| `detached` | The client detach. No kill. |
+| `kill_attempted` | The kill. |
+
+`partialEffect` is `true` whenever `stopPhase` isn't `none`, so on its own it doesn't say whether the kill went out. `selection_changed` with `stopPhase: "kill_attempted"` is a change seen after the kill, when the stop state is saved: the session was killed, so re-read the seat rather than offering the stop again.

docs/reference/community-harness-examples/dsh/dsh.yaml

@@ -7,11 +7,11 @@ runner: true
 launch_result: true
 first_input: true
 limits:
-  - written for dsh 0.2.1-rc.2 (npm @deepseek-ai/dsh) over ACP protocol 1, with Node 22 or later, and checked against a stand-in agent only
+  - written for dsh 0.2.1-rc.2 (npm @deepseek-ai/dsh) over ACP protocol 1, with Node 22 or later; run with a real dsh on two stacks only, dsh's Claude route with a Claude Haiku model in 0.7.8's release testing (with that release's copy of this runner, for an allow-once and a reject-once), and dsh's pi-ai OpenRouter chat-completions profile with the model deepseek/deepseek-v4-flash on OpenRig 0.7.8 (for queue work, activity through a quiet tool call, an allow-once and a reject-once, and cancel); in both, the permission answers came from a test answerer registered as a person, not a live person; dsh's DeepSeek Messages adapter pointed at OpenRouter failed with HTTP 400 in the second, cause not established; no other dsh route or model is checked
   - each seat gets its own dsh home under $OPENRIG_HOME/state/dsh/<node>/home, holding a patch that turns off dsh's session-log and plugin-list uploads to DeepSeek; give it credentials (.credentials.yaml there) or a provider key in the seat's env; set DSH_SEAT_HOME in the member's env to use a home you prepared, and DSH_SEAT_BIN for another dsh binary (a member's env can't set OPENRIG_* names, and the runner doesn't read dsh's own DSH_HOME)
   - DSH_PERMISSION_MODE defaults to workspace-write; a permission request goes to the default person or the seat's override (rig seat set-permissions-to), is permitted with dsh's allow_once (else allow_always) when no one can be asked or OpenRig can't answer, and is refused and held as needs input only after --deny
   - allow_always creates a standing grant in dsh, which OpenRig doesn't record or revoke; the runner prefers allow_once, and where dsh keeps the grant and for how long isn't checked
-  - resume isn't declared until a real dsh has reopened a conversation and recalled it; the runner reopens by the method dsh advertises and reports missing only on dsh's own not-found error, not when dsh's session list names the conversation; dsh answers -32602 for a missing conversation (and for one active elsewhere), so a missing dsh conversation reads unconfirmed
+  - resume isn't declared; a real dsh has reopened a conversation after down, snapshot and up and recalled it, with resume added to a copy of this preset, but a missing conversation's refusal hasn't been run; the runner reopens by the method dsh advertises and reports missing only on dsh's own not-found error, not when dsh's session list names the conversation; dsh answers -32602 for a missing conversation (and for one active elsewhere), so a missing dsh conversation reads unconfirmed
   - a prompt whose whole text is /abort can't be sent; a Ctrl-C byte inside a paste also aborts; without a TTY, input is read line by line; this isn't a binary-safe channel
   - one prompt at a time; a message that arrives mid-turn waits for the next turn
   - the seat's model isn't passed; dsh uses its own configured model

docs/reference/community-harness-presets.md

@@ -53,6 +53,7 @@ startup_assisted: true           # optional: startup submission needs a person's
 first_input: true                # optional: launcher takes the first turn from OPENRIG_FIRST_INPUT (below)
 # runner: true                  # optional, needs launch_result: the launcher reports the seat's activity (below)
 # passes_model: true            # optional: the launcher takes the seat's declared model from OPENRIG_MODEL (below)
+# launch_wrapper: true          # optional: a runner applies OPENRIG_LAUNCH_WRAPPER to its native child (below)
 ```
 
 - **`command`** is a single program and its arguments. Shell syntax such as `|`, `&&`, `;`, redirections or `$(…)`
@@ -164,6 +165,10 @@ and reports what it's doing, as OpenRig's own Pi runner does. The seat's activit
   unavailable.
 - **Input is as for any seat.** A send pastes the text and presses Enter, so a runner should submit everything up to that
   Enter as one prompt.
+- **`launch_wrapper: true` declares wrapper support and requires `runner: true`.** A member with a wrapper and a runner preset without this flag is refused at preflight and launch (not schema validation):
+  `launch_wrapper unsupported: the community preset "<runtime>" runs a runner without launch_wrapper: true; no agent started. Add launch_wrapper: true to the preset if its runner applies OPENRIG_LAUNCH_WRAPPER, or remove launch_wrapper from this member.` See the
+  [harness and runner contract](harness-runner-contract.md#2-declare-it-what-the-preset-says-fixed-at-launch)
+  for the environment handoff, native-child execution and refusal requirements.
 
 **What the runner posts.** The runner posts each event to `POST /api/activity/hooks` with the local activity hook token,
 one at a time and in order, without making the agent wait. It finds the address and token in `OPENRIG_URL` (or

docs/reference/daemon-http-clients.md

@@ -0,0 +1,210 @@
+# Writing a daemon client
+
+What a program that talks to the OpenRig daemon over HTTP needs and can't read from `rig --help`: how to find the
+daemon, its health route, which bearer token each route family checks, the event streams, and some route
+details. It describes the daemon as it is. It adds no promise beyond what each section says, and route shapes not
+listed here can change.
+
+## Finding the daemon
+
+`rig status --json` is the discovery surface. Its `daemon` object gives `state` and `reachable`, and `pid` and
+`port` when they are known: a stopped daemon's response has no `port`, and a running daemon found through
+`OPENRIG_URL` has no `pid`. `reachable` is `true` when the daemon answered its health probe, `false` when it is
+positively down, and `null` when that isn't confirmed (see `rig status` in
+[the command reference](../as-built/cli-reference.md#status)). It reads the recorded state and probes; it never
+starts a daemon.
+
+The CLI itself finds the daemon in these orders:
+
+- **Data commands** (everything that sends a request): `OPENRIG_URL` (or the legacy `RIGGED_URL`); else
+  `$OPENRIG_HOME/daemon.json` when it holds a valid `pid` and `port`; else the config keys `daemon.host` and
+  `daemon.port` (environment `OPENRIG_HOST` and `OPENRIG_PORT`), which default to `127.0.0.1:7433`.
+- **Status** (`rig status`, `rig daemon status`): `OPENRIG_URL` (or `RIGGED_URL`) first, probed directly; else
+  the state file below.
+- **Start and stop** (`rig daemon start`, `stop`): the state file, `$OPENRIG_HOME/daemon.json`, then the legacy
+  `~/.rigged/daemon.json`.
+
+`daemon.json` holds `pid`, `port`, `host`, `db` and `startedAt`. `rig daemon start` writes a new daemon's record
+only after its health check passes, so while a start is pending, the file holds no record of the new daemon: it
+may be missing, or hold an earlier daemon's stale record. It is the CLI's own state file, not a published
+contract: prefer `rig status --json`.
+
+## `GET /healthz`
+
+The liveness and identity read. It always returns `status: "ok"` and `pid`. Every other field is present only
+in the case named:
+
+| Field | When it is present |
+| --- | --- |
+| `euid` | On platforms with effective user ids. A self-report for the CLI's owner check, not authentication. |
+| `databaseWarnings` | When the daemon opened its database with warnings. |
+| `semver`, `commit`, `dirty`, `builtAt` | On a stamped (released) build; a development run has none. |
+| `selfHostId`, `selfHostIdSource` | After the boot reconcile has set this host's id. |
+| `eventLoop` | When the event-loop monitor runs: `lagMeanMs`, `lagP99Ms`, `utilization` (0 to 1), `lastTickAgeMs` and `healthy`. |
+| `routeTimings` | With `eventLoop`. |
+| `slowOperations` | When the slow-operation recorder runs. |
+| `bind` | The bind plan: `mode` (`explicit` or `default`), `hosts` (the addresses bound), `tailscaleDetected`, and `ignoredRoutingHost` when a routing value was ignored. |
+| `stuckSweep`, `wakeLadder` | The heartbeats of the stuck sweep and the wake ladder, when they run. |
+
+## Bearer tokens
+
+The daemon reads two tokens from its environment at start:
+
+- **`OPENRIG_AUTH_BEARER_TOKEN`** checks Mission Control's write routes (`POST /api/mission-control/action` and
+  `/notifications/test`), the `/api/hosts` writes (`add`, `update`, `remove`, `pair`), and the `/api/rig-mode`
+  binding writes. `rig host pair` hands this token to the paired machine, so a daemon that will be paired with
+  needs it set (see `rig host pair` in [the command reference](../as-built/cli-reference.md)).
+- **`OPENRIG_TERMINAL_BEARER_TOKEN`** checks every `/api/transport` and `/api/compaction` route, `POST
+  /api/terminal/follow` and `/viewer/ensure`, and every `GET /api/terminal/<name>` route, the terminal websocket
+  included. When the daemon is bound to an explicit address (`rig daemon start --host`) that isn't loopback or a
+  tailnet address, and this token isn't set, these routes check `OPENRIG_AUTH_BEARER_TOKEN` instead.
+
+A route whose token isn't set lets every request through. Send a token as `Authorization: Bearer <token>`; the
+`GET /api/terminal/<name>` routes, the websocket included, also take `?token=<token>`. The CLI sends the terminal token from `OPENRIG_TERMINAL_BEARER_TOKEN`,
+or else from the file `$OPENRIG_HOME/terminal-token`, which the person writes; the daemon doesn't create it.
+
+A wrong or missing token gets HTTP 401. The routes that check `OPENRIG_AUTH_BEARER_TOKEN`, and `/api/transport`
+and `/api/compaction`, answer `{ "error": "unauthorized", "message", "what_failed", "why_it_matters", "what_to_do" }`.
+The `GET /api/terminal/<name>` routes, the websocket included, answer `{ "error": "unauthorized", "hint" }`. A valid configured bearer token in the Authorization header can pass the browser boundary's target-name check
+for a well-formed but unaccepted Host. It cannot pass a malformed Host or a refused Origin; see
+[Browser access](browser-access.md).
+
+## Event streams
+
+All of them are server-sent events (`text/event-stream`):
+
+- **`GET /api/events`**: the daemon's event log, stored events only. `?rigId=` limits it to one rig. Each event
+  carries `id: <seq>`. `view.changed` is not stored, so it does not appear here, live or replayed; `view.changed`
+  rows stored by older versions still replay until an events retention (a separate change) removes them.
+  Without `Last-Event-ID`, `?replay=<n>` first sends the newest `n` stored events (`0`: none, live events only);
+  `?replay=all`, no `replay`, or any other value sends the whole log first. A `Last-Event-ID` header (what
+  `EventSource` sends when it reconnects) sends the events after that id instead, whatever `replay` says.
+  Live events that arrive during the replay follow it, without duplicates.
+- **`GET /api/activity/events`**: live only, no replay or ids. Each message is named after its event type:
+  `seat.activity_changed`, `seat.rung_health`, `proof.judged` or `proof.sources_changed`. It tells a view to read
+  again; it carries no derived state.
+- **`GET /api/views/sse`** (alias `/api/views/watch`) and **`GET /api/views/<viewName>/sse`**: `view.changed`, live
+  only. Its `id:` is the `seq` of the stored event that caused it, so it can repeat across views and is not a replay
+  cursor.
+- **`GET /api/queue/sse`** (alias `/api/queue/watch`): live only, with `id: <seq>`, for `queue.created`,
+  `queue.handed_off`, `queue.claimed`, `queue.unclaimed`, `qitem.fallback_routed`, `qitem.closure_overdue`,
+  `inbox.absorbed` and `inbox.denied`.
+
+## People, asks and updates without Slack
+
+`GET /api/gateway/participants` returns the same JSON projection as `rig gateway participants --json`:
+`{ ok, participants, default, organizations, groups, warnings, participantFile }`. Pick a person by `kind: "person"`
+and use their `id` for a Feed `view`, or their `address` for the queue's `destinationSession`. The effective
+`default` is `{ id, source }`, or `{ id: null, reason }`; it does not establish who is using the client.
+Participant records omit connector credential pointers. Check `warnings` and `participantFile.error` for a partial
+registry read; an unreadable human registry returns HTTP 503. See [Gateway participants](gateway-participants.md).
+
+| Read | Meaning |
+| --- | --- |
+| `GET /api/queue/list?destinationSession=<address>&state=pending,in-progress,blocked&limit=100` | Open rows for exactly that address. `humanIntent: "update"` identifies an FYI; other values are asks. Each row has `qitemId`, `state`, `summary` and `body`. |
+| `GET /api/queue/open-human-updates?view=<person-id>&limit=20` | Open human FYIs, even when no Slack delivery receipt exists. Returns `{ items, limit, truncated, perspective? }`. |
+| `GET /api/queue/<qitemId>` and `/<qitemId>/transitions` | The current row and its history, including closed state and recorded outcome. Closing or acknowledging a row is different from delivering it. |
+| `GET /api/queue/human-updates?view=<person-id>&limit=20` | Existing delivered-update history: explicit FYIs with a recorded Slack posting receipt, including closed rows. Items add `deliveredAt` and `deliveryReceipt`. It is not the open-update inbox. |
+
+Both update lists accept integer `limit` values from 1 to 100 (default 20); invalid values return HTTP 400.
+Open FYIs are ordered by priority, then newest first. `view=all` (also the default) shows all perspectives;
+`view=default` selects the effective default person, or all when no default is available. A named view that cannot
+resolve a person has `perspective.basis: "unknown"` and no items when a registry is available. With no registered
+people there is no perspective filter: inspect `perspective` rather than assuming the requested view was applied.
+A person's view also includes instance or unassigned items when that person is the default. These are views on a
+trusted daemon, not per-person authentication or access control; existing daemon and browser access rules still apply.
+
+Always display `truncated: true` as incomplete, even if fewer than `limit` items arrive. A person-view scan examines
+at most 1,001 open updates (101 for delivered history) before filtering; older matching items may be outside that
+window. These endpoints have no pagination cursor. Reading an open update never posts it to Slack, acknowledges it,
+or manufactures a delivery receipt.
+
+For example, this Node client takes a daemon URL, a registered person id and a rig name. It uses no Slack API and
+makes only reads; use the daemon discovery and access rules above. In an ES module, `process.argv.slice(1)` is the
+argument list for `node --input-type=module -e '<code>' <url> <person-id> <rig>`:
+
+```js
+// client-people-example
+const [base, personId, rig] = process.argv.slice(1);
+async function get(path) {
+  const response = await fetch(new URL(path, base));
+  if (!response.ok) throw new Error(`${path}: HTTP ${response.status}`);
+  return response.json();
+}
+const registry = await get("/api/gateway/participants");
+const person = registry.participants.find(p => p.kind === "person" && p.id === personId);
+if (!person) throw new Error("Choose a registered person id");
+const rows = await get(`/api/queue/list?destinationSession=${encodeURIComponent(person.address)}&state=pending,in-progress,blocked&limit=100`);
+const updates = await get(`/api/queue/open-human-updates?view=${encodeURIComponent(person.id)}&limit=20`);
+const receipts = await get(`/api/queue/human-updates?view=${encodeURIComponent(person.id)}&limit=20`);
+const rigHistory = await get(`/api/queue/list?rig=${encodeURIComponent(rig)}&limit=100`);
+const messages = await get(`/api/queue/outbox/list?rig=${encodeURIComponent(rig)}&limit=100`);
+console.log(JSON.stringify({ participants: registry.participants, default: registry.default,
+  warnings: registry.warnings, participantFile: registry.participantFile,
+  asks: rows.filter(row => row.humanIntent !== "update"), updates, receipts, rigHistory, messages }, null, 2));
+```
+
+The queue and outbox lists are bounded snapshots, not subscriptions. A full queue list may need a larger `limit`.
+An outbox entry's `deliveryState` reports the send attempt; only `delivered` confirms delivery. A queue transition,
+an open FYI, or a closed ask is not that receipt. The rig-wide reads and their limits are described below; the existing
+`outbox/list?senderSession=<session>` read remains available.
+
+## Route details
+
+- **A rig's queue history:** `GET /api/queue/list?rig=<rig>&updatedSince=<ISO time>&sort=updated` returns the
+  rig's queue items (any whose source or destination seat is in the rig) updated at or after that time: open items
+  (`pending`, `in-progress`, `blocked`) first, then the rest by latest update. `limit` defaults to 100 and has no
+  upper cap. The list is cut after `limit` rows in that order, so when it comes back full, raise `limit` to read
+  them all. Each item's own history is `GET /api/queue/<qitemId>/transitions`. This read scans the queue table, so
+  read it occasionally rather than polling it. For live changes, treat `GET /api/queue/sse` events as hints to read
+  `queue/list?rig=` again. Not every event names both seats (`queue.claimed` and `queue.unclaimed` carry only the
+  destination), and not every state change is on that stream, so the events alone don't track a rig's items.
+- **A rig's messages:** `GET /api/queue/outbox/list?rig=<rig>` returns the messages sent from or to the rig's
+  seats, newest first; `?senderSession=<session>` still returns one sender's. One of the two is required; with
+  both, `senderSession` is used and `rig` is ignored. `limit` defaults to 100. No event is sent when a message is delivered, so read this list to see delivery
+  states.
+- **`GET /api/queue/recent-transitions`** returns at most 20 rows. A positive `limit` above 20 is capped at 20;
+  a missing, zero, negative or unparseable `limit` means 20.
+- **`POST /api/transport/send`** takes `urgent: true`, the same as `rig send --urgent`. The daemon records who
+  marked it from the sender identity, not from the body.
+- **`POST /api/scope/approve`** is what `rig scope slice approve` and `rig scope mission approve` call. The body
+  takes `scopeTier` (`slice` or `mission`), `scopePath`, and optionally `approvalScope` (`spec` or `delivery`, the
+  default), `missionsRoot`, `onBehalfOf`, `reApprove`, `reason` and `lockedArtifacts`. The approver is the
+  request's sender identity (`actorSession` in the body is recorded beside it, and a different value is
+  superseded). Success is 201 with `scopeTier`, `scopeId`, `scopePath`, `approvalScope`, `approvedBy`,
+  `approvedAt`, `onBehalfOf`, `actionId`, `freezeFired` and `reApproved`. Refusals include 400 for an invalid
+  `scopeTier`, `scopePath` or `approvalScope`, or a missing actor (`actor_required`); 503 `action_log_unavailable`;
+  and, from the approval itself, 404 `scope_not_found`, 409 `already_approved` or `nothing_to_reapprove`, 503
+  `workspace_not_configured`, 500 `audit_write_failed`, and 400 for its other refusals. An unexpected error is 500
+  `internal`.
+- **`POST /api/proof/judge`** is what `rig proof judge` calls. The body takes `scope`, `item`, `verdict`
+  (`accept`, `reject` or `withdraw`), `reason`, `expectedRevision` and `expectedPrevious` (a judgment id or
+  `null`), and optionally `evidence`, `expectedEvidence`, `subject`, `operationId`, `replace`, `project` and
+  `projectRoot`. The judge is the request's sender identity. The response is `judgment`, the slice's `readiness`,
+  `replayed` and `notification`: 201 for a new judgment, 200 when an `operationId` replays an earlier one.
+- **`POST /api/workflow/instantiate`** requires `specPath`, `rootObjective` and `createdBySession`, and takes
+  `entryOwnerSession`, `targetRig` and `scope` (see [Project workspace](project-workspace.md) for `scope`).
+  Success is 201 with `instance`, `spec`, `entryQitemId` and `entryOwnerSession`, any `advisories`, and
+  `scopeIgnored` when a scope wasn't applied. The instance keeps `rootObjective`, which every instance read returns
+  (`/api/workflow/list`, `/api/workflow/<id>` and its `trace`). An instance started before it was kept reads `null`.
+- **`GET /api/provider/usage`** returns `{ hostUsage }`, this host's usage rows (`host: "local"`, a `provider`
+  and a `state`). A Claude row's `windows` are the seats' status-line readings, one entry per reading, each with
+  its `window`, `usedPercent`, `resetsAt`, `asOf` and `seatSession`. A Codex row has `windows: []`. Treating the
+  readings as one account rests on the deployment assumption the row's `provenance` names (one account per
+  host), not on a measured identity. Readings whose reset times disagree are reported as `anomalies`, and the
+  row's `state` can then be `explicit_unknown`. Show the row's `state`, `anomalies` and each reading's seat and
+  `asOf` as reported, rather than merging readings that happen to match.
+- **Slack channels and people:** `GET /api/gateway/slack/channels` and `GET /api/gateway/slack/users` each return
+  one page from Slack, read with the daemon's configured bot token, which never leaves the daemon. Channels are
+  public and unarchived, as `{ id, name, isMember }`; people are `{ id, name, realName, displayName, isBot,
+  deleted }`, with no email. Pass `cursor` (the previous page's `nextCursor`, `null` on the last page) and
+  `limit` (default 100, at most 200). Inbound delivery needn't be enabled. Failures: 503 `slack_not_configured`
+  when no bot token resolves, which is different from an empty list; 403 `slack_missing_scope` with the scope
+  Slack names in `needed` (an older app may lack `users:read`); 429 `slack_rate_limited` with Slack's
+  `Retry-After`; otherwise 502. Each request is a live Slack call, so page through a list when someone opens a
+  chooser rather than polling it. Private channels and direct messages aren't listed.
+- **Slices are found by folder name alone.** `GET /api/slices/:name`, `/:name/doc/*`, `/:name/proof-asset/*` and
+  `GET /api/review/slice/:name` take only the name; the detail route reads `?mission` only together with
+  `?project`. When two missions have a slice folder of the same name, the daemon keeps the first one it finds:
+  that one is listed and served, and the other doesn't appear in `GET /api/slices` at all. Keep slice names unique
+  across missions, as [Project workspace](project-workspace.md) says.

docs/reference/edge-types.md

@@ -44,13 +44,15 @@ All other edge kinds (`can_observe`, `collaborates_with`, `escalates_to`) do NOT
 
 `delegates_to` also names a seat's orchestrator for the queue: the source of the `delegates_to` edge into a seat. The
 wake ladder escalates an unacknowledged row to that orchestrator, and stuck-row findings and recovery go to it. A seat
-with no incoming `delegates_to` edge has no default orchestrator. `spawned_by` plays no part here.
+with no incoming `delegates_to` edge has no default orchestrator. A seat with several incoming `delegates_to` edges has
+no single orchestrator either: its recovery goes to its declared `escalates_to` destination, or keeps the existing
+fallback, and the recovery says the parents were ambiguous. `spawned_by` plays no part here.
 
 ### Fallback escalation across rigs
 
-A seat with **no declared `delegates_to` parent** may declare one `escalates_to` destination. The wake ladder and
-stuck sweep use it for new recovery rows; they do not move the original obligation or reassign an already-open recovery. A stopped or unbound declared
-parent still takes precedence, preserving the existing parent route and its fallback.
+A seat with **no declared `delegates_to` parent, or several**, may declare one `escalates_to` destination. The wake ladder and
+stuck sweep use it for new recovery rows; they do not move the original obligation or reassign an already-open recovery. A
+single stopped or unbound declared parent still takes precedence, preserving the existing parent route and its fallback.
 
 ```yaml
 edges:

docs/reference/gateway-participants.md

@@ -5,15 +5,34 @@ OpenRig's gateway can register **several people** and **shared channels**, group
 OpenRig doesn't enforce it. This page covers the commands, what the default is for, and what upgrading and rolling back
 do.
 
+For a remote client, `GET /api/gateway/participants` returns the same participant, organization, group, default and
+warning fields as `rig gateway participants --json`. Add `?organization=<id>` to filter the participant list; the default
+and organization and group maps still describe the whole registry. The result names the participant file and reports
+an unreadable side file without treating it as an empty one. If the human registry itself cannot be read, the route
+returns `503` with `error: "human_registry_unavailable"`. The result does not include connector secret pointers.
+
 ## Participants
 
 A participant is either:
-- **a person**, registered with `rig gateway human add` as before, with their own Slack binding; or
+- **a person**, registered with `rig gateway human add`, with or without a connector binding (such as Slack or an HTTP loopback [relay](gateway-relay.md)); or
 - **a shared channel**: one vetted Slack channel registered as one participant, `<id>@external`, not a person per
   sender.
 
 People and channels share one address space: a channel can't take a person's id.
 
+A registered person is addressed as `<id>@external`, including without Slack: for example,
+`rig gateway human add alex --display-name "Alex" --delivery-class B` creates `alex@external`, reached in the Feed and Steering.
+A person can also bind to a local bridge adapter using a `relay` binding:
+`rig gateway human add bob --display-name "Bob" --delivery-class B --binding "relay:local-bridge:file:/etc/openrig/relay.env:primary:handle=bob-ops"`
+(see [Gateway Relay Reference](gateway-relay.md) for endpoint, secret provisioning, and custody architecture).
+`human@host` is the unnamed person at the terminal; `human@kernel` is also accepted as a human/gate address.
+These aliases (including `human-<name>@host` and `human-<name>@kernel`) are not agent sessions: reply with
+`rig queue create --destination human@host --human-intent update --summary "<one line>" --evidence-ref "<path>" --body "..." --verify`,
+substituting the sender's address and your summary, evidence path and reply. A reply destination doesn't identify
+who acted or determine who may answer an ask. Clients acting as a registered person name that person's
+`actorSession`; a seat identity header takes precedence. The terminal's primary-user authority records the unnamed
+`human@host` actor, not the person whose queue is being viewed.
+
 ```bash
 # An organization and a group inside it (one command each, create or update)
 rig gateway org set acme --name "Acme" --policy "Answer Acme's customers within a working day. Never promise a date."
@@ -59,30 +78,26 @@ never changes `workspace.operator_seat_name`, which keeps its own meaning.
 
 ## In the TUI: whose queue, and who acts
 
-The person at the terminal is the **primary user**: the default person, or, with several people and none chosen, the
-person at the terminal unnamed. Whatever the TUI shows, its acts are theirs.
+The person at the terminal is the **primary user**. The TUI records their acts as unnamed (`human@host`),
+whatever person's queue it shows or who is configured as the default.
 - **Opening:** with a chosen default, the TUI opens on that person's Feed. With none chosen, including a one-person
   install that never chose one, it opens on the topology, as before.
 - **Whose queue:** `whose <person>` shows one person's asks, `whose all` everyone's, and `whose default` the default
   person's. The Feed's first line says whose queue it shows and who acts.
-- **Acting on someone else's ask** (answer, approve or resolve from the TUI) records both: the primary user as the actor,
-  and the person asked in the audit notes and the transition. The asking seat reads, for example, "Answered by
-  mike@external, the primary user at the terminal, on sarah@external's ask". With several people and none chosen, the
-  actor is `human@host`.
-- **Viewing Sarah isn't acting as Sarah.** Only Sarah, from her own channel, answers as Sarah. Without the primary
-  user's authority, one registered person can't answer, approve, deny or resolve what was asked of another; that's
-  refused (`answer_not_recorded`) and nothing changes.
-- With no one registered, the TUI and its acts are as before, except on a shared channel's ask or a row parked on a
-  channel: the TUI shows the act as `human@host`, the person at the terminal, and it's recorded that way, never as the
-  channel.
+- **Acting on someone else's ask** (answer, approve or resolve from the TUI) records both: the unnamed terminal
+  actor `human@host` and the person asked in the audit notes and transition. The asking seat reads, for example,
+  "Unattributed (human@host) answered sarah@external's ask".
+- **Viewing Sarah isn't acting as Sarah.** A registered person may answer, approve, deny or resolve another person's
+  ask as themselves; both the actor and person asked are recorded. No connector binding is required, and viewing
+  someone else's queue doesn't make its viewer that person.
 - **A shared channel never acts.** Its messages are correspondence, so nothing it says answers, approves, denies or
   resolves an ask, and a customer's earlier message in an ask's thread is never taken as its answer. A request without
   the primary user's authority that acts as a channel, with any Mission Control verb (including drop, route, hold, hand
   off and annotate), is refused (`answer_not_recorded`) and nothing changes.
-- **Where the authority applies:** only on a request that carries no seat identity header, which is how the TUI calls
-  the daemon. Seats and the CLI send that header, so their acts stay theirs. This is the existing trusted-network
-  distinction, not an authentication check: a request that leaves the header out is taken as the terminal's, as any
-  such request already is.
+- **Where the authority applies:** primary-user authority is accepted only on a request that carries no seat identity
+  header, which is how the TUI calls the daemon. A headerless client can instead name a person's `actorSession` without
+  that authority. Seats and the CLI send the identity header, so their acts stay theirs. This is the existing
+  trusted-network distinction, not an authentication check.
 
 ## Shared-channel conversations
 

docs/reference/gateway-relay.md

@@ -0,0 +1,268 @@
+# Gateway Relay: HTTP Loopback Transport Adapter & Custody Reference
+
+OpenRig's Gateway Relay is a local HTTP loopback transport adapter. It enables registered people to receive outbound notifications and questions (asks, updates, and operator digests) through an operator-owned local bridge process, and to post answers back into OpenRig with sender attribution resolved against registered human bindings.
+
+OpenRig does not embed third-party messaging platform SDKs (such as WhatsApp, Matrix, or Signal). An external bridge process translates between OpenRig loopback HTTP requests and external messaging platforms.
+
+```
++--------------------+           HTTP Loopback            +--------------------+
+|   OpenRig Daemon   |   POST <OPENRIG_RELAY_URL>  --->   |   Operator Bridge  |
+|  (Gateway Subsys)  |   <--- POST .../relay/inbound      |  (Loopback Server) |
++--------------------+                                    +--------------------+
+                                                                    |
+                                                      Platform-specific transport
+                                                                    v
+                                                          +--------------------+
+                                                          |  External Platform |
+                                                          |  (Mobile / Client) |
+                                                          +--------------------+
+```
+
+---
+
+## 1. System Topology & Custody Boundaries
+
+### Outbound Custody & Recovery
+- OpenRig enqueues outbound decisions to a local `DispatchBuffer` before issuing any HTTP `POST` to `OPENRIG_RELAY_URL` (in `push` mode) or holding the decision for connector pull (in `pull` mode).
+- Each logical decision freezes its complete `relayEnvelope` onto the parent `DispatchBuffer` entry before the first delivery pass, with no self-imposed outbound 64 KiB cap or multi-part splitting.
+- **Push mode (`OPENRIG_RELAY_MODE=push`, default)**:
+  - Transport is at-least-once: OpenRig sends the complete JSON envelope in a single HTTP `POST` per attempt with `deliveryId = <decisionId>` (preserving the full `body`, `humanDetail`, `humanQuestions`/options, `summary` overflow, and digest `items` without silent clipping).
+  - An HTTP `2xx` response from the bridge acknowledges durable bridge custody of that complete envelope (`deliveryId`). The bridge must persist each envelope by `deliveryId` and deduplicate replayed `deliveryId`s idempotently.
+  - OpenRig records the logical queue acceptance receipt (`relay-owner-notification-accepted` with `deliveryId`) and drains the decision from `DispatchBuffer` once the HTTP `POST` returns `2xx`.
+  - Failed dispatches remain in OpenRig's `DispatchBuffer` and record a relay-specific failure transition note (`relay-owner-notification-transport-failed` naming the `connectorRef`, distinct from Slack's `slack-owner-notification-transport-failed`). Replays after a lost ACK or daemon/bridge restart re-send the exact frozen `relayEnvelope` with the same `deliveryId` and bytes. If the relay subsystem stops while an outbound HTTP `POST` is in flight (or before it starts), the transport aborts cleanly (`stopped`) without recording a false failure transition note or false acceptance receipt, retaining the buffered decision for replay on restart. There is currently no periodic background timer redriving outbound dispatches on timeout.
+- **Pull mode (`OPENRIG_RELAY_MODE=pull`)**:
+  - Client-only connectors that do not bind a listening HTTP server set `OPENRIG_RELAY_MODE=pull`. A pull profile is ready with `OPENRIG_RELAY_INBOUND_TOKEN` alone (`OPENRIG_RELAY_URL` and `OPENRIG_RELAY_OUTBOUND_TOKEN` are optional and unused in pull mode).
+  - Outbound decisions remain in `DispatchBuffer` awaiting connector pull and are never treated as failed push destinations (`relay-owner-notification-transport-failed` is not recorded). While the outbox lists a notice (staged in `DispatchBuffer` and `relay-attempted-decisions.jsonl` for its `connectorRef`), it is not a missing post for up to **24 hours** (`PULL_OUTBOX_NEVER_POSTED_WINDOW_MS`) and reads `never-posted` (`undelivered-human-notice`) if still unacknowledged after that. A notice the outbox does not list (never staged, staging failed, or its `DispatchBuffer` or seen stores are unreadable, so the outbox answers `503`) uses the usual 120-second window (`NEVER_POSTED_WINDOW_MS`). This is reporting only: nothing is held, refused or re-sent.
+  - The connector lists its pending complete envelopes via `GET /api/gateway/relay/outbox` (read-only; no consumer leases, read receipts, or required polling cadences) and acknowledges durable custody via `POST /api/gateway/relay/outbox/ack`.
+  - An acknowledgment on `POST /api/gateway/relay/outbox/ack` has the exact same custody meaning as a push `2xx`: OpenRig records `relay-owner-notification-accepted` (closing informational `update` items as delivered), marks its local outbound seen/delivered guards, and drains the decision from `DispatchBuffer`. Duplicate acknowledgments reconcile idempotently (`alreadyAccepted: true`), and retries after a partial local write failure finish the remaining receipt and drain sequence.
+- **Custody boundary (both modes)**: Bridge acceptance (via push `2xx` or pull `POST /api/gateway/relay/outbox/ack`) confirms durable custody of the complete envelope by the connector only (projected as `accepted` / `"Accepted by bridge"`, whereas Slack posts use `posted` / `"Delivered"`). It does not mean posted, seen, answered, or approved.
+
+### Inbound Custody & Resolution
+- The operator bridge posts replies to `POST /api/gateway/relay/inbound`.
+- Requests require `Authorization: Bearer <OPENRIG_RELAY_INBOUND_TOKEN>` matching a configured, ready relay profile.
+- OpenRig verifies that the caller is a direct loopback peer (`127.0.0.1/8` or `::1`) over the underlying TCP socket. Proxy and forwarding headers (`Via`, `X-Forwarded-*`, `Forwarded`, `X-Real-Ip`) are ignored rather than rejected because caller identity is established directly from the loopback socket peer and the bearer token.
+- Sender attribution resolves by matching `sender.handle` against human fragments registered to the authenticated connector. Harmless extra fields on the top-level inbound JSON object or inside `sender` are ignored and can never override handle-to-person attribution or alter the event fingerprint. The handle is asserted by the trusted bridge; it does not provide independent proof of the physical human on the external platform.
+- When an inbound reply is accepted (`200 OK`), OpenRig reserves the event in `relay-inbound-seen.jsonl` (`received:<fingerprint>`), durably commits the answer/resolve action and its `mission_control_actions` audit record atomically through its Mission Control write contract (using existing `excludeOperationReply` revision checking without a second answer store or synthetic revision refresh), and marks the event `processed:<fingerprint>`. Exact duplicate replays of the same `eventId` and payload succeed idempotently (`deduplicated: true`).
+
+### Exclusions & Scope
+The relay adapter does not handle interactive UI buttons, message attachments, group rooms, unsolicited inbound triggers, multi-bridge failover clustering, or third-party platform SDK backends.
+
+---
+
+## 2. Secrets & Environment Configuration
+
+Each relay connector loads configuration from an absolute file referenced by the binding's `secretsRef` (`file:<absolute-path>`).
+
+### Security & File Requirements
+- **Absolute Path Only**: Must begin with `file:/`. Relative paths, bare paths, and `file://` authority URLs are rejected.
+- **File Permissions**: The file should restrict access to the owner only (no group or other permission bits set, i.e. `(mode & 0o077) === 0`, such as `0600` or `0400`). Owner-only modes (`0600`, `0400`) produce no warning. If group or world permission bits are set (e.g. `0644`), OpenRig logs a redacted warning in the profile's readiness `warnings` and `reason` (matching the Slack secret-file posture) while keeping the profile `ready`. Missing or unreadable secret files remain `not-ready`.
+- **No Environment Fallback**: OpenRig does not fall back to process environment variables (`process.env`). All variables must come directly from the referenced file.
+- **Profile Independence**: If a secrets file is missing, unreadable, or missing required keys, only connectors bound to that file report `not-ready`. Other connectors remain functional.
+
+### Required Secrets Keys
+
+**Push mode (`OPENRIG_RELAY_MODE=push`, default when omitted):**
+```ini
+OPENRIG_RELAY_URL=http://127.0.0.1:8088/webhook
+OPENRIG_RELAY_OUTBOUND_TOKEN=sec_outbound_token_value
+OPENRIG_RELAY_INBOUND_TOKEN=sec_inbound_token_value
+```
+
+**Pull mode (`OPENRIG_RELAY_MODE=pull`):**
+```ini
+OPENRIG_RELAY_MODE=pull
+OPENRIG_RELAY_INBOUND_TOKEN=sec_inbound_token_value
+```
+
+- `OPENRIG_RELAY_MODE`: Optional mode selector (`push` or `pull`; defaults to `push` when omitted). Any other value is rejected as `not-ready`.
+- `OPENRIG_RELAY_URL`: Required in `push` mode; ignored/unused in `pull` mode (`endpointUrl: null`). When required, must use `http:`, loopback host (`127.0.0.1/8`, `::1`, or `localhost`), without URL credentials or hash fragments. Redirects (3xx) are rejected.
+- `OPENRIG_RELAY_OUTBOUND_TOKEN`: Required in `push` mode; ignored/unused in `pull` mode. Non-empty Bearer token sent by OpenRig on outbound HTTP `POST` requests to the bridge.
+- `OPENRIG_RELAY_INBOUND_TOKEN`: Non-empty Bearer token required in both `push` and `pull` modes to authenticate connector requests (`GET /api/gateway/relay/outbox`, `POST /api/gateway/relay/outbox/ack`, and `POST /api/gateway/relay/inbound`). Must be unique across different configured relay connector profiles (`connectorRef`); sharing the same inbound token across multiple relay profiles is rejected as ambiguous (`not-ready`). Using the same token value for both `OPENRIG_RELAY_OUTBOUND_TOKEN` and `OPENRIG_RELAY_INBOUND_TOKEN` within a single `push` profile is allowed, with a redacted readiness warning explaining why separate directional tokens are recommended so an outbound capture cannot authenticate inbound replies.
+
+---
+
+## 3. Wire Protocol Specifications
+
+### Outbound Envelope (`POST <OPENRIG_RELAY_URL>` or `GET /api/gateway/relay/outbox`)
+
+```json
+{
+  "v": 1,
+  "deliveryId": "dec-019283-abc",
+  "recipient": { "handle": "alice-ops" },
+  "intent": "decision",
+  "outcome": "interrupt",
+  "summary": "Database Migration Approval",
+  "body": "Production migration script 004_add_indexes.sql is ready to execute. Reply YES to approve, NO to abort.",
+  "replyRef": "eyJ2IjoxLCJwYXlsb2FkSGFzaCI6...<base64url>"
+}
+```
+
+- `v`: Protocol version integer (`1`).
+- `deliveryId`: Stable whole-decision delivery identifier (`<decisionId>`).
+- `recipient.handle`: Recipient's registered safe alias for this connector.
+- `intent`: `"decision"` (actionable question) or `"update"` (informational notice).
+- `outcome`: Delivery classification (`"interrupt"`, `"notify"`, or `"digest"`).
+- `summary`: Plain text summary preview (up to 1,000 characters on the `summary` field, cut on a Unicode code-point boundary; any summary text exceeding 1,000 characters is preserved in full at the beginning of `body` or `item.body`).
+- `body`: Complete plain text content (including full `summary` overflow, `body`, `humanDetail`, and rendered `humanQuestions` headers, prompts, and option labels/descriptions) sent in a single complete JSON envelope without silent character clipping or self-imposed outbound byte splitting.
+- `replyRef`: Fixed-size opaque base64url token (`<= 4,096` characters for any supported ID lengths — `246`–`266` characters in protocol v1, hashing the canonical frozen `RelayReplyRefPayload` and binding `connectorHash`, `mode`, and optional `digest: true`) for actionable decisions; `null` for informational updates (and `null` at the top level of multi-item digests).
+- `items`: Present when `outcome` is `"digest"`. Array of objects with `qitemId`, `notificationKey`, `summary`, optional `body`, and per-item `replyRef` (or `null`).
+
+### Pull Outbox Endpoints (`GET /api/gateway/relay/outbox` & `POST /api/gateway/relay/outbox/ack`)
+
+Both endpoints require a direct loopback TCP peer (`127.0.0.1/8` or `::1`), reject browser cross-origin headers (`403`), and authenticate the caller via `Authorization: Bearer <OPENRIG_RELAY_INBOUND_TOKEN>` (`401`). If the human registry, `DispatchBuffer` (`pendingStrict()`), or outbound `SeenStore` cannot be read, both return `503` (`human_registry_unavailable`, `dispatch_buffer_unavailable`, `seen_store_unavailable`, `queue_store_unavailable`, or `receipt_write_failed`) rather than returning an empty `deliveries` array.
+
+- **`GET /api/gateway/relay/outbox?limit=<1..200>`** (default `50`): Returns `{ "ok": true, "connectorRef": "<ref>", "deliveries": [ <RelayOutboundEnvelope>, ... ] }` with complete pending envelopes from `DispatchBuffer` belonging to the authenticated `connectorRef` in FIFO buffer order. Strictly read-only (no leases or read receipts).
+- **`POST /api/gateway/relay/outbox/ack`**: Closed JSON object `{ "deliveryId": "<decisionId>" }` accepting any non-empty string `deliveryId` the outbox can list, within the 64 KiB request-body size bound (`413 payload_too_large`); unknown fields or an empty or non-string `deliveryId` return `400 invalid_envelope`. Finishes any remaining `relay-owner-notification-accepted` transition receipts, marks `slack-outbound-seen.jsonl` and `relay-delivered-decisions.jsonl`, releases the in-memory outbound poll guard, and drains `deliveryId` from `DispatchBuffer`. Returns `200 { "ok": true, "deliveryId": "<id>", "connectorRef": "<ref>", "alreadyAccepted": false }` (`"alreadyAccepted": true` on duplicate ack for the same `connectorRef`, or `404 unknown_delivery` if unknown or owned by another connector; duplicate and unknown acks are answered from `relay-delivered-decisions.jsonl` and `relay-attempted-decisions.jsonl`, and the transition receipt scan runs only when those disagree). On any `503` partial failure, a retry with the same `deliveryId` finishes the remaining receipt and drain steps.
+
+### Inbound Envelope (`POST /api/gateway/relay/inbound`)
+
+```json
+{
+  "eventId": "evt-relay-20261107-001",
+  "sender": { "handle": "alice-ops" },
+  "replyRef": "eyJ2IjoxLCJwYXlsb2FkSGFzaCI6...<base64url>",
+  "text": "Approved. Proceed with rollout.",
+  "ordering": "after-question"
+}
+```
+
+- `eventId`: Bridge-assigned identifier (`1..256` chars, matching `/^[A-Za-z0-9._:-]{1,256}$/`).
+- `sender.handle`: Sender's alias, matching `/^[A-Za-z0-9._-]+$/`. Must match a human fragment bound to the authenticated connector. Harmless extra fields inside `sender` or on the top-level JSON object are ignored and never override `sender.handle` resolution or affect the event fingerprint.
+- `replyRef`: Exact fixed-size opaque base64url token copied from the outbound envelope (maximum 4,096 characters). Bridges must not parse or alter `replyRef`.
+- `text`: Plain text response (maximum 32,768 characters, non-empty after trim, no NUL bytes). Processed as unstructured text without automatic natural-language approval parsing.
+- `ordering`: Optional enum (`"after-question"` | `"uncertain"`, default `"uncertain"`). When `"uncertain"`, text is accepted as recorded context rather than verified sequence relative to the question.
+
+#### Inbound Statuses & Responses
+- **200 OK**: Request accepted (`{ ok: true, deduplicated: false, reconciled: false, qitemId, verb, actorSession, actionId }`). Exact duplicate submissions of the same `eventId` return `ok: true` with `deduplicated: true`.
+- **400 Bad Request**: Malformed JSON, non-object payload, or invalid required field syntax/bounds (`eventId`, `sender.handle`, `replyRef`, `text`, or `ordering`). Harmless extra fields are ignored.
+- **401 Unauthorized**: Missing, malformed, unrecognized, unready, or cross-profile ambiguous inbound Bearer token, or no humans configured (`noHumans`).
+- **403 Forbidden**: Non-loopback TCP socket peer, browser cross-origin request, connector mismatch (`replyRef` emitted by another connector), or unregistered `sender.handle`. (`Via` and `X-Forwarded-*` headers are ignored).
+- **404 Not Found**: Queue item referenced in `replyRef` does not exist.
+- **409 Conflict**: Stale `decisionRevision` (question cancelled, answered, or re-asked), attempt to answer an informational update, unrecognized `replyRef` when both durable receipts and the pending dispatch buffer are readable, or conflicting payload submitted under an existing `eventId`.
+- **413 Payload Too Large**: Request body exceeds 64 KiB (65,536 bytes).
+- **503 Service Unavailable**:
+  - `human_registry_unavailable`: Local human registry (`gateway/humans/*.yaml`) failed to load due to a broken fragment or filesystem read error (normal `noHumans` returns `401`).
+  - `queue_store_unavailable`: Local SQLite queue/transition store read failed while resolving or verifying `replyRef`, returning `503` with zero seen-store reservation or Mission Control action side effects.
+  - `dispatch_buffer_unavailable`: An un-receipted `replyRef` required fallback verification against `<openrig-home>/gateway/dispatch-buffer.json`, and reading the buffer failed (`pendingStrict()`), returning `503` with zero seen-store reservation or Mission Control action side effects. (`replyRef`s already vouched for by durable accepted receipts in `queue_transitions` or `queue_transitions_archive` are verified before consulting the buffer and still succeed.)
+  - `seen_store_unavailable` (response includes `seenPath`): Local seen-ledger (`state/relay-inbound-seen.jsonl`) is unreadable, contains unparseable JSONL or unkeyable records (missing `id`), failed a reservation/processed mark, or has a corrupt/unvouched record for the requested `eventId`'s key. Keyed suspect records (`missing/invalid status` or conflicting recorded fingerprints for the same key; diagnostic `ts` or extra metadata fields are ignored) scope refusal to that `eventId` key only so unrelated `eventId`s continue to operate normally, while never treating unparseable or suspect history as unseen.
+
+---
+
+## 4. Identity & Binding Configuration
+
+Bindings are managed through human fragment files (`gateway/humans/<entityId>.yaml`) and CLI verbs (`rig gateway human`).
+
+### Binding Syntax
+```
+relay:<connectorRef>:file:<absolute-env-file>:<role>[:handle=<alias>]
+```
+- `kind`: `relay`.
+- `connectorRef`: Identifier for the connector instance (e.g. `local-bridge`).
+- `secretsRef`: Absolute file path reference (`file:/path/to/relay.env`).
+- `role`: `primary` or `secondary` (at least one primary binding required for outbound dispatch).
+- `handle`: Optional platform alias (`handle=<alias>`). Required for inbound reply resolution; omitted means outbound-only.
+
+Handles must match `/^[A-Za-z0-9._-]+$/` and are unique per `connectorRef`. Multiple people can share a connector, each registered with their own handle.
+
+---
+
+## 5. Operational Lifecycle
+
+### Step 1: Provision Secrets File
+Create an owner-only secrets file (no group/other access bits, e.g. `0600` or `0400`):
+```bash
+mkdir -p "$HOME/.config/openrig-relay"
+touch "$HOME/.config/openrig-relay/relay.env"
+chmod 600 "$HOME/.config/openrig-relay/relay.env"
+
+cat << 'EOF' > "$HOME/.config/openrig-relay/relay.env"
+OPENRIG_RELAY_URL=http://127.0.0.1:8088/webhook
+OPENRIG_RELAY_OUTBOUND_TOKEN=replace_with_outbound_token
+OPENRIG_RELAY_INBOUND_TOKEN=replace_with_inbound_token
+EOF
+```
+
+### Step 2: Register Human Binding
+```bash
+# Add a person with a primary relay binding
+rig gateway human add alice --display-name "Alice Smith" --delivery-class B \
+  --binding "relay:local-bridge:file:$HOME/.config/openrig-relay/relay.env:primary:handle=alice-ops"
+
+# Replace an existing binding slot (requires the full binding spec)
+rig gateway human set alice binding.0 "relay:local-bridge:file:$HOME/.config/openrig-relay/relay.env:primary:handle=alice-ops"
+```
+
+> [!NOTE]
+> `rig gateway human set alice binding.<n>` accepts either a full binding specification string or `unset` to remove that binding slot:
+> - **One-primary rule:** No replacement primary is chosen silently. If removing a `primary` binding would leave other bindings, the command refuses. To replace the primary in place, set its slot to the new primary spec; to promote an existing secondary binding, first save that binding's full spec (`kind:connectorRef:secretsRef:role[:handle=<id>]`) from `rig gateway human show <entityId> --json`, then unset that secondary slot, then find the primary's current slot in `remainingBindings` (or `rig gateway human show <entityId> --json`, since the primary's slot shifts down when a lower-index secondary is removed) and set it to the saved spec with role `primary` (promotion replaces the former primary binding; its connector reference and inbound handle are not retained as a secondary); to remove every binding, unset the secondary bindings first using their current slot indexes from `remainingBindings` or `rig gateway human show <entityId>`, then finish at `binding.0` (once all secondaries are gone, the primary is at `binding.0`).
+> - **Shifted slot indexes:** Removal returns `remainingBindings` with their new `binding.<n>` slot numbers (the primary's slot shifts down too when a lower-index secondary is removed). Because indexes shift down after removal, do not blindly retry `binding.0 unset`; inspect `remainingBindings`, the operation receipt, or `rig gateway human show <entityId>` before retrying.
+> - **What removal does:** Removing a binding keeps the person, their preferences, their other bindings, and their open questions. Removing the last binding makes the current delivery episode inactive, so new missing-post findings stop (an already-open recovery item for an undelivered notice may remain); the current delivery-outcome display may disappear too, while historical receipt notes remain. It does not answer, close, or cancel open questions, cancel or reroute already-prepared deliveries, or retract anything a bridge has already accepted. A person with no binding needs a daemon from the release that accepts one (v0.7.9+); do not roll back past it while any fragment has `connectorBindings: []`.
+
+Operator notices use the existing default person. If several people are registered, choose the recipient explicitly when needed:
+
+```bash
+rig gateway default alice
+```
+
+### Step 3: Activate Daemon
+When bindings or secrets files change, restart the daemon process:
+```bash
+rig daemon stop
+rig daemon start
+```
+*(OpenRig does not provide a `rig daemon restart` command).*
+
+### Step 4: Verify Readiness & Status
+```bash
+# Check delivery readiness for all participants
+rig gateway human list
+
+# Inspect detailed readiness status for an individual
+rig gateway human show alice
+
+# Check daemon process status
+rig daemon status
+```
+
+In `rig gateway human show <entityId> --json`, `deliveryReadiness.state` reports `ready` when the binding configuration and local secrets file are valid. This confirms local readiness, not whether the operator bridge process is currently running.
+
+### Step 5: Removing Bindings or Participants
+To remove a connector binding while keeping the participant, their preferences, their other bindings, and their open questions, unset the binding slot:
+```bash
+rig gateway human set alice binding.0 unset --reason "retire relay bridge"
+```
+Before removing a participant altogether, ensure any pending queue items and conversations assigned to them are resolved:
+```bash
+rig gateway human remove alice
+```
+*(The remove verb refuses to delete a participant if open conversations or non-terminal items exist, unless `--force` is specified).*
+
+---
+
+## 6. Troubleshooting Common Conditions
+
+| Condition | Observation | Resolution |
+| :--- | :--- | :--- |
+| **Missing Secrets File** | `deliveryReadiness.state: not-ready`<br>`reason: relay secret file ... is missing or unreadable` | Verify file path in `secretsRef` is absolute and readable. |
+| **Insecure Permissions** | `deliveryReadiness.state: ready` (with warning)<br>`reason: ... warning: ... insecure permissions (mode 0644; group/other access bits are set ...)` | Restrict permissions so group/other have no access (`chmod 600 <secrets-file>` or `chmod 400 <secrets-file>`). |
+| **Identical Directional Tokens** | `deliveryReadiness.state: ready` (with warning)<br>`reason: ... warning: ... uses the same token for OPENRIG_RELAY_OUTBOUND_TOKEN and OPENRIG_RELAY_INBOUND_TOKEN; separate tokens are recommended ...` | Provision distinct outbound and inbound tokens so outbound capture cannot authenticate inbound replies. |
+| **Invalid Loopback URL** | `deliveryReadiness.state: not-ready`<br>`reason: ... OPENRIG_RELAY_URL host must be loopback ...` | Ensure URL uses `http://` and host is `127.0.0.1/8`, `::1`, or `localhost`. |
+| **Reused Inbound Token Across Profiles** | `deliveryReadiness.state: not-ready`<br>`reason: ... reuses an ambiguous OPENRIG_RELAY_INBOUND_TOKEN shared with relay profile(s): ...` | Each distinct relay connector profile (`connectorRef`) must declare a unique inbound token. |
+| **Stale Question Reply** | HTTP 409 on inbound endpoint | Question was already answered, cancelled, or superseded by a newer revision. |
+| **Unregistered Handle** | HTTP 403 on inbound endpoint | The `sender.handle` is not bound to the authenticated `connectorRef`. |
+| **Non-Loopback Inbound** | HTTP 403 on inbound endpoint | Connection originated from a non-loopback TCP socket peer or included a cross-origin browser header. (`Via` and `X-Forwarded-*` headers are ignored). |
+| **Registry or Buffer Read Failure** | HTTP 503 on inbound endpoint (`human_registry_unavailable` or `dispatch_buffer_unavailable`) | Inspect local `gateway/humans/*.yaml` fragments or `<openrig-home>/gateway/dispatch-buffer.json` for syntax/permission corruption before retrying the bridge event. |
+| **Seen-Ledger Corrupt / Suspect** | HTTP 503 on inbound endpoint (`seen_store_unavailable`, includes `seenPath`) | Unreadable/unparseable/unkeyable `relay-inbound-seen.jsonl` refuses all inbound requests; a keyed suspect entry refuses only that specific `eventId`. Follow the conservative recovery procedure below. |
+
+### Conservative Seen-Ledger Recovery (`seen_store_unavailable`)
+
+When `POST /api/gateway/relay/inbound` returns `503` with `error: "seen_store_unavailable"`, the response body and message include the resolved `seenPath` (`<openrig-home>/state/relay-inbound-seen.jsonl`):
+
+1. **Quiesce writers first**: Pause the loopback bridge and stop the daemon (`rig daemon stop`) so no concurrent inbound request or retry appends to the ledger during inspection.
+2. **Preserve a full backup**: Copy `<seenPath>` to a timestamped backup outside `state/` before making any edit.
+3. **Correlate against durable evidence**: Each ledger entry keys `id` by `eventKey = SHA256(JSON.stringify([connectorRef, eventId]))` (64-character lowercase hex digest) and stores `status` as `received:<fingerprint>` or `processed:<fingerprint>`. Because `eventKey` is a one-way hash, correlate any damaged line or keyed suspect `id` against `mission_control_actions.audit_notes_json` (which records `operationId: "relay:<eventKey>:<fingerprint>"`, `connectorRef`, and `eventId`) and bridge delivery logs by computing `SHA256(JSON.stringify([connectorRef, eventId]))` for candidate events—never by guessing or assuming the hash is reversible.
+4. **Apply only evidence-verified, history-preserving repair**: Repair a damaged record only when its exact `id` (`<eventKey>`) and `status` (`received:<fingerprint>` or `processed:<fingerprint>`) are verified from `mission_control_actions.audit_notes_json` or bridge evidence, or restore only from a backup that preserves every subsequently acknowledged event. Do **not** delete, truncate, or skip unknown/unparseable records, and do **not** restore an older backup that omits acknowledged events; if an entry cannot be verified without losing history, leave the ledger in place (continuing to fail closed with `503`) and retain the backup.
+5. **Resume and retry**: After a verified, history-preserving repair is saved at `<seenPath>`, start the daemon (`rig daemon start`), resume the loopback bridge, and retry the original inbound delivery with the same `eventId`.
+
+For automated loopback bridge test reference, see `packages/daemon/test/fixtures/recording-relay.ts`.

docs/reference/getting-started.md

@@ -76,8 +76,17 @@ You need Node.js 22 or 24 and tmux, on macOS or Linux, and one working Claude
 Code or Codex account: one is enough, since the kernel operator adapts the
 shipped `starter` team to whichever you have. On Linux, the distribution's own
 Node.js can be older (Ubuntu 24.04's is 18); install a supported version with
-[nvm](https://github.com/nvm-sh/nvm) (`nvm install 22`) or NodeSource. On a Mac
-with Apple silicon, use Node.js 22 (see the [compatibility
+[nvm](https://github.com/nvm-sh/nvm) (`nvm install 22`) or NodeSource.
+
+For a tarball or manual Node install on Linux, put its bin directory on PATH in
+`~/.profile`, for example `export PATH="/path/to/node/bin:$PATH"`, rather than only
+in an interactive `~/.bashrc`. Bash reads the first of `~/.bash_profile`,
+`~/.bash_login` and `~/.profile`; if one of the first two exists, source `~/.profile`
+there. Other shells use their own login file (for example, zsh's `~/.zprofile`).
+Check `bash -lc 'command -v node rig'`: both paths must appear in the output.
+Use Codex's tool shell for this check if it differs from Bash.
+
+On a Mac with Apple silicon, use Node.js 22 (see the [compatibility
 history](../releases/v0.5.15.md#known-compatibility-limitation)). On a Mac where
 you aren't an administrator, or where Homebrew belongs to another account (its
 `brew install` is refused), install Node.js in your own home folder with
@@ -200,8 +209,9 @@ The same command without `--dry-run`,
 latest published `@openrig/cli` with `npm install -g`,
 runs the Node.js and SQLite check, then `rig setup --dry-run` and `rig setup`. `rig setup` checks both Claude Code and
 Codex and may install a missing one, as described below. A failed step prints
-`FAILED [n/4] <command or check> (exit <code>)`. Where a provider isn't signed in yet, step 4 ends that way and
-`rig setup` lists each sign-in under "Some steps need attention". If the only remaining failures are provider
+`FAILED [n/4] <command or check> (exit <code>)` for steps 1–3. If step 4 is not ready, it reports
+`INSTALLED, SETUP NOT COMPLETE` and points to the details from `rig setup` above. Follow that output and any Fix, then rerun
+`rig setup`; the CLI does not need reinstalling. If the only remaining failures are provider
 sign-ins, the install steps finished: sign in to each selected provider as below, then continue at
 [Start the kernel](#start-the-kernel-and-check-its-state). To install only what the selected providers need, go step
 by step instead:
@@ -260,6 +270,17 @@ opt in with
 `rig config set terminal.inherit_tmux_environment true`. Setup points out this option
 when it detects custom tmux configuration; it never enables it automatically. The setting
 applies to future launches and successors, not running agents or your personal tmux server.
+
+Managed Codex seats run tool commands in a non-login shell: OpenRig passes Codex
+`-c allow_login_shell=false`. Tool commands inherit the environment their pane's shell set up,
+OpenRig's command folder included, so plain `rig` works however OpenRig was installed.
+Inherited setup isn't re-run for each command. Setup that lives only in login or rc files, such as
+lines in `~/.bashrc` that your login file doesn't source, or a tmux `default-command` that doesn't
+start a login shell, may be missing from tool commands, and so may shell functions and aliases.
+To defer to Codex's own setting instead, run `rig config set launch.codex_login_shell true`. That
+omits OpenRig's override rather than forcing login shells, and a login file that sets `PATH`
+outright can then hide `rig`. It applies at each seat's next launch; a restart in place keeps
+the running choice.
 If the scoped environment cannot be prepared, launch keeps its previous environment and
 reports a warning in fresh launch results and the daemon log. For handovers, check
 that log for this warning. This is not a promise that shell startup cannot add or override values.
@@ -291,6 +312,17 @@ fine; if both accounts are authenticated, the kernel uses both even when your
 later project uses just one. A request to use only one provider for everything
 is a separate choice.
 
+`/api/kernel/status` retains the chosen template (`rig.yaml`,
+`rig-claude-only.yaml` or `rig-codex-only.yaml`) across daemon restarts and
+recovery, including starts with `--no-kernel`. It reads the rig's latest rig-spec
+bootstrap record, including a partial launch, so an older kernel's recorded
+selection is also available. This is the original selection; the rig's live spec
+describes its current configuration. The saved source must match a template under
+the current installation’s built-in specs directory. A custom project path with
+the same directory suffix, an old source path that can no longer be verified
+(for example, after a Node version change moves the global install), or a kernel
+without bootstrap history reports an unknown variant.
+
 **Started is not ready.** Read the `Kernel:` boot state in `rig status` and the
 individual seat state in `rig ps --nodes --rig kernel`. The view may open while
 agents finish starting; describe that state, without claiming they are ready.
@@ -483,7 +515,7 @@ different points in the same journey:
 | What you see | Why and what to do next |
 | --- | --- |
 | Codex offers to switch to a cheaper model near a rate limit | This is Codex's own menu. Ask the person to choose in the seat's terminal; do not silently change the team's configured model. OpenRig disables Codex's startup update check for managed launches, but does not answer this model-choice menu. |
-| A missing Node.js, npm or tmux; an install download or command fails | The named prerequisite or install step did not finish. Read its error and the [platform and provider instructions](#install-and-sign-in), resolve that step, then recheck it. `FAILED [4/4]` with only selected-provider sign-ins outstanding means the package installed but those logins still need the person. |
+| A missing Node.js, npm or tmux; an install download or command fails | The named prerequisite or install step did not finish. Read its error and the [platform and provider instructions](#install-and-sign-in), resolve that step, then recheck it. `INSTALLED, SETUP NOT COMPLETE` means the CLI is installed but setup needs attention; follow its output and any Fix, then rerun `rig setup`. |
 | The installing harness refuses a tool call or asks for permission | Its native rules, sandbox or automatic permission decision apply before OpenRig can run. Claude Code's [auto mode can deny calls](https://code.claude.com/docs/en/auto-mode-config) as well as approve them. Show the specific refusal and ask the person to review it in the harness permission controls or run that named step themselves. Keep their chosen restrictions; a refused call does not prove the OpenRig command failed. |
 | Herdr or Ghostty installation is unavailable or declined | Herdr is optional for the view: `--window` uses plain tmux when Herdr is absent, and its install failure is a warning. On macOS, Terminal.app is the fallback when supported Ghostty is absent. An explicitly requested Ghostty install can fail setup; resolve that install or decline it with `--no-ghostty`, preserving earlier choices. Address separate required-tool or login failures. Existing cmux remains supported. |
 | macOS asks for Automation access, or the window action is denied | The command asks the desktop terminal app to open a new surface. Let the person answer the system prompt. Read the error and inspect any new window before retrying an uncertain action; do not silently switch apps and open a duplicate. |
@@ -784,6 +816,19 @@ option, the seat's agent spec for the target runtime, or the runtime's own
 default), what is cleared, and the costs below. Nothing changes until you run it
 without `--dry-run`.
 
+Use `--runtime pi` to move a Claude Code or Codex seat to Pi; use `--runtime
+claude-code` or `--runtime codex` to move back. Pi uses a fresh managed state
+directory. Select a prepared Pi account with `--account-home /absolute/path`
+(the directory must contain readable `auth.json`), or use the target's configured
+login. A source account override or Claude token is cleared only at commit;
+credentials are never converted between runtimes. Pi-to-Pi account changes keep
+Pi settings, but cross-runtime switches resolve settings afresh for the target.
+
+A configured `launch_wrapper` carries across runtime switches among Pi, Codex
+and Claude Code. It must support each successor executable and preserve that
+runtime's credential environment. A wrapped Pi seat cannot switch to Claude
+Code; set `launch_wrapper: []` in the authored spec and re-apply it first.
+
 The new runtime takes over before the old one stops. Its successor starts beside
 the current occupant, reads a handover (the seat's recap, its open rows and the
 predecessor's screen), and confirms it with `rig seat set-runtime <seat>
@@ -815,7 +860,7 @@ What it costs:
   again afterwards to reach whoever holds the seat.
 - **Leftover files after a failed switch.** The target runtime's prepared seat
   storage and projections can remain.
-- **Codex and Claude Code, single-pane seats.** Startup slash commands aren't
+- **Codex, Claude Code and Pi, single-pane seats.** Startup slash commands aren't
   run by a switch; the dry run lists them.
 
 ## Stop your teams
@@ -850,6 +895,7 @@ the lifecycle API without booting a new kernel, then stop the remaining rigs.
 | --- | --- |
 | Tool missing or login fails | Use the specific setup/auth hint; recheck that executable in the launch shell. Do not send work to an unready seat. |
 | "Startup attention" or "Startup details" names a prompt | Answer the login, trust or bypass prompt in that seat's terminal, then run the `rig seat continue <seat>` it shows. |
+| A restored seat still counts as needing attention after its prompt was answered | Its restore outcome is still `attention_required`, and answering the prompt or running `rig seat continue` doesn't change it. Once the agent is up in its pane, run `rig seat clear-attention <seat>`; see [After a restored seat's prompt](agent-startup-guide.md#after-a-restored-seats-prompt). |
 | "Readiness timeout after 30s — harness did not become interactive" | The harness took longer than the readiness window. Raise it for the next launch with `rig config set runtime.readiness_timeout_seconds <1-600>`, then relaunch that seat. |
 | "Claude resume did not finish within Ns; Claude may still be starting, so the launch was kept" | A restored Claude seat was not interactive when the wait ended. Its pane and resume token are kept. Look at the pane; once Claude is up, run `rig seat clear-attention <seat>`. The wait is about `runtime.readiness_timeout_seconds` in total for each seat (default 30 s), counting the first 3.2 s, and the time adds up seat by seat in a restore. Raising the setting for slow fresh launches lengthens it. If `rig up` or `rig launch` times out at 120 s, that is not a rollback: check `rig ps` before running it again. |
 | Daemon is healthy, kernel is still starting | Read `rig status` and `rig ps --nodes --rig kernel`; kernel readiness is separate. |
@@ -1129,8 +1175,15 @@ rig up ./openrig-specs/rigs/launch/starter/rig.yaml --cwd .
 
 OpenRig's `permission_policy: builtin:yolo` setting selects
 `-s danger-full-access -a never` on fresh, resume and fork launches, replacing
-the named-profile argument. The profile recipe above remains useful when you
-want to maintain those choices in native configuration. The legacy
+the named-profile argument. That drops the whole profile, not only its sandbox
+and approval settings. To keep a named profile with full access, put full access
+in the profile itself and leave `permission_policy` absent, as in the recipe
+above. To keep it with a workspace-write or read-only posture that OpenRig
+records, use the member's `codex_permissions` (see
+[per-member Codex permissions](rig-spec.md#per-member-codex-permissions)):
+OpenRig passes `-p` and adds that envelope's flags, and for that seat the
+envelope replaces a policy posture such as `builtin:yolo`. Full access isn't an
+envelope. The legacy
 environment-only `OPENRIG_YOLO=1` path remains sandbox-only when no resolved
 policy is present. A standalone `codex --yolo` command is not an OpenRig setting.
 At full access, Codex can show its full-access and GPT-5.1 migration notices at

docs/reference/harness-runner-contract.md

@@ -6,13 +6,30 @@
 > - **Released in 0.7.8:** seat activity grants (section 4), and OpenRig's side of the answered permission path
 >   (section 7).
 > - **Runner-author behavior, not yet qualified:** the input, cancellation and permission recipe in sections 5-7, and
->   the runner-side shutdown obligation in section 9. A dsh port implementing this recipe is not qualified.
+>   the runner-side shutdown obligation in section 9. Two runners have shown parts of it natively (below); no
+>   runner is qualified beyond that.
 > - **Existing interfaces with author obligations:** launch-result identity, exact resume, first input and model
 >   handoff. These interfaces do not establish that a particular runner implements them correctly.
 > - **Unchanged:** plain community presets and the built-in runtimes (Claude Code, Codex, Pi, OMP, terminal).
 >
-> No harness has been run end to end under this contract yet, and the native verdict is pending. The first candidate
-> is dsh over ACP.
+> **What has run natively, in 0.7.8's release testing, against a Claude Haiku model:**
+> - **dsh-seat 0.3.0**, an outside runner for dsh 0.2.1-rc.2 over ACP: launch and first input, useful queue work, a
+>   multiline message with a `/abort` line inside it delivered as one prompt, activity that stayed working through a
+>   52-second quiet tool interval and then read idle, cancel by a bare `/abort` typed in the pane, and the same
+>   conversation resumed by `session/resume` after `rig down --snapshot` and `rig up`, with the runner and its child
+>   gone after the managed down.
+> - **The dsh example runner** (`community-harness-examples/dsh/`): dsh's native `session/request_permission` reached
+>   a person, whose allow-once took effect and whose reject-once didn't.
+>
+> **Not shown natively:** a DeepSeek model, the example runner's own resume (its preset declares none), the default
+> permit when nobody can be asked (checked without a model only), a typed or Slack answer, a second person's
+> answer, reboot recovery, and the model OpenRig observes. Any other runner, provider or path is unqualified until
+> it's shown.
+>
+> **Since 0.7.8, on one stack:** the dsh example runner has run with a real dsh 0.2.1-rc.2 on dsh's pi-ai OpenRouter
+> chat-completions profile with `deepseek/deepseek-v4-flash`: queue work, activity, an allow-once and a reject-once
+> answered by an automated test operator, cancel, and an exact return with resume added to a copy of the preset. A
+> missing conversation's refusal hasn't been run, and the preset still declares no resume.
 >
 > Release status and pinned references are in the [joined author guide](author-contracts.md). The
 > [community presets reference](community-harness-presets.md) describes the implemented fields, including the runner
@@ -61,7 +78,20 @@ limits:
 - **A preset without `runner` behaves exactly as today,** and so does a session recorded before the field existed.
 - **The other fields keep their existing meanings:**
   - `first_input: true` passes the seat's first startup turn in a file named by `OPENRIG_FIRST_INPUT`;
-  - `passes_model: true` passes the seat's declared model as `OPENRIG_MODEL`.
+  - `passes_model: true` passes the seat's declared model as `OPENRIG_MODEL`;
+  - `launch_wrapper: true` declares that the runner applies a member's configured wrapper to its native child.
+
+**Launch wrapper.** `launch_wrapper` is an optional boolean in format 1 and requires `runner: true`. When the member
+has a wrapper and its runner preset declares `launch_wrapper: true`, OpenRig passes `OPENRIG_LAUNCH_WRAPPER` as a JSON array of resolved argv strings.
+The runner must exec its native child as `<argv...> <agent> <args>`, preserving argument boundaries, and keep its own
+process outside the wrapper. The runner must remove `OPENRIG_LAUNCH_WRAPPER` from the environment it passes to the wrapper
+and native child; the handoff belongs to the runner only. The wrapper must `exec` the agent in place. If wrapping fails,
+the runner must never fall back to an unwrapped spawn. A runner preset without the flag is refused at preflight and launch (not schema validation) when a wrapper is
+configured: `launch_wrapper unsupported: the community preset "<runtime>" runs a runner without launch_wrapper: true; no agent started. Add launch_wrapper: true to the preset if its runner applies OPENRIG_LAUNCH_WRAPPER, or remove launch_wrapper from this member.`
+A plain preset needs no flag: OpenRig prefixes its command
+immediately before the preset's native command, after OpenRig's launcher assignments (for example,
+`/usr/bin/env OPENRIG_LAUNCH_RESULT=… <wrapper> <preset.command>`), at the same seam as Codex. OpenRig never sets
+`OPENRIG_LAUNCH_WRAPPER` for a plain preset and removes any inherited value.
 
 ## 3. Identity: report the conversation once, in the launch record
 
@@ -244,7 +274,7 @@ that runs without asking gains no approval step. Ordinary conversation and `rig
 |---|---|
 | **There's a default person** (chosen with `rig gateway default`, or the only registered person), and no override | They get an ordinary ask in the Feed and on Slack. No setup is needed. |
 | **The seat's operator chose someone** (`rig seat set-permissions-to <seat> --to <person> --reason "…"`) | That person gets the ask instead. |
-| **No one can be asked** | **Permitted at once**, with the agent's own allow option. No ask is opened. That covers: no registered person; several and none the default; the chosen person no longer registered; the people unreadable; or options a person can't be shown (bad ids, or not 2-4). The agent's `allow_once` is used, else its `allow_always`. If the agent offers only `allow_always`, the harness then remembers a permission nobody chose. |
+| **No one can be asked** | **Permitted at once**, with the agent's own allow option. No ask is opened, and OpenRig records the permit (below). That covers: no registered person; several and none the default; the chosen person no longer registered; the people unreadable; or options a person can't be shown (bad ids, or not 2-4). The agent's `allow_once` is used, else its `allow_always`. If the agent offers only `allow_always`, the harness then remembers a permission nobody chose. |
 | **The seat's operator opted in to refusal** (`rig seat set-permissions-to <seat> --deny --reason "…"`) | Refused at once and held as needs input, as in the first version. |
 
 **What the ask shows:**
@@ -263,7 +293,7 @@ that runs without asking gains no approval step. Ordinary conversation and `rig
 | replies with just a no ("no", "deny", "reject", "stop" …), or denies it in the Feed | the agent's reject (`reject_once`, else `reject_always`), or, with no reject option, cancelled. A person's no is never a grant |
 | routes or hands it to another person | the ask waits for that person, who gets the same options; their answer counts as if they'd been asked. The handoff itself never permits |
 | routes or hands it to an agent seat | that seat gets no options. The request waits until that row is closed, by the seat or a person. Closed as done without a choice (including a drop), or as failed, is no answer, so it's permitted. Closed as denied (the Feed's deny, `--closure-reason denied`, or the state `denied`) is a no. A row set to the state `canceled` cancels the request, and a further handoff is followed; a person it reaches gets the same options. Or the request waits until the agent's turn ends |
-| closes it without choosing: drop, approve, or other typed words | no answer, so the default applies: **permitted** with the agent's allow |
+| closes it without choosing: drop, approve, or other typed words | no answer, so the default applies: **permitted** with the agent's allow, and OpenRig records the permit (below) |
 | doesn't answer | the request **stays pending and visible** (the seat reads needs input) until the agent's turn ends. It's never timed out into a grant. |
 
 - **An answer reaches only the live request it was asked for,** once. A request from an earlier occupant reads
@@ -271,6 +301,17 @@ that runs without asking gains no approval step. Ordinary conversation and `rig
 - **Text typed into the runner's pane, an agent's own text, and `rig send` are never an answer.** While a request waits
   on a person, the seat reads needs input, and sends are normally refused. A send that does get through reaches the
   agent as a prompt.
+- **From 0.7.9, each permit OpenRig decides by default is recorded,** once, as a stream item from the seat. That covers
+  no one to ask, and an ask closed without a choice (the row passed to an agent seat included). The record names:
+  - the tool call;
+  - why no person answered, including, for a closed ask, who closed which row and how;
+  - the option the agent got, with its kind. An `allow_always` option is marked as remembered by the agent's harness;
+  - the conversation and request, or the ask.
+
+  For an ask opened before 0.7.9, the record may omit the option kind and whether the harness remembers it.
+
+  Read the records with `rig stream list --tag runner-permission-permit`, or `--source <seat>`. A record wakes no one,
+  and one that can't be written doesn't change the permit.
 
 **The runner's protocol:**
 1. **Submit** the agent's live request to `POST <base>/api/activity/permission-requests`, with the same token or seat
@@ -330,20 +371,21 @@ answer. Only the seat's opt-in refuses and holds attention.
 
 **When the runner can't get an answer from OpenRig,** nobody can be asked, so it permits with the agent's own allow
 option: `allow_once` when offered, else `allow_always`. Prefer `allow_once`, so the harness doesn't remember a
-permission nobody chose. That covers:
+permission nobody chose. Such a permit is the runner's own, so OpenRig has no record of it: the runner prints it in
+its pane. That covers:
 - an OpenRig without this path, which answers 404;
 - an OpenRig the runner can't reach, at submit or while reading (a runner may retry a read that got no answer a few
   times first);
 - a refusal the table marks "permits", and any status or code this section doesn't list.
 
-**Limits in 0.7.8:**
+**Limits of this path:**
 - **The answer owner is a registered person.** Choosing a seat (another agent) to answer is a later change.
 - **A colleague's answer is followed on this host.** If the ask is handed to another host, the request waits, shown as
   needs input, until the agent's turn ends.
 - **An override or `--deny` is set with the command above, never in a rig spec.** A spec field is a possible later
   change.
-- **Permits and answers aren't recorded beyond the runner's pane and the ask itself.** A permit with no one to ask
-  opens no ask.
+- **A person's answer is recorded on the ask itself,** and a default permit as the stream item above. That item lasts as
+  long as stream retention keeps it. A permit the runner decides on its own is in its pane only.
 
 ## 8. Resume: exactly the same conversation, or say honestly that you couldn't
 

docs/reference/help.md

@@ -97,6 +97,24 @@ missing tools or logins, kernel startup, a closed viewing terminal and recovery
 before choosing an action. A healthy daemon doesn't by itself mean the project's seats are ready: check
 `rig ps --nodes --rig <rig-name>`.
 
+If the whole team stopped when a herdr terminal closed, OpenRig or its tmux server was started from inside that
+terminal: herdr-core ends every process started there when the terminal closes. `rig daemon start` no longer passes a
+herdr terminal's scope on, but processes started before that keep it until they restart. OpenRig never restarts them
+for you; restart at a time you choose, from a terminal outside herdr.
+
+With `terminal.tmux_server` at its default (`inherit`), OpenRig puts seats on the tmux server of the shell that
+started it, usually your own default server. If that server was already running, and you had started it from a herdr
+terminal, it carries that terminal's scope. OpenRig can't prevent this, because the server is yours.
+The same applies when a command like `rig chatroom watch`, run in a herdr terminal, starts the default server. Start
+tmux outside herdr, or use `terminal.tmux_server isolated` (see `tmux-servers.md`), which gives OpenRig a server it
+starts itself.
+
+The daemon keeps `HC_SESSION` and `HC_SOCKET`. The herdr-core they select is the one the herdr-app provider starts
+viewers in; `rig terminal status --provider herdr-app` names it (`core` and `assumption`), as do open failures
+(`viewer_start_failed`, `herdr_app_unavailable`) and the note when a window has not confirmed showing a view. If your
+herdr-app window is attached to a different herdr-core than the daemon's, its viewers will not appear there; start the
+daemon with that core's `HC_SESSION` and `HC_SOCKET`.
+
 ### The agent is waiting for permission or can't reach the daemon
 
 Read [Have your agent configure permissions](getting-started.md#have-your-agent-configure-permissions)

docs/reference/herdr-app.md

@@ -0,0 +1,145 @@
+# OpenRig views in herdr-app
+
+herdr-app (Herdr 2, the desktop app on herdr-core) can show OpenRig's terminal views inside its own window. A view
+opens as a **Rig view**: an equal grid of the view's seats, each with a title row (`<agent> · <slice>`), padded to a
+full grid. Rig views live in OpenRig's own herdr-app layout, **Rig** (`layouts/rig.json`), beside your tabs. A strip
+above them switches between **Your tabs** and each view.
+
+Nothing here needs a desktop display or SSH: the view opens in the herdr-app window the request came from.
+herdr-app has to run on the same machine as the OpenRig daemon: OpenRig reaches it through its control socket
+(`<app home>/herdr-app.sock`) and writes its mods and layout into that app home. An app on another machine, connected
+to this machine's herdr-core over SSH, isn't supported.
+
+## Set up
+
+Nothing to install by hand: opening a view installs OpenRig's four mods (rig-strip, rig-cell-header, rig-pad,
+rig-ended) and updates them when OpenRig is updated.
+
+- The mods go to `<app home>/mods/rig-*`. The app home is herdr-app's own: `HERDR_APP_HOME`, else
+  `$XDG_CONFIG_HOME/herdr-app`, else `~/.config/herdr-app`. `--app-home <dir>` on `rig terminal open` names another.
+- OpenRig owns only the files it ships there (recorded in `<app home>/openrig/mods.json`). An update replaces those
+  and nothing else in the folder. Mods you install yourself, from any source, are never touched.
+- To move between a view's cells from the keyboard, bind the strip's actions in `keys.json`, section `layout:rig`
+  (your agent can add it, as for any herdr-app key):
+  `"layout:rig": { "alt-left": "rig-strip.left", "alt-right": "rig-strip.right", "alt-up": "rig-strip.up", "alt-down": "rig-strip.down" }`.
+  In the Rig layout these replace herdr-app's `global` defaults for the same keys (`pane.go`, `agent.prev`/`agent.next`),
+  which keep working on your other layouts. A key bound in `terminal` wins while a cell has focus, so pick keys that
+  aren't bound there. Once you edit `keys.json`, herdr-app treats it as yours and stops adding new default shortcuts.
+- An open never replaces a mod you changed: that folder stays as it is, the view opens anyway, and the open's notes
+  name the folder and why. Only a mod a view needs that is missing and can't be installed stops an open.
+
+## Open a view
+
+- **From the TUI inside herdr-app:** Terminals, choose a view (Enter previews it), then **Open terminals**. The TUI
+  recognises that it runs in a herdr-core terminal of a herdr-app window on this machine, and the preview and the
+  open both use that window, so the plan you previewed is the one that opens. The preview says the view opens in
+  this window; it shows no desktop or SSH recipe. A shared TUI viewed from a herdr-app terminal does the same. The
+  kernel view opens in its narrow arrangement there (one seat per page): the window's width isn't the TUI's.
+- **From a shell:** `--provider herdr-app` opens the view in the herdr-app window you run it in. Inside a herdr-app terminal, plain `rig terminal open <view>` does the same when that window is answering. `--window` always asks for a new desktop window.
+
+  | Command | Where the view opens |
+  |---|---|
+  | `rig terminal open <view> --provider herdr-app [--app-home <dir>]` | In the herdr-app window this runs in (requires the window to answer). |
+  | `rig terminal open <view>` | Inside a herdr-app terminal when the window is answering, in that herdr-app window; otherwise a desktop window. |
+  | `rig terminal open <view> --window` | Always asks for a new desktop terminal window, even inside a herdr-app terminal. |
+
+The herdr-app window must be answering on its socket (`<app home>/herdr-app.sock`); otherwise an explicit open fails and says
+so. Finding the socket file is how OpenRig guesses you are in a herdr-app window; whether the window is reachable is
+checked by asking it. When a plain `rig terminal open <view>` in a herdr-app terminal finds that the socket does not
+answer, it says so and opens a desktop window instead; `--provider herdr-app` never
+falls back to a desktop window or an SSH recipe.
+
+What an open does, in order:
+
+1. Each viewer is recorded first (`$OPENRIG_HOME/private/herdr-app/clients/`), then started by herdr-core outside
+   every tab (`hc run --bg --no-parent`, labelled `openrig.*`). A local viewer runs the view's tile for its seat, as in
+   [terminal sessions](terminal-sessions.md): it shows the seat. An interactive cell starts with the read-only tile and
+   takes no input until herdr-app has shown the view (step 3); then its writable tile takes over, and the first thing
+   you type attaches it. A read-only cell only shows. A seat on another host opens as its own ssh attachment, writable
+   or read-only as that view composes it. If the seat's session ends, the viewer says so and waits; it never leaves a
+   shell in the cell.
+2. One atomic write of `layouts/rig.json`: your current layout, with OpenRig's stack beside your tab block. Your own
+   layout file is never written. A `rig.json` OpenRig didn't write is never overwritten.
+3. The Rig layout is shown, the view is selected once herdr-app confirms it shows the new file, and the keyboard
+   goes to the first live cell. Once herdr-app has acknowledged the layout with the view, OpenRig writes the gate file
+   of each of the view's interactive cells, on every page (including pages the strip hasn't shown yet, since the strip
+   moves between pages on its own), and each cell takes input a fraction of a second later (it checks every 0.2 s).
+   Anything that reached a cell before that, typed or written to its terminal (`hc send`), is dropped, not delivered
+   late. The open reports a view herdr-app
+   hasn't confirmed as unconfirmed, not opened; its viewers keep running and it stays in the Rig layout, and its cells
+   take input once herdr-app shows it. OpenRig keeps watching for 10 minutes, across a daemon restart; if herdr-app
+   hasn't shown it by then, each interactive cell says so, takes no input, and names the command that opens the view
+   again (`rig terminal open <view> --provider herdr-app`).
+
+Other layouts and other herdr-app windows don't show Rig views. Each window (app home) has its own.
+
+## The keyboard
+
+- An open puts the keyboard on the new view's first live cell.
+- Entering a view again (a click in the strip, or herdr-app restarting with it shown) returns the keyboard to the cell
+  you were last in there, else the first live cell.
+- **Your tabs** gives the keyboard back to your own pane.
+- OpenRig's own rewrites (another view opening or closing, a cell ending) never move the keyboard, whether you are in
+  a view or in your tabs.
+- Typing in an interactive cell, once the view is shown, attaches its tile to the seat; after 30 seconds without a key
+  it goes back to showing it. When the seat's session is replaced (a handover, a fresh launch, a restart), the tile follows the same seat to
+  its new session, as every OpenRig tile does.
+
+## What OpenRig keeps in the app home
+
+All under `<app home>/openrig/`:
+
+| File | What it holds |
+|---|---|
+| `mods.json` | The files OpenRig shipped to `mods/rig-*`, so an update replaces only those. |
+| `ack.json` | Written by the strip: the last layout revision herdr-app shows and the terminals it references. A viewer ends only once this shows its cell is gone. |
+| `strip.json` | Written by the strip: the cell you were last in per view, and the view shown. herdr-app gives a mod new storage whenever its props change, which is every OpenRig write, so this is kept as a file. |
+
+## Close a view
+
+```sh
+rig terminal close <rv-id> [--app-home <dir>]     # the id is in the open result's `space`
+```
+
+The view's cells leave `rig.json` first. A viewer ends only after herdr-app confirms it shows a layout without it, or
+when the Rig layout isn't shown (herdr-app then reads the new file when it is). Only viewers OpenRig started (its
+record, its labels and its exact command all match) are ended. Seats are never touched. Anything uncertain is left
+running and named in the result.
+
+If a viewer ends some other way (killed outside OpenRig, herdr-core restarted), OpenRig rewrites its cell as an
+ended cell straight away. Views OpenRig recorded keep that owner across a daemon restart: when the daemon starts, it
+reconciles each app home it has views in (a viewer that ended while it was down is rewritten then) and finishes any
+pending ends. If `hc events` can't run (herdr-core missing or down), the watcher tries again after 5 s, then waits
+twice as long each time, up to 5 minutes. Stopping the daemon ends only its own `hc events` watcher and timers; viewers
+and views stay.
+
+## Which herdr-core
+
+```sh
+rig terminal status --provider herdr-app [--app-home <dir>]
+```
+
+`core` and `assumption` in the result name the herdr-core the daemon's `hc` runner starts and ends viewers in: the
+daemon's `HC_SESSION` (else `default`) and `HC_SOCKET` if set when the daemon started. The herdr-app window must be a
+client of that same core. Terminal addresses are per core, so a window attached to another core would show other
+terminals, or none, in the view's cells. Open failures (`herdr_app_unavailable`, `viewer_start_failed`) and the note
+when a window has not acknowledged a write also name the daemon's core (`HC_SESSION`/`HC_SOCKET`) so a mismatch is
+visible straight away.
+
+A status or a preview only reads: it starts no herdr-core process and creates nothing. OpenRig watches herdr-core's
+events (`hc events`) and creates its viewers folder only when it opens or closes a view, or when it starts with views
+already recorded.
+
+## Limits
+
+- A Rig view is a layout inside herdr-app, not a herdr tab: other windows and other layouts don't show it.
+- Following a task (`--follow`) isn't available in herdr-app yet.
+- When a viewer ends unexpectedly, herdr-app removes its block and collapses the layout, which briefly opens the Files
+  panel on that cell's folder until OpenRig's rewrite restores the layout. Viewers start in OpenRig's own empty
+  folder (`$OPENRIG_HOME/private/herdr-app/viewers`). Planned closes never trigger this.
+- OpenRig doesn't yet check that the window runs on the daemon's herdr-core (`core` in `rig terminal status`); it
+  relies on both using the same session on the same machine.
+- `POST /api/terminal/open` (and `POST /api/terminal/close`) carries no bearer middleware, matching the daemon's other
+  open routes; an open writes `mods/rig-*`, `openrig/mods.json` and `layouts/rig.json` into the named `--app-home` once
+  `herdr-app` answers a `status` request on `<app home>/herdr-app.sock`.
+- Runtime-tested on Linux only; macOS and Windows haven't been tried on real machines.

docs/reference/hosted-members.md

@@ -95,7 +95,7 @@ These facts stay separate on the row:
 - **Result:** the closing transition and its resolution, or the handoff.
 - **Notification:** where a person asked, Slack's posted receipt. A posted notice isn't the person's acknowledgment.
 
-**On the team:** `rig_team` lists the rig's hosted members (roster entries marked hosted that aren't one of its nodes), with their capabilities and caveat, no pane, and activity `unknown`.
+**On the team:** `rig_team` lists the rig's hosted members (roster entries marked hosted that aren't one of its nodes), with their capabilities and caveat, no pane, and activity `unknown`. With an external grant offering `team/list`, call it without `rigId` to discover the connected team. It selects the grant's sole rig, or returns only its granted names when a choice is needed.
 OpenRig doesn't observe a hosted member between its calls, so it never reads idle. `rig roster find` shows it the same way.
 - **Presence:** each entry carries `presence`, which says what OpenRig last saw of the member. It isn't pickup, and `null` means unknown:
   - `lastCallAt`: its last authenticated call to this host's `rig mcp serve --http`, recorded at most once a minute in
@@ -119,6 +119,7 @@ and nothing else is offered:
 | `queue`/`claim`, `queue`/`update` | `rig_queue_claim`, `rig_queue_update` |
 | `team`/`list` | `rig_team` |
 | `artifact`/`read` | `rig_artifact_read` |
+| `question`/`read`, `question`/`answer` | `rig_question_list`, `rig_question_answer`, only when the grant names its person (below) |
 
 - **One seat per connected assistant:** the service gives each connection its own seat and keeps it across that
   connection's reconnects. The seat must be one of the rig's nodes; a node created without a launch is enough. The caller
@@ -128,6 +129,10 @@ and nothing else is offered:
   - It reads the items it sent and the items addressed to it from those seats, never the team's internal items.
   - With `queue`/`claim` and `queue`/`update`, it claims and closes only items addressed to its own seat that it can read,
     and waits (blocked) only on an item it names and can read. Its answers hide what reading the item hides.
+- **A first task without a seat address:** with `--task-destination <member@rig>` on the same `rig mcp serve`, a
+  `rig_queue_create` that leaves out `destination` goes to that receiver. The operator sets it; a caller can't. It must be
+  exactly one existing local node within that request's grant, or nothing is written. An explicit `destination` works as
+  before.
 - **Two assistants together:** a grant with `team`/`list` sees its rigs' seats with `rig_team`, and may address and hear from
   another connected assistant's seat. A grant without it never reaches a seat in the service's caller form
   (`cus-<id>-<connection>@<rig>`), so an existing connection keeps the reach it had.
@@ -139,8 +144,45 @@ and nothing else is offered:
   - Reading or listing an item records nothing, and nothing wakes an assistant, so an item stays queued until its receiver
     claims it.
 
+### A person answering the team's questions through their assistant
+
+A grant can name the person who owns the connection: `person.entity_id`, their gateway entity id on this instance, so their
+address is `<entity_id>@external`. The service sends it only for a connection that person owns. With it and the `question`
+pair, the person's assistant can read the team's questions waiting on them and submit their answers, without opening
+Steering. The caller's seat is never taken for a person: without `person`, the `question` pair offers nothing, and an
+unusable `person` withholds only these two tools while the grant's other tools keep working.
+
+- **Register the person** on the instance, once:
+  `rig gateway human add <entity_id> --display-name "<name>" --delivery-class B --default`.
+  Without `--binding`, they're asked in the Feed and on Steering only. As the default person, today's asks addressed to no
+  registered person (such as `human@host`) are theirs too.
+- **`rig_question_list`** shows what Steering shows that person as answerable: an ask addressed to them, or a row parked on
+  them. Each comes with its options when it has any, and the `revision` an answer names. A question that moved to someone
+  else between Steering's view and the read of its row is left out.
+- **`rig_question_answer`** takes the question, the revision the person saw, and either the options they chose (one per
+  question) or exactly what they typed. It records through Steering's own write path (`answer` for an ask, `resolve` for a
+  park), as the person, so Steering shows the same answer and the asker's work continues once. The answer's audit names the
+  route it came by: the caller's seat, the grant's principal and the grant's id and version.
+- **What it says back:** a success names what was submitted (`submitted`), never a fresh read of the question's answers.
+  - `answered`;
+  - `recorded`, when parts of the question are still unanswered;
+  - `already-recorded`, for an identical retry whose submission is already on the question, recognised by its operation id
+    (never by an equal answer). That includes a settled ask or a park that's back at work;
+  - `question_changed`, with the current question, when someone else changed it since the revision the person saw;
+  - a retry of a submission that stopped partway confirms the parts already recorded and stops there (`question_changed`,
+    with `recorded` and `notRecorded`). The rest goes in a new answer to the question as `rig_question_list` shows it now,
+    so a stale retry never overwrites an answer someone gave in the meantime, including one recorded with no action (a
+    Slack click);
+  - a refusal partway names which parts were recorded (`recorded`, `notRecorded`). A send that got no answer, or a server
+    error, leaves that part `indeterminate`: send the same call again to confirm what was recorded;
+  - `not_waiting_on_person`, when it was answered or closed, or isn't theirs;
+  - `person_not_registered`, when the grant's person isn't registered on this instance. While no one is registered, Steering
+    shows every request, so nothing is answered for anyone.
+- **What it never does:** answer another person's question, reach a team-internal row, or record a decision of its own. A
+  delegated decision (`rig_queue_decide`) isn't offered to a grant-service caller.
+
 ## Limits today
 
 - **Webhook triggers only:** OpenRig rings a webhook. Email and other triggers aren't supported yet. A member without a trigger
   isn't rung, and its person is told instead.
-- **It can't act as a person:** it works rows addressed to its seat, a person's request to it included. It can't settle an ask addressed to a person, or hand work to one.
+- **It can't act as a person:** it works rows addressed to its seat, a person's request to it included. It can't settle an ask addressed to a person, or hand work to one, except through a grant service's grant that names its person, which records that person's own answers (above).

docs/reference/instance-layout.md

@@ -110,6 +110,17 @@ limitation: use the live spec-library commands to determine where a spec is
 served from, and do not infer convergence merely because the canonical
 directory exists.
 
+The spec library scans the built-in specs, `$OPENRIG_HOME/specs` and, when
+present, that legacy library. At any depth below them it skips directories
+named `.git`, `.worktrees`, `node_modules`, `dist`, `build`, `.turbo` and
+`.next`, which usually hold build output or dependencies (`SKIP_DIRS` in
+`packages/daemon/src/domain/spec-library-service.ts`). A spec inside one of
+them, such as `specs/build/rig.yaml`, isn't listed, and nothing reports that it
+was skipped. To list it, rename its directory and run `rig specs sync`.
+`rig specs add <directory>` copies a directory under its own name, so adding a
+directory named `build` doesn't help. `rig up <path-to-rig.yaml>` still launches
+a spec by its path, without the library.
+
 Config updates publish a complete replacement file atomically where directory
 permissions and the filesystem allow it. On POSIX systems, replacement preserves
 the existing owner, group, and read/write/execute permission bits. If creating or

docs/reference/keeping-a-watch.md

@@ -59,7 +59,9 @@ rig queue block <task> --on external:<what you check> --wake-after 1h \
 ```
 
 `--on` is always required; `external:<what>` names the outside thing you're waiting on. The timer first fires after
-the interval, not now.
+the interval, not now. Left without `--wake-after` (or `--wake-watchdog`), the task is still set aside, but nothing
+will wake you, and the reply says so with a `park_without_wake` advisory. Set it aside again with `--wake-after` to add
+the timer.
 
 **4. Say back what you set, in one or two plain lines:** what you'll check, how often, how they'll be told, and when
 it stops. For example: "I'll check the release PR every hour and message you here once it merges. Then I'll stop."
@@ -107,6 +109,28 @@ rig queue create --destination <entityId>@external --human-intent update \
 - **`posted` means the connector posted it, not that they read it.** Don't say they saw it.
 - **If delivery fails or is indeterminate,** inspect that task before anything else. Never resend blindly.
 
+## If the timer cannot reach its recipient
+
+A `periodic-reminder` occurrence that definitely fails before input remains owed. The watchdog
+retries through its existing loop, using `queue.wake_retry_interval_seconds` and
+`queue.wake_retry_cap` (by default, three retries five minutes apart). A failed attempt does not
+advance the reminder to its next regular cycle. A confirmed send advances the normal interval;
+recovery does not replay every elapsed cycle.
+
+`rig watchdog show <jobId> --json` includes the failed attempt, its retained outbox IDs, retry
+time and next action. If input may have reached the recipient, that occurrence ends unconfirmed
+and is not automatically sent again. Inspect the recipient and exact outbox record before any
+manual replay. A recurring reminder keeps its next normal cycle under a new occurrence identity. A confirmed receipt survives a daemon restart, so
+finishing its accounting does not type the same occurrence twice.
+
+Exhausted retries end only the failed occurrence, with its cause and next action; a recurring
+reminder keeps its next normal cycle. A permanent transport failure leaves a terminal job:
+correct the destination and register a replacement reminder. A one-shot park timer records its
+failed or unconfirmed wake on the parked row before it ends. A retryable failure is recorded
+there while the timer stays armed. Deliberately held
+input keeps its existing retained-message custody. Condition watches and parked-owner policy
+behavior are unchanged. This does not replay failed occurrences recorded by older versions.
+
 ## "What are you watching for me?"
 
 List **every active state**, not just set-aside ones. A watch being checked can be in progress, and a new one can be

docs/reference/pi-tool-selection-recovery.md

@@ -25,7 +25,8 @@ that restores nothing is not itself evidence of accidental loss.
 
 ## Choose the interface the seat actually has
 
-Pi 1.4.0 includes an opt-in selector at `examples/extensions/tools.ts`. Its
+Pi 1.4.0 includes an opt-in selector at
+`packages/coding-agent/examples/extensions/tools.ts` in the Pi repository. Its
 `/tools` command remembers selection per branch and needs Pi's own TUI. Use it
 where that extension and native TUI are available. **Do not send `/tools` to an
 agent OpenRig drives over RPC**: that interface cannot operate the selector.

docs/reference/project-workspace.md

@@ -7,6 +7,10 @@ without daemon-internal knowledge.
 This subtree is part of the wider [OpenRig Instance Layout](instance-layout.md).
 The instance initializer delegates these exact workspace bytes to this owner.
 
+To inspect or read an exact reconstruction under a selected work home, use
+[`rig context reconstruction`](reconstructions.md). It shares catalog and work-home
+resolution with `context work-install` and performs no installation or projection.
+
 ## Default Shape
 
 `rig config init-workspace` creates the default workspace at

docs/reference/queue-erasure.md

@@ -0,0 +1,161 @@
+# Queue erasure: closed-work retention and account deletion
+
+A daemon that serves several accounts sometimes needs to remove queue content for good: closed work after its retention window, or everything that belongs to an account when the account is deleted. `rig queue erase` and three retention settings do that.
+
+Erasure can't be undone. Run every scope as a dry run first, without `--apply`. The dry run prints exactly what `--apply` would erase and changes nothing.
+
+## What is erased
+
+A queue row is erased in one of two ways:
+
+- **Scrub.** The row's content is removed, and the row itself stays, marked `contentErasedAt`. A retried create with the same id still finds the row, so it makes no new work. If anything writes to the row after its scrub (a note, a summary, a reopen and close), the next run erases that content too.
+- **Delete.** The row and everything linked to it are removed. Only its id is kept, in an erased-id record with no endpoints and no content. A later create that reuses the id is refused with `qitem_erased` (409), so a retried or forwarded create can't bring the work back.
+
+| A scrub clears | A scrub keeps |
+|---|---|
+| body, summary, resolution, tags, blocker text, chain of record, evidence reference, human detail, questions and answers, delegated decision, the input-box reading | id, source and destination, state, priority, tier, times, closure reason, handoff links |
+| a closure's target when it's a reason or note, in the row, both transition tables and events. That's every target under a reason other than the two at right, and any target under those two that isn't a session address or a queue id | a closure's target that names a successor: under `handed_off_to` or `superseded`, shaped as a session address (`member@rig`, optionally `@host`) or a queue id (`qitem-…`) |
+| transition notes and recorded principals, live and archived | each transition's state, actor and time |
+| events' `summary`, `reason` and `error` (the event keeps its type and ids, marked `contentErased`) | |
+| mission-control audit text and snapshots | |
+
+**The row's own copies go with it:**
+- its transitions;
+- events whose payload names it;
+- its mission-control audit rows.
+
+**Copies held with someone else go only when they're the owner's.** These are outbox and inbox entries that point at the row or wake for it, and stream reader actions and dispositions that name it (including `duplicate_of`). For account deletion, the owner is the account: both of a copy's sessions must be members. Otherwise the owner is the row's own source and destination. Any other copy is kept and listed under `keptCopies`:
+- `mixed`: one of its sessions is the owner's;
+- `foreign`: none of them are.
+
+**A row a workflow still references is never erased.** That covers a step trail, a frontier binding, a failure record, or a live instance's frontier. It's reported under `gaps` with the code `workflow-reference`.
+
+## Scopes
+
+Name exactly one.
+
+| Scope | Erases | Action |
+|---|---|---|
+| `--members <file>` | Account deletion. Every row whose source and destination are both in the file, in any state, open work included. Add `--rig <id>` for the account rig's own records | delete |
+| `--ids <file>` | Exactly the listed rows. A report's `qitems` entries, or a ledger's lines, keep each entry's own action. A ledger's entries are replayed only for the batches that committed (below) | `--action delete` (default for plain ids) or `--action scrub` |
+| `--routine` | Closed rows past this daemon's retention windows (below) | scrub, then delete |
+
+A file is one entry per line (blank lines and `#` comments are skipped; a line starting with `{` is a JSON entry) or a JSON array.
+
+**Account deletion is exact.**
+- The members file is the account's recorded membership: the exact session names of its seats, including departed ones. List every address form the account used, such as a host-suffixed form. Names are compared as given, never by pattern, so a row addressed through an unlisted form is neither erased nor reported.
+- **Mixed rows:** a row with exactly one member endpoint is reported under `mixed` (with which endpoint is the member) and left untouched. Deciding who owns it is a separate decision. A row rerouted out of the account while the run is under way (by a fallback, say) is kept too, and reported as `changed-during-run`.
+- **After the rows:** account deletion also removes outbox and inbox entries between two members that no row owns, and the members' stream items. An entry with one member endpoint, or a stream item something still references, is counted and left.
+
+**The account's rig.** A rig's own records outlive the rig. Events are kept past rig and node deletion by design, and snapshots, node and seat records and watchdog jobs have no key that cascades. `--rig <id>`, given with `--members`, erases them:
+- the rig's events, including those of its nodes;
+- snapshots, operation correlations and bootstrap runs;
+- seat identity verdicts, delivery-guard changes, environment bindings and usage records for its nodes and member sessions;
+- watchdog jobs aimed at a member, with their history;
+- stream subscriptions and thread maps of member sessions.
+
+Records that name the account but also belong elsewhere are counted under `rig.left` and kept. These are a member's watchdog job aimed outside the account, a workflow instance or spec bound to the rig, classifier leases and custom views.
+- **While the rig exists,** its sessions and the members must match both ways, or the run is refused (`erase_rig_mismatch`). Every member's session name must carry the rig's name (departed members included), and every current session of the rig must be a member, so the rig holds no one else's seat.
+- **On that verified run,** the rig's id and name are recorded with a digest of the membership.
+- **Once the rig is removed,** the same command, with the same membership, uses that record to erase what's left under the rig's id. This includes the events the removal wrote, and the record still lets it count records bound to the rig's name.
+- **When nothing of the rig is left,** the record becomes a completion receipt: the rig's name is cleared, and its id, the membership digest and the time stay. An identical rerun, after a lost response say, still verifies and reports the same clean result, with `rig.completedAt`. The receipt is kept for good. It holds no content.
+- **A removed rig's id that was never verified** with this membership erases nothing keyed by it. Those records are counted under `rig.left`, the name-bound tables are listed under `rig.unknown`, and the run isn't clean.
+
+**Retention windows** (`rig config set …`, read at the daily sweep and at daemon start):
+
+| Setting | Default | Effect |
+|---|---|---|
+| `retention.closed_content_days` | `0` (off) | Scrub closed rows this many days after they closed |
+| `retention.closed_metadata_days` | `0` (off) | Delete closed rows, keeping only their ids, this many days after they closed |
+| `retention.erasure_ledger` | empty (none) | An absolute file path, on local disk, since each batch appends to it while it holds the daemon. Every applied erasure, by the sweep or the command, appends each row's id, action and batch id there, one JSON line each, before anything in its batch changes, then a commit line once the batch commits (an abort line if it rolls back). A path that can't be written stops the run before it changes anything |
+
+- **Closed** means `done`, `canceled`, `denied` or `handed-off`. `failed` and open work are never touched by the windows.
+- **Age:** a row's age counts from when it entered its final closed state. A note on a closed row doesn't restart it, and a row reopened and closed again counts from its latest close.
+- **Running them now:** `rig queue erase --routine --apply` runs the windows immediately. `retention.enabled false` turns off the whole sweep, these windows included.
+- **When the sweep's erasure stops on an error,** that's logged, and the sweep's other passes still run.
+
+## Account deletion, in order
+
+1. Revoke the account's access.
+2. Keep its exact membership list and rig id.
+3. Run `rig queue erase --members <file> --rig <id>` as a dry run, then with `--apply`. Rerun until `complete` is true.
+4. Remove the rig.
+5. Rerun the same command with `--apply`, to erase the records the removal wrote.
+6. The account is deleted only when the report's `clean` is true. Otherwise its `mixed`, `gaps`, `keptCopies`, `unlinked` and `rig.left` or `rig.unknown` entries need a decision. Rerunning won't resolve them.
+
+## Output
+
+The command prints one JSON report (`--json` makes it a single line):
+
+| Field | Meaning |
+|---|---|
+| `applied` | `false` for a dry run |
+| `complete` | the run finished. `false` when it stopped at its batch limit or on an error; rerun the same command to continue |
+| `clean` | `complete`, and nothing was left: no `mixed` rows, mixed `keptCopies`, `gaps`, `unreadableEvents`, left-over entries, or rig records left or unknown. Only a clean report means the scope is fully erased. It's about this run's scope: a routine or members run's `clean` isn't affected by the ledger's other batches, and only a replay (`--ids` with a ledger) reports a batch it can't show committed, as the gap `ledger-unresolved` |
+| `error` | why the run stopped (`stage` and `message`). Everything listed in `qitems` was committed before it; the command exits 2 |
+| `qitems` | each row erased (or that would be), with its `action` and `state` |
+| `mixed` | members scope: rows with one member endpoint, left untouched |
+| `keptCopies` | copies of erased rows held with someone else, kept: `table`, `id`, `qitemId`, `ownership` (`mixed` or `foreign`) |
+| `gaps` | rows not erased and why: `workflow-reference`; `not-found` for an id to scrub that doesn't exist; `changed-during-run` for a row that no longer qualified when its batch ran (reopened, rerouted out of the account, or written to so it left its window) or an absent id created since selection; `unreadable-event` for a row an unreadable event names; `ledger-unresolved` for a replayed entry whose batch can't be shown to have committed |
+| `skipped` | ids scope, a ledger replay: entries whose batch rolled back, with the batch id; not replayed |
+| `ledger` | an applied run with a ledger: batches whose missing commit line it wrote from the database's record (`reconciled`) |
+| `alreadyAbsent` | ids to delete with no row; their erased-id record is written (again), and they're listed once it's committed. They're rechecked as the record is written, so an id created during the run is reported as `changed-during-run` instead |
+| `unreadableEvents` | events naming a selected row whose payload can't be read: `seq` and the row ids found in its text. They're kept, since what they hold can't be inspected, and so are the rows they name, so every rerun reports them again until the event is repaired or removed. The run isn't clean |
+| `linked` | counts of the copies erased with the rows |
+| `unlinked` | members scope: entries and stream items erased that no row owned, plus those left (`mixedEntries`, `referencedStreamItems`) |
+| `rig` | members scope with `--rig`: `found`, `verified`, records erased per table (`erased`, counted as each chunk commits), records left (`left`), tables that can't be counted (`unknown`), and `completedAt` once the rig's erasure is complete |
+
+**How a run works:**
+- **Selection:** a run selects everything first and writes nothing.
+- **Events:** it reads the events table once, in small windows, to find each row's events.
+- **Batches:** each batch, of up to 100 rows, is one transaction. It rechecks its rows, writes their ledger lines, then changes their events, copies and rows.
+  - **The recheck drops** a row that a workflow has taken up, or that an unreadable event names. For account deletion, it also drops a row that is no longer between two members. For a routine run, it drops one that was reopened or no longer falls in the same window.
+  - **Each dropped row** is reported in `gaps`.
+  - **The current endpoints** decide whose its copies are.
+- **How long it holds the daemon:** the batch caps rows, not events or copies, so a row with many of either makes its batch longer. On a copy of a store with 2.4 million events, the longest single hold in a first routine run was about 0.25 s. With a ledger, a run doesn't read the ledger file: it reads the end of the file (one byte) and the database's record of batches still missing a commit line, so the ledger's size doesn't add to the hold.
+- **Stopping:** a run that stops leaves whole batches done, events included. A rerun picks up the rest, and erasing a row a second time changes nothing.
+- **Ledger lines and commits:** a batch's lines are written before its first change, so they can name rows that a rolled-back batch never erased. Each batch therefore also records its id in the database, in its own transaction, and appends a commit line (`{"batch": …, "committed": true}`) after it commits, or an abort line (`"committed": false`) when it rolls back.
+  - **A missing commit line:** a run that can't write one stops once the committed batch is reported. The database's record notes when each batch's commit line was written. Each applied run with a ledger writes the commit line for any recorded batch still without one, from the record alone, without reading the ledger file. It lists those under `ledger.reconciled`.
+  - **Neither line nor record:** a batch with neither, such as one whose transaction was cut off by a crash, is never decided from the database: on a restored database, a batch committed after the backup looks exactly like one that rolled back. A replay reports its entries as `ledger-unresolved` and doesn't apply them, until an operator resolves it by appending its commit or abort line once they know which.
+  - **A line cut short:** a full disk, say, can cut a line short. Each run starts the ledger on a fresh line first, so the cut line can't fuse with the next one. A replay refuses the file until that line is deleted. That's safe: an entry line's batch rolled back when its write failed, and a commit line is written again from the record.
+
+**Refusals**, with nothing erased:
+
+| Code | Cause |
+|---|---|
+| `erase_scope_required` | no scope named |
+| `erase_scope_conflict` | more than one scope |
+| `erase_members_invalid` | an empty list, or names with surrounding spaces |
+| `erase_rig_invalid` | a rig without `--members`, or an empty rig id |
+| `erase_rig_mismatch` | the rig exists, and its sessions and the members don't match both ways |
+| `erase_ids_invalid` | an empty ids list, or ids with surrounding spaces |
+| `erase_action_invalid` | an action other than `delete` or `scrub` |
+| `erase_routine_off` | both windows are `0` |
+
+The command records who ran it from `OPENRIG_SESSION_NAME`, or from `--actor <session>` when that isn't set.
+
+## Restores and backups
+
+A restored database can bring erased content back: rows, copies, rig records and the erased-id records themselves. It can even bring back a row that was still open when the backup was taken and was closed and erased later. Before a restored daemon serves anyone, replay what was erased after the backup was taken:
+
+1. **Keep a ledger.** Set `retention.erasure_ledger` to a file outside the backed-up data. Every applied erasure appends its ids, actions and batch there before changing anything, and a commit line once the batch commits. That covers the daily windows, `--routine`, `--ids` and account deletion's rows.
+2. **Replay the ledger:** `rig queue erase --ids <ledger> --apply`. It replays exactly the batches that committed, each line with its own action, so a scrub is replayed as a scrub, never promoted to a deletion.
+   - **What's replayed:** a row that was open in the backup and erased later is erased by its id, and other open work isn't touched.
+   - **What's skipped:** a batch with an abort line was never erased, and is listed under `skipped`.
+   - **What's left:** a batch with neither line, which the restored database doesn't record either, is reported as `ledger-unresolved` and not replayed. The run isn't clean until an operator decides it.
+3. **Replay account deletions:** rerun `rig queue erase --members <file> --rig <id> --apply` for each account deleted since the backup, from its membership list and rig id. Keep those outside the backed-up data until every backup taken before the deletion has expired. This also covers the copies and rig records the ledger doesn't list.
+4. **Rerun the windows:** `rig queue erase --routine --apply`.
+
+Backup expiry is the deployment's own schedule. Content erased from the live database stays in backups until they expire.
+
+## What this doesn't cover
+
+These stores need their own handling where a deployment uses them:
+- terminal transcripts;
+- daemon and service logs;
+- gateway and Slack stores and Slack's own copies;
+- downloaded files and artifacts;
+- chat messages (removed with their rig);
+- events tied neither to a row nor to the account's rig.
+
+The rig itself (its nodes, sessions and bindings) is removed by removing the rig, between the two runs above.

docs/reference/quickstart.md

@@ -90,7 +90,8 @@ Use `codex` or `both` in place of `claude` to match the coding agents you use. A
 - it adds an OpenRig block to `~/.tmux.conf` (mouse support and longer scrollback).
 
 It doesn't start anything. It ends with a summary of what's done and what still needs you. A missing sign-in shows
-there: run `claude auth login` or `codex login`, then carry on.
+there: run `claude auth login` or `codex login`, then carry on. Setup also tries to record the selected providers
+for a new kernel's first boot; its summary warns if it could not save the choice.
 
 ### 3. Start OpenRig
 
@@ -100,7 +101,10 @@ rig status
 rig ps --nodes --rig kernel
 ```
 
-The daemon starts and boots the kernel with the coding agents you're signed in to, or both if you're signed in to both.
+On the kernel's first boot, the daemon uses the providers selected by `rig setup --providers` that are signed in.
+If no provider choice was recorded, it considers every signed-in provider. An existing kernel keeps its seats;
+changing the setup choice does not replace them. An `OPENRIG_PROVIDER_SELECTED` value in the daemon's environment
+takes precedence over the saved choice.
 
 **What starting OpenRig changes, for the coding agents you use** (paths are the usual defaults):
 - **For Codex:** hooks and trust in `~/.codex/config.toml`, when OpenRig or a Codex agent starts.

docs/reference/reconstructions.md

@@ -0,0 +1,90 @@
+# Read a reconstruction of selected work
+
+A reconstruction records an author's working understanding, its evidence routes,
+constraints and gaps. `rig context reconstruction` reads an exact immutable
+`openrig.reconstruction/v0` revision locally. It does not start a daemon, fetch
+remote context, write records or transfer the author's authority.
+
+Inspect the purpose, depth, author-declared coverage, observation dates, complete
+source manifest and approximate size before reading a view:
+
+```bash
+rig context reconstruction inspect <full-sha256> --project example --mission task --slice 01-example
+rig context reconstruction view <full-sha256> --project example --mission task --slice 01-example --item decision --json
+```
+
+Without `--store`, the store is `reconstructions/` under the selected project,
+mission or slice home. Project selection follows `context work-install`: an
+explicit catalog project, the only entry, the calling rig's association, the
+deepest project containing `--cwd`, or the only unclaimed entry. Ambiguity is an
+error. The project's configured `missions.root`, declared mission membership,
+slice directory names and slice frontmatter IDs are honored. A workspace
+without a catalog can use its own project home. No skill installation or context
+projection happens during this lookup.
+
+Use `--store /absolute/local/store` for an imported store without resolving a
+project. Supplying work selectors as well still validates that work home.
+Revisions are full lowercase SHA-256 IDs; there is no implicit latest revision.
+
+## Faithful views and gaps
+
+Repeat `--item`, `--category` or `--tag` to select their union. A filtered view
+always retains authority, constraint and gap items, plus transitive item
+dependencies. It names omitted item IDs and uncited evidence routes, and keeps
+declared reconstruction coverage separate from actual view coverage. Inspect
+always provides the complete manifest; a filtered view provides cited routes.
+A full view also includes uncited registered routes.
+
+Base records, deltas and pinned compositions retain inherited observation dates.
+A revision's creation date does not refresh those observations. Composition
+members do not imply an atomic capture. Authors are claimed attribution, and
+coverage is authored description rather than proof of runtime state or acceptance.
+
+Unavailable records, missing selections, dependencies or citations make the
+result `complete: false` and set exit status 1. Available composition members
+remain visible with the missing member's exact store and revision. Unknown
+schemas, invalid content hashes and unsupported numeric representations are
+reported explicitly. The initial reader supports safe integer JSON numbers;
+floating-point/exponent and unsafe integer representations are rejected rather
+than assigned a different ID. Unicode keys follow v0's codepoint sorting.
+
+## Relocate a store without changing its records
+
+Parent and member pointers include absolute source-store paths in the hash.
+Do not edit those paths in immutable records. Copy their closure into local
+custody, then supply a separate JSON mapping:
+
+```json
+[
+  {
+    "sourceStore": "/original/work/reconstructions",
+    "localStore": "/local/import/reconstructions",
+    "sourceHost": "original-host"
+  }
+]
+```
+
+```bash
+rig context reconstruction inspect <full-sha256> --store /local/import/reconstructions --store-map /local/map.json
+```
+
+The mapping is read-time location information, not an identity registry or a
+remote fetch permission. It preserves source-host provenance. Since v0 pointers
+do not contain a host discriminator, duplicate mappings for the same source
+store are ambiguous and rejected. A store path that exists on this host can be
+read directly; a path that has moved needs a mapping to its local location.
+
+## Bounds and byte checks
+
+Reads are limited to 8 MiB per regular revision file, 32 MiB per closure, 64
+record references and 128 JSON nesting levels. Symlink revision files are not
+followed. Output defaults to 256 KiB; `--max-bytes` accepts 1 KiB through 8 MiB.
+Oversized output fails explicitly rather than silently truncating a purportedly
+complete view. Size estimates use UTF-8 bytes divided by four, not a model tokenizer.
+
+`--check-sources` hashes declared local regular files within 128 files and 64 MiB
+total. It reports changed, unchanged, unavailable or unchecked bytes without
+executing source commands, fetching opaque routes or updating observations.
+Discovery declarations are shown but not scanned. Unchanged known files do not
+establish semantic freshness or detect new queue decisions. There is no queue
+discovery cursor, native writer or continuity-consumer replacement in this reader.

docs/reference/rig-bundle.md

@@ -91,10 +91,10 @@ my-rig.rigbundle.sha256   — "a1b2c3d4..." (64-char hex SHA-256)
 The packer produces deterministic output:
 - Files are sorted alphabetically
 - A fixed mtime is used (`2026-01-01T00:00:00Z`)
-- Portable mode normalizes uid/gid/mode
+- Portable mode omits machine-specific owner metadata; file mode bits can be retained
 - Maximum gzip compression (level 9)
 
-Identical staged bytes (including manifest timestamps and provenance) produce the same archive hash. Use the package digest to compare packaged file content independently of manifest metadata.
+Identical staged bytes and file modes (including identical manifest timestamps and provenance) produce the same archive hash. A mode change can change the archive hash even when the file contents are unchanged. Use the package digest to compare packaged file content independently of manifest metadata and file modes.
 
 ---
 

docs/reference/rig-spec.md

@@ -2,7 +2,7 @@
 
 Version: 0.2 (pod-aware)
 Last validated against code: 2026-10-07, at the 0.6.6 cut `2620dea8` (whole document at `fcaf1f8e`; changes to its sources since then checked)
-Source of truth: `packages/daemon/src/domain/rigspec-schema.ts`, `packages/daemon/src/domain/types.ts`, `packages/daemon/src/domain/startup-validation.ts`, `packages/daemon/src/domain/permission-policy/policy-ref.ts`, `packages/daemon/src/domain/profile-resolver.ts`, `packages/daemon/src/domain/rigspec-preflight.ts`, `packages/daemon/src/adapters/kernel-authority.ts`, `packages/daemon/src/domain/native-permission-store.ts`, `packages/daemon/src/domain/codex-team-workspace.ts`
+Source of truth: `packages/daemon/src/domain/rigspec-schema.ts`, `packages/daemon/src/domain/types.ts`, `packages/daemon/src/domain/launch-wrapper.ts`, `packages/daemon/src/domain/startup-validation.ts`, `packages/daemon/src/domain/permission-policy/policy-ref.ts`, `packages/daemon/src/domain/profile-resolver.ts`, `packages/daemon/src/domain/rigspec-preflight.ts`, `packages/daemon/src/adapters/kernel-authority.ts`, `packages/daemon/src/domain/native-permission-store.ts`, `packages/daemon/src/domain/codex-team-workspace.ts`
 
 This is the canonical reference for the pod-aware RigSpec YAML format. Every field, validation rule, and default documented here was traced from the actual parser and validator code, not from prior documentation.
 
@@ -180,7 +180,7 @@ edges:
 | `non_interruptive` | boolean | no | — | Pod-aware rig launch-warning default. With full-bypass Claude/Codex seats, true accepts or hides supported harness warnings using launch flags. Explicit CLI choices override this field; this field overrides the machine default. It does not grant permissions, change native settings files, or bypass login. |
 | `permission_policy` | string | no | — | Permission policy attached to the rig. Either a built-in (`builtin:locked`, `builtin:standard`, `builtin:open`, `builtin:yolo`, `builtin:auto`) or a safe relative path to a custom policy file (resolved from this spec's directory; no `..`, no absolute, no empty segments, each segment `[A-Za-z0-9][A-Za-z0-9._-]*`), or `none`, a recorded choice of the floor. Absent uses the default floor with the scoped team launch allowances described below. A bare built-in name such as `yolo` is refused ("use 'builtin:yolo'"), and so is an explicit `null`. A custom file that is missing, unreadable or invalid isn't a validation error: it resolves to the floor, and preflight warns. A member may set its own `permission_policy` (not on a terminal member), which takes precedence over the rig-level one; there is no pod level. See "Attaching a permission policy" below. |
 | `env` | map | no | — | Launch environment for every seat in the rig: `NAME: value` pairs. A member's own `env` overrides it name by name. Names match `^[A-Z_][A-Z0-9_]*$`, names starting `OPENRIG_` are reserved, and values must be non-empty strings (quote numbers and booleans). Values are not secret. See "Launch environment" below. |
-| `managed_blocks` | map | no | `CLAUDE.md` | File that receives OpenRig's managed instruction blocks for Claude Code members. Only the `claude-code` key is accepted, with `CLAUDE.md` or `CLAUDE.local.md`. Codex, Pi and OMP members use `AGENTS.md`. See "Choosing the Claude instruction file" below. |
+| `managed_blocks` | map | no | `CLAUDE.md` | File that receives OpenRig's managed instruction blocks for Claude Code members. Only the `claude-code` key is accepted, with `CLAUDE.md` or `CLAUDE.local.md`. Codex, Pi and OMP members use `AGENTS.md`, except that a fresh Pi seat normally uses its own agent folder's `AGENTS.md` and a fresh Codex 0.164.0+ seat normally gets launch developer instructions; a link, a conflicting or unreadable Codex config layer, or a resume keeps the repository file. See [Codex guidance at launch](agent-startup-guide.md#codex-guidance-at-launch) and [Runtime Differences](agent-startup-guide.md#runtime-differences). See "Choosing the Claude instruction file" below. |
 | `workspace` | object | no | — | The rig's workspace: `workspace_root` (required), `repos[]` of `{name, path, kind}` with `kind` one of `user`, `project`, `knowledge`, `lab` or `delivery` and unique names, an optional `default_repo` naming one of them, and an optional `knowledge_root`. Relative repo paths resolve against `workspace_root`. |
 | `docs` | Doc[] | no | — | Documentation files that should travel with the rig. Included in rig bundles. Each entry has a `path` field (safe relative path). The engine does not consume these — they are for humans and agents setting up the environment before launch. |
 | `startup` | StartupBlock | no | — | Rig-level startup files and actions. Applied to all members via the startup layering model. |
@@ -404,6 +404,22 @@ ask rules; help can prompt in that fallback. A helper that disappears or times o
 after launch cannot supply a decision, leaving Claude's remaining native rules
 in effect. The hook is a convenience policy, not a containment boundary.
 
+Claude sessions launched with a named `team` or `kernel` hook policy read the
+installed built-in lists on each Bash hook call. Replacing those files at the
+same path changes the hook's lists on its next call; native inline allowances
+remain as launched. Sessions launched with an encoded policy keep their launch
+lists. A session retaining a versioned runtime path continues to use the files
+at that path, so installing another version elsewhere does not change its hook.
+This does not reload all permission settings in a running session.
+
+An in-place rollback to a hook from before named policies has a different effect:
+that helper cannot read `team` or `kernel` and returns no hook decision. Lifecycle
+asks can then disappear under the native `Bash(rig:*)` allowance. Relaunch affected
+Claude sessions before relying on those asks after such a rollback. Native
+personal, project and managed restrictions are not removed. This applies when
+the rollback replaces the hook at the path a running named-policy session uses;
+it does not describe every rollback or sessions using a retained versioned path.
+
 Codex keeps `workspace-write` with its existing approval policy and receives the
 configured OpenRig workspace root plus its pod's shared state directory as
 additional writable directories. The team workspace preparer creates missing
@@ -548,7 +564,7 @@ or local settings. Managed settings aren't inspected. `rig export` writes the se
 with an account home or an API-key selection (select those again after import), so a later switch shows there, and `rig provider switch --stored-login` removes it. Only
 `claude-code` members and legacy flat nodes accept it, and there's no rig-level default.
 
-**Switching to an isolated account home.** A running Codex or Claude Code seat can change its native account home
+**Switching to an isolated account home.** A running Codex, Claude Code or Pi seat can change its account home
 while staying in its rig:
 
 ```sh
@@ -560,7 +576,24 @@ rig seat set-runtime dev-impl@my-rig --runtime claude-code \
 
 The directory must already exist and contain the runtime's login or API settings. You choose the account and budget;
 OpenRig does not choose or create credentials. `--account-home` sets `CODEX_HOME` for Codex or `CLAUDE_CONFIG_DIR` for
-Claude Code before the successor starts.
+Claude Code before the successor starts. For Pi, the account directory must contain a readable `auth.json`; OpenRig links
+that file into a separate managed successor directory. The existing Pi process keeps its credentials and runner state
+until the successor confirms. Pi account changes keep the current model and effort unless explicitly overridden.
+
+```sh
+rig provider switch --seat pi-worker@my-rig --to-account work \
+  --account-home /home/me/accounts/pi-work --reason "Use the work account" --dry-run
+# Run again without --dry-run to stage the successor and request its confirmation.
+rig provider switch --seat pi-worker@my-rig --clear-account-home --reason "Return to configured login"
+```
+
+When switching into Pi with `--clear-account-home`, OpenRig restores the seat's original Pi login only if no effective member or rig Pi account home is configured. A member's `PI_CODING_AGENT_DIR` shadows the rig's value; if the member value is relative, it is ignored and the rig home does not take effect.
+
+Pi account changes work with or without `launch_wrapper`. The wrapper receives the successor's managed Pi directory;
+it must preserve the selected credentials if you want OpenRig's account choice to apply.
+Use `rig seat set-runtime <seat> --runtime <runtime> --reason "Change provider"` with `pi`, `codex` or `claude-code` for cross-runtime switches.
+Those resolve the target's settings afresh rather than keeping the source model or effort;
+see [runtime switching](getting-started.md#move-a-seat-to-the-other-runtime).
 
 This switch starts a fresh conversation with a handover, even when the runtime stays the same. The existing
 subscription-login switch preserves the conversation. Once the successor confirms, OpenRig commits its home, runtime,
@@ -740,7 +773,7 @@ directory's `CLAUDE.md`, including blocks written by other rigs.
 | `id` | string | yes | — | Member identifier. Must not contain dots or `@`. Must be unique within the pod. |
 | `agent_ref` | string | yes | — | Reference to an AgentSpec. Must start with `local:` (relative) or `path:` (absolute). Exception: `builtin:terminal` for infrastructure nodes. |
 | `profile` | string | yes | — | Profile name from the referenced AgentSpec. Use `default` for the default profile. Exception: `none` for terminal nodes. |
-| `codex_config_profile` | string | no | — | Codex-only native profile passed as `-p <name>`; letters, numbers, `_`, `.`, `-`. Separate from the AgentSpec `profile`. With the normal launch mode, this replaces OpenRig's explicit workspace-write sandbox flag. A full-bypass policy instead emits danger-full-access and omits this profile argument. |
+| `codex_config_profile` | string | no | — | Codex-only native profile passed as `-p <name>`; letters, numbers, `_`, `.`, `-`. Separate from the AgentSpec `profile`. With the normal launch mode, this replaces OpenRig's explicit workspace-write sandbox flag. A full-bypass policy instead emits danger-full-access and omits this profile argument. To keep the profile, put full access in the profile itself, or record a workspace-write or read-only `codex_permissions` envelope, which keeps it. |
 | `codex_permissions` | mapping | no | — | Codex-only sandbox and approval envelope, emitted as explicit flags on every managed launch and resume; see [per-member Codex permissions](#per-member-codex-permissions). |
 | `claude_token` | mapping | no | — | Claude-only `{ label, file }`: the seat's own subscription token file, by absolute path, used from its first launch; see "A seat's own subscription token" under [launch environment](#launch-environment). |
 | `runtime` | string | yes | — | Agent runtime. Current supported values: `claude-code`, `codex`, `pi`, `omp`, `terminal`, `stub`. `stub` is a deterministic test harness, not an agent: a Node runner in the seat's tmux pane goes through the normal launch, skill projection, startup-file and readiness path, follows `<cwd>/.openrig/stub/script.json` when present, and calls no model. See "What a stub can and cannot prove" in `docs/as-built/test-layers.md`. |
@@ -750,11 +783,13 @@ directory's `CLAUDE.md`, including blocks written by other rigs.
 | `effort` | string | no | — | Reasoning effort override, trimmed. Claude gets `--effort`, Codex `-c model_reasoning_effort=…`, Pi `--thinking` (see "Pi" below). A non-string value is ignored with an advisory. |
 | `advisor` | string | no | — | Claude-only Advisor model for this seat. Overrides the AgentSpec profile/default; blank or non-string values are ignored with an advisory. |
 | `claude_mcp` | `none` | no | normal loading | Claude members only. Disable MCP server loading for this agent on every launch and resume. Omit the field to keep Claude's normal configuration; see [Claude MCP loading](#claude-mcp-loading-runtime-claude-code). |
+| `launch_wrapper` | string[] | no | — | Local `pi`, `codex`, `claude-code` and community-harness members; runner presets must declare wrapper support. Literal argument prefix applied to the native executable and its arguments on fresh launch, relaunch, rebuild, resume, fork, restore and handover; `[]` explicitly clears a stored wrapper when re-applying an authored spec, while omitting the key leaves an existing seat's stored choice unchanged. See [Launch wrapper](#launch-wrapper-runtime-pi-and-codex). |
 | `permission_policy` | string | no | — | Member-level permission policy; overrides the rig's. Not valid on terminal members. |
 | `env` | map | no | — | Launch environment for this seat, laid over the rig's `env`. Same rules as the rig-level field. Valid on terminal members too. See "Launch environment" below. |
 | `agent_file` | string | no | — | Claude members only. A Claude agent definition (a `.md` file with `name` in its frontmatter) for this seat. It's copied into the seat's own storage and launched as `--agent openrig:<name>`. A path beside the spec: no `..`, not absolute. |
 | `append_system_prompt_file` | string | no | — | Claude members only. A prompt file appended to Claude's system prompt for this seat. It's copied into the seat's own storage and passed as `--append-system-prompt-file`. A path beside the spec. |
 | `plugin_dirs` | list | no | — | Claude members only. Claude plugin folders for this seat, each copied into the seat's own storage and loaded through `CLAUDE_CODE_PLUGIN_DIRS`. Paths beside the spec. Each is copied under its folder name, so two folders with the same name are refused; an empty list is none. A missing file or folder fails that seat's launch by name. `rig bundle create` carries all three; `rig export` names a seat whose additions it doesn't carry. |
+| `plugin_channels` | list | no | — | Claude members only. Plugin channels this agent registers for, as `plugin:<name>@<marketplace>` entries, passed as one `--channels` on every launch; see [Claude plugin channels](#claude-plugin-channels-runtime-claude-code). |
 | `role` | string | no | — | The seat's role, letters, digits, `_`, `.` and `-`. Not valid on terminal members. |
 | `restore_policy` | string | no | the AgentSpec's | Restore behavior. One of: `resume_if_possible`, `relaunch_fresh`, `checkpoint_only`. The default comes from the AgentSpec's `defaults.lifecycle` (else `resume_if_possible`) as narrowed by its profile; a member value may only narrow it further, in that order. |
 | `compaction_strategy` | string | no | — | `default-compaction`, `managed-compaction`, `handover` or `apprentice-handover`; `harness_native` and `pod_continuity` are accepted as deprecated aliases, and `custom_prompt` is refused. `managed-compaction` and `apprentice-handover` create their jobs only for `claude-code` seats; see [Codex compaction and sizing](agent-startup-guide.md#codex-compaction-and-sizing). |
@@ -810,6 +845,52 @@ OpenRig keeps the requested opt-out flags, so startup and resume remain subject
 to that native refusal; it does not fall back to loading servers. Use a host
 without that managed configuration, or omit the opt-out to use the managed set.
 
+### Claude plugin channels (`runtime: claude-code`)
+
+Set `plugin_channels` on a Claude member to register that agent for channel
+notifications (messages a plugin's MCP server pushes into the session):
+
+```yaml
+members:
+  - id: support
+    agent_ref: local:agents/support
+    profile: default
+    runtime: claude-code
+    cwd: .
+    plugin_channels:
+      - plugin:example-channel@example-marketplace
+```
+
+OpenRig passes `--channels plugin:example-channel@example-marketplace` to this
+agent's Claude on every launch, resume, fork, restore and handover. Several
+entries follow one `--channels`, space-separated, and OpenRig's own session
+options come after them. Claude Code 2.1.316 hides `--channels` from
+`claude --help`, so it isn't listed there. An agent without the field gets no
+`--channels`, and nothing is added for other agents or other runtimes; there is
+no rig-wide or environment default.
+
+Each entry is `plugin:<name>@<marketplace>`, both names made of letters, digits,
+`.`, `_` and `-`, starting with a letter or digit. `server:` entries and Claude's
+`--dangerously-load-development-channels` aren't supported. Combining a non-empty
+channel list with `claude_mcp: none` produces a validation warning: that setting
+starts Claude without MCP servers (plugin servers included), so no channel could
+register. OpenRig keeps both settings and launches the member normally. Remove
+`claude_mcp: none` to use the channels, or remove `plugin_channels` to keep the MCP
+opt-out. OpenRig doesn't install plugins or marketplaces, or change the settings
+that admit a channel: whether each channel registers is Claude's decision (the plugin installed, and
+admitted by your organisation's channel settings or Claude's approved list).
+
+Selecting a channel for one agent doesn't isolate what the plugin shares with
+other agents. A token or other credential the plugin keeps in its own
+configuration is the same for every agent that loads that plugin.
+
+The selection is saved with the seat's other launch additions when the seat
+starts from the spec, and starting it from a spec without the field removes it.
+`rig seat launch --plan` and `rig skill loadout --seat` list it, a fork keeps it,
+and `rig bundle create` and `rig export` carry it in the member; a bundle's
+before-install view names each channel as a need. A running Claude process is
+unaffected.
+
 ### Claude Advisor (`runtime: claude-code`)
 
 Set `advisor` independently of the primary `model`:
@@ -840,14 +921,42 @@ This requires a Claude Code version that supports `--advisor`. OpenRig never wri
 choose different Advisors. When Advisor is unset, no flag is emitted and Claude's
 own defaults/settings apply. Other runtimes ignore this field.
 
+<a id="launch-wrapper-runtime-pi-and-codex"></a>
+
+### Launch wrapper (`runtime: pi`, `codex` and `claude-code`)
+
+Set `launch_wrapper` on a local `pi`, `codex`, `claude-code` or community-harness member to prefix its native executable and arguments with a literal argv vector:
+
+```yaml
+members:
+  - id: impl
+    agent_ref: local:agents/impl
+    profile: default
+    runtime: codex
+    cwd: .
+    launch_wrapper: ["launcher", "exec", "--profile", "seat-a", "--"]
+```
+
+- **Scope and literal argv:** Local `pi`, `codex`, `claude-code` and supported community-harness members support `launch_wrapper`; legacy flat nodes support it only on `pi`, `codex` and `claude-code`. Other runtimes, including the maintained `openshell-claude` environment path, reject it, including `[]`. Community preset capability and execution requirements are defined in the [runner contract](harness-runner-contract.md#2-declare-it-what-the-preset-says-fixed-at-launch). Each fork uses its target member's own selection. Arguments are literal strings without NUL, CR or LF; spaces, quotes, metacharacters and empty non-executable arguments are preserved.
+- **Executable resolution:** `launch_wrapper[0]` must be an absolute path or bare command name. Preflight and launch resolve it using the daemon's launch `PATH` and working directory, before adding the seat's `rig` bin directory. Relative paths containing `/`, aliases and functions are unsupported for the wrapper itself.
+- **Native child and `exec` contract:** The wrapper prefixes the resolved Codex entry, prefixes the resolved Pi child executable inside the host runner, or prefixes the real Claude executable and native arguments. Managed Claude applies it after rebuilding the environment; pinned continuations retain their executable. The wrapper must `exec` its child in place, preserving working directory, environment, signals, exit status and TTY stdio (Claude/Codex) or RPC pipes (Pi). Supervising wrappers are unsupported.
+- **Classic Claude lookup:** Wrapped classic launch supports bash, zsh, sh and dash and requires `command -v claude` to resolve to an executable file; aliases and functions refuse with `launch_wrapper unavailable; no agent started.` To keep custom flags or environment, put them in a `#!` script named `claude` on `PATH`; the wrapper runs that script ([#1477](https://github.com/mvschwarz/openrig/issues/1477)).
+- **Fail-closed (never unwraps):** The selected wrapper applies across fresh launch, relaunch, rebuild, resume, fork, restore, handover and supported runtime switches. Invalid stored metadata, non-POSIX panes, unavailable wrapper support, missing session identity or an unusable wrapper refuses launch. Failures after dispatch remain visible in the pane or Pi runner's markers; OpenRig never runs unwrapped, including Pi's long-command (`launch_path_too_long`) host-runner retry.
+- **Persistence and clear:** On new seats, absent or `launch_wrapper: []` stores no wrapper. Applying an authored spec to an existing seat retains the stored selection when absent, clears it with `launch_wrapper: []`, and replaces it with a non-empty array. Recreation/import uses the incoming spec; older snapshots preserve the current selection. First-start retry refuses wrapper changes. Runtime switches carry it to supported successors; validation occurs at execution. There is no live edit verb: change the spec and use the supported create/recreate/apply flow. Export and node inventory include the configured value.
+- **Boundary:** An argv prefix is not proof of sandbox isolation, tooling availability, memory wiring or governance enforcement. Verify the wrapper and runtime independently.
+
 ### Pi (`runtime: pi`)
 
 `runtime: pi` launches the Pi coding agent (`pi --mode rpc`) through OpenRig's RPC runner.
 
-- **State:** Each seat uses `$OPENRIG_HOME/state/pi/<session>/agent` as its Pi agent directory, with `sessions/` beside it, instead of your default `~/.pi/agent`. The runner sets `PI_CODING_AGENT_DIR` to that directory. `<session>` is the seat's session name, unchanged; for a pod member it is `{podId}-{memberId}@{rigName}`. `$OPENRIG_HOME` defaults to `~/.openrig`. OpenRig creates these directories at launch and does not copy anything from `~/.pi/agent`.
+- **State:** Each seat uses `$OPENRIG_HOME/state/pi/<session>/agent` as its Pi agent directory, with `sessions/` beside it, instead of your default `~/.pi/agent`. The runner sets `PI_CODING_AGENT_DIR` to that directory. `<session>` is the seat's session name, unchanged; for a pod member it is `{podId}-{memberId}@{rigName}`. `$OPENRIG_HOME` defaults to `~/.openrig`. OpenRig creates these directories at launch and does not copy anything from `~/.pi/agent`. A confirmed account switch moves the seat into an isolated `state/pi/account-switches/<generation>/<session>/` directory; subsequent launches and resumes use that committed directory.
 - **Thinking level:** A Pi seat's effort (the choice saved with `rig seat set-effort`, else the spec's `effort`) is sent to Pi as `--thinking <level>` on every managed launch and resume: fresh launch, successor and restore. Pi accepts `off`, `minimal`, `low`, `medium`, `high`, `xhigh` and `max`; OpenRig passes the value as written and does not translate it, so Codex's `none` is not mapped to `off`. Only a value of the shape `set-effort` accepts (a letter, then up to 31 letters, digits, `-` or `_`) is sent; any other spec value, such as one with a space or one that starts with `--`, is left out, and `rig seat effort` says so. An unfamiliar level gets an advisory (a warning from `set-effort`), never a refusal, and Pi itself ignores an unknown level with a warning and keeps the level it would use without the flag: a model suffix, the resumed session's level, or its default. With no effort resolved, no flag is sent. The effort wins over a model suffix, and the model string is never changed. `rig seat effort` reports the level sent at launch and, on Pi 0.99.0 or later, the `thinkingLevel` Pi records on this launch's latest successful assistant turn (error and aborted turns are skipped); that is the session's level after Pi's own cap, not proof the provider applied it, and it is shown apart from the level sent when they differ. On older Pi it is NOT VERIFIED. Without an effort, you can still set the level as a model suffix, for example `model: openrouter/x-ai/grok-4:high`. Pi first looks for a model whose id is the whole pin, and uses it when there is one, suffix included. Only when there is none does Pi read the suffix as the level and record the model without it. Any other suffix, such as OpenRouter's `:free`, stays part of the model id. The model-divergence check reads the pin literally when the seat's `models.json` defines a model with that exact id; otherwise it accepts the model with or without the suffix. If the seat's `models.json` can't be read, a pin with a suffix stays unchecked, with the usual warning, when the recorded model matches the whole pin or the pin without its suffix. A record that matches neither is reported.
-- **Custom models:** Pi reads `models.json` from its agent directory, so a seat reads `$OPENRIG_HOME/state/pi/<session>/agent/models.json`, not `~/.pi/agent/models.json`. Custom provider and model definitions for a seat go in that file. A seat that uses only providers Pi already includes does not need one. OpenRig does not create or write `models.json`.
-- **Credentials:** The Pi process receives only `PATH`, `HOME`, `USER`, `LOGNAME`, `SSH_AUTH_SOCK`, `TERM`, `LANG`, `LC_ALL`, `SHELL`, `TMPDIR`, a fixed set of OpenRig seat and instance variables, `PI_CODING_AGENT_DIR`, `PI_CODING_AGENT_SESSION_DIR`, and at most one provider key. That key is `OPENROUTER_API_KEY`, `ZAI_API_KEY`, `KIMI_API_KEY` or `MINIMAX_API_KEY`, passed only when it is set in the seat's environment and the seat's `model` is written as `openrouter/<id>`, `zai/<id>`, `kimi-coding/<id>` or `minimax/<id>`. Naming that variable in `recovery.provider_auth_env_allowlist` (empty by default) makes the daemon add it, when set in the daemon's environment, to the seat's launch environment. No other variable reaches Pi, including other providers' keys and a custom provider's own key variable.
+- **Custom models:** Pi reads `models.json` from its agent directory, so a seat reads `$OPENRIG_HOME/state/pi/<session>/agent/models.json`, not `~/.pi/agent/models.json`. Custom provider and model definitions for a seat go in that file (or the corresponding committed account-switch directory after a switch). A seat that uses only providers Pi already includes does not need one. OpenRig does not create or write `models.json`.
+- **Credentials:** The Pi process receives only `PATH`, `HOME`, `USER`, `LOGNAME`, `SSH_AUTH_SOCK`, `TERM`, `LANG`, `LC_ALL`, `SHELL`, `TMPDIR`, a fixed set of OpenRig seat and instance variables, `PI_CODING_AGENT_DIR`, `PI_CODING_AGENT_SESSION_DIR`, `NODE_OPTIONS` for the selected-account Node preload, and at most one provider key. That key is `OPENROUTER_API_KEY`, `ZAI_API_KEY`, `KIMI_API_KEY` or `MINIMAX_API_KEY`, passed only when it is set in the seat's environment and the seat's `model` is written as `openrouter/<id>`, `zai/<id>`, `kimi-coding/<id>` or `minimax/<id>`. Naming that variable in `recovery.provider_auth_env_allowlist` (empty by default) makes the daemon add it, when set in the daemon's environment, to the seat's launch environment. No other variable reaches Pi, including other providers' keys and a custom provider's own key variable.
+
+- **Account homes:** Select an account with `rig provider switch --account-home`, or declare an absolute `PI_CODING_AGENT_DIR` in the seat’s `env:` as its configured account home; the persisted selection takes precedence. `~` is not expanded. A relative value is ignored for managed account selection (and reported by status), leaving the managed seat login in use when no persisted account is selected. The chosen directory must already contain a readable `auth.json`; env-only credentials and wrappers that create credentials later are outside this prepared-account path. Launches with neither a persisted selection nor a declared absolute account home retain their existing credential behaviour. The runner links that file into its managed agent directory before starting Pi, including launches through `launch_wrapper`; settings, models, extensions and sessions remain seat-local. Selected accounts use a Node preload shim that redirects only their auth-lock filesystem operations to the chosen account’s `auth.json.lock` pathname, matching direct Pi use even when the source `auth.json` is a symlink. The preload removes its own `NODE_OPTIONS` entry before Pi starts tools, preserving other Node options; descendants and unrelated agent directories keep their normal filesystem behaviour. This requires a Node-hosted Pi executable and a wrapper that preserves `NODE_OPTIONS`; compiled executables and wrappers that discard the preload are not covered. Removing the selection removes only the managed auth link, not the source account file. A missing account file stops the launch. Direct launch refuses to replace an existing seat-local login or an unmanaged auth link; use a confirmed account switch to preserve that login. An ambient pane-shell `PI_CODING_AGENT_DIR` does not select an account; only the resolved seat configuration or persisted account selection does. Persisted account selections survive resume and cannot be bypassed by the long-command fallback; explicit selections clear inherited provider-key overrides. Use `rig provider switch --account-home` or `--clear-account-home` to change an existing Pi seat through a confirmed successor with a fresh conversation and handover. The account reference and committed state directory survive daemon restart and resume; unconfirmed successors leave the old account and state intact. Clearing returns to the configured home, or the original managed seat login if no home was configured. Status reports a configured account home when present; without one, its home is null because Pi uses the managed seat directory rather than the operator's default login.
+
+**Upgrade note for prepared accounts:** An existing built-in Pi seat that declares an absolute `PI_CODING_AGENT_DIR` now selects that home on fresh launch, restart, resume and restore, even without a stored account override. Before upgrading, either prepare a readable `auth.json` there or remove that declaration to keep the managed seat login. Env-only accounts and wrappers that create the file after Pi starts are not supported by this path. Migration 151 preserves switch history; it does not create or copy credentials, or rewrite existing seat declarations. Relative values remain available to community harnesses and are not globally removed.
 
 - **Extension dialogs:** The RPC runner shows extension `select`, `confirm`, `input` and `editor` requests in the pane and cancels them because the managed pane cannot answer them. The seat needs attention through the end of the turn, until a new agent run starts. Review the cancelled request and send an instruction the extension can complete without a dialog. Extension `notify` text is shown without raising attention or sending a response.
 
@@ -912,7 +1021,7 @@ This is human-authored (you choose the pod/member IDs) and system-validated (the
 | `spawned_by` | Source was spawned by target; the target launches first. Constrains launch order. | Child → parent in hierarchical topologies |
 | `can_observe` | Source can observe target's output. Does NOT constrain launch order. | Reviewer → implementer, monitor → worker |
 | `collaborates_with` | Peer collaboration relationship. Does NOT constrain launch order. | Co-equal peers working together |
-| `escalates_to` | Fallback queue escalation for a seat without a declared delegation parent. Does NOT constrain launch order. | Local member or same-daemon `pod-member@rig` address |
+| `escalates_to` | Fallback queue escalation for a seat without a single declared delegation parent (none, or several). Does NOT constrain launch order. | Local member or same-daemon `pod-member@rig` address |
 
 A named `escalates_to` target may be in another rig on the same daemon and need not exist at import time.
 See [fallback escalation across rigs](edge-types.md#fallback-escalation-across-rigs) for precedence, single-hop delivery,
@@ -1030,7 +1139,7 @@ file; missing optional files do not trigger that warning.
 | Hint | Behavior |
 |------|----------|
 | `auto` | System chooses based on file type and context. |
-| `guidance_merge` | Merged as a managed block into the runtime's guidance file: for Claude, the `managed_blocks` file (`CLAUDE.md` or `CLAUDE.local.md`); for Codex, Pi and OMP, `AGENTS.md`. Delivered before harness boot. |
+| `guidance_merge` | Merged as a managed block into the runtime's guidance file: for Claude, the `managed_blocks` file (`CLAUDE.md` or `CLAUDE.local.md`); for Codex, Pi and OMP, `AGENTS.md`, except that a fresh Pi seat normally uses its own agent folder's and a fresh Codex 0.164.0+ seat normally gets launch developer instructions (the conditions are in [Codex guidance at launch](agent-startup-guide.md#codex-guidance-at-launch) and [Runtime Differences](agent-startup-guide.md#runtime-differences)). Delivered before harness boot. |
 | `skill_install` | Installed as a skill in the runtime's skill directory. Delivered before harness boot. |
 | `send_text` | Sent as text to the agent's terminal after the harness is ready. Requires the agent TUI to be active. |
 
@@ -1255,6 +1364,7 @@ spec can pass `rig spec validate` and then fail at preflight or import. Run
 24. Startup action `idempotent` is a required boolean.
 25. Non-idempotent actions must not include `restore` in `applies_on`.
 26. `applies_on` values must be from: `fresh_start`, `restore`.
+27. `launch_wrapper` is valid on `pi`, `codex`, `claude-code` and supported community-harness members (`[]` included); see the [runner contract](harness-runner-contract.md#2-declare-it-what-the-preset-says-fixed-at-launch) for community capability requirements. It must be an array of strings without NUL/CR/LF, and `launch_wrapper[0]` (when non-empty) must be an absolute path or a bare command name on the daemon `PATH`.
 
 ---
 

docs/reference/runner-conformance/README.md

@@ -13,6 +13,8 @@ node check-runner.mjs --runner 'my-runner --agent-command {agent}' --name my-har
 - **`--name`** is your preset's name, which your posts carry as `runtime`.
 - **`--resume` and `--first-input`** say your preset declares `resume: true` and `first_input: true`. Without them,
   those cases are skipped, because OpenRig wouldn't pass that input.
+- **`--launch-wrapper`** says your preset declares `launch_wrapper: true` and enables case 10 from the
+  [runner contract](../harness-runner-contract.md#2-declare-it-what-the-preset-says-fixed-at-launch).
 - **`--cases 2,4`** runs only some cases. `--json` prints a machine-readable report. `--keep` keeps each launch's
   folder: the agent's message log, the launch record and the scenario.
 - **`--wait-ms 20000`** waits longer for each step (default 8000), for a runner whose agent takes a while to start. The check's shorter settling times, and the fake agent's own delays and holds, grow in proportion.
@@ -62,6 +64,7 @@ In an installed OpenRig, the check is at `$(npm root -g)/@openrig/cli/daemon/doc
 | 7c | 9 | **Advisory.** Closing the pane ends the agent process. Section 9 marks this duty as proposed, and lets a runner list in `limits` what may keep running, so this result is shown as met or not met and never counts as a pass or a fail, or toward the exit status. The fake agent outlives a hangup here, so the runner has to end it |
 | 8 | 7 | An unknown request from the agent gets JSON-RPC `-32601`, and the next prompt still goes through |
 | 9 | 9 | `OPENRIG_FIRST_INPUT`'s exact bytes, newlines and spaces included, are the first prompt |
+| 10 | 2 | The native child runs through the configured wrapper, preserving literal arguments, without inheriting the runner's wrapper handoff |
 
 ## What it doesn't check
 

docs/reference/runner-conformance/check-runner.mjs

@@ -6,12 +6,12 @@
 //   - a small local endpoint stands in for OpenRig's activity route, and records what your runner posts;
 //   - input arrives the way `rig send` delivers it: one bracketed paste with raw newlines, then Enter.
 //
-//   node check-runner.mjs --runner '<command>' [--name <preset name>] [--resume] [--first-input]
+//   node check-runner.mjs --runner '<command>' [--name <preset name>] [--resume] [--first-input] [--launch-wrapper]
 //                         [--cases 1,2,4] [--wait-ms 8000] [--pass-env NAME] [--json] [--keep]
 //
 // `{agent}` in the command is replaced with the fake agent's path, an executable that takes any arguments your runner
 // adds. For example: --runner 'my-runner --agent-command {agent}'. --resume and --first-input say your preset declares
-// `resume: true` and `first_input: true`; without them those cases are skipped, since OpenRig wouldn't pass that input.
+// `resume: true` and `first_input: true`; --launch-wrapper declares `launch_wrapper: true`. Without a declaration, its case is skipped.
 // The runner gets a clean environment: PATH, HOME, USER, SHELL, LANG and TMPDIR, the OpenRig variables, and any
 // --pass-env names. Exit status: 0 when nothing fails, 1 when a case fails, 2 for a usage or setup problem.
 //
@@ -27,10 +27,10 @@ import { dirname, join } from "node:path";
 import { fileURLToPath } from "node:url";
 
 const here = dirname(fileURLToPath(import.meta.url));
-const usage = "usage: node check-runner.mjs --runner '<command with {agent}>' [--name <preset>] [--resume] [--first-input] [--cases 1,2] [--wait-ms 8000] [--pass-env NAME] [--json] [--keep]";
+const usage = "usage: node check-runner.mjs --runner '<command with {agent}>' [--name <preset>] [--resume] [--first-input] [--launch-wrapper] [--cases 1,2] [--wait-ms 8000] [--pass-env NAME] [--json] [--keep]";
 
 function parseArgs(argv) {
-  const opts = { name: "runner", cases: undefined, passEnv: [], resume: false, firstInput: false, json: false, keep: false, waitMs: 8000 };
+  const opts = { name: "runner", cases: undefined, passEnv: [], resume: false, firstInput: false, launchWrapper: false, json: false, keep: false, waitMs: 8000 };
   for (let i = 0; i < argv.length; i += 1) {
     const flag = argv[i];
     const value = () => { if (i + 1 >= argv.length) throw new Error(`${flag} needs a value`); return argv[++i]; };
@@ -41,6 +41,7 @@ function parseArgs(argv) {
     else if (flag === "--wait-ms") opts.waitMs = Number(value());
     else if (flag === "--resume") opts.resume = true;
     else if (flag === "--first-input") opts.firstInput = true;
+    else if (flag === "--launch-wrapper") opts.launchWrapper = true;
     else if (flag === "--json") opts.json = true;
     else if (flag === "--keep") opts.keep = true;
     else throw new Error(`unknown option ${flag}`);
@@ -178,7 +179,7 @@ await new Promise((resolve) => server.listen(0, "127.0.0.1", resolve));
 const base = `http://127.0.0.1:${server.address().port}`;
 
 const launches = [];
-async function launch(label, { scenario = {}, env = {}, asks } = {}) {
+async function launch(label, { scenario = {}, env = {}, asks, launchWrapper = false } = {}) {
   const dir = mkdtempSync(join(tmpdir(), `runner-check-${label}-`));
   const cwd = join(dir, "work");
   mkdirSync(cwd);
@@ -196,6 +197,13 @@ async function launch(label, { scenario = {}, env = {}, asks } = {}) {
   const agent = join(dir, "agent");
   writeFileSync(agent, `#!/bin/sh\nexec ${sh(process.execPath)} ${sh(join(here, "fake-acp-agent.mjs"))} --scenario ${sh(join(dir, "scenario.json"))} --log ${sh(h.agentLog)} "$@"\n`);
   chmodSync(agent, 0o755);
+  h.agent = agent;
+  if (launchWrapper) {
+    h.wrapperLog = join(dir, "wrapper.json");
+    const wrapper = join(dir, "wrapper.sh");
+    writeFileSync(wrapper, `#!/bin/sh\nrecord=$1\ntag=$2\nseparator=$3\nshift 3\n${sh(process.execPath)} -e 'const [file, tag, separator, ...argv] = process.argv.slice(1); require("node:fs").writeFileSync(file, JSON.stringify({tag, separator, argv, launchWrapper: process.env.OPENRIG_LAUNCH_WRAPPER ?? null}));' "$record" "$tag" "$separator" "$@"\nexec "$@"\n`);
+    env = { ...env, OPENRIG_LAUNCH_WRAPPER: JSON.stringify(["/bin/sh", wrapper, h.wrapperLog, "literal ' $ ; space", "--"]) };
+  }
   const vars = {
     PATH: process.env.PATH, HOME: process.env.HOME, USER: process.env.USER, SHELL: process.env.SHELL, LANG: process.env.LANG,
     TMPDIR: process.env.TMPDIR, TERM: "xterm-256color",
@@ -310,7 +318,7 @@ const cases = {
 
   async "2"() {
     const r = result("2", "Input framing, queued messages and cancel");
-    const h = await launch("2", { scenario: { turns: [{}, { holdMs: scale(15000) }, { holdMs: scale(2500), ignoreCancel: true }, {}] } });
+    const h = await launch("2", { scenario: { turns: [{}, { holdMs: scale(15000) }, { holdMs: scale(15000), ignoreCancel: true, endAfterCancelMs: scale(300) }, {}] } });
     if (!r.check("the runner opens its conversation", await ready(h), "no launch record")) return r;
     const block = message(h, "first line\n/abort\nthird line");
     h.paste(block);
@@ -339,10 +347,12 @@ const cases = {
     r.check("one /abort sends one session/cancel", h.received("session/cancel").length === 1, `got ${h.received("session/cancel").length}`);
     r.check("the waiting block becomes the next prompt", await waitFor(() => h.prompts()[2] === queued, scale(4000)), `prompts: ${JSON.stringify(h.prompts().slice(2))}`);
 
+    // The agent keeps this turn open until the cancel arrives and then ends it end_turn, so a runner that handles Ctrl-C
+    // late on a busy machine still has a turn to cancel. Only its delivery is timed, by the per-step bound.
     const interrupted = Date.now();
     await sleep(scale(200));
     h.keys("C-c");
-    r.check("Ctrl-C outside a paste sends session/cancel", await waitFor(() => h.received("session/cancel").length >= 2, scale(3000)), "no second session/cancel");
+    r.check("Ctrl-C outside a paste sends session/cancel", await waitFor(() => h.received("session/cancel").length >= 2), "no second session/cancel");
     const ended = await waitFor(() => nonExitStops(h, interrupted)[0], scale(6000));
     r.check("cancelled is reported only on the agent's own stop reason", ended?.body?.subtype === "end_turn",
       `the agent ended the turn end_turn; the runner reported ${ended ? `Stop/${ended.body.subtype}` : "no Stop"}`);
@@ -779,9 +789,22 @@ const cases = {
     if (!opts.keep) rmSync(dir, { recursive: true, force: true });
     return r;
   },
+  async "10"() {
+    const r = result("10", "Launch wrapper: the runner wraps its native child and scrubs the handoff environment");
+    if (!opts.launchWrapper) return r.skip("the preset doesn't declare launch_wrapper (--launch-wrapper): OpenRig passes no OPENRIG_LAUNCH_WRAPPER");
+    const h = await launch("10", { launchWrapper: true });
+    r.check("the wrapped child opens its conversation", await ready(h), "no launch record");
+    const seen = existsSync(h.wrapperLog) ? JSON.parse(readFileSync(h.wrapperLog, "utf8")) : undefined;
+    r.check("the wrapper receives literal arguments and the native child command", seen?.tag === "literal ' $ ; space" && seen?.separator === "--"
+      && seen?.argv?.[0] === h.agent, JSON.stringify(seen ?? null));
+    r.check("OPENRIG_LAUNCH_WRAPPER is absent from the wrapper and child environment", seen?.launchWrapper === null
+      && h.events("started")[0]?.launchWrapper === null, "wrapper handoff leaked to native child");
+    return r;
+  },
+
 };
 
-const order = ["1", "1b", "2", "3", "3b", "4", "4b", "4c", "5", "6", "7", "7b", "7c", "8", "9"];
+const order = ["1", "1b", "2", "3", "3b", "4", "4b", "4c", "5", "6", "7", "7b", "7c", "8", "9", "10"];
 const selected = opts.cases ?? order;
 const unknown = selected.filter((id) => !order.includes(id));
 if (unknown.length) { console.error(`unknown case ${unknown.join(", ")}; cases are ${order.join(", ")}`); process.exit(2); }

docs/reference/runner-conformance/fake-acp-agent.mjs

@@ -17,7 +17,8 @@
 //   turns            one entry per `session/prompt`, in order (the last one repeats):
 //     holdMs         keep the turn open this long; `session/cancel` ends it early as "cancelled"
 //     silent         send no message chunk for this turn, so a held turn is entirely quiet (a long tool or model call)
-//     ignoreCancel   on `session/cancel`, keep the turn open and end it with stopReason (default end_turn) anyway
+//     ignoreCancel   on `session/cancel`, keep the turn open and end it with stopReason (default end_turn) anyway, when
+//                    holdMs runs out or, with endAfterCancelMs, that long after the first cancel if that's sooner
 //     permission     ask `session/request_permission` first, then carry on whatever the answer. As an object:
 //                      options      the options offered (default: allow-once, allow_once; reject-once, reject_once)
 //                      waitMs       stop waiting for the answer after this long, and end the turn
@@ -41,7 +42,7 @@ const fail = (id, code, message) => send({ jsonrpc: "2.0", id, error: { code, me
 const exit = (status, why) => { log({ dir: "event", event: "exit", status, why }); process.exit(status); };
 const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
 
-log({ dir: "event", event: "started", pid: process.pid, cwd: process.cwd(), argv: process.argv.slice(2) });
+log({ dir: "event", event: "started", pid: process.pid, cwd: process.cwd(), argv: process.argv.slice(2), launchWrapper: process.env.OPENRIG_LAUNCH_WRAPPER ?? null });
 // Ctrl-C is the runner's to handle, so the agent never takes it as a signal. A stubborn agent also outlives a hangup and
 // its input closing, as some real agents do: only the runner ending it (SIGTERM or SIGKILL) stops it.
 process.on("SIGINT", () => log({ dir: "event", event: "signal", signal: "SIGINT" }));
@@ -90,7 +91,12 @@ async function prompt(id, params) {
   if (spec.holdMs) {
     const cancelled = await new Promise((resolve) => {
       const timer = setTimeout(() => { cancelTurn = null; resolve(false); }, spec.holdMs);
-      cancelTurn = () => { if (spec.ignoreCancel) return; clearTimeout(timer); cancelTurn = null; resolve(true); };
+      cancelTurn = () => {
+        if (!spec.ignoreCancel) { clearTimeout(timer); cancelTurn = null; resolve(true); return; }
+        if (spec.endAfterCancelMs === undefined) return;
+        cancelTurn = null;
+        setTimeout(() => { clearTimeout(timer); resolve(false); }, spec.endAfterCancelMs);
+      };
     });
     if (cancelled) stopReason = "cancelled";
   }

docs/reference/sdlc-conventions.md

@@ -144,6 +144,23 @@ changes either. `rig scope` reads an `established` or `canonical` slice whose
 `verified:` is missing, has no source, names a tool-scaffolded placeholder, or is
 older than 90 days as `provisional`.
 
+A slice's frontmatter may also carry `status:`. `GET /api/slices` reports the
+bucket each word falls in as the slice's `status`, and its
+`?filter=all|active|done|blocked` filters on that bucket:
+
+| `status:` word | Bucket |
+| --- | --- |
+| `active`, `in-flight`, `ratified` | `active` |
+| `draft`, `draft-pending-orch-ratification`, `placeholder` | `draft` |
+| `done`, `shipped`, `promoted`, `closed` | `done` |
+| `blocked`, `parked-with-evidence` | `blocked` |
+
+Words are matched ignoring case. Any other word is matched by what it contains:
+`done`, `ship` or `close` makes it `done`; `block` or `park`, `blocked`; `draft`
+or `pending`, `draft`. A word that matches none of these counts as `active`, so a
+slice marked `complete` counts as `active`. A slice with no `status:` counts as
+`draft`.
+
 When the owning scope selects `proofPolicy.judges`, use `rig proof judge` to
 record an attributed accept, reject or withdraw decision on a contract item,
 and `rig proof show` to read derived readiness up through slice, mission and
@@ -178,9 +195,11 @@ missions whose `mission.yaml` carries `metadata.status: active` or
 a mission does not change the project's readiness. A `planned` mission stays
 outside the rollup, even after its slices are accepted, until someone changes
 its status to `active` (or `release-candidate` for a release that is ready to
-judge). Nothing changes the status automatically. Missions created before this
-default carry no status and are not counted either; add the line by hand to
-include them. It also writes `metadata.name`, the mission's folder name, which
+judge). Nothing changes the status automatically. A mission can also carry
+`metadata.status: complete`, which only `POST /api/missions/:missionId/complete`
+writes (no `rig` verb does); a `complete` mission is outside the rollup too.
+Missions created before this default carry no status and are not counted either;
+add the line by hand to include them. It also writes `metadata.name`, the mission's folder name, which
 the project lifecycle compiler requires; a mission created without it refuses to
 compile with `lifecycle_field_missing` until you add that line.
 

docs/reference/seat-delivery.md

@@ -10,7 +10,7 @@ type: `openrig-urgent`, `urgent-only` and `sends-only`. Sibling seats keep their
 | `off` (default) | Existing delivery and prompt checks | Existing send result |
 | `draft-aware` | Paste into a recognized empty composer; check ownership again before Enter | Hold input and retain this attempt for inspection when bookkeeping is available |
 | `openrig-urgent` | OpenRig's own messages (watchdogs, scheduled wakes, notices) and anything marked urgent type, never over a draft | Other agents' messages are held, the sender is told, and held messages stay readable |
-| `urgent-only` | Only messages marked urgent type (`rig send --urgent`, or a queue item at priority urgent or critical), never over a draft | Everything else is held, the sender is told, and held messages stay readable |
+| `urgent-only` | Only messages marked urgent type (`rig send --urgent`, `urgent: true` on `POST /api/transport/send`, or a queue item at priority urgent or critical), never over a draft | Everything else is held, the sender is told, and held messages stay readable |
 | `sends-only` | Deliberate sends from people and agents type, as with `off`; a terminal member's typing isn't protected, because its shell input can't be read | OpenRig's own notices are held. A terminal member defaults to this level |
 | `hold` | No automatic terminal input, even at an empty prompt | Retain for inspection outside the pane |
 
@@ -41,6 +41,69 @@ nothing to send later.
 
 ## Configure and inspect
 
+### Latest permission observation
+
+`rig seat status <seat>` can show the latest observed harness permission request,
+with a fixed tool descriptor, source time and actor category. This is one latest
+observation across actors, not an aggregate of pending prompts. `request_observed`
+leaves the displayed actor's pending state unknown. `completed` applies only to
+that actor's later tool batch; other actors' pending state is still unknown. An
+unavailable observation proves neither a permission stop nor the absence of a
+waiting native prompt. With no observation, human status omits this section;
+JSON keeps the explicit unavailable state. The JSON field is `permission_stop`,
+and its `pending` value has the same observation-only scope.
+
+An observation may also carry Claude's **suggested rules** as `suggested_rules`.
+These are read-only evidence, not applied or approved, not guaranteed effective,
+not the exact command, and not a complete list of dialog options. The supported
+subset of [Claude's permission update schema](https://code.claude.com/docs/en/hooks#permission-update-entries)
+is `addRules`, `replaceRules` and `removeRules`, preserving type, ordered rules
+(`toolName`, optional `ruleContent`), behavior and destination exactly. Strings
+are neither trimmed nor broadened; absent `ruleContent` remains distinct from
+an empty string. Mode/directory updates and malformed entries are withheld.
+
+The envelope distinguishes a missing native field from a present empty array;
+`unsupported` and `size_cut` mark incomplete evidence. An older relay without
+this field, or a failed formatter, is unavailable rather than proof of native
+absence. Bounds are eight entries, sixteen rules per entry, 128 UTF-16 code
+units per tool name, 2,048 per rule-content string, and 16,384 UTF-8 bytes of
+serialized evidence. Oversized whole entries are dropped with `size_cut`,
+never shortened into different plausible rules. Both producer and ingress
+validate these bounds independently. Terminal JSON escapes controls and
+nonprinting format/separator characters while preserving decoded string values;
+its representation can expand, so the byte bound applies to stored/transported
+evidence, not the entire formatted status output.
+
+**Suggested rule strings can contain sensitive values.** The relay sends only
+this bounded rule subset alongside the fixed descriptor. Rule content can itself
+contain sensitive command-like text; raw `tool_input`, arguments outside that
+suggestion, and provider reasons are not forwarded. It posts to `/api/activity/hooks` at the
+configured `OPENRIG_URL` (or host/port), or the discovered endpoint in
+`OPENRIG_HOME/activity-endpoint.json`. Both `GET /api/seat/status/:seatRef` and
+`GET /api/seat/status/node/:nodeId` can return the evidence. The daemon adds no
+event, log, queue or telemetry persistence for it; a caller can still capture
+status output in a transcript or file. Evidence shares the observation's
+five-minute lifetime and existing seat/actor limits, is discarded lazily on the
+next observation/status read after expiry, and is lost on restart. No rule,
+permission/configuration write, hook decision, edited input, prompt submission
+or new attention/delivery behavior is produced.
+
+Actor attribution is a convention derived from the [Claude hook schema](https://code.claude.com/docs/en/hooks#common-input-fields): an omitted or null `agent_id`
+is treated as main, a valid ID identifies a subagent or in-process teammate, and
+an invalid ID is dropped rather than inferred as main. The fixtures are synthetic;
+no live PermissionRequest capture qualifies that convention on an installed native
+version. A hook denial without a source-bound guard receipt has unknown origin.
+This diagnostic adds no guard collector, permission decision or attention policy.
+
+The new PermissionRequest diagnostic is optional for activity-hook readiness.
+Every preexisting required event, relay asset and explicit hook-disable check
+still applies. An older installed plugin at the same manifest version retains
+its bytes and may not send this event: plugin-only delivery needs a future normal
+maintainer release refresh, and managed delivery needs normal reprojection. A
+manifest or green readiness check does not prove live hook execution. Raw tool
+arguments and provider reason text are withheld; diagnostic evidence is bounded,
+memory-only and expires after five minutes.
+
 ### Who may intervene
 
 `rig seat status <seat>` shows the seat's terminal interaction account alongside
@@ -90,12 +153,16 @@ drive a recognized native menu as before. Otherwise, with the delivery guard
 attached, it reads the visible composer under the seat's lease immediately before
 pasting, after buffer preparation. Empty input proceeds. A recognized dim
 suggestion is displayed separately and counts as empty; typed or pasted text
-(author unknown) stops the send with `override_input_present`. An unreadable input
-stops with `override_input_unknown`, never an empty-input claim. Both return single-send HTTP
+(author unknown) stops the send with `override_input_present`: single-send HTTP
 409, `sent: false`, and caller-only `existingInput` evidence (visible rows, suggestion
 rows, cursor/dimensions and target binding). A collapsed paste shows its label,
 not the hidden contents. OpenRig neither holds nor retries this stopped override.
 
+An input that can't be read doesn't stop the override. The answer is typed and
+nothing is cleared. The result reports `existingInput` with classification
+`unreadable`, never empty, and a warning that the answer may join or submit
+existing text.
+
 Inspect the shown input in the native terminal, deliberately clear the prompt with
 native/tmux controls, then issue the send again. Manual clearing runs outside
 OpenRig's lease and observation checks and is not recorded. For a deliberate
@@ -141,8 +208,21 @@ to finish. The API may return HTTP 202 with `pending: true`; wait for
 and effective modes and retains the text `Typing guard: on (hold)` for hold.
 
 A held send immediately returns `outcome: "retained"`, `sent: false`, its exact
-reason (such as `draft_input_busy`) and the saved attempt's `outboxIds`. No automatic
-retry is scheduled. Each explicit draft-aware request is a new attempt, including
+reason and the saved attempt's `outboxIds`. No automatic
+retry is scheduled. The reasons are:
+
+- `typing_guard_enabled`: the seat's mode holds this send (`hold`, or a level that
+  holds this kind of message).
+- `runtime_switch_in_progress`: the seat is switching runtime.
+- In `draft-aware` mode, a refusal before any input was written:
+  `draft_input_busy` (the input holds unsubmitted text), `draft_input_unknown` (the
+  input can't be read), `draft_input_changed` (the input or its target changed during
+  the delivery), `draft_wake_superseded`, `target_needs_input`,
+  `wait_for_idle_timeout`, `target_activity_unknown`, `target_runtime_unverified`,
+  `target_runtime_not_running`, `target_runtime_conflict`, `invalid_wait_for_idle`,
+  `invalid_dangerously_interact`, `dangerously_interact_requires_reason`,
+  `prompt_override_audit_unavailable`, `transport_unavailable`, `session_missing` or
+  `tmux_unavailable`. Each explicit draft-aware request is a new attempt, including
 one with a reused `deliveryId`; the result warns older callers that this field no
 longer deduplicates input. The CLI does not generate a deduplication key.
 
@@ -160,6 +240,26 @@ false` and the protection reason, rather than claiming durable custody.
 Reading never sends. Retirement frees that record's active retention quota while
 preserving evidence; it does not deliver a message or close queue work.
 
+A seat holds one message per sender, so one sender can't fill its held quota and
+drop everyone else's messages. A sender's newest held message retires its earlier
+ones at that seat into itself (`retired_by: outbox@system`, reason `coalesced into
+<id>`), in every mode, `draft-aware` included. The newest names them: `rig seat
+held-messages` shows `coalesced: { count, outboxIds }` on it, and each coalesced
+record stays readable with `--id`. Held messages are never replayed, so this changes
+no delivery. A `draft-aware` attempt takes its sender's place only once it settles
+as held; one that is then delivered leaves the held messages as they are. A queue
+wake to a `draft-aware` seat counts its sender's held message there as room, by the
+same rule. At a full quota, a sender with nothing held there is still refused, and
+nothing from another sender is touched. A committed wake is never refused, and the
+sender's next held message coalesces it like any other. A message that settles as
+held late, behind a newer held send, folds into that send's newest part and leaves
+its other parts held. An urgent message coalesced under a routine one passes its
+marking to the newest. Never coalesced: the other parts of the same send, a
+watchdog's reminders (one per job), and an unattributed sender.
+
+A backlog held before this change clears at that sender's next held message to the
+seat, or at once with `rig seat retire-held-message`.
+
 ## Input evidence and outcomes
 
 The cursor-aware reader lives in `session-transport.ts` and shares the composer
@@ -260,7 +360,10 @@ accounting. Historical outbox receipts do not suppress a lone reminder or preven
 message tells its recipient that delivery may repeat after an unconfirmed attempt. Its job history records the
 current transport result; an unknown write is not proof that nothing arrived.
 
-The transport still protects approval/question prompts, human drafts and recipient identity. Scheduler timing,
+The transport still protects approval/question prompts and recipient identity. Under the default
+`transport.draft_delivery` (`deliver`), an ordinary send or work notice that meets a person's draft at a Claude Code or
+Codex prompt is pasted into it and the combined input is submitted, unless the seat's typing guard holds input
+(`draft-aware`, `openrig-urgent`, `urgent-only` or `hold`). With `refuse`, a draft refuses the send. Scheduler timing,
 current parked-row eligibility and stop checks still apply. This changes message retries, not the idempotence of
 structured continuity actions such as creating a cutover baton.
 

docs/reference/slack-app-setup.md

@@ -66,7 +66,14 @@ not fill in; if so, follow the form.
     asked at RECORD stays on Steering and the Feed without a post. Running it again reuses each ask's
     one decision rather than making another, so an ask whose post is confirmed isn't posted again. As
     with any Slack post, one whose earlier attempt Slack couldn't confirm may post once more, marked as
-    a possible repeat. `--wait <seconds>` (default 10) is how long it waits for Slack receipts.
+    a possible repeat. An ask OpenRig can't fully check still posts, once per gateway run, through the
+    person's delivery rules, so a readable quiet window still holds it. That's an ask whose attempt
+    record can't be read, whose state can't be checked, or whose person's quiet hours can't be read.
+    The Slack post can show "Delivery may repeat after an earlier unconfirmed attempt.", "OpenRig
+    couldn't check that this ask is still open." and "OpenRig couldn't read your quiet hours." These
+    supplementary notes are omitted if they do not fit. The row and command also report when the
+    person's quiet-hours setting couldn't be read. An ask that is no longer open isn't posted. `--wait <seconds>` (default 10) is how long it waits for
+    Slack receipts.
 
 ## What the app asks for
 
@@ -99,8 +106,24 @@ So after installing, compare the granted scopes Slack shows for the app with all
 `rig slack manifest --json` lists.
 
 The app subscribes to messages in public channels it is a member of (`message.channels`), to
-mentions of the app (`app_mention`), and to emoji reactions (`reaction_added`). It does not request
-direct-message or private-channel access.
+mentions of the app (`app_mention`), and to emoji reactions (`reaction_added`). The manifest does
+not request direct-message or private-channel access.
+
+### Using a private channel
+
+The connector doesn't check whether a channel is public or private, so a private channel can be
+the default channel or a mapped one once the app can read it. OpenRig doesn't test this setup. In
+the app's settings:
+
+1. Under **OAuth & Permissions**, add the bot scopes `groups:history` (read messages in private
+   channels the app is in) and `groups:read` (read a private channel's details; `rig slack verify`
+   needs it to confirm the app is a member).
+2. Under **Event Subscriptions**, subscribe to the bot event `message.groups`.
+3. Reinstall the app to the workspace, then invite it to the private channel.
+4. To have `rig slack verify` require the new scopes too, run
+   `rig slack setup --required-scopes chat:write,channels:history,channels:read,groups:history,groups:read`.
+
+Direct messages are not supported.
 
 A reaction added to an ask reaches the agent that asked, as a note naming who reacted and with
 which emoji; the agent decides what it means. Every message OpenRig posted for the ask counts:
@@ -166,6 +189,9 @@ rig slack channel-map list
 rig slack channel-map remove pr@my-rig
 ```
 
+A client can offer the channels to choose from with `GET /api/gateway/slack/channels`, which lists the public
+channels the bot token can see, with their ids ([Writing a daemon client](daemon-http-clients.md#route-details)).
+
 - **Which channel an item posts to.** The item's seat is matched exactly first (`pr@my-rig`),
   then its rig (`my-rig`). Anything unmapped posts to the default channel, so a config
   without a map behaves exactly as before. Aggregate delivery digests have no single seat and
@@ -221,8 +247,9 @@ configuration matches the file.
 
 The map is stored as `channelMap` in `slack-connector.json`, as a list of `{ "match", "channel" }`
 entries, each with an optional `"inbound"` seat. Saving refuses any other field. A field written by a newer OpenRig is ignored when the
-connector reads the file, and `rig slack status` names it. Private channels and direct messages
-are not supported targets.
+connector reads the file, and `rig slack status` names it. A private channel can be a target once
+the app has the private-channel scopes ([Using a private channel](#using-a-private-channel)).
+Direct messages are not supported targets.
 
 ## Named bots (optional)
 
@@ -329,8 +356,7 @@ decisions, and posts its own replies.
 - **Changing forwarding entries is not a drain or cancellation:** relays already retained keep
   their original host and remain visible in `rig slack status`, even after the last entry is
   removed. An enabled connector retries them; disabling it pauses retry without deleting them.
-  Let relays finish before removing a forward entry. A zero dead-letter backlog alone does not
-  prove they have drained: it does not count relay HTTP requests still in flight.
+  Let relays finish before removing a forward entry; `rig slack status` shows local in-flight attempts separately from retained records, but zero does not prove global quiescence.
 - **Who answered:** every post starts `From <seat> on <machine>`, using the machine's host id (set
   it with `host.name`).
 - **The hub's own channels** are handled on the hub, as before. A channel can be relayed to one
@@ -358,14 +384,53 @@ Quiet hours hold a person's routine asks while they're on, and post them once th
   wake-ladder alerts. These reach the person exactly as they would outside quiet. An ask below the posting level is only
   logged, as at any time.
 - **When quiet ends,** each held ask that never started posts once, on its own, with no mention. One that no longer applies
-  isn't posted, and its row says why: it closed, a newer notification replaced it, its recipient changed, or it's now below
-  the posting level.
+  isn't posted, and its row says why: it closed, a newer notification replaced it, its recipient changed, its person has no
+  connector any more (it stays on Steering and the Feed), or it's now below the posting level.
 - **A post that may already be in Slack**, confirmed or only partly sent, isn't held: it finishes as itself, in its own
   thread.
-- **If quiet hours can't be read,** nothing is posted to that person until they can.
+- **If quiet hours can't be read,** that isn't a quiet window. An ask that would post goes to its usual destination with
+  no mention, and the post says "OpenRig couldn't read your quiet hours."; the ask's row records it. The person's other
+  delivery settings still decide: an escalation, a digest, an ask below the posting level, and a quiet window that can be
+  read work as before. This covers ordinary posts and digests, and asks already held for quiet hours: while the quiet
+  hours can't be read, the next release check (within about a minute) posts each of them, with no mention and the same
+  line, even if the window may still be on. The line is supplementary: a post at Slack's text or block limit leaves it
+  out, and the ask's row and `rig slack post-open-asks` still say the quiet hours couldn't be read. A person reached
+  through a relay waits until their quiet hours can be read, as before, because that post can't carry the line.
 
 ## Reconnect recovery and status
 
+When a posted ask is answered, closed, canceled or replaced elsewhere, the connector edits its
+original primary message with the status, time, actor and closure note. Buttons become option
+text and the answer hint disappears; the remaining content and reconciliation marker stay as
+posted. Supplemental messages are unchanged. This adds no post or reaction and needs no agent
+turn. Separate resolution notices retain their existing behavior.
+
+A deliberately reopened direct ask says **Open again**; its options remain as text. If it has
+no committed answer, it tells the person to reply in its thread. Otherwise it says **Earlier
+answer retained**: reopening does not accept a second answer. Earlier replaced parks remain
+historical; **Asked again** does not claim that the later ask has been posted. Edits reconcile
+the current state again after awaited Slack calls, so an obsolete close is not recorded as the
+current result after a reopen.
+
+Edits follow committed queue changes, with a startup catch-up for posted asks on retained queue
+rows. Each primary costs one exact-message read and one `chat.update`. The serial edit worker
+uses the recovery budget helpers (four calls, 100 entries, 15 seconds) and waits for the next
+budget window when exhausted. Slack's per-method `Retry-After` is retained across restarts.
+Only an acknowledged edit gets a success receipt. `rig slack status` shows queued reconciliation
+and the last edit failure. Transient failures retry with increasing waits, capped at an hour,
+behind other queued work. Permanent Slack failures (such as missing scope, a deleted original,
+an archived channel or an uneditable message) remain unfinished. A deleted original is recorded
+on the row and never asked about again; a new post of the same ask is tried afresh. The others,
+which a scope, membership, a different bot token or an upgrade can fix, are retried on the next
+connector start. A message is uneditable when the configured token didn't post it. They do not keep consuming edit capacity. These edit calls have their own budget,
+separate from missed-message scanning below.
+
+The baseline already includes public-channel history access and `chat:write`. Private-channel
+reads may require `groups:history`, which the baseline does not grant; a refused read never
+counts as an edit. An unavailable/deleted original cannot be replaced by a new post. Startup
+catch-up cannot edit rows already removed by queue retention or posts without their original
+notification receipt and channel stamp.
+
 The daemon scans available top-level messages and replies broadcast with **Also send to
 channel** in the channels this machine owns: its default channel and its channel map's, except a
 channel it relays to another machine. It also scans replies in the threads it tracks in those
@@ -405,9 +470,18 @@ not covered by that margin. A corrupt or unreadable checkpoint leaves recovery u
 and preserves the existing file; live inbound continues.
 
 Each pass admits at most four history calls, 100 entries and 15 seconds of work, with a
-five-second request ceiling. Channels and threads take turns within that one budget, so with many
-tracked threads each thread is scanned every few passes; `rig slack status` shows how far behind
-the oldest waiting thread is. Work already started on a message, such as
+five-second request ceiling. Channels and threads take turns within that one budget. Each tracked
+thread is scanned once after each start and after each Socket Mode reconnect that followed a gap
+(not a refresh Slack overlaps with the old connection), which catches replies sent while no
+connection was up; replies while connected arrive live. A thread the pass at the connect itself
+reaches is scanned once more, since that pass stops five seconds short of the connect. Those scans go in need order: a thread
+leaving the seven-day window within six hours first, since once it leaves a missed reply is never
+recovered; then threads whose ask still waits on the person (its queue row is pending, in progress
+or blocked) before the rest, each newest first. The order comes from the saved checkpoints, so a
+restart starts with the newest open ask, not the oldest thread. After that a thread is scanned again
+only as a backstop, hourly while its ask waits on the person and daily otherwise. A machine the hub
+relays to can't see a gap in relayed events, so there a settled thread's backstop is three hours. `rig slack status` shows how far behind the oldest waiting thread is and how
+many threads are settled. Work already started on a message, such as
 an attachment transfer, is not cancelled by these limits or by a stop, and a new
 pass waits for it to finish. Rate limits retain Slack's Retry-After across restart, and hold back
 only the method Slack limited (channel history or thread replies).
@@ -416,8 +490,8 @@ When Slack reports a plan history limit, recovery still lands the available page
 and advances normally, retaining an older-history limitation in status across restarts.
 Missing bot credentials, history scope, membership, retention limits and API
 failures appear as recovery limitations. A tracked thread Slack can't find stays incomplete in
-status, with no further call, until its thread mapping changes or the daemon restarts or rewires
-the connector; that isn't proof the thread was deleted. No additional scope is required to keep
+status and is asked again at most once a day, across restarts; that isn't proof the thread was
+deleted. No additional scope is required to keep
 live delivery working. Threads nobody tracks, older open threads, threads closed more than seven
 days ago and other channels remain outside this recovery scope; global chronological order is
 not promised. This change does not replay intervals already checkpointed before broadcast-reply

docs/reference/telemetry.md

@@ -115,3 +115,67 @@ successful empty history. The CLI preserves the error body in JSON mode and exit
 1 for request errors, 2 for unavailable reads. Existing SSE and usage APIs retain
 their behavior. Collector deployment, exporters and durable store epochs are
 separate work.
+
+## Independent exporter mapping and limits
+
+OpenRig's telemetry guarantee ends at the finite HTTP/CLI read envelope above. Any
+projection of those finite reads (or of cached rig/node inventory) into
+OpenTelemetry logs, custom metrics, or spans is defined and owned by an external
+consumer or exporter—not guaranteed by OpenRig's core API.
+
+The external OTLP exporter at `smota/openrig-telemetry` applies the following
+exporter-specific mapping choices to OpenRig's source facts (other consumers may
+make different choices):
+
+- **Recorded queue-state intervals (not execution time)**: In OpenRig, queue
+  transition rows record state changes (`pending`, `in-progress`, `blocked`, and
+  departure states), not agent execution duration or model latency; the first
+  observation of an item has no prior boundary; repeated same-state rows remain
+  distinct source records; and any page carrying an ID-range or withheld-record gap
+  (`before_retained_floor`, `retained_sequence_gap`, `conflicting_transition_id`,
+  `record_byte_limit`, `historical_transitions_not_read`, truncated gap details, or
+  a boot/watermark discontinuity) cannot prove what happened in the missing range.
+  In `smota/openrig-telemetry`, the exporter pairs an entering state with a later
+  departure state to measure only the recorded interval between two observed rows
+  with valid, ordered UTC timestamps, suppresses or drops open intervals across gap
+  boundaries rather than inventing durations, and reports cumulative transition
+  counts only for rows observed from its checkpoint (including historical rows when
+  started with `start=retained`), never lifetime daemon totals.
+- **No current-owner, seat, tenure, or native-session join**: In OpenRig's finite
+  API, the `queue-transitions` stream does not join transition rows to current seat
+  owners, mutable node or rig names, occupant tenures (`tenures`), or native
+  harness sessions, and it never selects queue bodies or transition notes; each
+  transition row does include `qitemId`, `actorSession`, `closureTarget`, and
+  `originalTimestamp`. In its exported OTLP output, `smota/openrig-telemetry`
+  chooses to emit derived queue spans as root/internal observations without
+  invented parent/child trace hierarchy, cross-stream causal links, or `gen_ai.*`
+  model/token/cost attributes; places `qitemId` only on opt-in log/span attributes
+  rather than metric dimensions; and omits the source fields `actorSession`,
+  `closureTarget`, and raw `originalTimestamp` from exported OTLP logs, metrics,
+  and spans.
+- **Sampling unknowns on cached seat inventory**: OpenRig's cached rig and seat
+  inventory endpoints (`GET /api/rigs`, `GET /api/rigs?archived=only`,
+  `GET /api/rigs/:rigId/nodes`) return point-in-time daemon projections
+  (`activityState.display`, `sessionStatus`, `arbitratedNeedsInput`, and
+  `activityState.lastSwap`), not a continuous event stream or independent OS
+  process probe; intermediate activity changes or multiple occupant handovers
+  between polls are not observable from those snapshots, and initial
+  observation has no backfilled history. In OpenRig source,
+  `arbitratedNeedsInput` is a tri-state boolean (`true`, `false`, or `null`):
+  `true` means the seat needs input, `false` means it was observed clear of
+  needing input, and `null` means no evidence exists either way (it does not
+  indicate `false`). Separately, external consumers or metric mappings must
+  choose how to interpret or project this tri-state value into their own
+  metrics or counters. `smota/openrig-telemetry` polls these snapshots as
+  point-in-time observations and relies on downstream metric storage to enforce
+  freshness/staleness TTL when polling stops or a seat departs.
+
+### External OTLP exporter repository and version compatibility
+
+An independent, standalone OTLP/HTTP exporter is maintained outside OpenRig core at:
+
+- **Repository**: <https://github.com/smota/openrig-telemetry>
+
+Its repository documents installation, configuration, and supported OpenRig versions; OpenRig does not test or guarantee the exporter.
+
+OpenRig's finite telemetry reads are versioned (`schemaVersion: 1`), while cached rig and node inventory endpoints are operational projections rather than a versioned telemetry contract.

docs/reference/terminal-sessions.md

@@ -72,11 +72,11 @@ verified viewer mapping.
 
 New tiles keep their place when a seat's tmux session is replaced. They wait up
 to 30 seconds for the same session name and managed seat identity to return,
-then reattach with the original read-only setting. This uses the shared tile
-command for Herdr, cmux, Rex and plain tmux, including SSH tiles. In-place
-handover and restart keep the existing attachment.
+then show the new session with the original read-only setting. This uses the
+shared tile command for Herdr, cmux, Rex and plain tmux, including SSH tiles.
+In-place handover and restart keep showing the same session.
 
-A deliberate detach closes the tile after about two seconds. A different seat
+Detaching a tile you typed in returns it to showing the seat. A different seat
 under the old name is never followed. If no replacement returns in time, or a
 session has no managed identity, reopen the rig view with
 `rig terminal open rig:<name> --window` (or your explicit provider). Views
@@ -153,6 +153,10 @@ timed out, OpenRig closes nothing: the result says `rex_outcome_unknown`, or
 adds a note when other pages opened, and names the session to check in Rex
 before retrying. `rig terminal status` gives Rex about a second to answer.
 
+## herdr-app
+
+`--provider herdr-app` opens the view in the herdr-app window you run it in. Inside a herdr-app terminal, plain `rig terminal open <view>` does the same when that window is answering (finding the socket file is how OpenRig guesses you are in a herdr-app window, and whether the window is reachable is checked by asking it). `--window` always asks for a new desktop window. See [OpenRig views in herdr-app](herdr-app.md).
+
 ## HTTP API
 
 - `POST /api/terminal/open`: `{ view, provider?: "herdr", session?: string, expectedPlan?: string }`.
@@ -222,11 +226,26 @@ after a grace period; the agent and its conversation keep running. The
 explanation strip names joins, departures, unavailable terminals and limits.
 The first workspace on an empty provider may become focused.
 
-This increment supports the qualified **Linux Herdr 0.9.5/protocol 22**
-endpoint. Other platforms/providers and other selector kinds retain ordinary
-one-shot open. Use `--provider herdr --json` to operate without opening a
-desktop window, and `--session <name>` to select that endpoint. Opening a tile
-can change the agent's terminal size; `ignore-size` does not guarantee isolation.
+Live following supports **Linux and macOS with Herdr protocol 22**. Admission
+checks the live protocol and selected socket, then confirms each pane's IDs,
+focus state and foreground attachment. A shell startup `cd` does not change
+pane identity; filesystem aliases of the same existing path are accepted.
+If a creation response is lost and shell startup changes the directory, Core
+may be unable to attribute that pane. It reports uncertainty and does not
+repeat the creation.
+The detected Herdr version remains visible in provider status. The native
+qualification version is 0.9.5; a matching release label alone is not proof of
+compatibility, and future protocol numbers are not assumed compatible.
+
+If a new view's platform or protocol cannot support following, Core opens a
+labelled **snapshot** instead and disables automatic following. It will not
+track later owner changes. Incomplete membership prevents that snapshot from
+opening; the result explains why. A failed endpoint/identity read or an existing
+canvas stays unconfirmed, preserving it instead of creating a fallback view.
+Other providers and selector kinds retain ordinary one-shot open. Use
+`--provider herdr --json` to operate without opening a desktop window, and
+`--session <name>` to select that endpoint. Typing in a tile can change the
+agent's terminal size while you type.
 
 ```bash
 rig terminal open qitem:example --follow --provider herdr --json
@@ -284,11 +303,17 @@ operation uses the daemon's existing terminal bearer and browser boundary;
 a bridge must retain its signed-in owner, Origin and expiry controls. It does
 not grant authority based on task membership or accept a caller terminal ID.
 
-On the qualified Linux Herdr 0.9.5/protocol 22 endpoint, it creates one viewer
+On a supported Linux or macOS Herdr protocol-22 endpoint, it creates one viewer
 or reuses an owned, verified `{workspaceId,paneId,terminalId,readOnly}` mapping.
-A repeated/concurrent request does not create another viewer. A lost create
-reply is reconciled before starting anything else; a lost run reply is checked
-against the actual foreground process, never blindly repeated. Unknown or
+A repeated/concurrent request does not create another viewer. Each viewer's
+private `attach.sh` is kept for provider restarts. `--no-follow` preserves the
+workspace and its scripts; stopping Herdr is not proof that its panes are gone.
+Core removes only an exact, unedited owned script after confirming its viewer
+is absent. Old views that are never reconciled again can therefore accumulate
+scripts; there is no cleanup based only on an unreachable provider.
+
+A lost create reply is reconciled before starting anything else; a lost run
+reply is checked against the actual foreground process, never blindly repeated. Unknown or
 changed terminal identity remains unconfirmed. The operation does not relayout,
 close departed viewers, take over control, focus existing panes or launch/stop
 agents. A first workspace on an otherwise empty Herdr server may become focused.
@@ -300,12 +325,27 @@ small presentation provenance, separate from queue membership. Herdr display
 labels are not that provenance. Other providers/platforms remain unsupported
 for ensure-viewer until qualified; ordinary open retains its existing behavior.
 
-Writable tile clients use `tmux attach -f ignore-size`; read-only clients keep
-`-r` (which also implies ignore-size). This can avoid contributing to sizing
-when other clients establish the agent window's dimensions. It is **not resize
-isolation**: opening a tile can change the agent's terminal size when it is the
-only sizing client. Observers of a Herdr terminal do not themselves resize it;
-a controlling browser can. No cross-version resize guarantee is implied.
+A tmux window has one size, and a tmux client narrower than it sees only the
+left part of each line. So a tile shows its seat without attaching to it: about
+once a second it reads the seat's screen with `tmux capture-pane`, redraws only
+after tmux records new activity, and lets the tile's own terminal wrap each
+line. A tile narrower than the agent's window still shows whole lines, and
+showing a seat never changes its size. A rule drawn across the window is cut to
+the tile's width. A read-only tile only shows the seat. In a writable tile, the
+first thing you type or paste attaches the tile and goes through tmux as it
+would in any attached client: a paste arrives as one paste, and tmux's prefix
+works (prefix then `[` to scroll back, for example). The tile sizes the window
+while you use it, as any client you type in does; after 30 seconds without a
+key it detaches and shows the seat again. A tile that's only showing has no
+scrollback of its own. If the seat's tmux server restarts, a tile shows the
+seat again only when the same managed seat returns. The printed manual attach commands
+keep `-f ignore-size`. Observers of a Herdr terminal do not themselves resize
+it; a controlling browser can. No cross-version resize guarantee is implied.
+
+To type into a seat without changing its size, send the text through OpenRig
+(`rig send`, or `POST /api/transport/send`): it is written to the pane with tmux
+buffers and keys, and no client attaches. There is no supported way to attach a
+writable client that is sure never to resize the agent's window.
 
 
 ## Following a slice or mission's current work

docs/reference/tmux-servers.md

@@ -87,6 +87,37 @@ server, start the upgraded daemon from that original server to establish its
 record, or stop those seats before changing servers. Earlier versions did not
 persist enough information to find an arbitrary old socket from a different shell.
 
+## Moving an existing team onto tmux control mode
+
+`OPENRIG_TMUX_CONTROL=1` in the daemon's environment sends its routine tmux commands
+through one long-lived control-mode client instead of a new tmux client each time,
+which saves a process start per command on a busy host. It takes effect only on a
+tmux server at 3.7c or later. The version of the server that is running decides,
+not the tmux binary on `PATH`, and a server keeps the version it started with until
+it exits. So installing a newer tmux changes nothing for a server that is already up.
+On an older server the daemon uses a spawned client for every command. It says so
+once in its log (`[tmux] OPENRIG_TMUX_CONTROL=1 is set, but this daemon's tmux server
+runs tmux 3.4 …`), and `rig doctor` reports it as `tmux_control_mode`.
+
+To move a team that runs on an older server:
+
+1. Install tmux 3.7c or later where the daemon finds it first on its `PATH`. For a
+   daemon run as a service, that's the service's environment.
+2. Choose the server the team will use. `isolated` gives this OpenRig home a server of
+   its own, started by the daemon with that tmux, and leaves your own tmux sessions
+   alone: `rig config set terminal.tmux_server isolated`. In `inherit` mode the team
+   shares your default server, which runs the new version only after every session
+   on it, yours included, has ended.
+3. Stop each rig deliberately with `rig down <rig>`. Its seats' sessions end, and a
+   turn in progress stops. Queue rows, rigs and seats stay as they are.
+4. Restart the daemon with `OPENRIG_TMUX_CONTROL=1` in its environment.
+   `rig terminal status --json` should show `migrationPending: false`. If it's still
+   `true`, `blockingSessions` names the sessions still on the old server.
+5. Bring each rig back with `rig up <rig> --existing`. Each seat resumes its
+   conversation where its harness allows, under the restoration rules above.
+6. Check: the daemon's log says `[tmux] Control mode is on`, and `rig doctor` passes
+   `tmux_control_mode`.
+
 ## Attachments, discovery and cleanup
 
 `rig capture`, managed launch/stop/recovery, terminal previews and desktop viewers
@@ -164,3 +195,20 @@ whose version cannot be recognized as a fixed release, OpenRig uses spawned
 clients instead: starting a control client while another session is created can
 crash those servers (#2304). This avoids creating OpenRig's own control client;
 it cannot protect the server from other tools' control clients.
+
+On a large fleet, the daemon's two seat-activity pollers (the activity sampler and the
+structural pane cache) are a large share of its tmux traffic: each runs once a second.
+The setting `daemon.seat_poll_interval_ms` slows both down, from the 1000 ms default up to
+2000 ms. Set it with `rig config set daemon.seat_poll_interval_ms 2000` (saved in
+`config.json`; `rig config` shows it with its source) or with `OPENRIG_SEAT_POLL_INTERVAL_MS`,
+which overrides the saved value. It is startup-only: the daemon reads it once when it starts,
+so a change takes a daemon restart. `rig config set` refuses a value outside 1000..2000; an
+environment value outside it is dropped with a warning and the daemon keeps the saved value or
+the default.
+A slower cadence makes activity and idle changes show up later, and a Codex permission prompt
+can also show up as much as one interval later. Keep the default if a structural sweep of your
+tmux server can take more than about 2 s (about 3 s at the default): past that, a sweep that
+runs over skips the next tick, the Codex working-frame proof expires, and a working Codex seat
+can briefly read idle. At 2000 ms the remaining 1 s margin is also what a lagging event loop
+uses up. The two settings compose: `OPENRIG_TMUX_CONTROL=1` makes each tmux command cheaper,
+and this one makes there be fewer of them.

docs/reference/whats-new.md

@@ -89,6 +89,25 @@ after an upgrade it can still hold the previous version's note.
 
 ## Behaviour that changed
 
+- **Old history-only events can be pruned, if you choose to.** `retention.events_days` (default `0`, which keeps every
+  event) turns on a pass in the retention sweep (at daemon start and daily) that deletes `view.changed`,
+  `snapshot.created`, `watchdog.evaluation_fired` and the `queue.created`, `queue.updated`, `queue.claimed`,
+  `queue.unclaimed` and `queue.handed_off` events older than that many days; 14 is a reasonable value
+  (`rig config set retention.events_days 14`). Every other type is kept at any age, because restore, startup proof and
+  the seat readers fold from them, and the newest event is always kept. What you give up: pruned events no longer
+  replay from `/api/events` (an older cursor or a large `?replay`), no longer appear in `rig telemetry events` (which
+  reports the gap) or in older log and feed entries, and a later change of the setting can't bring them back. A new
+  value is read at the next sweep: the daily one, or the next daemon start. The first sweep after turning it on works
+  through the old history in short steps while the daemon keeps answering. `retention.enabled false` stops every
+  retention pass from the next daemon start.
+- **The database file does not shrink by itself.** Pruned rows free pages that new events reuse, which slows the file's
+  growth; other tables still grow. To give the space back to the disk: find the file with `rig config get db.path`, stop the daemon
+  (`rig daemon stop`), copy the file aside, run `sqlite3 <that file> "VACUUM;"` (it needs free disk about the size of
+  the file and can take a few minutes), then `rig daemon start`. Never vacuum the file of a running daemon.
+- **With pruning on, `rig telemetry events` reports the pruned ranges.** A page that spans pruned events inside the
+  retained window has a `retained_sequence_gap` for each missing seq range and `coverage.status` `partial` (it used to be
+  `retained_window`); events pruned from the start of the log show up as `before_retained_floor`.
+
 - **Team skills live in each seat's own storage** (`$OPENRIG_HOME/state/seats/<rig id>/<logical seat>/<runtime>/`), not
   in the project's `.claude/skills` or `.agents/skills`. Claude seats get them as a generated plugin (`openrig:<id>`),
   Codex seats as their own catalog. If a link sits in that path, the seat starts without them and the result names

package-lock.json

@@ -4260,6 +4260,13 @@
         "url": "https://github.com/sponsors/ljharb"
       }
     },
+    "node_modules/graceful-fs": {
+      "version": "4.2.11",
+      "resolved": "https://registry.npmjs.org/graceful-fs/-/graceful-fs-4.2.11.tgz",
+      "integrity": "sha512-RbJ5/jmFcNNCcDV5o9eTnBLJ/HszWV0P73bc+Ff4nS/rJj+YaS6IGyiOL0VoBYX+l1Wrl3k63h/KrH+nhJ0XvQ==",
+      "dev": true,
+      "license": "ISC"
+    },
     "node_modules/graphlib": {
       "version": "2.1.8",
       "resolved": "https://registry.npmjs.org/graphlib/-/graphlib-2.1.8.tgz",
@@ -5230,6 +5237,18 @@
         "node": "^10.13.0 || ^12.13.0 || ^14.15.0 || >=15.0.0"
       }
     },
+    "node_modules/proper-lockfile": {
+      "version": "4.1.2",
+      "resolved": "https://registry.npmjs.org/proper-lockfile/-/proper-lockfile-4.1.2.tgz",
+      "integrity": "sha512-TjNPblN4BwAWMXU8s9AEz4JmQxnD1NNL7bNOY/AKUzyamc379FWASUhc/K1pL2noVb+XmZKLL68cjzLsiOAMaA==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "graceful-fs": "^4.2.4",
+        "retry": "^0.12.0",
+        "signal-exit": "^3.0.2"
+      }
+    },
     "node_modules/proxy-addr": {
       "version": "2.0.8",
       "resolved": "https://registry.npmjs.org/proxy-addr/-/proxy-addr-2.0.8.tgz",
@@ -5531,6 +5550,16 @@
         "url": "https://github.com/privatenumber/resolve-pkg-maps?sponsor=1"
       }
     },
+    "node_modules/retry": {
+      "version": "0.12.0",
+      "resolved": "https://registry.npmjs.org/retry/-/retry-0.12.0.tgz",
+      "integrity": "sha512-9LkiTwjUh6rT555DtE9rTX+BKByPfrMzEAtnlEtdEwr3Nkffwiihqe2bWADg+OQRjt9gl6ICdmB/ZFDCGAtSow==",
+      "dev": true,
+      "license": "MIT",
+      "engines": {
+        "node": ">= 4"
+      }
+    },
     "node_modules/reusify": {
       "version": "1.1.0",
       "resolved": "https://registry.npmjs.org/reusify/-/reusify-1.1.0.tgz",
@@ -5839,6 +5868,13 @@
       "dev": true,
       "license": "ISC"
     },
+    "node_modules/signal-exit": {
+      "version": "3.0.7",
+      "resolved": "https://registry.npmjs.org/signal-exit/-/signal-exit-3.0.7.tgz",
+      "integrity": "sha512-wnD2ZE+l+SPC/uoS0vXeE9L1+0wuaMqKlfz9AMUo38JsyLSBWSFcHR1Rri62LZc12vLr1gb3jl7iwQhgwpAbGQ==",
+      "dev": true,
+      "license": "ISC"
+    },
     "node_modules/smol-toml": {
       "version": "1.8.0",
       "resolved": "https://registry.npmjs.org/smol-toml/-/smol-toml-1.8.0.tgz",
@@ -7317,6 +7353,7 @@
         "@types/better-sqlite3": "^7.6.13",
         "@types/node": "^22.0.0",
         "@types/tar": "^6.1.13",
+        "proper-lockfile": "^4.1.2",
         "tsx": "^4.19.0",
         "typescript": "^5.7.0",
         "vitest": "^3.0.0"

package.json

@@ -19,6 +19,7 @@
     "typecheck:prep": "npm run build -w packages/daemon",
     "lint": "npm run typecheck:prep && tsc --noEmit --project packages/daemon/tsconfig.json && tsc --noEmit --project packages/ui/tsconfig.json && tsc --noEmit --project packages/cli/tsconfig.json && tsc --noEmit --project packages/tui/tsconfig.json",
     "build:package": "bash scripts/build-package.sh",
+    "pack:release": "node scripts/pack-release.mjs",
     "test:tui-package": "node scripts/test-tui-package.mjs",
     "mirror-skills": "node scripts/mirror-skills.mjs",
     "mirror-skills:check": "node scripts/mirror-skills.mjs --check",

packages/acp/LICENSE

@@ -0,0 +1,190 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to the Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by the Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding any notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   Copyright 2026 Mike Schwarz
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

packages/acp/README.md

@@ -0,0 +1,36 @@
+# @openrig/acp
+
+OpenRig as an agent in apps that speak the [Agent Client Protocol](https://agentclientprotocol.com) (ACP), such as
+Zed. Your prompt goes to your OpenRig team as a queue row. The app shows who has the work
+and their notes as it moves, any question the team asks you as a choice, and the team's result as the reply.
+
+This package has one command, `openrig-acp`. It starts `rig acp serve` from the `@openrig/cli` version it depends on,
+and warns on stderr if it finds another, so ACP apps and the ACP registry can launch it with `npx`.
+
+## What you need
+
+- OpenRig on this computer, with its daemon running (`rig daemon start`) and a team.
+- The seat that receives your team's new work, set once with `rig config set acp.to <member@rig>`, or passed as
+  `--to <member@rig>`.
+- You, as a person OpenRig knows: register with
+  `rig gateway human add <id> --display-name "<your name>" --delivery-class <A|B|C|D>` (the only registered person
+  is the default; with several, choose one with `rig gateway default <id>`), or pass `--as <address>`. The team's
+  questions reach an `<id>@external` address only once it's registered.
+
+The app's sign-in step, "OpenRig on this machine", checks that the daemon answers and that it can tell who you are.
+This agent talks to the same OpenRig daemon as `rig` on this computer: the local one, unless `OPENRIG_URL` names another.
+
+## In an app, without the registry
+
+Add a custom agent whose command is `npx` with the arguments `@openrig/acp@<version>`, plus `--to <member@rig>` if
+you haven't set `acp.to`. In Zed's `settings.json`:
+
+```json
+"agent_servers": {
+  "OpenRig": { "type": "custom", "command": "npx", "args": ["@openrig/acp@<version>", "--to", "lead@my-team"] }
+}
+```
+
+If `rig` is already installed, `rig acp serve` is the same agent without this package.
+
+Guide: `rig context get reference/acp-agents.md#openrig-as-an-agent-in-acp-apps`.

packages/acp/bin/openrig-acp.js

@@ -0,0 +1,43 @@
+#!/usr/bin/env node
+// OpenRig as an agent in ACP apps. The ACP registry starts an agent with `npx <package>@<version>`, and npx needs a
+// package with one executable: @openrig/cli has two. This one starts `rig acp serve` from the @openrig/cli this package
+// depends on, as Node resolves it (never a `rig` on the PATH), with the app's stdin and stdout, arguments, environment
+// and working directory, and exits as it exits. Only ACP messages may reach stdout, so anything this file says goes to
+// stderr.
+import { spawn } from "node:child_process";
+import { readFileSync } from "node:fs";
+import { createRequire } from "node:module";
+import { dirname, join } from "node:path";
+
+let rig;
+try {
+  const manifestPath = createRequire(import.meta.url).resolve("@openrig/cli/package.json");
+  const manifest = JSON.parse(readFileSync(manifestPath, "utf8"));
+  // Module resolution can reach a different @openrig/cli further up when this package's own copy is missing. That copy
+  // most likely still works, and one that doesn't fails visibly, so say which version runs and run it.
+  const wanted = JSON.parse(readFileSync(new URL("../package.json", import.meta.url), "utf8")).dependencies["@openrig/cli"];
+  if (manifest.version !== wanted) {
+    console.error(`openrig-acp: found @openrig/cli ${manifest.version} at ${dirname(manifestPath)}, not the ${wanted} this package pins; running it anyway. Reinstall @openrig/acp to get ${wanted}.`);
+  }
+  rig = join(dirname(manifestPath), manifest.bin.rig);
+} catch (error) {
+  console.error(`openrig-acp: @openrig/cli isn't installed beside this package (${error.message}). Reinstall @openrig/acp.`);
+  process.exit(1);
+}
+
+const child = spawn(process.execPath, [rig, "acp", "serve", ...process.argv.slice(2)], { stdio: "inherit" });
+const forward = (signal) => child.kill(signal);
+for (const signal of ["SIGINT", "SIGTERM", "SIGHUP"]) process.on(signal, forward);
+child.on("error", (error) => {
+  console.error(`openrig-acp: couldn't start rig acp serve: ${error.message}`);
+  process.exit(1);
+});
+child.on("exit", (code, signal) => {
+  if (signal) {
+    // End the same way the server did, so the app sees the signal rather than an exit code.
+    process.removeListener(signal, forward);
+    process.kill(process.pid, signal);
+    return;
+  }
+  process.exit(code ?? 1);
+});

packages/acp/package.json

@@ -0,0 +1,38 @@
+{
+  "name": "@openrig/acp",
+  "version": "0.7.1",
+  "description": "OpenRig as an agent in ACP apps: starts rig acp serve from the @openrig/cli it pins.",
+  "type": "module",
+  "bin": {
+    "openrig-acp": "bin/openrig-acp.js"
+  },
+  "files": [
+    "bin",
+    "LICENSE",
+    "README.md"
+  ],
+  "engines": {
+    "node": ">=22"
+  },
+  "dependencies": {
+    "@openrig/cli": "0.7.1"
+  },
+  "license": "Apache-2.0",
+  "author": {
+    "name": "mvschwarz",
+    "url": "https://esoteric.run"
+  },
+  "homepage": "https://openrig.dev",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/mvschwarz/openrig",
+    "directory": "packages/acp"
+  },
+  "keywords": [
+    "ai",
+    "agents",
+    "acp",
+    "agent-client-protocol",
+    "openrig"
+  ]
+}

packages/acp/registry/agent.json

@@ -0,0 +1,18 @@
+{
+  "id": "openrig",
+  "name": "OpenRig",
+  "version": "0.7.1",
+  "description": "Send work to your OpenRig team of agents and follow it through to the result",
+  "repository": "https://github.com/mvschwarz/openrig",
+  "website": "https://openrig.dev",
+  "authors": [
+    "mvschwarz"
+  ],
+  "license": "Apache-2.0",
+  "license_url": "https://github.com/mvschwarz/openrig/blob/main/LICENSE",
+  "distribution": {
+    "npx": {
+      "package": "@openrig/[email protected]"
+    }
+  }
+}

packages/acp/registry/icon.svg

@@ -0,0 +1 @@
+<svg width="16" height="16" xmlns="http://www.w3.org/2000/svg" viewBox="8 -754 596 808" fill="currentColor"><g transform="scale(1,-1)"><path d="M306.0 -14.0Q426.0 -14.0 495.0 59.5Q564.0 133.0 564.0 272.0V428.0Q564.0 563.0 495.0 638.5Q426.0 714.0 306.0 714.0Q186.0 714.0 117.0 638.5Q83.32444763183594 601.652099609375 66.08428192138672 550.631591796875L195.21009826660156 523.18505859375Q207.8548583984375 554.14111328125 233.0 573.0Q261.0 594.0 306.0 594.0Q351.0 594.0 378.5 573.0Q406.0 552.0 419.0 516.0Q432.0 480.0 432.0 436.0V264.0Q432.0 194.0 405.0 150.0Q378.0 106.0 306.0 106.0Q234.0 106.0 207.0 150.0Q180.0 194.0 180.0 264.0V436.0Q180.0 461.6716613769531 184.25511169433594 484.6199951171875L55.7750358581543 511.9292907714844Q48.0 473.3169250488281 48.0 428.0V272.0Q48.0 133.0 117.0 59.5Q186.0 -14.0 306.0 -14.0ZM138.0 176.0C138.0 162.7451629638672 127.25483703613281 152.0 114.0 152.0C100.74516296386719 152.0 90.0 162.7451629638672 90.0 176.0C90.0 189.2548370361328 100.74516296386719 200.0 114.0 200.0C127.25483703613281 200.0 138.0 189.2548370361328 138.0 176.0Z"/></g></svg>

packages/cli/src/acp/agent-server.ts

@@ -0,0 +1,540 @@
+import { randomBytes } from "node:crypto";
+import type { JsonRpcPeer } from "./json-rpc-stdio.js";
+
+// OpenRig as an ACP agent (rig acp serve): an ACP app's prompt becomes a queue row to the team's receiver, the row's
+// transitions and handoffs come back as progress, and its closing resolution is the reply. A question the team asks the
+// caller while it works comes to the app as a permission request, and the choice goes back through the person's answer
+// path. The queue is the only work store: a session is the rows tagged with its id, so it can be listed and reopened.
+// A turn's end_turn says only that the turn ended: the reply says whether the team closed the item, and a cancel stops
+// this process waiting, never the team.
+
+export const ACP_PROTOCOL_VERSION = 1;
+export const AUTH_METHOD_ID = "openrig-local";
+export const SESSION_TAG = "acp-session:";
+export const CWD_TAG = "acp-cwd:";
+/** Where the caller's prompt ends in a row's body; the rest is the note rig acp serve adds for the team. */
+const FOOTER = "\n\n---\nSent by ";
+
+/** One queue row as this adapter reads it. */
+export interface TeamRow {
+  qitemId: string;
+  state: string;
+  destinationSession: string;
+  handedOffTo: string | null;
+  closureReason: string | null;
+  closureTarget: string | null;
+  resolution: string | null;
+  /** Who closed the row, from its closing transition (a note written after the close doesn't change it); null while open. */
+  closer: string | null;
+}
+
+export interface TeamTransition {
+  transitionId: number;
+  ts: string;
+  state: string;
+  transitionNote: string | null;
+  actorSession: string;
+}
+
+/** A row to the caller that asks for a decision, with the choices it offers (none: words only). */
+export interface TeamAsk {
+  qitemId: string;
+  sourceSession: string;
+  state: string;
+  tsCreated: string;
+  summary: string | null;
+  body: string;
+  tags: string[];
+  questions: Array<{ id: string; question: string; options: Array<{ id: string; label: string; recommended?: boolean }> }>;
+  /** What's recorded so far, by question id: the chosen option's id, or the words of a typed reply. */
+  answers: Record<string, string>;
+}
+
+/** A row the caller sent through rig acp serve. */
+export interface SentRow {
+  qitemId: string;
+  tsCreated: string;
+  tsUpdated: string;
+  summary: string | null;
+  body: string;
+  tags: string[];
+}
+
+/** The daemon answered a create and refused it, so nothing was written. Any other failure of a create leaves it unknown. */
+export class TeamRefused extends Error {}
+
+/** The queue verbs the adapter uses, acting as the caller. */
+export interface TeamQueue {
+  /** Create the row under the id the adapter chose, so a create whose answer is lost can be read back by that id. Throws
+   *  TeamRefused only for a refusal the daemon answered with. */
+  create(row: { qitemId: string; destination: string; body: string; summary: string; tags: string[] }): Promise<void>;
+  show(qitemId: string): Promise<TeamRow>;
+  /** The row, or null when this machine has no such row; throws when it can't tell. */
+  find(qitemId: string): Promise<TeamRow | null>;
+  transitions(qitemId: string): Promise<TeamTransition[]>;
+  /** The row a handoff of `qitemId` created on this machine, if any. */
+  successor(qitemId: string): Promise<string | undefined>;
+  /** Open asks addressed to `person`. */
+  openAsks(person: string): Promise<TeamAsk[]>;
+  /** Every ask to `person` tagged for one chat, open or settled, to replay with it. */
+  sessionAsks(person: string, sessionId: string): Promise<TeamAsk[]>;
+  /** Record the caller's choice on an ask, through the person's answer path; a refusal comes back with its reason. */
+  answer(qitemId: string, questionId: string, optionId: string): Promise<{ ok: true } | { ok: false; reason: string }>;
+  /** Rows the caller sent through rig acp serve. */
+  sent(caller: string): Promise<SentRow[]>;
+}
+
+/** What a session works through: the queue as the caller, and the seat that receives the caller's tasks (when set). */
+export interface TeamConnection {
+  queue: TeamQueue;
+  caller: string;
+  receiver: string | undefined;
+  /** Something about this connection the person should know, said once in each session before its first prompt's work. */
+  notice?: string;
+}
+
+export interface AcpAgentOptions {
+  /** Reaches the daemon (`rig`'s: the local one, or OPENRIG_URL's) and resolves the caller; returns guidance when it can't.
+   *  Called lazily, never at initialize. */
+  connect: () => Promise<TeamConnection | { error: string }>;
+  version: string;
+  pollMs: number;
+  log?: (line: string) => void;
+  sleep?: (ms: number) => Promise<void>;
+  /** Settles when the app's connection ends: open turns stop following their work (nothing could carry a reply). */
+  closed?: Promise<void>;
+}
+
+const ACTIVE_STATES = new Set(["pending", "in-progress", "blocked"]);
+/** How long a cancel waits for the app to settle questions it was showing (ACP has the client answer them cancelled). */
+const QUESTION_SETTLE_MS = 2000;
+const AUTH_REQUIRED = -32000;
+const INVALID_PARAMS = -32602;
+
+class AcpError extends Error {
+  constructor(readonly code: number, message: string, readonly data?: unknown) { super(message); }
+}
+
+interface Session {
+  id: string;
+  cwd: string;
+  /** The first row each prompt created, oldest first, with the last row its work reached. */
+  sent: Array<{ root: string; tail: string }>;
+  turn?: { cancel: () => void };
+  noticed?: boolean;
+}
+
+type ContentBlock = { type: string; text?: string; uri?: string; name?: string; resource?: { uri?: string; text?: string } };
+
+/** The questions one turn has put to the app. Once the turn has ended, nothing from them reaches the app or the queue. */
+interface TurnQuestions {
+  finished: boolean;
+  /** The person cancelled the turn: no further question opens, though one the app is showing still settles. */
+  cancelled: boolean;
+  /** Questions the app is still showing, by tool call id, each with how the chat names it. */
+  outstanding: Map<string, string>;
+  tasks: Set<Promise<void>>;
+}
+type Update = Record<string, unknown>;
+
+/** The prompt as plain text for the row's body: text as written, links and embedded text resources named. */
+export function promptText(prompt: ContentBlock[]): string {
+  const parts: string[] = [];
+  for (const block of prompt) {
+    if (block.type === "text" && block.text) parts.push(block.text);
+    else if (block.type === "resource_link" && block.uri) parts.push(`[${block.name ?? "link"}](${block.uri})`);
+    else if (block.type === "resource" && block.resource?.uri) {
+      parts.push(block.resource.text !== undefined ? `${block.resource.uri}:\n\`\`\`\n${block.resource.text}\n\`\`\`` : block.resource.uri);
+    }
+  }
+  return parts.join("\n\n").trim();
+}
+
+/** A queue row id in the form the CLI mints (`qitem-<UTC yyyymmddhhmmss>-<16 hex>`). */
+export const newQitemId = () => `qitem-${new Date().toISOString().replace(/\D/g, "").slice(0, 14)}-${randomBytes(8).toString("hex")}`;
+
+function oneLine(text: string, max: number): string {
+  const line = text.split("\n").find((l) => l.trim())?.trim() ?? "";
+  return line.length > max ? `${line.slice(0, max - 1)}…` : line;
+}
+
+/** The person's own words in a prompt, for the row's summary and the session's title. An app that wraps them in its
+ *  own instructions marks them (T3 Code sends several blocks of harness text, then `<user_request>…</user_request>`);
+ *  the row's body keeps everything the app sent. Only a prompt with one opening and one closing mark is read this
+ *  way: with more, an example could be taken for the request, so the prompt stands as sent. */
+function personsWords(text: string): string {
+  if (text.split("<user_request>").length !== 2 || text.split("</user_request>").length !== 2) return text;
+  const marked = /<user_request>([\s\S]*?)<\/user_request>/.exec(text)?.[1];
+  return marked?.trim() ? marked : text;
+}
+
+const toolStatus = (state: string) =>
+  state === "pending" ? "pending" : ACTIVE_STATES.has(state) ? "in_progress" : state === "done" || state === "handed-off" ? "completed" : "failed";
+
+const transitionLine = (t: TeamTransition) =>
+  `${t.ts.slice(11, 19)}Z ${t.actorSession} · ${t.state}${t.transitionNote ? `: ${t.transitionNote}` : ""}`;
+
+const textContent = (text: string) => [{ type: "content", content: { type: "text", text } }];
+
+/** Where one prompt's work stands after a read of its rows. */
+type ChainState = { kind: "open"; tail: TeamRow } | { kind: "ended"; tail: TeamRow; reply: string };
+
+/** Follows one prompt's work through its rows: each row is a step, and its new notes are emitted as they appear. */
+class ChainFollower {
+  private readonly chain: string[];
+  private readonly seen = new Map<string, { lines: string[]; lastId: number; status: string; title: string }>();
+  /** The seats the work has reached; a question from one of them during the turn belongs to it. */
+  readonly seats = new Set<string>();
+
+  constructor(private readonly queue: TeamQueue, private readonly emit: (update: Update) => void, readonly root: string) {
+    this.chain = [root];
+  }
+
+  async read(): Promise<ChainState> {
+    for (;;) {
+      let tail!: TeamRow;
+      for (const qitemId of this.chain) {
+        const row = await this.queue.show(qitemId);
+        const transitions = await this.queue.transitions(qitemId);
+        tail = row;
+        this.seats.add(row.destinationSession);
+        const status = toolStatus(row.state);
+        let entry = this.seen.get(qitemId);
+        if (!entry) {
+          const title = qitemId === this.root ? `Sent to ${row.destinationSession} (${qitemId})` : `Handed to ${row.destinationSession} (${qitemId})`;
+          entry = { lines: [], lastId: 0, status, title };
+          this.seen.set(qitemId, entry);
+          this.emit({ sessionUpdate: "tool_call", toolCallId: qitemId, kind: "other", status, title, rawInput: { qitemId, destination: row.destinationSession } });
+        }
+        const fresh = transitions.filter((t) => t.transitionId > entry!.lastId);
+        if (fresh.length > 0 || status !== entry.status) {
+          for (const t of fresh) { entry.lines.push(transitionLine(t)); entry.lastId = Math.max(entry.lastId, t.transitionId); }
+          entry.status = status;
+          // An app that shows a step collapsed still shows its title, so the latest note leads it: a narrow panel cuts the end.
+          const noted = fresh.filter((t) => t.transitionNote).at(-1);
+          this.emit({
+            sessionUpdate: "tool_call_update", toolCallId: qitemId, status,
+            ...(noted ? { title: `${noted.actorSession}: ${oneLine(noted.transitionNote!, 100)} · ${entry.title}` } : {}),
+            ...(entry.lines.length > 0 ? { content: textContent(entry.lines.join("\n")) } : {}),
+          });
+        }
+      }
+      // Both handoff forms pass the work on: `handoff` leaves the row handed-off, `handoff-and-complete` closes it done
+      // with closure reason handed_off_to. Either way the answer is the successor's, not this row's.
+      if (tail.state === "handed-off" || (tail.state === "done" && tail.closureReason === "handed_off_to")) {
+        const next = await this.queue.successor(tail.qitemId);
+        if (next) { this.chain.push(next); continue; }
+        return { kind: "ended", tail, reply: `${tail.qitemId} was handed to ${tail.handedOffTo ?? tail.closureTarget ?? "another seat"}, whose row isn't on this machine, so this app can't follow it further. Ask your team for its result.` };
+      }
+      if (ACTIVE_STATES.has(tail.state)) return { kind: "open", tail };
+      // A done row's closure reason says how it ended (canceled is a done close too); the ordinary one goes unsaid.
+      const reason = tail.closureReason && tail.closureReason !== "no-follow-on" ? ` (${tail.closureReason})` : "";
+      const how = tail.state === "done" ? `done${reason}` : `as ${tail.state}${reason}`;
+      return { kind: "ended", tail, reply: `${tail.closer ?? tail.destinationSession} closed ${tail.qitemId} ${how}.\n\n${tail.resolution?.trim() || "(No resolution was written.)"}` };
+    }
+  }
+}
+
+export function serveAcpAgent(peer: JsonRpcPeer, options: AcpAgentOptions): void {
+  const sessions = new Map<string, Session>();
+  const sleep = options.sleep ?? ((ms: number) => new Promise<void>((resolve) => setTimeout(resolve, ms)));
+  const log = options.log ?? (() => {});
+  let clientName = "an ACP app";
+  /** The app asks forms (ACP `elicitation/create`, mode form), so a question's choices go to it as a form, each by its
+   *  label. A permission request shows them only in apps that list each option: T3 Code shows a fixed Approve, which
+   *  sends the first. */
+  let asksForms = false;
+  let connection: TeamConnection | undefined;
+  void options.closed?.then(() => { for (const session of sessions.values()) session.turn?.cancel(); });
+
+  const connected = async (): Promise<TeamConnection> => {
+    if (connection) return connection;
+    const result = await options.connect();
+    if ("error" in result) throw new AcpError(AUTH_REQUIRED, result.error);
+    connection = result;
+    return result;
+  };
+
+  const update = (sessionId: string, body: Update) => peer.notify("session/update", { sessionId, update: body });
+  const say = (sessionId: string, text: string) => update(sessionId, { sessionUpdate: "agent_message_chunk", content: { type: "text", text } });
+  /** A question the app showed ends unanswered: its card says so, and so does the chat, which every app shows, since an
+   *  app need not show a card's later text. */
+  const unanswered = (sessionId: string, toolCallId: string, named: string, why: string) => {
+    update(sessionId, { sessionUpdate: "tool_call_update", toolCallId, content: textContent(`${why} The question is still open in the OpenRig TUI and Slack.`) });
+    say(sessionId, `\n\n${why} ${named} is still open in the OpenRig TUI and Slack.\n\n`);
+  };
+  const newSessionId = () => `acps-${new Date().toISOString().replace(/\D/g, "").slice(0, 14)}-${randomBytes(4).toString("hex")}`;
+
+  /** Each of the caller's prompts in a session, oldest first, by the rows tagged with its id. */
+  const sessionRoots = async (team: TeamConnection, sessionId: string) =>
+    (await team.queue.sent(team.caller)).filter((row) => row.tags.includes(`${SESSION_TAG}${sessionId}`))
+      .sort((a, b) => a.tsCreated.localeCompare(b.tsCreated));
+
+  /** Put one ask's open questions to the person, as a form where the app asks forms and otherwise as permission
+   *  requests, one at a time, and record each choice. A reply that arrives after the turn ended answers nothing here:
+   *  the question stays open in the TUI and Slack. */
+  const askPerson = async (team: TeamConnection, sessionId: string, ask: TeamAsk, turn: TurnQuestions) => {
+    const emit = (body: Update) => { if (!turn.finished) update(sessionId, body); };
+    if (ask.questions.length === 0) {
+      emit({ sessionUpdate: "agent_message_chunk", content: { type: "text", text: `\n\n${ask.sourceSession} asks you: ${ask.summary ?? oneLine(ask.body, 200)} (${ask.qitemId}). It wants an answer in words, which this app can't send: answer it in the OpenRig TUI or Slack.\n\n` } });
+      return;
+    }
+    for (const q of ask.questions.filter((question) => !(question.id in ask.answers))) {
+      if (turn.finished || turn.cancelled) return;
+      const toolCallId = `${ask.qitemId}#${q.id}`;
+      emit({ sessionUpdate: "tool_call", toolCallId, kind: "other", status: "pending", title: `${ask.sourceSession} asks: ${q.question}`, content: textContent(ask.body) });
+      const label = (o: (typeof q.options)[number]) => (o.recommended ? `${o.label} (recommended)` : o.label);
+      const named = `${ask.sourceSession}'s question "${q.question}" (${ask.qitemId})`;
+      let chosenId: string | undefined;
+      turn.outstanding.set(toolCallId, named);
+      try {
+        if (asksForms) {
+          // Each choice is offered as its label, the one form every app shows as choices: T3 Code 0.1.7 builds them only
+          // from `enum`, a value serving as both label and answer, and reads `oneOf` as free text. Labels that repeat
+          // carry their option's id, so a value names one choice; the answer is mapped back by that exact value.
+          const shown = q.options.map(label);
+          const values = new Set(shown).size === shown.length ? shown : q.options.map((o, i) => `${shown[i]} (${o.id})`);
+          const result = (await peer.request("elicitation/create", {
+            sessionId, toolCallId, mode: "form", message: `${ask.sourceSession} asks: ${q.question}`,
+            requestedSchema: { type: "object", required: [q.id], properties: { [q.id]: { type: "string", title: q.question, enum: values } } },
+          })) as { action?: string; content?: Record<string, unknown> | null } | undefined;
+          const value = result?.action === "accept" ? result.content?.[q.id] : undefined;
+          chosenId = typeof value === "string" ? q.options[values.indexOf(value)]?.id : undefined;
+        } else {
+          const result = (await peer.request("session/request_permission", {
+            sessionId,
+            toolCall: { toolCallId, title: `${ask.sourceSession} asks: ${q.question}`, kind: "other", status: "pending" },
+            options: q.options.map((o) => ({ optionId: o.id, name: label(o), kind: "allow_once" })),
+          })) as { outcome?: { outcome?: string; optionId?: string } } | undefined;
+          chosenId = result?.outcome?.outcome === "selected" ? result.outcome.optionId : undefined;
+        }
+      } catch (error) {
+        if (!turn.finished) unanswered(sessionId, toolCallId, named, `This app couldn't show the question (${(error as Error).message}).`);
+        return;
+      } finally {
+        turn.outstanding.delete(toolCallId);
+      }
+      if (turn.finished) return;
+      // Only one of the question's own options is recorded: the app's reply is checked again here.
+      const chosen = q.options.find((o) => o.id === chosenId);
+      if (!chosen) {
+        unanswered(sessionId, toolCallId, named, "Not answered here.");
+        return;
+      }
+      const recorded = await team.queue.answer(ask.qitemId, q.id, chosen.id).catch((error: Error) => ({ ok: false as const, reason: error.message }));
+      emit(recorded.ok
+        ? { sessionUpdate: "tool_call_update", toolCallId, status: "completed", content: textContent(`You answered: ${chosen.label}. ${ask.sourceSession} gets it as your answer.`) }
+        : { sessionUpdate: "tool_call_update", toolCallId, status: "failed", content: textContent(`Your choice (${chosen.label}) wasn't recorded: ${recorded.reason}`) });
+      log(`session ${sessionId}: ${recorded.ok ? "answered" : "couldn't answer"} ${ask.qitemId} ${q.id}`);
+    }
+  };
+
+  peer.onRequest("initialize", async (params) => {
+    const p = (params ?? {}) as { clientInfo?: { name?: string; title?: string }; clientCapabilities?: { elicitation?: { form?: unknown } | null } };
+    clientName = p.clientInfo?.title ?? p.clientInfo?.name ?? clientName;
+    // ACP counts a mode as supported only when its field is present and not null.
+    asksForms = p.clientCapabilities?.elicitation?.form != null;
+    return {
+      protocolVersion: ACP_PROTOCOL_VERSION,
+      agentCapabilities: {
+        loadSession: true,
+        promptCapabilities: { image: false, audio: false, embeddedContext: true },
+        mcpCapabilities: { http: false, sse: false },
+        sessionCapabilities: { list: {} },
+      },
+      authMethods: [{
+        id: AUTH_METHOD_ID,
+        name: "OpenRig on this machine",
+        description: "Sends your work to your OpenRig team, as you, through the same OpenRig daemon as `rig` on this computer: the local one (start it with `rig daemon start`), unless OPENRIG_URL names another.",
+      }],
+      agentInfo: { name: "openrig", title: "OpenRig", version: options.version },
+    };
+  });
+
+  peer.onRequest("authenticate", async (params) => {
+    const methodId = (params as { methodId?: unknown } | undefined)?.methodId;
+    if (methodId !== AUTH_METHOD_ID) throw new AcpError(INVALID_PARAMS, `Unknown auth method ${String(methodId)}; this agent offers ${AUTH_METHOD_ID}`);
+    const team = await connected();
+    log(`authenticated as ${team.caller}${team.receiver ? `, sending to ${team.receiver}` : ", with no receiver set"}`);
+    return {};
+  });
+
+  peer.onRequest("session/new", async (params) => {
+    await connected();
+    const id = newSessionId();
+    sessions.set(id, { id, cwd: String((params as { cwd?: unknown } | undefined)?.cwd ?? ""), sent: [] });
+    return { sessionId: id };
+  });
+
+  peer.onRequest("session/list", async (params) => {
+    const team = await connected();
+    const cwd = (params as { cwd?: unknown } | undefined)?.cwd;
+    const byId = new Map<string, { sessionId: string; cwd: string; title: string; updatedAt: string }>();
+    for (const row of (await team.queue.sent(team.caller)).sort((a, b) => a.tsCreated.localeCompare(b.tsCreated))) {
+      const sessionId = row.tags.find((t) => t.startsWith(SESSION_TAG))?.slice(SESSION_TAG.length);
+      if (!sessionId) continue;
+      const known = byId.get(sessionId);
+      if (known) { if (row.tsUpdated > known.updatedAt) known.updatedAt = row.tsUpdated; continue; }
+      byId.set(sessionId, {
+        sessionId, cwd: row.tags.find((t) => t.startsWith(CWD_TAG))?.slice(CWD_TAG.length) ?? "",
+        title: oneLine(personsWords(row.body.split(FOOTER)[0] ?? ""), 100), updatedAt: row.tsUpdated,
+      });
+    }
+    const listed = [...byId.values()].filter((s) => typeof cwd !== "string" || s.cwd === cwd).sort((a, b) => b.updatedAt.localeCompare(a.updatedAt));
+    return { sessions: listed };
+  });
+
+  peer.onRequest("session/load", async (params) => {
+    const p = (params ?? {}) as { sessionId?: string; cwd?: string };
+    const team = await connected();
+    const sessionId = String(p.sessionId ?? "");
+    const roots = await sessionRoots(team, sessionId);
+    if (roots.length === 0) throw new AcpError(INVALID_PARAMS, `No session ${sessionId} from ${team.caller} on this machine`);
+    const session: Session = { id: sessionId, cwd: String(p.cwd ?? ""), sent: [] };
+    const asks = (await team.queue.sessionAsks(team.caller, sessionId)).sort((a, b) => a.tsCreated.localeCompare(b.tsCreated));
+    // Replay: each prompt as the caller wrote it, then its work as it stands now, the questions the team asked during
+    // it with what's recorded, and the reply. Nothing is asked again here; an open question comes up in the next turn.
+    for (const [i, row] of roots.entries()) {
+      update(sessionId, { sessionUpdate: "user_message_chunk", content: { type: "text", text: row.body.split(FOOTER)[0] ?? "" } });
+      const follower = new ChainFollower(team.queue, (u) => update(sessionId, u), row.qitemId);
+      const state = await follower.read();
+      session.sent.push({ root: row.qitemId, tail: state.tail.qitemId });
+      const next = roots[i + 1]?.tsCreated;
+      for (const ask of asks.filter((a) => (i === 0 || a.tsCreated >= row.tsCreated) && (next === undefined || a.tsCreated < next))) {
+        for (const q of ask.questions) {
+          const answer = ask.answers[q.id];
+          const outcome = answer !== undefined ? `You answered: ${q.options.find((o) => o.id === answer)?.label ?? answer}.`
+            : ACTIVE_STATES.has(ask.state) ? "Not answered yet. It's still open in the OpenRig TUI and Slack." : `Not answered: ${ask.qitemId} was closed (${ask.state}).`;
+          update(sessionId, {
+            sessionUpdate: "tool_call", toolCallId: `${ask.qitemId}#${q.id}`, kind: "other", status: answer === undefined && ACTIVE_STATES.has(ask.state) ? "pending" : "completed",
+            title: `${ask.sourceSession} asks: ${q.question}`, content: [...textContent(ask.body), ...textContent(outcome)],
+          });
+        }
+      }
+      say(sessionId, state.kind === "ended" ? state.reply : `${state.tail.qitemId} is still ${state.tail.state} with ${state.tail.destinationSession}.`);
+    }
+    sessions.set(sessionId, session);
+    return {};
+  });
+
+  peer.onNotification("session/cancel", (params) => {
+    const sessionId = (params as { sessionId?: string } | undefined)?.sessionId;
+    if (sessionId) sessions.get(sessionId)?.turn?.cancel();
+  });
+
+  peer.onRequest("session/prompt", async (params) => {
+    const p = (params ?? {}) as { sessionId?: string; prompt?: ContentBlock[] };
+    const session = p.sessionId ? sessions.get(p.sessionId) : undefined;
+    if (!session) throw new AcpError(INVALID_PARAMS, `No session ${String(p.sessionId)} in this agent; start one with session/new or session/load`);
+    if (session.turn) throw new AcpError(INVALID_PARAMS, "This session already has a turn running; wait for it or cancel it first");
+    const team = await connected();
+    const text = promptText(Array.isArray(p.prompt) ? p.prompt : []);
+    if (!text) {
+      say(session.id, "Nothing to send: the prompt had no text.");
+      return { stopReason: "end_turn" };
+    }
+    if (!team.receiver) {
+      say(session.id, "This OpenRig agent has no team receiver, so nothing was sent. Choose the seat that takes your team's new work with `rig config set acp.to <member@rig>`, or add `--to <member@rig>` to the agent's command in your app's settings.");
+      return { stopReason: "end_turn" };
+    }
+    if (team.notice && !session.noticed) {
+      session.noticed = true;
+      say(session.id, `${team.notice}\n\n`);
+    }
+
+    let wake: () => void = () => {};
+    const questions: TurnQuestions = { finished: false, cancelled: false, outstanding: new Map(), tasks: new Set() };
+    session.turn = { cancel: () => { questions.cancelled = true; wake(); } };
+    let root: string | undefined;
+    try {
+      // Each earlier item as it stands; one whose create was never confirmed may not exist, or can't be read: say which.
+      const earlier = await Promise.all(session.sent.map(async (s) => {
+        const row = await team.queue.find(s.tail).catch(() => undefined);
+        return { ...s, row, state: row === undefined ? "state unknown" : row === null ? "not found here" : row.state };
+      }));
+      const open = earlier.filter((s) => s.row && ACTIVE_STATES.has(s.row.state));
+      if (open.length > 0) say(session.id, `${open.map((s) => `${s.row!.qitemId} is still ${s.row!.state} with ${s.row!.destinationSession}`).join("; ")}. Sending this as a follow-up.\n\n`);
+      const lineage = earlier.length === 0 ? "" : `\nFollows up on ${earlier.map((s) => `${s.root} (${s.tail === s.root ? "" : `reached ${s.tail}, `}${s.state})`).join(", ")}.`;
+      const body = `${text}${FOOTER}${team.caller} from ${clientName} through rig acp serve (session ${session.id}).${lineage}\n` +
+        `The app shows this item's notes and handoffs as progress, and its closing resolution as the answer: close it with a resolution written for ${team.caller}. ` +
+        `A question with choices you ask ${team.caller} while this is open, tagged for this chat, appears in the app as a choice: ` +
+        `rig queue create --destination ${team.caller} --human-questions-file … --tags ${SESSION_TAG}${session.id}.`;
+      const turnStarted = new Date().toISOString();
+      const qitemId = newQitemId();
+      try {
+        await team.queue.create({
+          qitemId, destination: team.receiver, body, summary: `ACP: ${oneLine(personsWords(text), 100)}`,
+          tags: ["acp-serve", `${SESSION_TAG}${session.id}`, ...(session.cwd ? [`${CWD_TAG}${session.cwd}`] : [])],
+        });
+      } catch (error) {
+        const reason = (error as Error).message;
+        // A refusal the daemon answered with wrote nothing, so it can be said plainly.
+        if (error instanceof TeamRefused) throw new AcpError(-32603, `The team's queue refused the task: ${reason}`);
+        // No answer: the request may still land after any readback, so only a row that reads back settles it.
+        const landed = await team.queue.find(qitemId).catch(() => undefined);
+        if (!landed) {
+          session.sent.push({ root: qitemId, tail: qitemId });
+          say(session.id, `Sending ${qitemId} to ${team.receiver} got no answer (${reason}), and ${landed === null ? "it wasn't found on the immediate readback" : "it couldn't be read back"}: delivery is unconfirmed, so it may still reach the team. A new prompt in this session names it, and \`rig queue show ${qitemId}\` shows it if it landed.`);
+          throw new AcpError(-32603, `Couldn't confirm ${qitemId}: ${reason}`);
+        }
+        log(`session ${session.id}: ${qitemId}'s create got no answer, but it reads back`);
+      }
+      root = qitemId;
+      const sent = { root, tail: root };
+      session.sent.push(sent);
+      log(`session ${session.id}: sent ${root} to ${team.receiver}`);
+
+      // ponytail: polls the daemon for the chain's rows, their transitions and the caller's open asks; the
+      // daemon's event stream is the upgrade path if a poll per open turn ever costs too much.
+      const follower = new ChainFollower(team.queue, (u) => update(session.id, u), root);
+      const seen = new Set<string>();
+      /** Close this turn's questions: any the app still shows are named as open elsewhere, then nothing more comes from them. */
+      const finishQuestions = (why: string) => {
+        for (const [toolCallId, named] of questions.outstanding) unanswered(session.id, toolCallId, named, why);
+        questions.finished = true;
+      };
+      for (;;) {
+        const state = await follower.read();
+        sent.tail = state.tail.qitemId;
+        if (state.kind === "ended") {
+          finishQuestions("The work closed before this was answered here.");
+          say(session.id, state.reply);
+          return { stopReason: "end_turn" };
+        }
+        const asks = await team.queue.openAsks(team.caller);
+        // After a cancel, even one that landed during that read, no new question goes to the person.
+        for (const ask of questions.cancelled ? [] : asks) {
+          if (seen.has(ask.qitemId)) continue;
+          // An ask tagged for a chat belongs to that chat alone. An untagged one from a seat this work reached might be
+          // another chat's, so it isn't asked here; it's named, with where to answer it.
+          const chats = ask.tags.filter((tag) => tag.startsWith(SESSION_TAG));
+          if (chats.includes(`${SESSION_TAG}${session.id}`)) {
+            seen.add(ask.qitemId);
+            const task: Promise<void> = askPerson(team, session.id, ask, questions)
+              .catch((error: Error) => log(`session ${session.id}: asking about ${ask.qitemId} failed: ${error.message}`))
+              .finally(() => questions.tasks.delete(task));
+            questions.tasks.add(task);
+          } else if (chats.length === 0 && follower.seats.has(ask.sourceSession) && ask.tsCreated >= turnStarted) {
+            seen.add(ask.qitemId);
+            say(session.id, `\n\n${ask.sourceSession}, which this work reached, has a question for you that isn't tagged for this chat (${ask.qitemId}: ${ask.summary ?? oneLine(ask.body, 120)}). Answer it in the OpenRig TUI or Slack.\n\n`);
+          }
+        }
+        if (questions.cancelled) {
+          // ACP has the app answer the questions it was showing as cancelled; let those answers land before the turn ends.
+          if (questions.tasks.size > 0) await Promise.race([Promise.allSettled([...questions.tasks]), sleep(QUESTION_SETTLE_MS)]);
+          finishQuestions("Not answered here: the turn was cancelled.");
+          say(session.id, `Stopped waiting. ${state.tail.qitemId} is still ${state.tail.state} with ${state.tail.destinationSession}: cancelling here doesn't stop the team's work. A new prompt in this session goes to the team as a follow-up and shows where this stands.`);
+          return { stopReason: "cancelled" };
+        }
+        await Promise.race([sleep(options.pollMs), new Promise<void>((resolve) => { wake = resolve; })]);
+      }
+    } catch (error) {
+      if (error instanceof AcpError || !root) throw error;
+      say(session.id, `Lost track of ${root}: ${(error as Error).message}. The item stays with the team; its state is unknown from here.`);
+      throw new AcpError(-32603, `Couldn't follow ${root}: ${(error as Error).message}`);
+    } finally {
+      questions.finished = true;
+      session.turn = undefined;
+    }
+  });
+}

packages/cli/src/acp/connector.ts

@@ -1,7 +1,7 @@
 import { spawn } from "node:child_process";
 import { appendFileSync, mkdirSync } from "node:fs";
 import { dirname } from "node:path";
-import { JsonRpcRequestError, JsonRpcTransportClosed, StdioJsonRpcPeer } from "./json-rpc-stdio.js";
+import { JsonRpcRequestError, JsonRpcTransportClosed, StdioJsonRpcPeer, type JsonRpcErrorObject } from "./json-rpc-stdio.js";
 import { findRunProcesses, RUN_MARKER } from "./run-processes.js";
 
 // `rig acp run`: one queue row's task to one ACP agent turn, with the outcome back on that row. OpenRig carries the
@@ -48,10 +48,22 @@ export interface AcpRunOptions {
   detachRequested?: Promise<void>;
   /** How long to wait for the agent to exit after its stdin closes (default 5 s). */
   exitWaitMs?: number;
+  /** A harness's own evidence that a `session/load` error means that session doesn't exist. Without it, or when it
+   *  doesn't match, the load error reads `failed` with the agent's code and message: no code means missing by itself. */
+  missingSession?: (error: JsonRpcErrorObject, sessionId: string) => boolean;
+  /** Called once with what the stop found: the agent's exit and whether any process carrying the run's marker remains. */
+  onSettled?: (report: string) => void;
 }
 
-export type AcpOutcomeKind = "completed" | "stopped" | "cancelled" | "failed" | "unknown";
-export interface AcpOutcome { kind: AcpOutcomeKind; summary: string; text: string; sessionId?: string }
+export type AcpOutcomeKind = "completed" | "stopped" | "cancelled" | "failed" | "unknown" | "missing";
+/** What the agent itself reported; OpenRig measures none of it. */
+export interface AcpUsage {
+  /** The prompt response's `usage` (an unstable ACP field), as the agent counted it. */
+  tokens?: { total?: number; input?: number; output?: number; thought?: number; cachedRead?: number; cachedWrite?: number };
+  /** The last `usage_update` cost: the agent's own cumulative figure for the whole session, in its unit. */
+  cost?: { amount: number; currency: string };
+}
+export interface AcpOutcome { kind: AcpOutcomeKind; summary: string; text: string; sessionId?: string; usage?: AcpUsage }
 
 /** Agent text that reports a failed turn while still ending it `end_turn` (OpenClaw #178057; its #44294 is open). */
 export const INTERRUPTION_MARKERS = ["[OpenClaw interruption]"];
@@ -87,6 +99,7 @@ export async function runAcpAssignment(queue: AcpQueue, options: AcpRunOptions):
   let sessionId: string | undefined;
   let text = "";
   let usage: { used?: number; size?: number; cost?: unknown } | undefined;
+  const reported: AcpUsage = {};
   let toolCalls = 0;
   let lastTool = "";
   let lastProgressNote = Date.now();
@@ -104,6 +117,8 @@ export async function runAcpAssignment(queue: AcpQueue, options: AcpRunOptions):
       progress(`\n[tool] ${lastTool}\n`);
     } else if (update.sessionUpdate === "usage_update") {
       usage = { used: update.used as number, size: update.size as number, cost: update.cost };
+      const cost = update.cost as { amount?: unknown; currency?: unknown } | null | undefined;
+      if (typeof cost?.amount === "number" && typeof cost.currency === "string") reported.cost = { amount: cost.amount, currency: cost.currency };
     }
     if (Date.now() - lastProgressNote >= 60_000) {
       lastProgressNote = Date.now();
@@ -153,10 +168,11 @@ export async function runAcpAssignment(queue: AcpQueue, options: AcpRunOptions):
 
   const record = async (outcome: AcpOutcome): Promise<AcpOutcome> => {
     const shown = outcome.text.length > NOTE_TEXT_LIMIT ? `${outcome.text.slice(0, NOTE_TEXT_LIMIT)}\n… (truncated; the full transcript is in ${options.logPath})` : outcome.text;
+    const withUsage: AcpOutcome = { ...outcome, usage: { ...reported } };
     await queue.update(row.qitemId, {
-      note: `ACP run ${options.runId} outcome ${outcome.kind}: ${outcome.summary}${shown ? `\n\n${shown}` : ""}\nThis is the agent's report, not a judgment of the work; the row stays open for you to judge and close with your own resolution.`,
+      note: `ACP run ${options.runId} outcome ${outcome.kind}: ${outcome.summary}${shown ? `\n\n${shown}` : ""}\nThis is the agent's report, not a judgment of the work; the row stays open for you to judge and close with your own resolution.\n${describeUsage(withUsage.usage)}`,
     });
-    return outcome;
+    return withUsage;
   };
 
   try {
@@ -166,7 +182,12 @@ export async function runAcpAssignment(queue: AcpQueue, options: AcpRunOptions):
     const loads = init.agentCapabilities?.loadSession === true;
     if (options.loadSession) {
       if (!loads) return await record({ kind: "failed", summary: `the agent doesn't advertise loadSession, so session ${options.loadSession} wasn't loaded and no new session was started`, text: "" });
-      await peer.request("session/load", { sessionId: options.loadSession, cwd: options.cwd, mcpServers: [] });
+      try {
+        await peer.request("session/load", { sessionId: options.loadSession, cwd: options.cwd, mcpServers: [] });
+      } catch (error) {
+        if (!(error instanceof JsonRpcRequestError) || !options.missingSession?.(error.error, options.loadSession)) throw error;
+        return await record({ kind: "missing", text: "", summary: `session ${options.loadSession} doesn't exist for this agent (session/load returned ${describeError(error.error)}); no new session was started` });
+      }
       sessionId = options.loadSession;
     } else {
       sessionId = (await peer.request("session/new", { cwd: options.cwd, mcpServers: [] }) as { sessionId: string }).sessionId;
@@ -190,13 +211,18 @@ export async function runAcpAssignment(queue: AcpQueue, options: AcpRunOptions):
       void options.detachRequested?.then(() => { void peer.end(0); reject(new DetachedBeforeTurnEnd()); });
     });
     detached.catch(() => {});
-    const response = await Promise.race([peer.request("session/prompt", { sessionId, prompt: [{ type: "text", text: row.body }] }), detached]) as { stopReason?: string };
+    const response = await Promise.race([peer.request("session/prompt", { sessionId, prompt: [{ type: "text", text: row.body }] }), detached]) as { stopReason?: string; usage?: Record<string, unknown> };
     turnOver = true;
+    const tokens = response.usage;
+    if (tokens && typeof tokens === "object") {
+      const count = (key: string) => (typeof tokens[key] === "number" ? tokens[key] as number : undefined);
+      reported.tokens = { total: count("totalTokens"), input: count("inputTokens"), output: count("outputTokens"), thought: count("thoughtTokens"), cachedRead: count("cachedReadTokens"), cachedWrite: count("cachedWriteTokens") };
+    }
     return await record(classify(response.stopReason, text, usage, cancelRequestedAt, sessionId));
   } catch (error) {
     turnOver = true;
     if (error instanceof JsonRpcRequestError) {
-      return await record({ kind: "failed", sessionId, text, summary: `${error.method} returned error ${error.error.code} ${error.error.message}${error.error.data === undefined ? "" : ` ${JSON.stringify(error.error.data)}`}` });
+      return await record({ kind: "failed", sessionId, text, summary: `${error.method} returned error ${describeError(error.error)}` });
     }
     if (error instanceof DetachedBeforeTurnEnd) {
       return await record({ kind: "unknown", sessionId, text, summary: "detached before the turn ended; the agent may still be working" });
@@ -212,7 +238,9 @@ export async function runAcpAssignment(queue: AcpQueue, options: AcpRunOptions):
     const settled = left === undefined ? "unknown (the process table couldn't be read)"
       : left.length === 0 ? "settled: no process carrying this run's marker remains"
         : `incomplete: ${left.length} process(es) carrying this run's marker still run (${left.map((p) => p.pid).join(", ")}); stop them with the runtime's own stop, or rig acp stop ${options.runId} -- <its stop command>`;
-    await queue.update(row.qitemId, { note: `ACP run ${options.runId} detached: agent ${exit === "still-running" ? "still running after stdin closed" : `exited (code ${exit.code ?? "none"}, signal ${exit.signal ?? "none"})`}; ${settled}` }).catch(() => {});
+    const report = `agent ${exit === "still-running" ? "still running after stdin closed" : `exited (code ${exit.code ?? "none"}, signal ${exit.signal ?? "none"})`}; ${settled}`;
+    options.onSettled?.(report);
+    await queue.update(row.qitemId, { note: `ACP run ${options.runId} detached: ${report}` }).catch(() => {});
     peer.release(); // a still-running agent keeps running, but holds nothing that keeps this process alive
   }
 }
@@ -231,6 +259,22 @@ export function classify(stopReason: string | undefined, text: string, usage: {
   return { kind: "completed", sessionId, text, summary: "the agent ended its turn (end_turn)" };
 }
 
+const describeError = (error: JsonRpcErrorObject) => `${error.code} ${error.message}${error.data === undefined ? "" : ` ${JSON.stringify(error.data)}`}`;
+
+/** Usage and cost as the agent reported them, with where each came from; a figure it didn't report reads unknown. */
+export function describeUsage(usage: AcpUsage | undefined): string {
+  const t = usage?.tokens;
+  const parts = t ? [["in", t.input], ["out", t.output], ["thought", t.thought], ["cache read", t.cachedRead], ["cache write", t.cachedWrite]]
+    .filter(([, n]) => n !== undefined).map(([label, n]) => `${n} ${label}`) : [];
+  const tokens = t?.total !== undefined || parts.length > 0
+    ? `${t?.total !== undefined ? `${t.total} tokens` : "tokens"}${parts.length ? ` (${parts.join(", ")})` : ""}, from the agent's prompt response`
+    : "tokens unknown (the agent's prompt response reported none)";
+  const cost = usage?.cost
+    ? `cost ${usage.cost.amount} ${usage.cost.currency}, the agent's own cumulative figure for the session (usage_update)`
+    : "cost unknown (the agent reported none)";
+  return `Usage as the agent reported it: ${tokens}; ${cost}.`;
+}
+
 /** An owner answers with the offered option's id; its kind (allow_once …) works when only one option has it. */
 export function chooseOption(offered: PermissionOption[], answer: string | null | undefined): PermissionOption | undefined {
   const value = answer?.trim();

packages/cli/src/acp/json-rpc-stdio.ts

@@ -1,9 +1,10 @@
 import type { ChildProcessWithoutNullStreams } from "node:child_process";
 import { createInterface } from "node:readline";
 
-// ACP v1 is JSON-RPC 2.0 over newline-delimited JSON on the agent's stdio. This peer covers what an OpenRig client
-// needs (a few outgoing requests, incoming requests and notifications); the official SDK would add a dependency for
-// a handful of methods, so swap to it only if the method set grows.
+// ACP v1 is JSON-RPC 2.0 over newline-delimited JSON on the agent's stdio. This peer covers what OpenRig needs on either
+// side (a few outgoing requests, incoming requests and notifications): StdioJsonRpcPeer is the client end of an agent
+// process OpenRig launched, and JsonRpcPeer over this process's own stdin and stdout is the agent end (rig acp serve).
+// The official SDK would add a dependency for a handful of methods, so swap to it only if the method set grows.
 
 export interface JsonRpcErrorObject {
   code: number;
@@ -33,27 +34,18 @@ export type NotificationHandler = (params: unknown) => void;
 /** Every packet in both directions, in order; lines that aren't JSON arrive as `in-text`. */
 export type PacketRecorder = (direction: "out" | "in" | "in-text", packet: unknown) => void;
 
-export class StdioJsonRpcPeer {
+/** One JSON-RPC connection over a pair of streams: incoming lines on `input`, outgoing packets written to `output`. */
+export class JsonRpcPeer {
   private nextId = 0;
   private readonly pending = new Map<JsonRpcId, { method: string; resolve: (value: unknown) => void; reject: (error: Error) => void }>();
   private readonly requestHandlers = new Map<string, IncomingRequestHandler>();
   private readonly notificationHandlers = new Map<string, NotificationHandler>();
   private ended = false;
-  private readonly lines: ReturnType<typeof createInterface>;
-  readonly exited: Promise<{ code: number | null; signal: NodeJS.Signals | null }>;
+  protected readonly lines: ReturnType<typeof createInterface>;
 
-  constructor(private readonly child: ChildProcessWithoutNullStreams, private readonly record: PacketRecorder = () => {}) {
-    this.lines = createInterface({ input: child.stdout });
+  constructor(input: NodeJS.ReadableStream, private readonly output: NodeJS.WritableStream, private readonly record: PacketRecorder = () => {}) {
+    this.lines = createInterface({ input });
     this.lines.on("line", (line) => this.receive(line));
-    this.exited = new Promise((resolve) => child.once("close", (code, signal) => {
-      this.ended = true;
-      for (const [id, request] of this.pending) {
-        this.pending.delete(id);
-        request.reject(new JsonRpcTransportClosed(request.method));
-      }
-      resolve({ code, signal });
-    }));
-    child.stdin.on("error", () => { /* EPIPE after the agent exits; pending requests reject on close */ });
   }
 
   request(method: string, params: unknown): Promise<unknown> {
@@ -77,25 +69,18 @@ export class StdioJsonRpcPeer {
     this.notificationHandlers.set(method, handler);
   }
 
-  /** Close the agent's stdin (detach) and wait for the process to exit, up to `timeoutMs`. */
-  async end(timeoutMs: number): Promise<{ code: number | null; signal: NodeJS.Signals | null } | "still-running"> {
-    this.child.stdin.end();
-    let timer: NodeJS.Timeout | undefined;
-    const timeout = new Promise<"still-running">((resolve) => { timer = setTimeout(() => resolve("still-running"), timeoutMs); });
-    try { return await Promise.race([this.exited, timeout]); } finally { clearTimeout(timer); }
-  }
-
-  /** Drop this side's hold on the agent (its pipes and process handle) without signalling it, so a detached agent that
-   *  keeps running doesn't keep the caller's process alive. */
-  release(): void {
-    this.lines.close();
-    for (const stream of [this.child.stdin, this.child.stdout, this.child.stderr]) stream.destroy();
-    this.child.unref();
+  /** The connection is gone: nothing more is sent, and every request still waiting rejects as closed. */
+  protected close(): void {
+    this.ended = true;
+    for (const [id, request] of this.pending) {
+      this.pending.delete(id);
+      request.reject(new JsonRpcTransportClosed(request.method));
+    }
   }
 
   private send(packet: Record<string, unknown>): void {
     this.record("out", packet);
-    this.child.stdin.write(`${JSON.stringify(packet)}\n`);
+    this.output.write(`${JSON.stringify(packet)}\n`);
   }
 
   private receive(line: string): void {
@@ -127,7 +112,7 @@ export class StdioJsonRpcPeer {
   private async answer(method: string, id: JsonRpcId, params: unknown): Promise<void> {
     const handler = this.requestHandlers.get(method);
     if (!handler) {
-      this.send({ jsonrpc: "2.0", id, error: { code: -32601, message: `Client does not handle ${method}` } });
+      this.send({ jsonrpc: "2.0", id, error: { code: -32601, message: `Method not found: ${method}` } });
       return;
     }
     try {
@@ -135,7 +120,48 @@ export class StdioJsonRpcPeer {
       if (!this.ended) this.send({ jsonrpc: "2.0", id, result });
     } catch (error) {
       const code = typeof (error as { code?: unknown }).code === "number" ? (error as { code: number }).code : -32603;
-      if (!this.ended) this.send({ jsonrpc: "2.0", id, error: { code, message: (error as Error).message } });
+      const data = (error as { data?: unknown }).data;
+      if (!this.ended) this.send({ jsonrpc: "2.0", id, error: { code, message: (error as Error).message, ...(data !== undefined ? { data } : {}) } });
     }
   }
 }
+
+/** The client end of an ACP agent process OpenRig launched. */
+export class StdioJsonRpcPeer extends JsonRpcPeer {
+  readonly exited: Promise<{ code: number | null; signal: NodeJS.Signals | null }>;
+
+  constructor(private readonly child: ChildProcessWithoutNullStreams, record: PacketRecorder = () => {}) {
+    super(child.stdout, child.stdin, record);
+    this.exited = new Promise((resolve) => child.once("close", (code, signal) => {
+      this.close();
+      resolve({ code, signal });
+    }));
+    child.stdin.on("error", () => { /* EPIPE after the agent exits; pending requests reject on close */ });
+  }
+
+  /** Close the agent's stdin (detach) and wait for the process to exit, up to `timeoutMs`. */
+  async end(timeoutMs: number): Promise<{ code: number | null; signal: NodeJS.Signals | null } | "still-running"> {
+    this.child.stdin.end();
+    let timer: NodeJS.Timeout | undefined;
+    const timeout = new Promise<"still-running">((resolve) => { timer = setTimeout(() => resolve("still-running"), timeoutMs); });
+    try { return await Promise.race([this.exited, timeout]); } finally { clearTimeout(timer); }
+  }
+
+  /** Drop this side's hold on the agent (its pipes and process handle) without signalling it, so a detached agent that
+   *  keeps running doesn't keep the caller's process alive. */
+  release(): void {
+    this.lines.close();
+    for (const stream of [this.child.stdin, this.child.stdout, this.child.stderr]) stream.destroy();
+    this.child.unref();
+  }
+}
+
+/** The agent end over this process's stdin and stdout (or any stream pair): `closed` settles when the input ends. */
+export class StreamJsonRpcPeer extends JsonRpcPeer {
+  readonly closed: Promise<void>;
+
+  constructor(input: NodeJS.ReadableStream, output: NodeJS.WritableStream, record: PacketRecorder = () => {}) {
+    super(input, output, record);
+    this.closed = new Promise((resolve) => this.lines.once("close", () => { this.close(); resolve(); }));
+  }
+}

packages/cli/src/acp/run-processes.ts

@@ -9,9 +9,10 @@ export const RUN_MARKER = "OPENRIG_ACP_RUN";
 /** Only the pid: on macOS the listing carries the environment, which may hold secrets, so nothing else leaves here. */
 export interface RunProcess { pid: number }
 
-/** The caller's processes carrying this run's marker. Linux reads /proc; elsewhere `ps e` shows own environments. */
-export async function findRunProcesses(runId: string): Promise<RunProcess[]> {
-  const marker = `${RUN_MARKER}=${runId}`;
+/** The caller's processes carrying this run's marker (`rig exec` names its connector's own). Linux reads /proc;
+ *  elsewhere `ps e` shows own environments. */
+export async function findRunProcesses(runId: string, markerName = RUN_MARKER): Promise<RunProcess[]> {
+  const marker = `${markerName}=${runId}`;
   if (process.platform === "linux") {
     const found: RunProcess[] = [];
     for (const entry of await readdir("/proc")) {

packages/cli/src/bin-wrapper.ts

@@ -12,7 +12,8 @@ const REEXEC_GUARD_ENV = "OPENRIG_BIN_REEXEC";
 // Each `rig` is a new process: Node's on-disk compile cache lets the next call skip compiling the modules
 // again. The cache dates from Node v22.1.0, but this JavaScript API from v22.8.0: before it, `?.()` does
 // nothing. NODE_COMPILE_CACHE and NODE_DISABLE_COMPILE_CACHE still control it.
-module.enableCompileCache?.();
+// A piped bare invocation only prints usage; cold-cache writes can otherwise delay its exit.
+if (process.argv.length > 2 || (process.stdin.isTTY && process.stdout.isTTY)) module.enableCompileCache?.();
 
 export function resolveBinEntry(invokedPath = process.argv[1], moduleUrl = import.meta.url): string {
   const wrapperPath = invokedPath ? realpathSync(invokedPath) : realpathSync(fileURLToPath(moduleUrl));

packages/cli/src/commands/acp.ts

@@ -4,16 +4,20 @@ import { randomBytes } from "node:crypto";
 import { mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { join, resolve } from "node:path";
 import { DaemonClient, type DaemonResponse } from "../client.js";
-import { daemonStatusGuard, getDaemonStatus, getDaemonUrl } from "../daemon-lifecycle.js";
+import { daemonMayAnswer, daemonStatusGuard, getDaemonStatus, getDaemonUrl, statusGuardMessage } from "../daemon-lifecycle.js";
 import { processIdentity } from "../local-daemon-owner.js";
 import { getOpenRigHome, readOpenRigEnv } from "../openrig-compat.js";
 import { answerCommand, runAcpAssignment, type AcpQueue } from "../acp/connector.js";
 import { findRunProcesses } from "../acp/run-processes.js";
+import { serveAcpAgent, SESSION_TAG, TeamRefused, type TeamAsk, type TeamConnection, type TeamQueue } from "../acp/agent-server.js";
+import { StreamJsonRpcPeer } from "../acp/json-rpc-stdio.js";
+import { CLI_VERSION } from "../version.js";
+import { ConfigStore } from "../config-store.js";
 import { realDeps } from "./daemon.js";
 import type { StatusDeps } from "./status.js";
 
-const runsDir = () => join(getOpenRigHome(), "acp", "runs");
-const newId = (prefix: string, bytes: number) => `${prefix}-${new Date().toISOString().replace(/\D/g, "").slice(0, 14)}-${randomBytes(bytes).toString("hex")}`;
+export const runsDir = () => join(getOpenRigHome(), "acp", "runs");
+export const newId = (prefix: string, bytes: number) => `${prefix}-${new Date().toISOString().replace(/\D/g, "").slice(0, 14)}-${randomBytes(bytes).toString("hex")}`;
 
 interface RunFile { runId: string; qitemId: string; pid: number; processStart?: string; agent: string[]; startedAt: string; endedAt?: string }
 
@@ -44,6 +48,127 @@ export function daemonAcpQueue(client: DaemonClient): AcpQueue {
   };
 }
 
+/** rig acp serve's queue verbs over the daemon API, acting as the caller the client was built for. */
+export function daemonTeamQueue(client: DaemonClient): TeamQueue {
+  const ok = <T>(response: DaemonResponse<T>, what: string): T => {
+    if (response.status >= 400) throw new Error(`${what} returned HTTP ${response.status}: ${JSON.stringify(response.data)}`);
+    return response.data;
+  };
+  const face = (data: Record<string, unknown>) => {
+    const row = (data.item ?? data) as Record<string, unknown>;
+    const text = (value: unknown) => (typeof value === "string" ? value : null);
+    return {
+      qitemId: String(row.qitemId), state: String(row.state), destinationSession: String(row.destinationSession),
+      handedOffTo: text(row.handedOffTo), closureReason: text(row.closureReason), closureTarget: text(row.closureTarget), resolution: text(row.resolution),
+      closer: text((row.closingTransition as { actorSession?: unknown } | null | undefined)?.actorSession),
+    };
+  };
+  const list = async (query: string, what: string): Promise<Array<Record<string, unknown>>> => {
+    const data = ok(await client.get<unknown>(`/api/queue/list?${query}`), what);
+    return (Array.isArray(data) ? data : ((data as { items?: unknown[] }).items ?? [])) as Array<Record<string, unknown>>;
+  };
+  /** The rows that ask the person for a decision (questions with choices, or a decision in words). */
+  const asks = (rows: Array<Record<string, unknown>>): TeamAsk[] =>
+    rows.filter((row) => row.humanIntent !== "update" && (Array.isArray(row.humanQuestions) || row.humanIntent === "decision")).map((row) => ({
+      qitemId: String(row.qitemId), sourceSession: String(row.sourceSession), state: String(row.state), tsCreated: String(row.tsCreated),
+      summary: typeof row.summary === "string" ? row.summary : null, body: typeof row.body === "string" ? row.body : "",
+      tags: Array.isArray(row.tags) ? (row.tags as string[]) : [],
+      questions: Array.isArray(row.humanQuestions) ? (row.humanQuestions as TeamAsk["questions"]) : [],
+      // A click records the option's id; a typed reply records its words.
+      answers: Object.fromEntries(Object.entries((row.humanAnswers as Record<string, unknown> | null) ?? {})
+        .map(([id, value]) => [id, typeof value === "string" ? value : String((value as { text?: unknown } | null)?.text ?? "")])),
+    }));
+  return {
+    async create(row) {
+      const response = await client.post("/api/queue/create", { qitemId: row.qitemId, destinationSession: row.destination, body: row.body, summary: row.summary, tags: row.tags });
+      // A 4xx is the daemon's answered refusal, with nothing written; a 5xx leaves the write unknown, like no answer.
+      if (response.status >= 400 && response.status < 500) throw new TeamRefused(`HTTP ${response.status}: ${JSON.stringify(response.data)}`);
+      ok(response, `queue create for ${row.destination}`);
+    },
+    async show(qitemId) {
+      return face(ok(await client.get<Record<string, unknown>>(`/api/queue/${encodeURIComponent(qitemId)}`), `queue show ${qitemId}`));
+    },
+    async find(qitemId) {
+      const response = await client.get<Record<string, unknown>>(`/api/queue/${encodeURIComponent(qitemId)}`);
+      if (response.status === 404) return null;
+      return face(ok(response, `queue show ${qitemId}`));
+    },
+    async transitions(qitemId) {
+      const data = ok(await client.get<Array<Record<string, unknown>>>(`/api/queue/${encodeURIComponent(qitemId)}/transitions`), `queue transitions ${qitemId}`);
+      return data.map((t) => ({
+        transitionId: Number(t.transitionId), ts: String(t.ts), state: String(t.state),
+        transitionNote: typeof t.transitionNote === "string" ? t.transitionNote : null, actorSession: String(t.actorSession),
+      }));
+    },
+    async successor(qitemId) {
+      return (await list(`handedOffFrom=${encodeURIComponent(qitemId)}&limit=1`, `queue successor of ${qitemId}`))[0]?.qitemId as string | undefined;
+    },
+    async openAsks(person) {
+      return asks(await list(`destinationSession=${encodeURIComponent(person)}&state=pending&limit=500`, `open asks to ${person}`));
+    },
+    async sessionAsks(person, sessionId) {
+      const rows = await list(`destinationSession=${encodeURIComponent(person)}&limit=2000`, `rows to ${person}`);
+      return asks(rows.filter((row) => Array.isArray(row.tags) && (row.tags as string[]).includes(`${SESSION_TAG}${sessionId}`)));
+    },
+    async answer(qitemId, questionId, optionId) {
+      // The person's answer path, as a click in Slack records one: the daemon takes the actor from this client's caller.
+      const bearer = process.env.OPENRIG_AUTH_BEARER_TOKEN;
+      const response = await client.post<Record<string, unknown>>("/api/mission-control/action", { verb: "answer", qitemId, questionId, optionId },
+        bearer ? { headers: { Authorization: `Bearer ${bearer}` } } : undefined);
+      if (response.status < 400) return { ok: true };
+      const data = (response.data ?? {}) as { message?: unknown; error?: unknown };
+      return { ok: false, reason: String(data.message ?? data.error ?? `HTTP ${response.status}`) };
+    },
+    async sent(caller) {
+      const rows = await list(`sourceSession=${encodeURIComponent(caller)}&limit=2000`, `rows from ${caller}`);
+      return rows.filter((row) => Array.isArray(row.tags) && (row.tags as string[]).includes("acp-serve")).map((row) => ({
+        qitemId: String(row.qitemId), tsCreated: String(row.tsCreated), tsUpdated: String(row.tsUpdated),
+        summary: typeof row.summary === "string" ? row.summary : null, body: typeof row.body === "string" ? row.body : "", tags: row.tags as string[],
+      }));
+    },
+  };
+}
+
+/** The seat that takes this agent's tasks: --to, else this machine's acp.to setting (an ACP registry launch has no args).
+ *  A setting that can't be read comes back as the reason, for the app's sign-in step to show. */
+export function acpReceiver(to: string | undefined, store: Pick<ConfigStore, "get"> = new ConfigStore()): { receiver?: string } | { error: string } {
+  if (to?.trim()) return { receiver: to.trim() };
+  try {
+    return { receiver: String(store.get("acp.to") ?? "").trim() || undefined };
+  } catch (error) {
+    return { error: `${(error as Error).message} The agent reads acp.to from it; adding --to <member@rig> to the agent's command skips it.` };
+  }
+}
+
+/** Why the team's questions can't reach `address`, when it's a person this machine's OpenRig doesn't know. Says nothing
+ *  for a seat's address, a registered person, or a registry that can't be read (unknown isn't unregistered). */
+export async function unregisteredPersonNotice(address: string, home?: string): Promise<string | undefined> {
+  const suffix = "@external";
+  if (!address.endsWith(suffix)) return undefined;
+  try {
+    const registry = await import("@openrig/daemon/gateway-human-registry");
+    const people = registry.projectHumans(home);
+    if (!people.ok || people.entities.some((person) => person.address === address)) return undefined;
+  } catch {
+    return undefined;
+  }
+  const id = address.slice(0, -suffix.length);
+  return `Your team can't ask you questions here yet: ${address} isn't a registered person on this machine, so OpenRig refuses a question addressed to it. Your work is still sent. To be asked, register yourself with \`rig gateway human add ${id} --display-name "<your name>" --delivery-class <A|B|C|D>\`.`;
+}
+
+/** The person this machine's OpenRig names as its default (`<id>@external`), or why there isn't one. */
+async function defaultPersonAddress(): Promise<{ address: string } | { reason: string }> {
+  try {
+    const registry = await import("@openrig/daemon/gateway-human-registry");
+    const people = registry.projectHumans();
+    if (!people.ok) return { reason: people.error };
+    const chosen = registry.resolveDefaultPerson(people.entities);
+    return chosen.person ? { address: chosen.person.address } : { reason: chosen.source === "none" ? chosen.reason : "no default person" };
+  } catch (error) {
+    return { reason: (error as Error).message };
+  }
+}
+
 function readRun(runId: string): RunFile | undefined {
   try { return JSON.parse(readFileSync(join(runsDir(), `${runId}.json`), "utf8")) as RunFile; } catch { return undefined; }
 }
@@ -62,7 +187,7 @@ function liveConnector(run: RunFile): boolean {
 }
 
 export function acpCommand(depsOverride?: StatusDeps): Command {
-  const cmd = new Command("acp").description("Hand a queue row's task to an ACP (Agent Client Protocol) agent and keep its outcome on the row");
+  const cmd = new Command("acp").description("ACP (Agent Client Protocol): hand a queue row's task to an ACP agent, or serve OpenRig as an agent to ACP apps");
   const getDeps = () => depsOverride ?? { lifecycleDeps: realDeps(), clientFactory: (url: string) => new DaemonClient(url) };
   const connect = async (): Promise<DaemonClient | undefined> => {
     const deps = getDeps();
@@ -135,6 +260,62 @@ Guide: rig context get reference/acp-agents.md`)
       }
     });
 
+  cmd
+    .command("serve")
+    .description("Serve OpenRig as an ACP agent on stdio, for an ACP app (Zed and others) to launch: its prompts go to your team as queue rows")
+    .option("--to <member@rig>", "The seat that receives your tasks: your team's entry point (default: rig config get acp.to)")
+    .option("--as <address>", "Who the tasks come from (default: OPENRIG_SESSION_NAME, else this machine's default person)")
+    .option("--poll-seconds <n>", "How often to read the work's progress while a turn is open", "2")
+    .addHelpText("after", `
+An ACP app launches this command and talks to it on stdin and stdout. Each prompt becomes a queue row from you to
+--to. The app shows the row's notes and handoffs as progress, and the closing resolution as the reply. A later
+prompt in the same session goes to --to as a follow-up that names the earlier items. A question the working seats
+ask you meanwhile comes to the app as a choice, recorded as your answer. The app can list and reopen your sessions.
+
+The turn ends when the team closes the item, or when you cancel. Cancelling stops this process waiting; the team's
+work carries on. A turn that ended doesn't mean the work was accepted: the reply says how the item was closed.
+
+Without --to, the receiver is the acp.to setting (rig config set acp.to <member@rig>), which is how an app that
+starts the agent from the ACP registry, with no arguments, reaches your team.
+
+It uses the same OpenRig daemon as rig: the one on this computer, started with rig daemon start, unless OPENRIG_URL
+names another. The app's sign-in step ("OpenRig on this machine") checks that the daemon answers and that you
+resolve to an address: --as, OPENRIG_SESSION_NAME, or the default person: the only one registered
+(rig gateway human add <id> ...), or with several, the one chosen with rig gateway default <id>.
+
+In Zed's settings.json:
+  "agent_servers": { "OpenRig": { "type": "custom", "command": "rig", "args": ["acp", "serve", "--to", "lead@my-team"] } }
+
+Guide: rig context get reference/acp-agents.md#openrig-as-an-agent-in-acp-apps`)
+    .action(async (opts: { to?: string; as?: string; pollSeconds: string }) => {
+      // Only ACP messages may reach stdout: anything a helper prints goes to stderr instead.
+      console.log = console.error;
+      console.info = console.error;
+      process.stdout.on("error", () => { /* EPIPE once the app has gone; the input closing ends the server */ });
+      const connectTeam = async (): Promise<TeamConnection | { error: string }> => {
+        const status = await getDaemonStatus(getDeps().lifecycleDeps);
+        if (!daemonMayAnswer(status)) {
+          const err = statusGuardMessage(status);
+          return { error: `${err.fact} ${err.action}` };
+        }
+        let caller = opts.as?.trim() || self()?.trim();
+        if (!caller) {
+          const person = await defaultPersonAddress();
+          if ("reason" in person) return { error: `OpenRig doesn't know who you are on this machine (${person.reason}). Register yourself with rig gateway human add <id> --display-name "<your name>" --delivery-class <A|B|C|D> (the only registered person is the default; with several, choose one with rig gateway default <id>), or add --as <address> to the agent's command.` };
+          caller = person.address;
+        }
+        const target = acpReceiver(opts.to);
+        if ("error" in target) return target;
+        return { queue: daemonTeamQueue(new DaemonClient(getDaemonUrl(status), { sessionName: caller })), caller, receiver: target.receiver, notice: await unregisteredPersonNotice(caller) };
+      };
+      const peer = new StreamJsonRpcPeer(process.stdin, process.stdout);
+      serveAcpAgent(peer, {
+        connect: connectTeam, version: CLI_VERSION, pollMs: Math.max(1, Number(opts.pollSeconds) || 2) * 1000,
+        log: (line) => console.error(`[rig acp serve] ${line}`), closed: peer.closed,
+      });
+      await peer.closed;
+    });
+
   cmd
     .command("cancel <runId>")
     .description("Ask a running rig acp run to cancel its agent's turn (the same as Ctrl-C there)")

packages/cli/src/commands/agent.ts

@@ -45,7 +45,7 @@ export function agentCommand(depsOverride?: AgentDeps): Command {
 
       const client = deps.clientFactory(getDaemonUrl(status));
 
-      const res = await client.postText<{ valid?: boolean; errors?: string[] }>("/api/agents/validate", yaml);
+      const res = await client.postText<{ valid?: boolean; errors?: string[]; advisories?: string[] }>("/api/agents/validate", yaml);
 
       if (opts.json) {
         console.log(JSON.stringify(res.data));
@@ -65,6 +65,8 @@ export function agentCommand(depsOverride?: AgentDeps): Command {
       }
 
       const data = res.data;
+      // Advisories are printed whether or not the spec is valid; they never make it invalid.
+      for (const advisory of data.advisories ?? []) console.error(`⚠ agent spec advisory: ${advisory}`);
       if (data.valid) {
         const name = yamlScalar(yaml, "name") ?? "unknown";
         const version = yamlScalar(yaml, "version") ?? "unknown";

packages/cli/src/commands/attach.ts

@@ -62,7 +62,7 @@ export function attachCommand(depsOverride?: StatusDeps): Command {
     .option("--member <name>", "Member name for pod attach mode")
     .option("--runtime <runtime>", "Runtime for pod attach mode; optional guard for node mode")
     .option("--cwd <path>", "Working directory to record", process.cwd())
-    .option("--display-name <name>", "Display label inside tmux; external session name outside tmux")
+    .option("--display-name <name>", "Display label inside tmux; external session address outside tmux (new bare names gain @rig)")
     .option("--print-env", "Print shell exports for OPENRIG_NODE_ID and OPENRIG_SESSION_NAME")
     .option("--json", "JSON output, including the activity env the attached app needs. Its token is a credential: don't print or paste it, and use rig whoami for identity")
     .action(async (opts: {

packages/cli/src/commands/broadcast.ts

@@ -222,10 +222,13 @@ async function runCrossHostBroadcast(
  *  local route does, so per-target honesty prints identically. A partial
  *  fan-out sets a non-zero exit either way. */
 function renderBroadcastResults(data: Record<string, unknown>): void {
-  const results = (data["results"] as Array<{ sessionName: string; ok: boolean; error?: string; managedCompaction?: { notice?: unknown } }>) ?? [];
+  const results = (data["results"] as Array<{ sessionName: string; ok: boolean; error?: string; inputPhase?: string; sent?: boolean; managedCompaction?: { notice?: unknown } }>) ?? [];
   for (const r of results) {
     if (r.ok) {
       console.log(`${r.sessionName}: sent`);
+    } else if (r.inputPhase === "attempted" || r.inputPhase === "written" || r.sent === true) {
+      // The daemon records this recipient's receipt as indeterminate (#2507): its text or Enter may have reached it.
+      console.log(`${r.sessionName ? `${r.sessionName}: ` : ""}UNCONFIRMED — ${r.error ?? "unknown error"}; the text may have reached it, so check with rig capture before resending`);
     } else {
       // A target that resolved to no sessions has no name to report.
       console.log(`${r.sessionName ? `${r.sessionName}: ` : ""}FAILED — ${r.error ?? "unknown error"}`);

packages/cli/src/commands/compact.ts

@@ -55,8 +55,20 @@ manual request. Cancel an active preparation before retrying or using --skip-map
 Automatic preparation allows 25 minutes from prep delivery for the map. A completed
 map stays armed for idle through the next retry interval (10 minutes by default),
 without another preparation prompt; --state includes idleDeadlineAt. A completed
-map that never reaches idle reports map_complete_seat_never_idle; an unfinished
-map reports preparation_incomplete. Manual requests still share one 180-second budget.
+automatic map whose /compact is refused before input (a draft in the box, the pane in
+a tmux mode) or that never reaches idle in that time is held instead: --state shows
+prepared with preparedAt and preparedReason, and /compact goes in at the next idle
+moment with an empty box, with the same map and no new preparation prompt. A held
+attempt stops on a fresh sample below the threshold or from another conversation
+(below_threshold_while_prepared), once usage has risen 15 points since its request
+or two hours after its map was ready (prepared_map_stale), or when its map no longer
+reads ready (preparation_incomplete).
+After an attempt stops with its request delivered, the next automatic request for the
+same conversation waits for an idle seat, across a daemon restart too, for up to the
+restart-recovery window (6 hours by default) after that attempt stopped. An unfinished
+map reports preparation_incomplete. Manual requests still
+share one 180-second budget and stop as before; while an attempt is held, a manual
+request reads already_in_progress until --cancel.
 Stopped attempts may retry automatically after the cooldown; cancellation does not.
 Cancellation disarms future input; it cannot undo a command already submitted. Raw provider /compact and normal messages are unchanged.`);
 

packages/cli/src/commands/config.ts

@@ -11,7 +11,7 @@ function summarizeSettings(store: ConfigStore): Record<ValidKey, ResolvedSetting
 }
 
 // These settings remain boot-only; capture interval/line settings reload live.
-const BOOT_ONLY_KEYS = ["daemon.port", "daemon.host", "daemon.db_conflict", "db.path", "terminal.tmux_server"];
+const BOOT_ONLY_KEYS = ["daemon.port", "daemon.host", "daemon.db_conflict", "daemon.seat_poll_interval_ms", "db.path", "terminal.tmux_server"];
 const BOOT_ONLY_PREFIXES = ["transcripts.enabled", "transcripts.path"];
 
 function isBootOnlyKey(key: string): boolean {
@@ -55,7 +55,7 @@ Examples:
   rig config init-workspace                  # scaffold a repo-ready project workspace
 
 Keys:
-  daemon.*               port, host
+  daemon.*               port, host, seat_poll_interval_ms (startup-only seat-pollers cadence, 1000–2000; default 1000)
   db.path
   transcripts.*          enabled, path, lines, poll_interval_seconds
                          (lines/poll_interval set scrollback CAPTURE depth + cadence.
@@ -103,7 +103,26 @@ Keys:
                            an unreadable registry does not suppress eligibility; delivery still needs a ready route)
   retention.*            enabled, transitions_days, watchdog_days,
                          watchdog_keep_per_job, batch_size
+  retention.events_days  delete the history-only event types (views, snapshots, watchdog firings, queue events) older
+                         than N days; 0 keeps them (default). Removed events no longer replay from /api/events, leave
+                         the telemetry export, the log and the feed, and can't be brought back. Read at the next sweep
+                         (daily and at daemon start)
+  retention.closed_content_days  remove the content of closed queue work (done, canceled, denied, handed-off) N
+                         days after it closed; the row stays, without content. 0 keeps it (default). Can't be undone
+  retention.closed_metadata_days  delete closed queue rows N days after they closed, keeping only their ids so a retry
+                         can't recreate them. 0 keeps them (default). Both read at the next sweep; "rig queue erase
+                         --routine --apply" runs them now
+  retention.erasure_ledger  an absolute file where queue erasure appends each erased id and its action, one JSON
+                         line each, before it commits them; keep it outside backups and replay it with "rig queue
+                         erase --ids <file> --apply" after a restore. Empty keeps no ledger (default)
   launch.non_interruptive  accept harness first-launch warnings at full bypass for new rigs (default off)
+  launch.codex_login_shell  defer to Codex's own login-shell setting in managed Codex seats (default off)
+                         Off adds allow_login_shell=false: tool commands inherit the seat's environment, so plain rig
+                         works wherever OpenRig is installed. Inherited setup isn't re-run per command: setup only in
+                         login or rc files (including a tmux default-command that isn't a login shell) and shell
+                         functions or aliases may be absent. On omits that override; it doesn't force login shells,
+                         and a login file that sets PATH outright can then hide rig. Applies at each seat's next
+                         launch; a restart in place keeps the running choice.
   terminal.inherit_tmux_environment  inherit stored tmux global values for new agents and successors (default off)
                          POSIX launches otherwise start from daemon + seat environment, preserving shell startup.
                          Turn on for custom setups that intentionally rely on tmux globals. Running agents are unchanged.
@@ -123,10 +142,13 @@ Keys:
   usage.pi_openrouter_charge  with usage.pi_cost_facts on, Pi seats look up OpenRouter's actual charge for their requests, shown beside the estimate (default off; changes take effect when a Pi seat next launches or resumes; earlier outcomes stay visible)
   transport.verify_input_box  with --verify, a send to a Claude Code or Codex seat (also queue wakes) reads verified only when the seat was proven idle before the paste and its input box showed the exact text before Enter and was empty after it, on the same pane process; a busy or unknown seat, and anything else, reads unconfirmed, not failed. Reporting only (default off)
   transport.local_user_guard  on a Mac, the CLI refuses a local daemon whose reported user, checked with ps, isn't yours; an explicit OPENRIG_URL skips the check. A mitigation for reaching the wrong daemon, not authentication (default off)
+  transport.draft_delivery  deliver (default): a rig send or work notice to a seat whose prompt holds a person's draft goes through: the message is pasted at the current cursor into the draft (normally its end), then Enter is sent to submit the combined input, and the send's result says whether that was confirmed. OpenRig can't read every prompt layout: where it doesn't detect the draft, the result doesn't mention it, and refuse doesn't stop the send. refuse: such a send is refused, and a work notice retried once the prompt is clear, where OpenRig detects the draft. A draft found only by reading the input box may leave that retry waiting for the seat's next activity change or the stuck sweep, rather than the moment the draft clears. Either way, menus and permission prompts are refused, a per-seat typing guard still holds, managed compaction waits for an empty prompt, and a send with --wait-for-idle still waits
   decisions.enabled      typed decisions (OpenRouter + Jev) for the consumers in decisions.consumers, with OPENROUTER_API_KEY from the
                          rig process that runs them; off makes no provider calls (default off). rig project decisions status shows the state
   decisions.consumers    comma-separated consumers that may call (supported: stream; default none). An unknown name makes no calls
   decisions.template_dir absolute folder of <name>.yaml templates that replace the shipped ones (default empty: shipped templates only)
+  acp.to                 the seat (member@rig) that receives tasks from rig acp serve when it isn't given --to, as when an ACP app
+                         starts it from the ACP registry (default empty: none, so nothing is sent until one is set)
 
 Precedence: CLI flag > environment variable > config file > default`)
     .action((opts: { json?: boolean; withSource?: boolean }) => {
@@ -193,10 +215,17 @@ Precedence: CLI flag > environment variable > config file > default`)
     .action(async (key: string, value: string) => {
       try {
         store.set(key, value);
-        console.log(`${key} = ${store.get(key)}`);
+        const resolved = store.resolveWithSource(key);
+        console.log(resolved.source === "env"
+          ? `${key} = ${resolved.value} (source: env; file override saved but masked)`
+          : resolved.source === "default"
+            ? `${key} file override saved; effective value ${resolved.value} (source: default)`
+            : `${key} = ${resolved.value}`);
         // SWEEP-c: boot-only key + running daemon = stale-until-restart; say so loudly.
         if (isBootOnlyKey(key) && (await probeDaemonRunning())) {
-          console.error(`note: '${key}' is read at daemon BOOT — the running daemon keeps its current value; this takes effect on the next daemon restart (rig daemon stop && rig daemon start).`);
+          console.error(resolved.source === "env"
+            ? `note: '${key}' is read at daemon BOOT — the running daemon keeps its current value; the saved value remains masked after restart while the environment override is set.`
+            : `note: '${key}' is read at daemon BOOT — the running daemon keeps its current value; this takes effect on the next daemon restart (rig daemon stop && rig daemon start).`);
           if (key === "terminal.tmux_server") console.error("note: surviving sessions keep their previous server until deliberately stopped; rig terminal status shows the effective selection and any pending migration.");
         }
       } catch (err) {
@@ -212,9 +241,22 @@ Precedence: CLI flag > environment variable > config file > default`)
       try {
         store.reset(key);
         if (key) {
-          console.log(`${key} reset to default (${store.get(key)}).`);
+          const resolved = store.resolveWithSource(key);
+          console.log(resolved.source === "env"
+            ? `${key} file override cleared; effective value ${resolved.value} (source: env; default: ${resolved.defaultValue}).`
+            : `${key} reset to default (${resolved.value}).`);
         } else {
-          console.log("Config reset to defaults.");
+          let envKeys: string[] = [];
+          try {
+            envKeys = Object.entries(store.resolveAllWithSource())
+              .filter(([, resolved]) => resolved.source === "env")
+              .map(([name]) => name);
+          } catch (err) {
+            console.error(`note: could not read effective configuration after reset: ${(err as Error).message}`);
+          }
+          console.log(envKeys.length
+            ? `Config reset to defaults, except where the environment overrides them: ${envKeys.join(", ")}.`
+            : "Config reset to defaults.");
         }
       } catch (err) {
         console.error((err as Error).message);

packages/cli/src/commands/context-reconstruction.ts

@@ -0,0 +1,69 @@
+import { Command } from "commander";
+import { join, resolve } from "node:path";
+import { ConfigStore } from "../config-store.js";
+import { readOpenRigEnv } from "../openrig-compat.js";
+import { resolveWorkContextHome } from "../lib/work-install.js";
+import { ReconstructionError, readReconstruction, readStoreMappings, reconstructionOutput } from "../lib/reconstruction-reader.js";
+
+interface Options {
+  store?: string; storeMap?: string; project?: string; mission?: string; slice?: string; cwd?: string;
+  item: string[]; category: string[]; tag: string[]; checkSources?: boolean; maxBytes: string; json?: boolean;
+}
+const collect = (value: string, previous: string[]) => [...previous, value];
+
+/** A local config/filesystem consumer. This command never starts or contacts a
+ * daemon, invokes a helper, writes revisions, or takes over the selected work. */
+export function reconstructionCommand(): Command {
+  const cmd = new Command("reconstruction")
+    .description("Inspect or read an exact immutable reconstruction locally (read-only v0 consumer)");
+  for (const mode of ["inspect", "view"] as const) {
+    cmd.command(`${mode} <revision>`)
+      .description(mode === "inspect" ? "Inspect provenance, dates, size and the complete source manifest" : "Read a faithful selected view with dependencies and explicit omissions")
+      .option("--store <absolute-path>", "Explicit local v0/import store; default: selected work home/reconstructions")
+      .option("--store-map <file>", "JSON array of {sourceStore, localStore, sourceHost?}; read-time mapping only")
+      .option("--project <id>", "Catalog project, with the same inference as context work-install")
+      .option("--mission <id>", "Mission in the selected project's configured missions root")
+      .option("--slice <id>", "Slice under the selected mission")
+      .option("--cwd <path>", "Working directory used for project inference")
+      .option("--item <id>", "Include item (repeatable; selectors form a union)", collect, [])
+      .option("--category <name>", "Include category (repeatable)", collect, [])
+      .option("--tag <name>", "Include tag (repeatable)", collect, [])
+      .option("--check-sources", "Hash declared local files within 128 files / 64 MiB; never refresh observation dates")
+      .option("--max-bytes <bytes>", "Maximum complete output, 1024..8388608 UTF-8 bytes", "262144")
+      .option("--json", "Structured output; inspect always uses JSON")
+      .addHelpText("after", "\nSupports openrig.reconstruction/v0 with safe integer JSON numbers; other numeric representations are explicitly unsupported. No implicit remote reads. Context does not transfer authority.\n")
+      .action((revision: string, opts: Options) => {
+        try {
+          let store = opts.store;
+          let resolution: object = { mode: "explicit-store", project: null, home: null };
+          if (store === undefined || opts.project !== undefined || opts.mission !== undefined || opts.slice !== undefined) {
+            const config = new ConfigStore();
+            const home = resolveWorkContextHome({
+              workspaceRoot: String(config.resolveWithSource("workspace.root").value),
+              catalogPath: String(config.resolveWithSource("workspace.catalog_path").value),
+              cwd: resolve(opts.cwd ?? process.cwd()),
+              ...(readOpenRigEnv("OPENRIG_SESSION_NAME", "RIGGED_SESSION_NAME") ? { sessionName: readOpenRigEnv("OPENRIG_SESSION_NAME", "RIGGED_SESSION_NAME")! } : {}),
+              ...(opts.project === undefined ? {} : { project: opts.project }),
+              ...(opts.mission === undefined ? {} : { mission: opts.mission }),
+              ...(opts.slice === undefined ? {} : { slice: opts.slice }),
+            });
+            if ("error" in home) throw new ReconstructionError(home.error.code, home.error.message);
+            resolution = { mode: store === undefined ? "work-home" : "explicit-store", ...home };
+            store ??= join(home.home, "reconstructions");
+          }
+          const result = readReconstruction({ store, revision,
+            ...(opts.storeMap ? { mappings: readStoreMappings(resolve(opts.storeMap)) } : {}),
+            items: opts.item, categories: opts.category, tags: opts.tag, checkSources: opts.checkSources });
+          const response = { ...result, resolution };
+          console.log(reconstructionOutput(response, mode, Boolean(opts.json), Number(opts.maxBytes)));
+          if (!result.complete) process.exitCode = 1;
+        } catch (error) {
+          const code = error instanceof ReconstructionError ? error.code : "reconstruction_unavailable";
+          if (opts.json || mode === "inspect") console.log(JSON.stringify({ complete: false, error: { code, message: (error as Error).message } }, null, 2));
+          else console.error(`${code}: ${(error as Error).message}`);
+          process.exitCode = 1;
+        }
+      });
+  }
+  return cmd;
+}

packages/cli/src/commands/context.ts

@@ -43,6 +43,7 @@ import { realDeps } from "./daemon.js";
 import { prepareDaemonAutoStart } from "../daemon-auto-start.js";
 import { readOpenRigEnv } from "../openrig-compat.js";
 import type { StatusDeps } from "./status.js";
+import { reconstructionCommand } from "./context-reconstruction.js";
 
 const contextRuntimeArg = enumArg(["claude-code", "claude", "codex"]);
 
@@ -194,6 +195,7 @@ async function installPackFromUrl(
     // Re-validate the on-disk staged pack, then publish atomically.
     validateContextPackManifestForInstall(join(staging, "manifest.yaml"));
     assertTreeHasNoSymlinks(staging);
+    mkdirSync(dirname(targetDir), { recursive: true });
     renameSync(staging, targetDir);
     return { targetDir, installName };
   } catch (err) {
@@ -251,6 +253,8 @@ Examples:
   rig context trace --rig factory --pod dev --seat dev-qa --name LEARNED.md
 `);
 
+  cmd.addCommand(reconstructionCommand());
+
   const getDeps = (): StatusDeps => depsOverride ?? {
     lifecycleDeps: realDeps(),
     clientFactory: (url: string) => new DaemonClient(url),

packages/cli/src/commands/daemon.ts

@@ -57,10 +57,12 @@ export function createProcessLiveness(deps: {
   readProcessState: (pid: number) => string | null;
 }): (pid: number) => ProcessLiveness {
   return (pid: number) => {
-    if (deps.signal(pid) === "missing") return "dead";
+    const signal = deps.signal(pid);
+    if (signal === "missing") return "dead";
     const state = deps.readProcessState(pid)?.trim();
     if (!state) return "unknown";
-    return state.startsWith("Z") ? "dead" : "alive";
+    if (state.startsWith("Z")) return "dead";
+    return signal === "not-permitted" ? "unknown" : "alive";
   };
 }
 

packages/cli/src/commands/doctor.ts

@@ -6,6 +6,7 @@ import { execSync } from "node:child_process";
 import { resolveDaemonPath } from "../daemon-lifecycle.js";
 import { DaemonClient, ForeignLocalDaemonError, formatDaemonHostForUrl } from "../client.js";
 import { ConfigStore } from "../config-store.js";
+import { fetchWithTimeout } from "../fetch-with-timeout.js";
 import { buildWritableHomeCheck } from "../system-preflight.js";
 import {
   CMUX_SETTINGS_DISCLOSURE_PATH,
@@ -45,7 +46,7 @@ export interface DoctorDeps {
   /** The caller's effective user id, compared with the one the port's daemon reports. Absent: this process's own;
    *  null: this platform has none. */
   callerUid?: number | null;
-  /** How long the port's /healthz body may take to arrive (default 2000 ms); after that its owner is unknown. */
+  /** How long each part of /healthz and the cmux status response may take (default 2000 ms). */
   ownerReportTimeoutMs?: number;
   /**
    * Build B — path to a rig spec to check against the RUNNING rig of the same name (`--spec`).
@@ -61,6 +62,7 @@ export interface DoctorDeps {
 }
 
 const DEFAULT_PORT = 7433;
+const DOCTOR_HTTP_TIMEOUT_MS = 2000;
 const CMUX_SOCKET_CONTROL_FIX = 'In cmux Settings > Automation, select "Automation mode" for Socket Control Mode, or set automation.socketControlMode to "automation" in ~/.config/cmux/cmux.json (older versions use ~/.config/cmux/settings.json; current cmux reads that legacy file only for keys absent from cmux.json). cmux reloads saved settings automatically; rerun `rig doctor` to re-probe without restarting the daemon.';
 
 /** The port's OpenRig daemon, judged by the OS user it reports. /healthz's pid and euid are the daemon's own report,
@@ -301,6 +303,15 @@ export function runDoctorChecks(deps: DoctorDeps, providerChecks: DoctorCheck[]
   const daemonPort = config.daemon.port ?? DEFAULT_PORT;
   const daemonBase = `http://${formatDaemonHostForUrl(daemonHost)}:${daemonPort}`;
   const fetchFn = deps.fetch ?? globalThis.fetch;
+  const healthHeaders = async (controller: AbortController) => {
+    let timer: ReturnType<typeof setTimeout> | undefined;
+    try {
+      return await Promise.race([
+        fetchFn(`${daemonBase}/healthz`, { signal: controller.signal }),
+        new Promise<null>(resolve => { timer = setTimeout(() => { resolve(null); controller.abort(); }, deps.ownerReportTimeoutMs ?? 2000); }),
+      ]);
+    } finally { clearTimeout(timer); }
+  };
   const portCheck = deps.checkPort(daemonPort, daemonHost).then(async (available): Promise<DoctorCheck> => {
     if (available) {
       return { name: "port", status: "pass", message: `Port ${daemonHost}:${daemonPort} available.` };
@@ -308,7 +319,13 @@ export function runDoctorChecks(deps: DoctorDeps, providerChecks: DoctorCheck[]
     // Port in use — check if it's an OpenRig daemon via healthz, and whose.
     const owner = new AbortController();
     try {
-      const res = await fetchFn(`${daemonBase}/healthz`, { signal: owner.signal });
+      const res = await healthHeaders(owner);
+      if (!res) return {
+        name: "port", status: "warn",
+        message: `Port ${daemonHost}:${daemonPort} is in use by an unknown listener.`,
+        reason: "Its /healthz response headers did not arrive in time, so doctor cannot identify it.",
+        fix: `Check the listener on port ${daemonPort} or configure a free port for the daemon.`,
+      };
       if (res.ok) return await listenerOwnerCheck(`${daemonHost}:${daemonPort}`, daemonBase, res, () => owner.abort(), deps);
     } catch { /* not our daemon */ }
     finally { owner.abort(); }
@@ -325,19 +342,27 @@ export function runDoctorChecks(deps: DoctorDeps, providerChecks: DoctorCheck[]
   const asyncChecks: Promise<DoctorCheck>[] = [portCheck];
   if (shellCmuxPassed) {
     const daemonCmuxCheck = (async (): Promise<DoctorCheck> => {
+      const health = new AbortController();
       try {
-        const healthRes = await fetchFn(`${daemonBase}/healthz`);
-        if (!healthRes.ok) {
+        const healthRes = await healthHeaders(health);
+        if (!healthRes?.ok) {
           return { name: "cmux_daemon", status: "skipped", message: "Daemon healthz not ok. Skipping daemon cmux check." };
         }
       } catch {
         return { name: "cmux_daemon", status: "skipped", message: "Daemon not reachable. Skipping daemon cmux check." };
-      }
+      } finally { health.abort(); }
 
       try {
-        const cmuxRes = await fetchFn(`${daemonBase}/api/adapters/cmux/status`);
-        if (cmuxRes.ok && cmuxRes.json) {
-          const data = (await cmuxRes.json()) as { available?: boolean };
+        let data: { available?: boolean } = {};
+        const cmuxRes = await fetchWithTimeout(fetchFn as typeof fetch, `${daemonBase}/api/adapters/cmux/status`, {}, {
+          timeoutMs: deps.ownerReportTimeoutMs ?? DOCTOR_HTTP_TIMEOUT_MS,
+          timeoutMessage: "Doctor's cmux status read timed out.",
+          consumeResponse: async res => {
+            if (res.ok && res.json) data = (await res.json()) as { available?: boolean };
+            else await res.body?.cancel();
+          },
+        });
+        if (cmuxRes.ok) {
           if (data.available) {
             return { name: "cmux_daemon", status: "pass", message: "Daemon cmux control available." };
           }
@@ -350,11 +375,72 @@ export function runDoctorChecks(deps: DoctorDeps, providerChecks: DoctorCheck[]
     asyncChecks.push(daemonCmuxCheck);
   }
 
+  asyncChecks.push(buildTmuxControlModeCheck(deps, daemonBase));
   asyncChecks.push(buildSpecConformanceCheck(deps));
 
   return { checks, portCheck, asyncChecks };
 }
 
+type TmuxControlReport = { requested?: boolean; state?: "on" | "allowed" | "off" | "undecided"; serverVersion?: string };
+
+/**
+ * #161 — does the daemon drive tmux through one control-mode client (OPENRIG_TMUX_CONTROL=1)? Asked for on a server
+ * older than tmux 3.7c, it uses a spawned client for every command, and nothing else says so.
+ */
+async function buildTmuxControlModeCheck(deps: DoctorDeps, daemonBase: string): Promise<DoctorCheck> {
+  const name = "tmux_control_mode";
+  let control = null as TmuxControlReport | null;
+  try {
+    await fetchWithTimeout((deps.fetch ?? globalThis.fetch) as typeof fetch, `${daemonBase}/api/adapters/tmux/control`, {}, {
+      timeoutMs: deps.ownerReportTimeoutMs ?? DOCTOR_HTTP_TIMEOUT_MS,
+      timeoutMessage: "Doctor's tmux control mode read timed out.",
+      consumeResponse: async res => {
+        if (res.ok && res.json) control = (await res.json()) as TmuxControlReport;
+        else await res.body?.cancel();
+      },
+    });
+  } catch { /* no daemon, or one from before this route */ }
+  if (!control) {
+    return { name, status: "skipped", message: "The daemon didn't say whether it uses tmux control mode (it isn't running, or it's older than this check)." };
+  }
+  if (!control.requested) {
+    if ((deps.env ?? process.env)["OPENRIG_TMUX_CONTROL"] !== "1") {
+      return { name, status: "skipped", message: "tmux control mode isn't asked for (OPENRIG_TMUX_CONTROL=1 in the daemon's environment)." };
+    }
+    return {
+      name, status: "warn",
+      message: "OPENRIG_TMUX_CONTROL=1 is set in this shell, but the running daemon wasn't started with it, so it spawns a tmux client for every command.",
+      fix: "Set OPENRIG_TMUX_CONTROL=1 in the daemon's own environment (the shell or service that starts it), then restart the daemon.",
+    };
+  }
+  const version = control.serverVersion;
+  if (control.state === "on") {
+    return { name, status: "pass", message: `tmux control mode is on: the daemon's tmux server runs tmux ${version}.` };
+  }
+  if (control.state === "undecided") {
+    return { name, status: "skipped", message: "tmux control mode is asked for, and the daemon hasn't read its tmux server's version yet. It does at its first tmux command while a session runs." };
+  }
+  if (control.state === "allowed") {
+    return {
+      name, status: "warn",
+      message: `tmux control mode is asked for and the server's tmux ${version} allows it, but no control client is connected right now, so tmux commands use a spawned client.`,
+      reason: "The daemon starts a control client at a tmux command while another session runs on the server. One that can't start or attach is tried again after a while, and commands are spawned meanwhile.",
+    };
+  }
+  if (!version) {
+    return { name, status: "warn", message: "tmux control mode is asked for but this daemon doesn't use a control client: it runs on Windows (psmux) or with OPENRIG_TMUX_SHELL=1." };
+  }
+  const known = /^\d+\.\d+[a-z]?$/.test(version);
+  return {
+    name, status: "warn",
+    message: known
+      ? `tmux control mode is asked for but off: the daemon's tmux server runs tmux ${version}, and control mode needs tmux 3.7c or later.`
+      : `tmux control mode is asked for but off: the daemon's tmux server runs tmux ${version}, which isn't a tmux release known to be 3.7c or later.`,
+    reason: "Starting a control client can crash a tmux server older than 3.7c, so the daemon spawns a tmux client for every command instead. The running server's version decides this, not the tmux binary on PATH.",
+    fix: "Run the daemon's sessions on a tmux 3.7c or later server: rig context get reference/tmux-servers.md#moving-an-existing-team-onto-tmux-control-mode",
+  };
+}
+
 /**
  * Build B — does the rig spec still describe the rig that is running?
  *
@@ -398,6 +484,12 @@ async function buildSpecConformanceCheck(deps: DoctorDeps): Promise<DoctorCheck>
   let liveIds: string[] | null;
   try { liveIds = deps.fetchLiveLogicalIds ? await deps.fetchLiveLogicalIds(rigName) : null; }
   catch (error) {
+    if (error instanceof AmbiguousLiveRigError) return {
+      name, status: "skipped",
+      message: `Multiple live rigs match '${rigName}': ${error.ids.join(", ")}; conformance unknown.`,
+      reason: "The spec name does not identify one live rig, so no topology was compared.",
+      fix: "Inspect the matching rig IDs with rig ps --json; use rig ps --nodes -A --json to inspect nodes.",
+    };
     if (error instanceof ForeignLocalDaemonError) return { name, status: "fail", message: error.message };
     throw error;
   }
@@ -482,22 +574,43 @@ function readTmuxMouseMode(deps: DoctorDeps): "on" | "off" | null {
   }
 }
 
+class AmbiguousLiveRigError extends Error {
+  constructor(readonly ids: string[]) { super("Multiple live rigs share the spec name"); }
+}
+
 /** Raw topology reads retain their existing headers; the owner policy is shared with the client. */
 export async function readLiveLogicalIds(base: string, rigName: string, ownerClient = new DaemonClient(base)): Promise<string[] | null> {
   try {
     await ownerClient.assertLocalOwner();
-    const rigsRes = await fetch(`${base}/api/rigs`);
+    let rigs: Array<{ id?: string; name?: string }> = [];
+    const rigsRes = await fetchWithTimeout(fetch, `${base}/api/rigs`, {}, {
+      timeoutMs: DOCTOR_HTTP_TIMEOUT_MS,
+      timeoutMessage: "Doctor's live rig read timed out.",
+      consumeResponse: async res => {
+        if (res.ok) rigs = (await res.json()) as typeof rigs;
+        else await res.body?.cancel();
+      },
+    });
     if (!rigsRes.ok) return null;
-    const rigs = (await rigsRes.json()) as Array<{ id?: string; name?: string }>;
-    const rig = rigs.find((r) => r.name === rigName);
+    const matches = rigs.filter((r) => r.name === rigName);
+    if (matches.length > 1) throw new AmbiguousLiveRigError(matches.map(r => r.id ?? "(missing id)"));
+    if (matches.length === 0) return null;
+    const rig = matches[0];
     if (!rig?.id) return null;
     await ownerClient.assertLocalOwner();
-    const nodesRes = await fetch(`${base}/api/rigs/${rig.id}/nodes`);
+    let nodes: Array<{ logicalId?: string | null }> = [];
+    const nodesRes = await fetchWithTimeout(fetch, `${base}/api/rigs/${rig.id}/nodes`, {}, {
+      timeoutMs: DOCTOR_HTTP_TIMEOUT_MS,
+      timeoutMessage: "Doctor's live node read timed out.",
+      consumeResponse: async res => {
+        if (res.ok) nodes = (await res.json()) as typeof nodes;
+        else await res.body?.cancel();
+      },
+    });
     if (!nodesRes.ok) return null;
-    const nodes = (await nodesRes.json()) as Array<{ logicalId?: string | null }>;
     return nodes.map((n) => n.logicalId ?? "");
   } catch (error) {
-    if (error instanceof ForeignLocalDaemonError) throw error;
+    if (error instanceof ForeignLocalDaemonError || error instanceof AmbiguousLiveRigError) throw error;
     return null;
   }
 }
@@ -529,7 +642,7 @@ export function doctorCommand(depsOverride?: DoctorDeps): Command {
             const base = `http://${formatDaemonHostForUrl(cfg.daemon.host ?? "127.0.0.1")}:${cfg.daemon.port ?? DEFAULT_PORT}`;
             return await readLiveLogicalIds(base, rigName);
           } catch (error) {
-            if (error instanceof ForeignLocalDaemonError) throw error;
+            if (error instanceof ForeignLocalDaemonError || error instanceof AmbiguousLiveRigError) throw error;
             return null;
           }
         },

packages/cli/src/commands/down.ts

@@ -147,6 +147,11 @@ Before running it, check who is working (rig ps --nodes --rig <rig>) and that th
 don't stop a team you didn't start unless asked. OpenRig snapshots the rig first (best-effort), and
 rig up <rig> --existing brings it back, resuming each seat's conversation where its harness allows. With
 --delete the rig record is removed too, so it can't be brought back that way.
+Once no live seat uses a working folder, rig down also takes back what OpenRig added there: its
+guidance blocks, and in a Claude seat's folder the status line, activity hooks and context collector.
+It keeps and lists an edited skills-line block, edited team skill files, and settings or a collector
+that don't match what OpenRig wrote or that another seat shares by a link. Any other guidance block
+goes even if someone edited it. rig up <rig> --existing puts the guidance back.
 When your resolved seat belongs to this rig, confirm before stopping it. Without a terminal,
 the command warns and exits without stopping; re-run with --yes to confirm explicitly.`);
   const getDepsF = () => depsOverride ?? { lifecycleDeps: realDeps(), clientFactory: (url: string) => new DaemonClient(url) };

packages/cli/src/commands/env.ts

@@ -30,9 +30,14 @@ Examples:
       throw new Error(error ?? `Failed to list rigs (HTTP ${summaries.status})`);
     }
     if (!Array.isArray(summaries.data)) throw new Error("The daemon returned an invalid rig summary.");
-    const match = summaries.data.find((r) => r.name === rigRef || r.id === rigRef);
-    if (!match) throw new Error(`Rig '${rigRef}' not found`);
-    return match.id;
+    const exactId = summaries.data.find((r) => r.id === rigRef);
+    if (exactId) return exactId.id;
+    const matches = summaries.data.filter((r) => r.name === rigRef);
+    if (matches.length > 1) {
+      throw new Error(`'${rigRef}' matches ${matches.length} rigs (${matches.map((r) => r.id).join(", ")}). Nothing was done; pass one of these ids instead.`);
+    }
+    if (matches.length === 0) throw new Error(`Rig '${rigRef}' not found`);
+    return matches[0]!.id;
   }
 
   // rig env status <rig>

packages/cli/src/commands/exec.ts

@@ -0,0 +1,581 @@
+import { Command } from "commander";
+import { spawn } from "node:child_process";
+import { closeSync, linkSync, mkdirSync, openSync, readdirSync, readFileSync, realpathSync, renameSync, rmSync, statSync, writeFileSync } from "node:fs";
+import { delimiter, dirname, join, resolve } from "node:path";
+import { fileURLToPath } from "node:url";
+import { DaemonClient } from "../client.js";
+import { daemonStatusGuard, getDaemonStatus, getDaemonUrl } from "../daemon-lifecycle.js";
+import { readOpenRigEnv } from "../openrig-compat.js";
+import { describeUsage, runAcpAssignment, type AcpOutcomeKind, type AcpQueue, type AcpUsage } from "../acp/connector.js";
+import type { JsonRpcErrorObject } from "../acp/json-rpc-stdio.js";
+import { findRunProcesses, RUN_MARKER } from "../acp/run-processes.js";
+import { daemonAcpQueue, newId, processStart, runsDir } from "./acp.js";
+import { realDeps } from "./daemon.js";
+import type { StatusDeps } from "./status.js";
+
+// `rig exec`: one queue row's task as one turn on a named, version-qualified ACP harness (Claude Code or Codex through
+// their maintained adapters), run by a detached connector that outlives the calling CLI. The connector owns the run,
+// the run file is its registry, and the row stays the custody record: the outcome is the agent's report, and accepting
+// the work stays with the row's custodian. `rig acp run` is the synchronous form and is unchanged.
+
+/** Set only in the connector's own environment, so a run's connector can be found before it records its pid. */
+export const CONNECTOR_MARKER = "OPENRIG_EXEC_CONNECTOR";
+
+export interface Harness {
+  name: string;
+  /** The adapter package, and the version this configuration was checked with. */
+  package: string;
+  qualified: string;
+  bin: string;
+  native: { package: string; label: string; versionField?: string };
+  /** Environment the caller may set to choose what runs or what the agent asks; passed through untouched. */
+  env: string[];
+  reject: string;
+  /** The adapter's own evidence that a `session/load` error means the session doesn't exist. */
+  missingSession: (error: JsonRpcErrorObject, sessionId: string) => boolean;
+}
+
+/** The ACP SDK's `sessionNotFound`: -32002 with `data.sessionId` naming the session asked for. A bare -32002 isn't. */
+const sessionNotFound = (error: JsonRpcErrorObject, sessionId: string) =>
+  error.code === -32002 && (error.data as { sessionId?: unknown } | null | undefined)?.sessionId === sessionId;
+
+export const HARNESSES: Record<string, Harness> = {
+  claude: {
+    name: "claude", package: "@agentclientprotocol/claude-agent-acp", qualified: "0.101.0", bin: "claude-agent-acp",
+    native: { package: "@anthropic-ai/claude-agent-sdk", label: "Claude Code", versionField: "claudeCodeVersion" },
+    env: ["CLAUDE_CODE_EXECUTABLE", "ANTHROPIC_MODEL", "CLAUDE_CONFIG_DIR"],
+    reject: "a reject refuses that one action, and the turn continues",
+    missingSession: sessionNotFound,
+  },
+  codex: {
+    name: "codex", package: "@agentclientprotocol/codex-acp", qualified: "2.5.0", bin: "codex-acp",
+    native: { package: "@openai/codex", label: "Codex" },
+    env: ["INITIAL_AGENT_MODE", "CODEX_HOME", "CODEX_CONFIG"],
+    reject: "its only refusal, cancel, ends the whole turn",
+    // codex-acp 2.5.0 passes a failed thread resume through as -32603 carrying Codex's own text, which the adapter's
+    // own code reads as a missing rollout. Any other -32603 is an internal error, not a missing session.
+    missingSession: (error, sessionId) => sessionNotFound(error, sessionId) || (error.code === -32603
+      && JSON.stringify(error.data ?? "").includes("no rollout found for thread id") && JSON.stringify(error.data).includes(sessionId)),
+  },
+};
+
+export interface ResolvedHarness {
+  harness: Harness;
+  /** The adapter's program as found on PATH; it runs as-is. */
+  command?: string;
+  adapterVersion?: string;
+  native: string;
+  problem?: string;
+}
+
+function readJson(file: string): Record<string, unknown> | undefined {
+  try { return JSON.parse(readFileSync(file, "utf8")) as Record<string, unknown>; } catch { return undefined; }
+}
+
+function onPath(bin: string, path = ""): string | undefined {
+  for (const dir of path.split(delimiter)) {
+    if (!dir) continue;
+    try { if (statSync(join(dir, bin)).isFile()) return join(dir, bin); } catch { /* not in this folder */ }
+  }
+  return undefined;
+}
+
+/** The installed package a file belongs to: the nearest folder above it whose package.json has that name. */
+function packageRoot(file: string, name: string): string | undefined {
+  for (let dir = dirname(file); dir !== dirname(dir); dir = dirname(dir)) {
+    if (readJson(join(dir, "package.json"))?.name === name) return dir;
+  }
+  return undefined;
+}
+
+/** Node's lookup from the adapter's folder, reading package.json directly (a package's exports map may not list it). */
+function dependencyManifest(from: string, name: string): Record<string, unknown> | undefined {
+  for (let dir = from; dir !== dirname(dir); dir = dirname(dir)) {
+    const manifest = readJson(join(dir, "node_modules", name, "package.json"));
+    if (manifest) return manifest;
+  }
+  return undefined;
+}
+
+export function resolveHarness(harness: Harness, env: NodeJS.ProcessEnv = process.env): ResolvedHarness {
+  const found = onPath(harness.bin, env.PATH);
+  if (!found) {
+    return { harness, native: "unknown", problem: `${harness.bin} isn't on PATH; install it with npm install --prefix <dir> ${harness.package}@${harness.qualified} and put <dir>/node_modules/.bin on PATH` };
+  }
+  const root = packageRoot(realpathSync(found), harness.package);
+  const version = root ? readJson(join(root, "package.json"))?.version : undefined;
+  const manifest = root ? dependencyManifest(root, harness.native.package) : undefined;
+  const field = harness.native.versionField ? manifest?.[harness.native.versionField] : undefined;
+  const bundled = !manifest ? `${harness.native.label}: ${harness.native.package} wasn't found beside the adapter`
+    : field !== undefined ? `${harness.native.label} ${String(field)}, from ${harness.native.package} ${String(manifest.version)}`
+      : `${harness.native.label} ${String(manifest.version)} (${harness.native.package})`;
+  const native = harness.name === "claude" && env.CLAUDE_CODE_EXECUTABLE
+    ? `the Claude Code at ${env.CLAUDE_CODE_EXECUTABLE} (CLAUDE_CODE_EXECUTABLE; its version isn't read here), not ${bundled}`
+    : bundled;
+  return { harness, command: found, adapterVersion: typeof version === "string" ? version : undefined, native };
+}
+
+/** The harness's caller-set environment, by name; CODEX_CONFIG's contents aren't shown. */
+function callerEnv(harness: Harness, env: NodeJS.ProcessEnv = process.env): string[] {
+  return harness.env.filter((name) => env[name] !== undefined).map((name) => (name === "CODEX_CONFIG" ? `${name} (set)` : `${name}=${env[name]}`));
+}
+
+export function describeHarness(resolved: ResolvedHarness): string {
+  const { harness } = resolved;
+  const installed = resolved.problem ? `not installed: ${resolved.problem}`
+    : `${harness.package} ${resolved.adapterVersion ?? "(version unreadable)"} at ${resolved.command}${resolved.adapterVersion === harness.qualified ? "" : `; checked with ${harness.qualified}`}`;
+  const set = callerEnv(harness);
+  return [
+    `${harness.name}: ${installed}`,
+    `  runs: ${resolved.native}`,
+    `  answers: ${harness.reject}`,
+    `  permissions: the agent's own mode decides what it asks; rig exec adds no policy. Read here: ${harness.env.join(", ")}${set.length ? ` (set: ${set.join(", ")})` : " (none set)"}`,
+  ].join("\n");
+}
+
+type ExecState = "requested" | "started" | "finished" | "interrupted";
+
+/** A run's registry entry. `pid` is whichever process owns the run now: the requesting CLI until the connector records itself. */
+export interface ExecRun {
+  runId: string;
+  qitemId: string;
+  harness: string;
+  agent: string[];
+  cwd: string;
+  self: string;
+  permissionsTo?: string;
+  loadSession?: string;
+  pollMs: number;
+  adapterVersion?: string;
+  native: string;
+  state: ExecState;
+  pid: number;
+  processStart?: string;
+  requestedAt: string;
+  startedAt?: string;
+  endedAt?: string;
+  outcome?: AcpOutcomeKind;
+  summary?: string;
+  usage?: AcpUsage;
+  settle?: string;
+  transcript: string;
+  connectorLog: string;
+}
+
+const now = () => new Date().toISOString();
+const runPath = (runId: string) => join(runsDir(), `${runId}.json`);
+/** A row's claims: one file per generation, `000001.json` upward, each published whole and never removed. */
+const rowDir = (qitemId: string) => join(runsDir(), "rows", qitemId.replace(/[^A-Za-z0-9._-]/g, "_"));
+
+/** A run's two transitions, each published once: `start` by the connector that runs it, or by whoever settles a run
+ *  that never started; `end` by whichever records how a started run ended. The slots are the run's state; the record
+ *  file is a copy written after each, for readers like `rig acp cancel`, and may lag a crash. */
+const slotPath = (runId: string, slot: "start" | "end") => join(runsDir(), `${runId}.${slot}.json`);
+type StartSlot = { by: "connector"; pid: number; processStart?: string; startedAt: string } | { by: "settled"; ended: Partial<ExecRun> };
+type EndSlot = { by: "connector" | "settled"; ended: Partial<ExecRun> };
+
+let drafts = 0;
+/** Publishes `content` at `target` whole, and only if nothing is there: a hard link fails if the target exists. */
+function publishOnce(target: string, content: unknown): boolean {
+  const draft = `${target}.${process.pid}.${++drafts}.tmp`;
+  writeFileSync(draft, `${JSON.stringify(content)}\n`);
+  try {
+    linkSync(draft, target);
+    return true;
+  } catch (error) {
+    if ((error as NodeJS.ErrnoException).code === "EEXIST") return false;
+    throw error;
+  } finally { rmSync(draft, { force: true }); }
+}
+
+export function readRun(runId: string): ExecRun | undefined {
+  const base = readJson(runPath(runId)) as ExecRun | undefined;
+  if (!base || typeof base.state !== "string") return undefined;
+  const start = readJson(slotPath(runId, "start")) as StartSlot | undefined;
+  if (start?.by === "settled") return { ...base, ...start.ended };
+  let run: ExecRun = start?.by === "connector"
+    ? { ...base, state: "started", pid: start.pid, processStart: start.processStart, startedAt: start.startedAt }
+    : base;
+  const end = readJson(slotPath(runId, "end")) as EndSlot | undefined;
+  if (end) run = { ...run, ...end.ended };
+  return run;
+}
+
+function writeRun(run: ExecRun): void {
+  const tmp = `${runPath(run.runId)}.${process.pid}.tmp`;
+  writeFileSync(tmp, `${JSON.stringify(run)}\n`);
+  renameSync(tmp, runPath(run.runId));
+}
+
+function topClaim(qitemId: string): { generation: number; runId?: string } | undefined {
+  let names: string[];
+  try { names = readdirSync(rowDir(qitemId)); } catch { return undefined; }
+  const generation = Math.max(0, ...names.filter((name) => /^\d{6}\.json$/.test(name)).map((name) => Number(name.slice(0, 6))));
+  if (generation === 0) return undefined;
+  const runId = readJson(join(rowDir(qitemId), `${String(generation).padStart(6, "0")}.json`))?.runId;
+  return { generation, runId: typeof runId === "string" ? runId : undefined };
+}
+
+export const rowRun = (qitemId: string): string | undefined => topClaim(qitemId)?.runId;
+
+/** Makes `runId` the row's run: when the row has none, or when its current run is exactly `previous`. The next
+ *  generation's file is published with a hard link, which fails if it exists, so of two callers that read the same
+ *  current run only one claims, and nothing is ever moved aside. `afterRead` lets a test act between the read and
+ *  the publish. */
+export function claimRow(qitemId: string, runId: string, previous?: string, afterRead?: () => void): boolean {
+  const dir = rowDir(qitemId);
+  mkdirSync(dir, { recursive: true });
+  const top = topClaim(qitemId);
+  if (top?.runId !== previous) return false;
+  afterRead?.();
+  return publishOnce(join(dir, `${String((top?.generation ?? 0) + 1).padStart(6, "0")}.json`), { runId, qitemId, replaces: previous ?? null });
+}
+
+type Liveness = "live" | "dead" | "unknown";
+
+/** A started run is live while its recorded process is still the one that recorded itself. A requested run is live
+ *  only while a process carries its connector marker: the record then names the requester, which isn't the connector
+ *  and may itself be waiting. */
+async function liveness(run: ExecRun): Promise<Liveness> {
+  if (run.endedAt) return "dead";
+  if (run.state !== "requested" && run.processStart !== undefined && processStart(run.pid) === run.processStart) return "live";
+  const found = await findRunProcesses(run.runId, CONNECTOR_MARKER).catch(() => undefined);
+  return found === undefined ? "unknown" : found.length > 0 ? "live" : "dead";
+}
+
+/** Ends `run` as `ended`, exclusively: a run that hasn't started by claiming its start slot, so its connector can no
+ *  longer start; a started run by claiming its end slot, so its connector can no longer record an outcome over this.
+ *  Only the winner writes the copy and the row note; a loser gets the run as whoever won left it. */
+export async function settleRun(run: ExecRun, ended: Partial<ExecRun>, note?: string, queue?: () => Promise<AcpQueue | undefined>): Promise<{ won: boolean; run: ExecRun }> {
+  const won = run.state === "requested"
+    ? publishOnce(slotPath(run.runId, "start"), { by: "settled", ended } satisfies StartSlot)
+    : publishOnce(slotPath(run.runId, "end"), { by: "settled", ended } satisfies EndSlot);
+  const after = readRun(run.runId) ?? run;
+  if (!won) return { won, run: after };
+  writeRun(after);
+  const q = note ? await queue?.().catch(() => undefined) : undefined;
+  await q?.update(run.qitemId, { note: note! }).catch(() => {});
+  return { won, run: after };
+}
+
+/** A run whose connector is gone without recording an outcome reads interrupted, outcome unknown, with one row note. */
+export async function reconcileRun(run: ExecRun, queue?: () => Promise<AcpQueue | undefined>, connectorExit?: string): Promise<ExecRun> {
+  if (run.endedAt || (await liveness(run)) !== "dead") return run;
+  const current = readRun(run.runId) ?? run;
+  if (current.endedAt || (current.state !== run.state && (await liveness(current)) !== "dead")) return current;
+  const left = await findRunProcesses(run.runId, RUN_MARKER).catch(() => undefined);
+  const agent = left === undefined ? "whether an agent process remains is unknown (the process table couldn't be read)"
+    : left.length === 0 ? "no process carrying the run's marker remains"
+      : `${left.length} process(es) carrying the run's marker still run (${left.map((p) => p.pid).join(", ")})`;
+  const summary = current.state === "requested"
+    ? `interrupted: its connector ${connectorExit ?? "isn't running"} and never recorded itself, so no turn was sent; ${agent}; its log is ${current.connectorLog}`
+    : `interrupted: its connector ended without recording an outcome, so the outcome is unknown; ${agent}`;
+  const settled = await settleRun(current, { state: "interrupted", outcome: "unknown", summary, endedAt: now() },
+    `rig exec run ${run.runId} ${summary}. Its transcript: ${run.transcript}`, queue);
+  return settled.run;
+}
+
+export function describeRun(run: ExecRun): string {
+  const state = run.state === "requested" ? "requested; its connector hasn't recorded itself yet"
+    : run.state === "started" ? `running (connector pid ${run.pid}, since ${run.startedAt})`
+      : run.state === "interrupted" ? run.summary ?? "interrupted; outcome unknown"
+        : `finished, outcome ${run.outcome}: ${run.summary}`;
+  return [
+    `${run.runId} for ${run.qitemId} on ${run.harness} (${run.agent[0]}${run.adapterVersion ? ` ${run.adapterVersion}` : ""}; runs ${run.native})`,
+    `  ${state}`,
+    run.state === "finished" ? `  ${describeUsage(run.usage)}` : undefined,
+    run.endedAt ? "  The outcome is the agent's report, not acceptance: the row stays open for its custodian to judge and close." : undefined,
+    run.settle ? `  After the turn: ${run.settle}` : undefined,
+    `  transcript ${run.transcript}; connector log ${run.connectorLog}`,
+  ].filter((line) => line !== undefined).join("\n");
+}
+
+/** Tests only: points where a test acts as a concurrent reconciler would. */
+export interface ConnectorHooks { afterGuard?: () => Promise<void>; afterStart?: () => Promise<void>; beforeEnd?: () => Promise<void> }
+
+/** The detached connector: claims the run's start, runs the turn as `rig acp run` does, then claims its end. */
+export async function runExecConnector(runId: string, queue: AcpQueue | undefined, hooks: ConnectorHooks = {}): Promise<number> {
+  const run = readRun(runId);
+  if (!run) { console.error(`No rig exec run ${runId} on this machine`); return 1; }
+  // Only the connector `rig exec run` started for this record runs it, once: anything else would send the row's task a
+  // second time. The record is left untouched.
+  if (run.state !== "requested" || process.env[CONNECTOR_MARKER] !== runId) {
+    console.error(`${runId} is ${run.state}${run.state === "requested" ? ", and this process wasn't started for it" : ""}; rig exec connector runs only the run rig exec run started it for`);
+    return 1;
+  }
+  delete process.env[CONNECTOR_MARKER]; // the agent doesn't inherit it; this process's own environment still shows it
+  // Before the run names this process, so a `rig exec cancel` that reads the record can't reach a default SIGINT.
+  let cancel!: () => void;
+  let detach!: () => void;
+  const cancelRequested = new Promise<void>((done) => { cancel = done; });
+  const detachRequested = new Promise<void>((done) => { detach = done; });
+  let interrupts = 0;
+  process.on("SIGINT", () => { interrupts += 1; if (interrupts === 1) cancel(); else detach(); });
+  process.on("SIGHUP", () => { /* no terminal owns this process; a hangup isn't a cancel */ });
+  await hooks.afterGuard?.();
+  // Claiming the start is what lets this process send the turn. A reconciler that judged the run dead may have claimed
+  // it first; then the run stays settled and nothing is sent.
+  if (!publishOnce(slotPath(runId, "start"), { by: "connector", pid: process.pid, processStart: processStart(process.pid), startedAt: now() } satisfies StartSlot)) {
+    console.error(`${runId} was settled before this connector started (${readRun(runId)?.summary ?? "no summary"}); no turn is sent`);
+    return 1;
+  }
+  const started = readRun(runId)!;
+  writeRun(started);
+  await hooks.afterStart?.();
+  /** Records the end only if nothing recorded one first; a reconciler's record stands, and the row hears what happened. */
+  const end = async (ended: Partial<ExecRun>): Promise<void> => {
+    await hooks.beforeEnd?.();
+    if (publishOnce(slotPath(runId, "end"), { by: "connector", ended } satisfies EndSlot)) { writeRun(readRun(runId)!); return; }
+    const recorded = readRun(runId);
+    console.error(`${runId} was recorded ${recorded?.state} before this connector finished; its own outcome was ${ended.outcome}`);
+    await queue?.update(run.qitemId, { note: `rig exec run ${runId}: its connector finished with outcome ${ended.outcome} (${ended.summary}) after the run had been recorded ${recorded?.state}; that record stands` }).catch(() => {});
+  };
+  if (!queue) {
+    await end({ state: "finished", outcome: "failed", summary: "the OpenRig daemon couldn't be reached, so no turn was sent", endedAt: now() });
+    return 1;
+  }
+  let settle: string | undefined;
+  try {
+    const outcome = await runAcpAssignment(queue, {
+      qitemId: run.qitemId, agent: run.agent, cwd: run.cwd, self: run.self, permissionsTo: run.permissionsTo,
+      loadSession: run.loadSession, runId, logPath: run.transcript, pollMs: run.pollMs,
+      progress: (text) => process.stdout.write(text), cancelRequested, detachRequested,
+      missingSession: HARNESSES[run.harness]?.missingSession, onSettled: (report) => { settle = report; },
+    });
+    await end({ state: "finished", outcome: outcome.kind, summary: outcome.summary, usage: outcome.usage, settle, endedAt: now() });
+    return outcome.kind === "completed" ? 0 : 2;
+  } catch (error) {
+    await end({ state: "finished", outcome: "failed", summary: (error as Error).message, settle, endedAt: now() });
+    return 1;
+  }
+}
+
+export interface ExecDeps {
+  status?: StatusDeps;
+  /** The queue as the calling seat (default: the daemon's). */
+  queue?: () => Promise<AcpQueue | undefined>;
+  /** The detached connector's command line (default: this CLI's own entry, `exec connector <runId>`). */
+  connectorCommand?: (runId: string) => string[];
+  pollMs?: number;
+  startWaitMs?: number;
+  /** Tests only: acts between the row-claim read and its publish, as a concurrent caller would. */
+  beforeClaim?: () => void;
+  /** Tests only: runs right after the CLI has told its caller about the run, as a process descheduled there would. */
+  afterAnnounce?: () => void;
+}
+
+const sleep = (ms: number) => new Promise((done) => setTimeout(done, ms));
+const exitFor = (run: ExecRun) => (run.outcome === "completed" ? 0 : 2);
+type Waiter = { left: boolean; leave: () => void };
+
+export function execCommand(deps: ExecDeps = {}): Command {
+  const cmd = new Command("exec").description("Run a queue row's task on a named harness (Claude Code or Codex over ACP) in a connector that outlives the command, with the outcome kept on the row");
+  const queueFor = deps.queue ?? (async () => {
+    const d = deps.status ?? { lifecycleDeps: realDeps(), clientFactory: (url: string) => new DaemonClient(url) };
+    const status = await getDaemonStatus(d.lifecycleDeps);
+    return daemonStatusGuard(status) ? daemonAcpQueue(d.clientFactory(getDaemonUrl(status))) : undefined;
+  });
+  const connectorCommand = deps.connectorCommand
+    ?? ((runId: string) => [process.execPath, fileURLToPath(new URL("../index.js", import.meta.url)), "exec", "connector", runId]);
+  const pollMs = deps.pollMs ?? 1000;
+
+  /** The wait's own Ctrl-C, armed before the CLI tells its caller anything about the run, so a Ctrl-C from then on
+   *  only stops the wait: armed later, one landing just after the announcement met the default and killed the CLI. The
+   *  connector is in its own session either way. */
+  const armWait = (wait: boolean | undefined): Waiter | undefined => {
+    if (!wait) return undefined;
+    const waiter: Waiter = { left: false, leave: () => { waiter.left = true; } };
+    process.on("SIGINT", waiter.leave);
+    return waiter;
+  };
+  /** Polls the run file until it ends, or until the waiter leaves. */
+  const waitForEnd = async (runId: string, waiter: Waiter): Promise<ExecRun | "left"> => {
+    for (let polls = 1; ; polls++) {
+      const run = readRun(runId);
+      if (!run) throw new Error(`the record of ${runId} is gone`);
+      if (run.endedAt) return run;
+      if (waiter.left) return "left";
+      if (polls % 5 === 0) { const checked = await reconcileRun(run, queueFor); if (checked.endedAt) return checked; }
+      await sleep(pollMs);
+    }
+  };
+  const finish = (result: ExecRun | "left", runId: string) => {
+    if (result === "left") {
+      console.log(`\n[rig exec] stopped waiting; ${runId} keeps running. rig exec status ${runId} shows it; rig exec cancel ${runId} cancels it.`);
+      process.exitCode = 3;
+      return;
+    }
+    console.log(describeRun(result));
+    process.exitCode = exitFor(result);
+  };
+  /** Every path that holds the row's run ends here: a run that has ended exits by its outcome (0 completed, 2 any
+   *  other), --wait waits for one still going (3 if the waiter leaves), and without --wait a run still going exits 0. */
+  const report = async (run: ExecRun, waiter: Waiter | undefined) => {
+    try {
+      if (run.endedAt) process.exitCode = exitFor(run);
+      else if (waiter) finish(await waitForEnd(run.runId, waiter), run.runId);
+    } finally { if (waiter) process.removeListener("SIGINT", waiter.leave); }
+  };
+
+  cmd
+    .command("harnesses")
+    .description("List the named harnesses: the adapter and native versions installed here, and how each one differs")
+    .option("--json", "JSON output")
+    .action((opts: { json?: boolean }) => {
+      const all = Object.values(HARNESSES).map((harness) => resolveHarness(harness));
+      if (opts.json) {
+        console.log(JSON.stringify(all.map((r) => ({
+          name: r.harness.name, package: r.harness.package, checkedWith: r.harness.qualified, installed: r.adapterVersion ?? null,
+          command: r.command ?? null, runs: r.native, answers: r.harness.reject, reads: r.harness.env, set: callerEnv(r.harness), problem: r.problem ?? null,
+        })), null, 2));
+        return;
+      }
+      console.log(all.map(describeHarness).join("\n\n"));
+    });
+
+  cmd
+    .command("run <qitemId>")
+    .description("Start one row's task as a single turn on a named harness, in a connector that outlives this command")
+    .requiredOption("--harness <name>", `The harness: ${Object.keys(HARNESSES).join(" or ")} (rig exec harnesses)`)
+    .option("--cwd <dir>", "The agent session's working directory", process.cwd())
+    .option("--permissions-to <session>", "The seat that answers the agent's permission requests (default: this seat)")
+    .option("--load-session <id>", "Continue an earlier session; nothing new starts when it can't be loaded")
+    .option("--wait", "Wait here for the outcome; leaving the wait (Ctrl-C) never cancels the run")
+    .option("--plan", "Show what would run, with its versions, and start nothing")
+    .option("--again", "Start another turn on a row whose earlier run has ended")
+    .option("--poll-seconds <n>", "How often the connector looks for a permission answer", "2")
+    .addHelpText("after", `
+A connector owns the run and outlives this command: closing this terminal, or leaving --wait, doesn't stop it.
+It runs one turn of the row's task on the harness's ACP adapter, as rig acp run does, and records the binding,
+progress, permission requests and the outcome on the row. The outcome is the agent's report: a completed turn isn't
+accepted work, so the row stays open for you to judge and close.
+
+One run per row: running it again returns the row's run, still going or ended, instead of sending the task twice.
+--again starts another turn once that run has ended. A run whose connector died without recording an outcome reads
+interrupted, outcome unknown.
+
+The agent keeps its own permission mode, and rig exec adds no policy: your environment passes through (rig exec
+harnesses lists what each harness reads). A permission request becomes a row to --permissions-to.
+
+  rig exec status <runId|qitemId>   state, outcome, usage and cost as the agent reported them, and what still runs
+  rig exec cancel <runId>           asks the agent to cancel its turn; the acknowledgement isn't settlement
+
+Guide: rig context get reference/acp-agents.md`)
+    .action(async (qitemId: string, opts: { harness: string; cwd: string; permissionsTo?: string; loadSession?: string; wait?: boolean; plan?: boolean; again?: boolean; pollSeconds: string }) => {
+      const harness = HARNESSES[opts.harness];
+      if (!harness) { console.error(`No harness ${opts.harness}; the named ones are ${Object.keys(HARNESSES).join(", ")}`); process.exitCode = 1; return; }
+      const resolved = resolveHarness(harness);
+      if (opts.plan) {
+        console.log(`${describeHarness(resolved)}\n  would run: ${resolved.command ?? harness.bin} for ${qitemId} in ${resolve(opts.cwd)}${opts.loadSession ? `, continuing session ${opts.loadSession}` : ", in a new session"}; permission requests to ${opts.permissionsTo ?? "this seat"}`);
+        if (resolved.problem) process.exitCode = 1;
+        return;
+      }
+      const seat = readOpenRigEnv("OPENRIG_SESSION_NAME", "RIGGED_SESSION_NAME");
+      if (!seat) { console.error("rig exec run acts as the seat that owns the row, but OPENRIG_SESSION_NAME isn't set"); process.exitCode = 1; return; }
+      if (resolved.problem) { console.error(resolved.problem); process.exitCode = 1; return; }
+      const queue = await queueFor();
+      if (!queue) { process.exitCode = 1; return; }
+      mkdirSync(runsDir(), { recursive: true });
+
+      const previousId = rowRun(qitemId);
+      const previous = previousId ? readRun(previousId) : undefined;
+      if (previous) {
+        const run = await reconcileRun(previous, async () => queue);
+        if (opts.again && !run.endedAt) {
+          console.error(`${run.runId} is still ${run.state} on ${qitemId}; cancel it (rig exec cancel ${run.runId}) or let it end before --again`);
+          process.exitCode = 1;
+          return;
+        }
+        if (!opts.again) {
+          const waiter = armWait(opts.wait);
+          console.log(`${qitemId} already has a run, so the task isn't sent twice:\n${describeRun(run)}\nOnce it has ended, --again starts another turn on this row.`);
+          await report(run, waiter);
+          return;
+        }
+      }
+
+      const row = await queue.show(qitemId).catch((error: Error) => error);
+      if (row instanceof Error) { console.error(`rig exec run: ${row.message}`); process.exitCode = 1; return; }
+      if (row.destinationSession !== seat) { console.error(`${qitemId} belongs to ${row.destinationSession}; run rig exec from that seat, or hand the row to ${seat} first`); process.exitCode = 1; return; }
+      if (row.state !== "pending" && row.state !== "in-progress") { console.error(`${qitemId} is ${row.state}; rig exec runs an in-progress or pending row`); process.exitCode = 1; return; }
+
+      const runId = newId("exec", 4);
+      const run: ExecRun = {
+        runId, qitemId, harness: harness.name, agent: [resolved.command!], cwd: resolve(opts.cwd), self: seat,
+        permissionsTo: opts.permissionsTo, loadSession: opts.loadSession, pollMs: Math.max(1, Number(opts.pollSeconds) || 2) * 1000,
+        adapterVersion: resolved.adapterVersion, native: resolved.native, state: "requested", pid: process.pid,
+        processStart: processStart(process.pid), requestedAt: now(),
+        transcript: join(runsDir(), `${runId}.jsonl`), connectorLog: join(runsDir(), `${runId}.connector.log`),
+      };
+      writeRun(run);
+      if (!claimRow(qitemId, runId, previousId, deps.beforeClaim)) {
+        rmSync(runPath(runId), { force: true });
+        const winnerId = rowRun(qitemId);
+        const winner = winnerId ? readRun(winnerId) : undefined;
+        const waiter = winner ? armWait(opts.wait) : undefined;
+        console.log(`${qitemId} was started by another caller as ${winnerId ?? "another run"}, so the task isn't sent twice${winner ? `:\n${describeRun(winner)}` : ""}`);
+        if (winner) await report(winner, waiter);
+        else process.exitCode = 1;
+        return;
+      }
+
+      const log = openSync(run.connectorLog, "a");
+      const [file, ...args] = connectorCommand(runId);
+      let spawnError: Error | undefined;
+      let connectorExit: string | undefined;
+      try {
+        const child = spawn(file!, args, { detached: true, stdio: ["ignore", log, log], env: { ...process.env, [CONNECTOR_MARKER]: runId } });
+        child.once("error", (error) => { spawnError = error; });
+        child.once("exit", (code, signal) => { connectorExit = `exited (code ${code ?? "none"}, signal ${signal ?? "none"})`; });
+        child.unref();
+      } finally { closeSync(log); }
+      const waiter = armWait(opts.wait);
+      let current: ExecRun = run;
+      for (const deadline = Date.now() + (deps.startWaitMs ?? 10_000); Date.now() < deadline; await sleep(50)) {
+        current = readRun(runId) ?? current;
+        if (current.state !== "requested" || spawnError || connectorExit || waiter?.left) break;
+      }
+      if (spawnError && current.state === "requested") {
+        current = (await settleRun(current, { state: "finished", outcome: "failed", summary: `the connector couldn't be started: ${spawnError.message}`, endedAt: now() })).run;
+      } else if (connectorExit && current.state === "requested") {
+        current = await reconcileRun(current, async () => queue, connectorExit);
+      }
+      console.log(`[rig exec] ${runId} for ${qitemId} on ${harness.name}: ${current.endedAt ? `ended, outcome ${current.outcome}: ${current.summary}` : current.state === "started" ? "started" : "requested; its connector hasn't recorded itself yet"}`);
+      console.log(`  rig exec status ${runId} shows it; rig exec cancel ${runId} cancels it.`);
+      deps.afterAnnounce?.();
+      await report(current, waiter);
+    });
+
+  cmd
+    .command("status <runIdOrQitemId>")
+    .description("Show a run's state, the outcome as the agent reported it, usage and cost, and what still runs")
+    .option("--json", "JSON output")
+    .action(async (target: string, opts: { json?: boolean }) => {
+      const runId = readRun(target) ? target : rowRun(target);
+      const run = runId ? readRun(runId) : undefined;
+      if (!run) { console.error(`No rig exec run for ${target} on this machine`); process.exitCode = 1; return; }
+      const current = await reconcileRun(run, queueFor);
+      console.log(opts.json ? JSON.stringify(current, null, 2) : describeRun(current));
+    });
+
+  cmd
+    .command("cancel <runId>")
+    .description("Ask a running rig exec run's agent to cancel its turn")
+    .action((runId: string) => {
+      const run = readRun(runId);
+      if (!run || run.endedAt || run.state !== "started" || run.processStart === undefined || processStart(run.pid) !== run.processStart) {
+        console.error(`No running rig exec run ${runId}`);
+        process.exitCode = 1;
+        return;
+      }
+      process.kill(run.pid, "SIGINT");
+      console.log(`Cancel requested for ${runId}; its row ${run.qitemId} records what the agent acknowledges. An acknowledgement isn't settlement: rig exec status ${runId} shows the outcome and what still runs.`);
+    });
+
+  cmd
+    .command("connector <runId>", { hidden: true })
+    .description("The detached connector that owns a run (started by rig exec run)")
+    .action(async (runId: string) => { process.exitCode = await runExecConnector(runId, await queueFor()); });
+
+  return cmd;
+}

packages/cli/src/commands/gateway.ts

@@ -175,7 +175,7 @@ export function gatewayCommand(deps: GatewayCommandDeps = {}): Command {
     .requiredOption("--display-name <name>", "Human-readable display name")
     .option(
       "--binding <kind:connectorRef:secretsRef:role[:handle=<id>]>",
-      "Connector binding (repeatable); leave it out for a person with no connector, who's reached in the Feed and Steering only. With bindings, role primary|secondary, EXACTLY ONE primary. connectorRef and secretsRef are config/vault labels (e.g. for Slack: slack:openrig-personal:file:<env-path>:primary:handle=<id>; tokens come from 'rig slack setup --secrets-env-file', never inline). Optional handle=<platform id> makes the binding inbound-resolvable (unique per kind); omit it for outbound-only.",
+      "Connector binding (repeatable); leave it out for a person with no connector, who's reached in the Feed and Steering only. With bindings, role primary|secondary, EXACTLY ONE primary. connectorRef and secretsRef are config/vault labels (e.g. for Slack: slack:openrig-personal:file:<env-path>:primary:handle=<id>; tokens come from 'rig slack setup --secrets-env-file', never inline; or file:<path> for relay: relay:<connectorRef>:file:<absolute-env-file>:primary:handle=<safe-alias>). Optional handle=<platform id> makes the binding inbound-resolvable (unique per kind for Slack, unique per connectorRef+handle for relay); omit it for outbound-only.",
       (v: string, acc: string[] = []) => { acc.push(v); return acc; },
     )
     .requiredOption("--delivery-class <A|B|C|D>", "Notification loudness class (the notifications register selection)")
@@ -253,16 +253,24 @@ export function gatewayCommand(deps: GatewayCommandDeps = {}): Command {
     .addHelpText("after", `
 Binding specification details:
   Format: <kind:connectorRef:secretsRef:role[:handle=<id>]>
-  - kind: Connector kind (e.g. "slack")
-  - connectorRef: Configuration label identifying the connector instance (e.g. "openrig-personal" or "main")
-  - secretsRef: Pointer or label identifying where secrets are resolved (e.g. "file:/path/to/slack.env" or "vault://slack/user").
-    NOTE: Never pass credentials or tokens inline here; credentials (like SLACK_BOT_TOKEN) are loaded from the connector's secrets env file (configured via 'rig slack setup --secrets-env-file').
+  - kind: Connector kind: "slack" or "relay"
+  - connectorRef: Configuration label identifying the connector instance (e.g. "openrig-personal", "main", or "local-bridge")
+  - secretsRef: Pointer or label identifying where secrets are resolved:
+    * For Slack: "file:/path/to/slack.env" or "vault://slack/user". Credentials (like SLACK_BOT_TOKEN) are loaded from the connector's secrets env file (configured via 'rig slack setup --secrets-env-file').
+    * For Relay: "file:<absolute-env-file>" pointing to a chmod 600 file holding OPENRIG_RELAY_URL (HTTP loopback only, e.g. http://127.0.0.1:8088), OPENRIG_RELAY_OUTBOUND_TOKEN, and OPENRIG_RELAY_INBOUND_TOKEN (distinct non-empty values; no environment fallback).
+    NOTE: Never pass credentials or tokens inline here.
   - role: "primary" (exactly one binding must be primary) or "secondary"
-  - handle: Optional platform user identifier (e.g. "handle=U12345678") for inbound routing
+  - handle: Optional platform user identifier (e.g. "handle=U12345678" for Slack, or "handle=<safe-alias>" matching /^[A-Za-z0-9._-]+$/ for relay) for inbound routing. For relay, handles are unique per connectorRef.
 
-Example:
+Examples:
+  # Slack binding
   rig gateway human add alice --display-name "Alice" \\
     --binding "slack:openrig-personal:file:/etc/openrig/slack.env:primary:handle=U12345678" \\
+    --delivery-class B
+
+  # Relay binding (loopback bridge adapter)
+  rig gateway human add bob --display-name "Bob" \\
+    --binding "relay:local-bridge:file:/etc/openrig/relay.env:primary:handle=bob-ops" \\
     --delivery-class B`);
 
   // ── S12 (OPR.0.5.5.12): the fragment lifecycle beyond add. Every verb operates through
@@ -559,9 +567,14 @@ Examples:
 
   human
     .command("set <entityId> <field> <value>")
-    .description("Edit one field through the verb (same validation as add; re-projection immediate). Fields: display-name, delivery-class, away, availability (available|focus|away|off, or unset), slack-subscription (the person's Slack subscription: every-ask, the default, posts every ask to Slack too; urgent-only posts only asks created with --alert, the rest wait on Steering and the TUI; or unset), binding.<n>; and, kept in the participant side file: organization (an id, or unset), groups (comma-separated ids, or unset), policy (text, @<path> to read a file, or unset)")
+    .description("Edit one field through the verb (same validation as add; re-projection immediate). Fields: display-name, delivery-class, away, availability (available|focus|away|off, or unset), slack-subscription (the person's Slack subscription: every-ask, the default, posts every ask to Slack too; urgent-only posts only asks created with --alert, the rest wait on Steering and the TUI; or unset), binding.<n> (a full kind:connectorRef:secretsRef:role[:handle=<id>] spec, or unset to remove that binding); and, kept in the participant side file: organization (an id, or unset), groups (comma-separated ids, or unset), policy (text, @<path> to read a file, or unset)")
     .option("--reason <reason>", "Reason recorded with connector binding changes", "update human binding")
     .option("--actor <actor>", "Named operator when outside a managed seat")
+    .addHelpText("after", `
+Removing a connector binding (binding.<n> unset):
+  - The one-primary rule stays: no replacement primary is chosen silently. If removing a primary binding would leave other bindings, the command refuses. To replace the primary in place, set its slot to the new primary spec; to promote an existing secondary binding, first save that binding's full spec (kind:connectorRef:secretsRef:role[:handle=<id>]) from rig gateway human show <entityId> --json, then unset that secondary slot, then find the primary's current slot in remainingBindings (or rig gateway human show <entityId> --json, since the primary's slot shifts down when a lower-index secondary is removed) and set it to the saved spec with role primary (promotion replaces the former primary binding; its connector reference and inbound handle are not retained as a secondary); to remove every binding, unset the secondary bindings first using their current slot indexes from remainingBindings or rig gateway human show <entityId>, then finish at binding.0 (once all secondaries are gone, the primary is at binding.0).
+  - Remaining bindings are printed with their shifted binding.<n> slot indexes (the primary's slot shifts down too when a lower-index secondary is removed). Because indexes shift after removal, do not blindly retry 'binding.0 unset'; check remainingBindings, the receipt, or 'rig gateway human show <entityId>' before retrying.
+  - Removing a binding keeps the person, their preferences, their other bindings and their open questions. Removing the last binding makes the current delivery episode inactive, so new missing-post findings stop (an already-open recovery item for an undelivered notice may remain); the current delivery-outcome display may disappear too, while historical receipt notes remain. It does not answer, close or cancel open questions, cancel or reroute already-prepared deliveries, or retract anything a bridge has already accepted.`)
     .action(async (entityId: string, field: string, value: string, opts: { reason: string; actor?: string }) => {
       const registry = await import("@openrig/daemon/gateway-human-registry");
       if (field === "organization" || field === "groups" || field === "policy") {
@@ -594,7 +607,19 @@ Examples:
               return { value: res, after: { digest }, effect: digest === prior ? "no-op" : "applied" };
             },
           });
-          console.log(JSON.stringify({ ok: true, entityId, field, path: result.value.path, receipt: result.receipt }));
+          console.log(JSON.stringify({
+            ok: true,
+            entityId,
+            field,
+            path: result.value.path,
+            receipt: result.receipt,
+            ...(value === "unset"
+              ? {
+                  remainingBindings: result.value.remainingBindings ?? [],
+                  note: result.value.note,
+                }
+              : {}),
+          }));
         } catch (error) { console.error(`refused: ${(error as Error).message}`); process.exitCode = 1; }
         return;
       }
@@ -732,17 +757,16 @@ Examples:
     .option("--json", "Complete records as JSON, including organizations and groups")
     .action(async (opts: { organization?: string; json?: boolean }) => {
       const registry = await import("@openrig/daemon/gateway-human-registry");
+      if (opts.json) {
+        const result = registry.participantsJsonView(undefined, opts.organization);
+        if (!result.ok) { console.error(`refused: ${result.error}`); process.exitCode = 1; return; }
+        console.log(JSON.stringify(result));
+        return;
+      }
       const view = registry.participantsView();
       if (!view.ok) { console.error(`refused: ${view.error}`); process.exitCode = 1; return; }
       const shown = opts.organization ? view.participants.filter((p) => p.organization?.id === opts.organization) : view.participants;
       const defaultPerson = view.default.person ? { id: view.default.person.entityId, source: view.default.source } : { id: null, reason: view.default.source === "none" ? view.default.reason : null };
-      if (opts.json) {
-        const side = registry.readParticipantsFile();
-        console.log(JSON.stringify({ ok: true, participants: shown, default: defaultPerson,
-          organizations: side.ok ? side.file.organizations : null, groups: side.ok ? side.file.groups : null,
-          warnings: view.warnings, participantFile: view.sideFile }));
-        return;
-      }
       if (!shown.length) console.log(opts.organization ? `no participants in organization "${opts.organization}"` : "no participants yet: rig gateway human add … registers a person, rig gateway channel add … a shared channel");
       for (const p of shown) {
         console.log(`${p.id}  ${p.kind}  "${p.displayName}"  ${p.address}${p.channel ? `  channel ${p.channel.channelId}` : ""}`);

packages/cli/src/commands/host.ts

@@ -24,8 +24,9 @@ import { existsSync, mkdirSync, writeFileSync, rmSync } from "node:fs";
 import { join } from "node:path";
 import { hostname as osHostname, userInfo } from "node:os";
 import { connect } from "node:net";
+import { execFile } from "node:child_process";
 import { getOpenRigHome } from "../openrig-compat.js";
-import { addHostEntry, updateHostEntry, removeHostEntry, defaultHostRegistryPath, loadHostRegistry, validateHostRegistry, hostDisplayTarget, resolveHost, resolveRemoteBearer, bearerAuthHeaders, type HostEntry, type HttpHostEntry } from "../host-registry.js";
+import { addHostEntry, updateHostEntry, removeHostEntry, defaultHostRegistryPath, loadHostRegistry, validateHostRegistry, hostDisplayTarget, resolveHost, resolveRemoteBearer, bearerAuthHeaders, type HostEntry, type HttpHostEntry, type SshHostEntry } from "../host-registry.js";
 import { runCrossHostCommand, type CrossHostResult } from "../cross-host-executor.js";
 import { recordHostObservation, loadHostBindings, describeBindingConflict, defaultHostBindingsPath, type HostObservationOutcome } from "../host-bindings.js";
 import { DaemonClient } from "../client.js";
@@ -167,15 +168,35 @@ function defaultDoctorDeps(): DoctorDeps {
   };
 }
 
+/** Where ssh would connect for this host, from `ssh -G`, which reads ssh config without connecting.
+ *  Null when ssh can't resolve the target or a ProxyJump/ProxyCommand sits in between, since no
+ *  direct dial stands for that transport. */
+export function resolveSshEndpoint(host: SshHostEntry, timeoutMs = 1500): Promise<{ hostname: string; port: number } | null> {
+  const args = [...(host.user ? ["-l", host.user] : []), "-G", host.target];
+  return new Promise((resolve) => {
+    execFile("ssh", args, { timeout: timeoutMs }, (error, stdout) => {
+      if (error) return resolve(null);
+      const conf = new Map(stdout.split("\n").map((line) => {
+        const i = line.indexOf(" ");
+        return [line.slice(0, i).toLowerCase(), line.slice(i + 1).trim()] as const;
+      }));
+      const proxied = ["proxyjump", "proxycommand"].some((k) => (conf.get(k) ?? "none") !== "none");
+      const hostname = conf.get("hostname");
+      const port = Number(conf.get("port"));
+      resolve(hostname && Number.isInteger(port) && !proxied ? { hostname, port } : null);
+    });
+  });
+}
+
 /** FR-1 doctor legs — each failing step maps to a DISTINCT actionable error:
  *  "SSH works but daemon is down" ≠ "daemon works but registry is wrong" ≠
  *  "remote rig binary missing/old". */
 /** OPR.0.4.6.MH1 FR-3 — the coarse ls status word: a BOUNDED tcp dial
- *  against the host's transport endpoint (ssh: target:22; http: the
- *  URL's host:port). Coarse by design (the kubectl STATUS / docker
- *  context ls precedent): reachable | unreachable | unknown — never a
- *  hang (hard timeout), never a guess (dial failure = unreachable,
- *  probe error/unparseable endpoint = unknown). */
+ *  against the host's transport endpoint (ssh: the host and port ssh itself
+ *  resolves, see resolveSshEndpoint; http: the URL's host:port). Coarse by
+ *  design (the kubectl STATUS / docker context ls precedent): reachable |
+ *  unreachable | unknown — never a hang (hard timeout), never a guess (dial
+ *  failure = unreachable, probe error/unparseable endpoint = unknown). */
 export async function probeHostStatus(
   host: HostEntry,
   tcpProbe: DoctorDeps["tcpProbe"],
@@ -185,8 +206,9 @@ export async function probeHostStatus(
     let target: string;
     let port: number;
     if (host.transport === "ssh") {
-      target = host.target;
-      port = 22;
+      const endpoint = await resolveSshEndpoint(host, timeoutMs);
+      if (!endpoint) return "unknown";
+      ({ hostname: target, port } = endpoint);
     } else {
       const u = new URL(host.url);
       target = u.hostname.replace(/^\[|\]$/g, "");
@@ -761,9 +783,9 @@ export function hostCommand(doctorDepsOverride?: DoctorDeps): Command {
   cmd
     .command("pair")
     .description("Pair with a remote host from one pasted address — one approval on the target, done (FR-6)")
-    .argument("<url>", "The target daemon's address (http[s]://host:port, or host:port)")
+    .argument("<url>", "The target daemon's address (http[s]://host:port, or host:port); that daemon must run with OPENRIG_AUTH_BEARER_TOKEN set")
     .option("--id <id>", "Registry id for the new host (default: derived from the hostname)")
-    .option("--timeout <seconds>", "How long to wait for the target-side approval", "600")
+    .option("--timeout <seconds>", "How long to wait for the target-side approval; the request stays open that long on the target (a target older than this CLI keeps it 10 minutes)", "600")
     .option("--human <address>", "Registered @external approval recipient on the target (required when several exist)")
     .option("--json", "JSON output")
     .action(async (rawUrl: string, opts: { id?: string; timeout?: string; json?: boolean; human?: string }) => {
@@ -835,11 +857,17 @@ export function hostCommand(doctorDepsOverride?: DoctorDeps): Command {
         try { return JSON.parse(s) as Record<string, unknown>; } catch { return {}; }
       };
 
+      // The wait is BOUNDED: a non-finite or overflowing --timeout is clamped to
+      // MAX_PAIR_TIMEOUT_MS, so a target that stops answering can never spin the
+      // retry loop forever. Values that work today keep their wait. The target
+      // keeps the request open as long as this side waits, and its approval says so.
+      const timeoutMs = resolvePairTimeoutMs(opts.timeout);
       let issued: { pairId?: string; code?: string; approvalQitemId?: string; error?: string; message?: string };
       try {
         const res = await httpPost(`${targetBase}/api/hosts/pair-request`, {
           requester: `${userInfo().username}@${osHostname()}`,
           ...(opts.human ? { human: opts.human } : {}),
+          timeoutMs,
         });
         issued = parseJson(res.body) as typeof issued;
         if (res.status !== 200 || !issued.pairId || !issued.code) {
@@ -859,25 +887,25 @@ export function hostCommand(doctorDepsOverride?: DoctorDeps): Command {
         console.log(`On the target: rig queue update ${issued.approvalQitemId ?? "<qitem-id>"} --state done --closure-reason no-follow-on`);
       }
 
-      // The wait is BOUNDED: a non-finite or overflowing --timeout is clamped to
-      // MAX_PAIR_TIMEOUT_MS, so a target that stops answering can never spin the
-      // retry loop forever. Values that work today keep their wait. A note names
-      // the bound actually used so a clamped value is not silently absorbed.
-      const timeoutMs = resolvePairTimeoutMs(opts.timeout);
+      // A note names the bound actually used so a clamped value is not silently absorbed.
       if (!opts.json && timeoutMs === MAX_PAIR_TIMEOUT_MS) {
         console.log(`Waiting up to ~${(MAX_PAIR_TIMEOUT_MS / 86_400_000).toFixed(1)} days for approval on the target (--timeout capped at the maximum).`);
       }
+      // Poll at once, then every 2 s, never sleeping past the deadline: the target keeps the request open only as long
+      // as this side waits, so a short --timeout still sees an approval that's already there.
       const deadline = Date.now() + timeoutMs;
       let outcome: { status?: string; token?: string } = {};
-      while (Date.now() < deadline) {
-        await new Promise((r) => setTimeout(r, 2000));
+      for (;;) {
         try {
           const res = await deps.httpGet(`${targetBase}/api/hosts/pair-request/${issued.pairId}`);
           outcome = parseJson(res.body) as typeof outcome;
+          if (outcome.status && outcome.status !== "pending") break;
         } catch {
-          continue; // transient poll failure: keep waiting until the deadline
+          // transient poll failure: keep waiting until the deadline
         }
-        if (outcome.status && outcome.status !== "pending") break;
+        const remaining = deadline - Date.now();
+        if (remaining <= 0) break;
+        await new Promise((r) => setTimeout(r, Math.min(2000, remaining)));
       }
 
       if (outcome.status !== "approved" || !outcome.token) {

packages/cli/src/commands/launch.ts

@@ -35,7 +35,7 @@ type LaunchResponse = {
 // operator must act). These are NOT successful launches — never print "Launched" for
 // them; exit non-zero. Mirrors the daemon's NON_RUNNING_LAUNCH_STATUSES.
 const NON_RUNNING_LAUNCH_STATUSES = new Set(["awaiting-decision", "attention_required", "failed", "left_attached"]);
-function launchStatusRunning(status: string): boolean {
+export function launchStatusRunning(status: string): boolean {
   return !NON_RUNNING_LAUNCH_STATUSES.has(status);
 }
 

packages/cli/src/commands/mcp.ts

@@ -2,7 +2,7 @@ import { Command } from "commander";
 import { serveStdio } from "@modelcontextprotocol/server/stdio";
 import { DaemonClient } from "../client.js";
 import { getDaemonStatus, getDaemonUrl, type LifecycleDeps , daemonStatusGuard} from "../daemon-lifecycle.js";
-import { createMcpServer } from "../mcp-server.js";
+import { createMcpServer, dropUnfilledSeat, unfilledSeatText, type UnfilledSeat } from "../mcp-server.js";
 import { addGrant, McpHttpConfigError, mcpTokenFilePath, readGrantServiceHeader, readGrants, revokeGrant, startMcpHttpServer } from "../mcp-http.js";
 import { readOpenRigEnv } from "../openrig-compat.js";
 import { realDeps } from "./daemon.js";
@@ -25,12 +25,20 @@ export function mcpCommand(depsOverride?: StatusDeps): Command {
     .option("--bind <host:port>", "With --http: where to listen (default 127.0.0.1:7437). Beyond loopback needs a grant (rig mcp token add)")
     .option("--grant-service <url>", "With --http: check each request's key with the grant service at this URL instead of the token file")
     .option("--grant-service-header-file <path>", "With --grant-service: a file holding one 'Name: value' header to send with each grant request")
+    .option("--task-destination <seat>", "With --http --grant-service: operator-configured local team seat for external requests that omit destination")
     .option("--oauth-resource <url>", "With --grant-service and --oauth-issuer: the public https URL clients reach this server as (its path ends in /mcp), for OAuth discovery")
     .option("--oauth-issuer <url>", "With --oauth-resource: the OAuth authorization server that issues the keys --grant-service checks")
+    .option("--front-door", "Serve a person's agent in their own app: it hands the team work and brings the team's questions to the person, instead of working its inbox as a member")
     .addHelpText("after", `
 Both MCP protocol revisions are served, over stdio and --http alike: clients on 2026-07-28 (which may start with
 server/discover) and clients that open with 2025's initialize handshake.
 
+An empty OPENRIG_SESSION_NAME, or one an app passed on unfilled (such as \${user_config.seat}), means no seat: the
+server acts as none, never as a seat named after the placeholder or one inherited as RIGGED_SESSION_NAME. With no
+seat, messages, new tasks and chat still go through as an unknown sender, each with a note on what is known (why there's
+no seat, and how the MCP client describes itself); a task is followed by its qitemId, since there's no return address.
+Claiming, updating, handing off or deciding a task needs a seat.
+
 With --http and no token file, the server stays on loopback and acts as OPENRIG_SESSION_NAME, declared and not
 authenticated, like stdio. Once a grant exists (rig mcp token add), every request needs its bearer credential, loopback
 included, and each caller acts as its grant's seat within its rigs and roots. Add a grant before reaching the port
@@ -43,14 +51,23 @@ for this daemon's instance (its self id); the token file isn't read. A caller ac
 must already exist here (created is enough; it needn't have been launched), and reaches only what the grant names. If
 the grant service can't answer, the request is refused.
 
+With --task-destination <member@rig>, external callers can send their first task without knowing a seat address.
+Supply the intended receiver from trusted operator configuration for this server, never from a caller's input.
+On every omitted destination, the receiver must be exactly one existing local node and within that request's grant.
+Missing, ambiguous or forbidden receivers cause no write. This option doesn't grant access or select a receiver from
+topology, names or task history. Explicit destinations keep their existing behavior; local and member calls still
+require destination. A server with multiple teams needs a receiver valid for the calling grant, or explicit destinations.
+
 With --grant-service, --oauth-resource and --oauth-issuer (both or neither) add OAuth discovery (RFC 9728) for clients
 that sign in with OAuth. The resource is the public URL this server is reached as, and the issuer is the authorization
 server that issues the keys --grant-service checks. GET /.well-known/oauth-protected-resource<resource path> then answers
 with both, exactly as given, without a credential or a call to the grant service, and every 401 points there. Both must
 be https with no userinfo, query or fragment, written as a URL writes them (a lowercase host, no default port). /mcp
 stays the endpoint; a proxy in front forwards the metadata path unchanged.`)
-    .action(async (opts: { port?: string; http?: boolean; bind?: string; grantService?: string; grantServiceHeaderFile?: string; oauthResource?: string; oauthIssuer?: string }) => {
+    .action(async (opts: { port?: string; http?: boolean; bind?: string; grantService?: string; grantServiceHeaderFile?: string; taskDestination?: string; oauthResource?: string; oauthIssuer?: string; frontDoor?: boolean }) => {
       const deps = getDepsF();
+      const unfilledSeat = dropUnfilledSeat();
+      if (unfilledSeat) console.error(`rig mcp serve: ${unfilledSeatText(unfilledSeat)}; serving with no seat until it names this app's address on the team.`);
 
       let daemonUrl: string;
       if (opts.port) {
@@ -80,6 +97,11 @@ stays the endpoint; a proxy in front forwards the metadata path unchanged.`)
         process.exitCode = 1;
         return;
       }
+      if (opts.taskDestination !== undefined && !(opts.http && opts.grantService)) {
+        console.error("--task-destination needs --http --grant-service");
+        process.exitCode = 1;
+        return;
+      }
       if ((opts.oauthResource === undefined) !== (opts.oauthIssuer === undefined)) {
         console.error("--oauth-resource and --oauth-issuer go together: give both, or neither");
         process.exitCode = 1;
@@ -92,7 +114,7 @@ stays the endpoint; a proxy in front forwards the metadata path unchanged.`)
         return;
       }
       if (opts.http) {
-        await serveHttp(daemonUrl, opts.bind ?? "127.0.0.1:7437", opts.grantService, opts.grantServiceHeaderFile, oauth);
+        await serveHttp(daemonUrl, opts.bind ?? "127.0.0.1:7437", opts.grantService, opts.grantServiceHeaderFile, oauth, opts.taskDestination, opts.frontDoor === true, unfilledSeat);
         return;
       }
       const client = deps.clientFactory(daemonUrl);
@@ -102,7 +124,7 @@ stays the endpoint; a proxy in front forwards the metadata path unchanged.`)
       process.once("SIGINT", finish);
       process.once("SIGTERM", finish);
       // Serves the protocol revision the client opens with: 2026-07-28 per request, or 2025's initialize handshake.
-      const served = serveStdio(() => createMcpServer(client), { onerror: (err) => console.error(`rig mcp serve: ${err.message}`) });
+      const served = serveStdio(() => createMcpServer(client, { frontDoor: opts.frontDoor === true, ...(unfilledSeat ? { unfilledSeat } : {}) }), { onerror: (err) => console.error(`rig mcp serve: ${err.message}`) });
       try {
         // EOF is a normal stdio disconnect, not an unresolved top-level await.
         if (process.stdin.readableEnded) finish();
@@ -169,7 +191,7 @@ stays the endpoint; a proxy in front forwards the metadata path unchanged.`)
 }
 
 async function serveHttp(daemonUrl: string, bind: string, grantService?: string, grantServiceHeaderFile?: string,
-  oauth?: { resource: string; issuer: string }): Promise<void> {
+  oauth?: { resource: string; issuer: string }, taskDestination?: string, frontDoor = false, unfilledSeat?: UnfilledSeat): Promise<void> {
   const match = /^(\[[^\]]+\]|[^:]+):(\d+)$/.exec(bind);
   if (!match) {
     console.error(`--bind takes host:port (for example 127.0.0.1:7437 or [::1]:7437), not ${bind}`);
@@ -181,7 +203,8 @@ async function serveHttp(daemonUrl: string, bind: string, grantService?: string,
   let served;
   try {
     const header = grantServiceHeaderFile ? readGrantServiceHeader(grantServiceHeaderFile) : undefined;
-    served = await startMcpHttpServer({ daemonUrl, host, port: Number(match[2]), tokenFile,
+    served = await startMcpHttpServer({ daemonUrl, host, port: Number(match[2]), tokenFile, taskDestination, frontDoor,
+      ...(unfilledSeat ? { unfilledSeat } : {}),
       ...(grantService ? { grantService: { url: grantService, ...(header ? { header } : {}) } } : {}), ...(oauth ? { oauth } : {}) });
   } catch (err) {
     console.error(err instanceof McpHttpConfigError ? err.message : `Can't serve on ${bind}: ${(err as Error).message}`);

packages/cli/src/commands/proof.ts

@@ -163,9 +163,17 @@ checkboxes do not accept an item under the selected proof policy.
     if (cmd.opts().workspace) throw new Error("--workspace applies to proof add. Select the judgment daemon using OPENRIG_URL; proof show reports its scope basis.");
     return new DaemonClient();
   };
-  const response = async (r: { status: number; data: unknown }): Promise<Record<string, any>> => {
+  // A path_escape response names the supported routes to another workspace's tree.
+  const projectSelected = (project: unknown, scope: unknown) => Boolean(project) || /^[^/:]+:/.test(typeof scope === "string" ? scope : "");
+  const escapeHint = (project: boolean) => project
+    ? "\nThe scope is outside that project's missions root; name a mission/slice path inside it."
+    : "\nThe scope is outside the selected daemon workspace. For a catalog project, pass --project <id> or prefix the scope with <project>:."
+      + " To use another daemon's workspace, select that daemon with OPENRIG_URL.";
+  const response = async (r: { status: number; data: unknown }, project = false): Promise<Record<string, any>> => {
     const data = r.data as Record<string, any>;
-    if (r.status >= 400) throw new Error(`${data.error ?? r.status}: ${data.message ?? "Read the named source and retry"}`);
+    if (r.status >= 400) {
+      throw new Error(`${data.error ?? r.status}: ${data.message ?? "Read the named source and retry"}${data.error === "path_escape" ? escapeHint(project) : ""}`);
+    }
     return data;
   };
   const projectHelp = "Workspace-catalog project id; resolves the scope under that project's missions root (same as a <project>:<scope> prefix)";
@@ -175,7 +183,7 @@ checkboxes do not accept an item under the selected proof policy.
     .action(async (scope, opts) => {
       try {
         const query = new URLSearchParams({ ...(opts.project ? { project: opts.project } : {}), ...(scope ? { scope } : {}) }).toString();
-        const data = await response(await client().get(`/api/proof${query ? `?${query}` : ""}`));
+        const data = await response(await client().get(`/api/proof${query ? `?${query}` : ""}`), projectSelected(opts.project, scope));
         console.log(opts.json ? JSON.stringify(data) : JSON.stringify(data, null, 2));
       } catch (e) { console.error(e instanceof Error ? e.message : String(e)); process.exitCode = 1; }
     });
@@ -196,7 +204,7 @@ checkboxes do not accept an item under the selected proof policy.
         const refs = [...new Set<string>([...opts.evidence, ...(opts.comparison ? [opts.comparison] : [])])];
         const query = new URLSearchParams({ ...(opts.project ? { project: opts.project } : {}), scope });
         for (const ref of refs) query.append("evidence", ref);
-        const c = client(), view = await response(await c.get(`/api/proof?${query}`));
+        const c = client(), view = await response(await c.get(`/api/proof?${query}`), projectSelected(opts.project, scope));
         const items = view.items as Array<{ id: string; text: string; index: number; revision: string; judgment: { id: string } | null }> | undefined;
         const item = selector ? items?.find(i => i.id === selector || i.text === selector || String(i.index) === selector) : items?.length === 1 ? items[0] : undefined;
         if (!item) throw new Error("Select one current item with scope#item; rig proof show lists IDs, text and indices");
@@ -209,7 +217,7 @@ checkboxes do not accept an item under the selected proof policy.
           operationId: opts.operationId, replace: opts.replace === true, ...(opts.project ? { project: opts.project } : {}),
           // Pin the project root the prepared read resolved; the daemon refuses (409 project_changed) if it moved.
           ...((view.project as { root?: string } | undefined)?.root ? { projectRoot: (view.project as { root: string }).root } : {}) };
-        const result = await response(await c.post("/api/proof/judge", body));
+        const result = await response(await c.post("/api/proof/judge", body), projectSelected(opts.project, scope));
         console.log(opts.json ? JSON.stringify(result) : JSON.stringify(result, null, 2));
       } catch (e) { console.error(e instanceof Error ? e.message : String(e)); process.exitCode = 1; }
     });

packages/cli/src/commands/provider.ts

@@ -603,6 +603,9 @@ API-key approval before the first turn. OpenRig does not enforce API budgets or
 Codex can select a saved API-key auth profile with --to-account just like a subscription profile. Its shared-home effects
 and native credential-store requirements are unchanged. No fresh account home or conversation is created.
 Do not combine --api-key-file with --account-home or --clear-account-home; select a home and switch credentials separately.
+Pi seats support --account-home with an existing readable auth.json, or --clear-account-home, through a fresh successor
+with a handover. Their current model and effort are retained unless overridden. Pi account changes work with or without
+a launch_wrapper. To change runtime, use rig seat set-runtime with --runtime pi, codex or claude-code.
 --clear-account-home returns to the configured member/rig/daemon/native home through a fresh successor with a handover.
 Use it without --to-account, --stored-login, --account-home or --token-file. It removes an account-home token reference
 only after confirmation; it does not restore historical login bytes. Anthropic active/default profiles live outside
@@ -638,7 +641,7 @@ CLAUDE_CONFIG_DIR and are not isolated by an account home; check them before rel
         try {
           const status = await client.get<{ runtime?: string }>(`/api/seat/status/${encodeURIComponent(opts.seat)}`);
           const runtime = status.data.runtime;
-          if (status.status >= 400 || (runtime !== "codex" && runtime !== "claude-code")) {
+          if (status.status >= 400 || (runtime !== "codex" && runtime !== "claude-code" && runtime !== "pi")) {
             const refusal = { ok: false, code: "account_home_preflight_failed", message: "The seat runtime could not be read; no switch was requested." };
             if (opts.json) console.log(JSON.stringify(refusal)); else console.error(refusal.message);
             process.exitCode = 1; return;

packages/cli/src/commands/ps.ts

@@ -88,6 +88,8 @@ interface NodeEntry {
   storedSessionStatus?: string | null;
   /** A community harness's seat: its label, maintainer and declared limits. */
   communityHarness?: { harness: string; maintainer: string | null; limits: string[]; identity: string; needsInput: string; presetError?: string };
+  /** Configured literal argv prefix wrapping local built-in Pi/Codex launches; omitted when absent. */
+  launchWrapper?: string[];
   /** Slice 15 — `has-work-to-do` primitive. Derived from queue_items;
    *  NEVER derived from terminalActive. */
   hasAssignedWork?: boolean;
@@ -99,6 +101,8 @@ interface NodeEntry {
   blockedWorkCount?: number;
   /** OPR.0.3.4.11 — held reason from node.held event. */
   heldReason?: string | null;
+  /** #2456: the required startup files this occupant's handover started it without. */
+  startedWithoutRequired?: string[];
   agentActivity?: {
     state: "running" | "needs_input" | "idle" | "unknown";
     reason: string;
@@ -215,6 +219,7 @@ const ALLOWED_NODE_FIELDS = new Set([
   "nodeKind",
   "runtime",
   "communityHarness",
+  "launchWrapper",
   "model",
   "modelObservation",
   "sessionStatus",
@@ -238,6 +243,7 @@ const ALLOWED_NODE_FIELDS = new Set([
   "needsInputEvidence",
   "contextUsage",
   "heldReason",
+  "startedWithoutRequired",
   "storedSessionStatus",
   "identityVerdict",
 ]);
@@ -705,6 +711,11 @@ export function formatDeclaredModel(model: string | null | undefined): string {
   return m.length > 0 ? m : "not-declared";
 }
 
+/** #2456: the attention reason for an occupant started without a required startup file. */
+function startedWithoutLine(files: string[]): string {
+  return `missing required ${files.map((file) => file.split("/").pop()).join(", ")}`;
+}
+
 export function compactNodeProjection(nodes: NodeEntry[]): Array<Record<string, unknown>> {
   return nodes.map((n) => {
     const keepDiagnostics = n.needsAttention !== false;
@@ -725,6 +736,7 @@ export function compactNodeProjection(nodes: NodeEntry[]): Array<Record<string,
       // 15 claude-code seats — see formatDeclaredModel for the render rule.
       runtime: n.runtime ?? null,
       ...(n.communityHarness ? { communityHarness: n.communityHarness } : {}),
+      ...(n.launchWrapper?.length ? { launchWrapper: n.launchWrapper } : {}),
       model: n.model ?? null,
       // Compact tuple: [literal model, evidence source, native record time].
       // Full JSON / --fields retain the complete labelled object and occupant generation.
@@ -768,6 +780,7 @@ export function compactNodeProjection(nodes: NodeEntry[]): Array<Record<string,
     }
     // held/attention reason (short) when present.
     if (n.heldReason) compact.heldReason = n.heldReason;
+    if (n.startedWithoutRequired?.length) compact.startedWithoutRequired = n.startedWithoutRequired;
     if (keepDiagnostics && n.latestError) {
       compact.latestError = n.latestError;
     }
@@ -921,12 +934,12 @@ contextUsage.percent, contextUsage.state (with --nodes). Other keys are rejected
 activeCount, hasWorkCount, attentionCount, status, lifecycleState, uptime,
 latestSnapshot.
 --fields accepts (node-level, with --nodes): rigId, rigName, logicalId, podId,
-podNamespace, hostSelfId, canonicalSessionName, nodeKind, runtime, communityHarness, model, sessionStatus,
+podNamespace, hostSelfId, canonicalSessionName, nodeKind, runtime, communityHarness, launchWrapper, model, sessionStatus,
 intentionallyStopped, startupStatus, restoreOutcome, lifecycleState, tmuxAttachCommand,
 resumeCommand, latestError, terminalActive, hasAssignedWork,
 assignedWorkCount, pendingWorkCount, inProgressWorkCount, blockedWorkCount,
 agentActivity, needsAttention, arbitratedNeedsInput, needsInputEvidence,
-contextUsage, heldReason, storedSessionStatus, identityVerdict.
+contextUsage, heldReason, startedWithoutRequired, storedSessionStatus, identityVerdict.
 
 --host runs on a remote host declared in ~/.openrig/hosts.yaml (no current-rig
 default applied; the remote host's rigs are shown).
@@ -1358,7 +1371,8 @@ async function handleNodes(
     console.log(padCompactNodeRow("RIG", "SESSION", "LIFECYCLE", "ACTIVITY", "WORK", "REASON", "ATTN"));
     for (const n of humanList as NodeEntry[]) {
       const reason = missingSessionGuidance(n) ?? (n.needsAttention !== false || hasCurrentUsageLimit(n)
-        ? (n.latestError ? truncate(n.latestError, 40) : n.agentActivity?.reason ?? "—")
+        ? (n.latestError ? truncate(n.latestError, 40) : n.startedWithoutRequired?.length ? truncate(startedWithoutLine(n.startedWithoutRequired), 40)
+          : n.agentActivity?.reason ?? "—")
         : livenessUnconfirmed(n) ? "liveness unconfirmed" : "—");
       console.log(padCompactNodeRow(
         n.rigName,
@@ -1395,6 +1409,7 @@ async function handleNodes(
         formatContextUsage(n.contextUsage),
         n.restoreOutcome,
         missingSessionGuidance(n) ?? (n.latestError ? truncate(n.latestError, 30) : n.heldReason ? `held: ${truncate(n.heldReason, 25)}`
+          : n.startedWithoutRequired?.length ? truncate(startedWithoutLine(n.startedWithoutRequired), 30)
           : livenessUnconfirmed(n) ? "liveness unconfirmed" : "—"),
       ));
     }

packages/cli/src/commands/queue.ts

@@ -1,8 +1,8 @@
 import fs from "node:fs";
 import { randomBytes } from "node:crypto";
 import { Command, InvalidArgumentError } from "commander";
-import { DaemonClient, DaemonConnectionError, DaemonTimeoutError, DaemonResponseError, type DaemonResponse } from "../client.js";
-import { getDaemonStatus, getDaemonUrl , daemonStatusGuard} from "../daemon-lifecycle.js";
+import { DaemonClient, DaemonConnectionError, DaemonTimeoutError, DaemonResponseError, remoteDaemonClient, type DaemonResponse } from "../client.js";
+import { getDaemonStatus, getDaemonUrl , daemonStatusGuard, resolveOriginSelfHostId } from "../daemon-lifecycle.js";
 import { readOpenRigEnv } from "../openrig-compat.js";
 import { sessionRigOf, isHumanSeatSessionRef } from "../session-name.js";
 import { realDeps } from "./daemon.js";
@@ -10,11 +10,13 @@ import { enumArg, positiveIntArg } from "../cli-error.js";
 import type { StatusDeps } from "./status.js";
 import { resolveContextRef } from "../context-resolve.js";
 import { shellQuote } from "../cross-host-executor.js";
-import { loadHostRegistry, resolveHost, resolveRemoteBearer } from "../host-registry.js";
+import { bearerAuthHeaders, loadHostRegistry, resolveHost, resolveRemoteBearer } from "../host-registry.js";
 import { loadHostBindings } from "../host-bindings.js";
+import { eligibleQueueBindings, ineligibleBindingDetail } from "@openrig/daemon/learned-host-bindings";
 import { omittedReadField, readView } from "../read-view.js";
 import { buildWriteReceipt, compactOutputEnabled, type ReceiptContext } from "../queue-receipt.js";
 import { TRANSITIONS_SHOWN, compactShow, compactTransitions, showEmptyNotice, transitionsNotice } from "../queue-read-compact.js";
+import { codeResultOf, planItemArtifact, type FileRead, type SharedRoot } from "../queue-artifact.js";
 
 /**
  * `rig queue` — coordination primitive L3/inbox/outbox commands (PL-004 Phase A).
@@ -36,6 +38,8 @@ export interface DeliveryVerifyDeps {
 
 export interface QueueDeps extends StatusDeps {
   deliveryVerify?: DeliveryVerifyDeps;
+  /** The host registry `--host` resolves against on the verbs that name an existing row; tests give their own. */
+  hostRegistryLoader?: () => ReturnType<typeof loadHostRegistry>;
 }
 
 interface DeliveryResultBase {
@@ -558,6 +562,45 @@ async function queueReadTarget(deps: QueueDeps, client: DaemonClient, hostId: st
   return { reader, options };
 }
 
+/** Where a verb that names an existing row (show, transitions, list, claim, update, block) sends its request with
+ *  `--host`: the remote daemon itself, called by this CLI with the host's bearer, as `send --host` does (#2637). The
+ *  selected daemon has no route that reads another host's queue, and a direct call lets the remote record the actor
+ *  as `<seat>@<rig>@<this host>`, the provenance `send --host` gives. Without `--host`, `local`, or a `--host` naming
+ *  the selected daemon's own host id, the selected daemon, unchanged. Explicit-only: never `rig host select`. */
+interface QueueHostTarget { client: DaemonClient; options: { headers?: Record<string, string>; timeoutMs?: number }; remote?: string }
+
+async function queueHostTarget(deps: QueueDeps, local: DaemonClient, hostId: string | undefined, write: boolean): Promise<QueueHostTarget | { error: string; message: string }> {
+  const remote = remoteHostOf(hostId);
+  if (!remote) return { client: local, options: {} };
+  // A --host naming the selected daemon's own id stays local, as create's does (the daemon's resolvesToLocalHost).
+  if (await daemonSelfHostId(local) === remote) return { client: local, options: {} };
+  const registry = (deps.hostRegistryLoader ?? loadHostRegistry)();
+  if (!registry.ok) return { error: "host_registry_unreadable", message: `--host ${remote}: this CLI's host registry can't be read (${registry.error}).` };
+  // The queue forwarder's rule for learned bindings (one policy): a self-id resolves only through a binding learned at
+  // its entry's current URL, with no recorded conflict, so a stale binding never sends a write to a replaced endpoint.
+  const learned = loadHostBindings();
+  const match = resolveHost(registry.registry, remote, eligibleQueueBindings(registry.registry, learned));
+  if (!match.ok) {
+    const unused = ineligibleBindingDetail(registry.registry, learned, remote);
+    return { error: "unknown_host", message: `--host ${remote}: no such host in this CLI's host registry (see rig host list).${unused ? ` ${unused}` : ""}` };
+  }
+  if (match.host.transport !== "http") return { error: "host_not_http", message: `--host ${remote}: its registry entry is SSH, not a daemon URL; queue verbs need an http host.` };
+  const bearer = resolveRemoteBearer(match.host);
+  if (!bearer.ok) return { error: "host_bearer_unavailable", message: `--host ${remote}: ${bearer.error}` };
+  const client = remoteDaemonClient(deps.clientFactory, match.host.url, await resolveOriginSelfHostId(deps.lifecycleDeps));
+  const headers = bearerAuthHeaders(bearer.token);
+  return {
+    client,
+    options: { ...(Object.keys(headers).length > 0 ? { headers } : {}), ...(write ? { timeoutMs: CROSS_HOST_QUEUE_WRITE_TIMEOUT_MS } : {}) },
+    remote,
+  };
+}
+
+/** The ` --host <id>` a printed follow-up command needs to reach the same row again. */
+function hostFlag(target: QueueHostTarget): string {
+  return target.remote ? ` --host ${shellQuote(target.remote)}` : "";
+}
+
 /** A successful forward isn't a delivery receipt. Never accept a different remote row as this create's receipt. */
 function createReceiptReader(deps: QueueDeps, client: DaemonClient, hostId: string | undefined, qitemId: string, sent: SentCreate): Pick<DaemonClient, "get"> {
   if (!remoteHostOf(hostId)) return client;
@@ -604,8 +647,10 @@ async function readBackCreate(deps: QueueDeps, client: DaemonClient, hostId: str
   }
 }
 
-async function printWriteResult(client: DaemonClient, json: boolean, body: unknown, status: number, receipt: ReceiptContext | null): Promise<void> {
-  const settled = receipt && status < 400 ? await settleReceiptHost(client, receipt) : null;
+async function printWriteResult(client: DaemonClient, json: boolean, body: unknown, status: number, receipt: ReceiptContext | null, hostUsed = false): Promise<void> {
+  // A direct --host write already settled its host when it chose the target: its receipt keeps the host it used.
+  // create and handoff, which the selected daemon forwards, are settled here as before.
+  const settled = receipt && status < 400 ? (hostUsed ? receipt : await settleReceiptHost(client, receipt)) : null;
   const projected = settled ? buildWriteReceipt(body, settled) : null;
   if (!projected) {
     printResult(json, body, status);
@@ -622,6 +667,7 @@ async function printQueueItemResult(
   body: unknown,
   status: number,
   receipt: ReceiptContext | null = null,
+  remoteHost?: string,
 ): Promise<void> {
   if (status >= 400 && body && typeof body === "object"
     && (body as { error?: unknown }).error === "qitem_not_found") {
@@ -636,10 +682,11 @@ async function printQueueItemResult(
     body = {
       ...body,
       hint: `Queue item ${JSON.stringify(qitemId)} was not found on ${daemon}.\n`
-        + `If it came from another host, run rig host list; replace <daemon-url> with its registered daemon URL and run: OPENRIG_URL='<daemon-url>' rig queue show ${shellQuote(qitemId)} --full --json`,
+        + `If it came from another host, run rig host list; replace <daemon-url> with its registered daemon URL and run: OPENRIG_URL='<daemon-url>' rig queue show ${shellQuote(qitemId)} --full --json`
+        + (remoteHost ? "" : `\nOr name that host: rig queue show ${shellQuote(qitemId)} --host <host-id> --full --json`),
     };
   }
-  await printWriteResult(client, json, body, status, receipt);
+  await printWriteResult(client, json, body, status, receipt && remoteHost ? { ...receipt, remoteHost } : receipt, remoteHost !== undefined);
 }
 
 // OPR.0.4.3.03 — `rig queue show` body preview.
@@ -922,6 +969,9 @@ function emitHostResolutionError(res: { error: string; message: string }, json:
 const QUEUE_HOST_OPTION_HELP =
   "OPR.0.4.6.MH3: route this queue write to a REGISTERED remote host (see rig host ls). EXPLICIT-ONLY — queue verbs never follow the persisted 'rig host select' selection. Equivalent to the host-qualified destination form member@rig@<host>.";
 
+const QUEUE_HOST_ROW_OPTION_HELP =
+  "#2637: read or act on a row held by a REGISTERED remote host (see rig host ls). This CLI calls that host's daemon directly with its bearer, as send --host does; a write is recorded there as <seat>@<rig>@<this host>. EXPLICIT-ONLY — queue verbs never follow the persisted 'rig host select' selection. A blocker, watchdog or row id given with it must exist on that host.";
+
 export function queueCommand(depsOverride?: QueueDeps): Command {
   const cmd = new Command("queue").description("Coordination L3 — owned-work queue + inbox/outbox");
   const getDeps = (): QueueDeps => depsOverride ?? {
@@ -1206,6 +1256,7 @@ export function queueCommand(depsOverride?: QueueDeps): Command {
       // no body claim. Verify the env (the header source) or the daemon returns 400 actor_required (no seat identity to record; P18 retired the 401 refusal).
       if (!requireSeatSession("destination")) return;
       const deps = getDeps();
+      // No --host here (#2678 review): a row is claimed by its destination seat, on the host that holds it.
       await withClient(deps, async (client) => {
         const res = await client.post<unknown>(`/api/queue/${encodeURIComponent(qitemId)}/claim`, {});
         await printQueueItemResult(client, qitemId, opts.json ?? false, res.data, res.status, writeReceipt(opts.full, { verb: "claim" }));
@@ -1240,13 +1291,15 @@ export function queueCommand(depsOverride?: QueueDeps): Command {
     .option("--reopen", "Explicitly acknowledge a deliberate terminal-to-active repair; requires --state and --note")
     .option("--closure-reason <reason>", "Required for state=done. --state canceled takes no reason, except 'superseded' (with --closure-target = the successor), which records a supersession, distinct from an abandoned cancel")
     .option("--closure-target <target>", "Required for handed_off_to, escalation, and superseded; blocked_on park records use state=blocked with --blocked-on")
+    .option("--closure-required-at <iso>", "Accept a revised closure deadline on an in-progress item; keeps the claim and records the old/new time and accepting actor")
     .option("--resolution <text>", "The outcome, stored on the row and shown by rig queue show. Needs --state done, canceled, denied, failed or handed-off, and non-blank text. Stored when the update closes an open row, after anything the row already has (a fallback line, an earlier outcome); a row that is already closed keeps what it has")
     .option("--blocked-on <blocker>", "For state=blocked: the blocker — a qitem id (must exist and be live), a human seat (FR-6 park; requires summary + evidence_ref), or a typed non-qitem gate 'fold:<what>' / 'auth:<what>' / 'external:<what>'")
     .option("--wake-watchdog <jobId>", "For state=blocked: attach an existing live watchdog id targeting the row owner")
     .option("--wake-after <duration>", "For state=blocked: atomically arm a timer (for example 90s, 15m, 2h)", wakeDurationSeconds)
     .option("--summary <text>", "Summary persisted onto the item (human-seat parks, or standalone summary update without --state)")
-    .option("--evidence-ref <path>", "Pointer to the durable artifact, persisted onto the item on a human-seat park or a done close (--state done)")
+    .option("--evidence-ref <path>", "Pointer to the durable artifact, persisted onto the item on a human-seat park or a done close (--state done). For a code result, a JSON file of kind openrig.code-result/v1 (repo, base and revision as full commit ids, paths) lets a reader diff it exactly")
     .option("--note <text>", "Transition note for the audit log")
+    .option("--host <id>", QUEUE_HOST_ROW_OPTION_HELP)
     .option("--full", "Print the whole response row(s) as today, even when output.compact is on (the receipt's waiting keeps only blocker and nextBackstop)")
     .option("--json", "JSON output for agents")
     .action(async (qitemId: string, opts: {
@@ -1255,6 +1308,7 @@ export function queueCommand(depsOverride?: QueueDeps): Command {
       reopen?: boolean;
       closureReason?: string;
       closureTarget?: string;
+      closureRequiredAt?: string;
       resolution?: string;
       blockedOn?: string;
       wakeWatchdog?: string;
@@ -1262,6 +1316,7 @@ export function queueCommand(depsOverride?: QueueDeps): Command {
       summary?: string;
       evidenceRef?: string;
       note?: string;
+      host?: string;
       json?: boolean;
       full?: boolean;
     }) => {
@@ -1285,12 +1340,15 @@ export function queueCommand(depsOverride?: QueueDeps): Command {
       // header source) or the daemon returns 400 actor_required (no seat identity to record; P18 retired the 401 refusal). Matches `resolve`.
       if (!requireSeatSession("actor")) return;
       const deps = getDeps();
-      await withClient(deps, async (client) => {
-        const res = await client.post<unknown>(`/api/queue/${encodeURIComponent(qitemId)}/update`, {
+      await withClient(deps, async (local) => {
+        const target = await queueHostTarget(deps, local, opts.host, true);
+        if ("error" in target) return emitHostResolutionError(target, opts.json ?? false);
+        const res = await target.client.post<unknown>(`/api/queue/${encodeURIComponent(qitemId)}/update`, {
           state: opts.state,
           reopen: opts.reopen,
           closureReason: opts.closureReason,
           closureTarget: opts.closureTarget,
+          ...(opts.closureRequiredAt !== undefined ? { closureRequiredAt: opts.closureRequiredAt } : {}),
           resolution: opts.resolution,
           blockedOn: opts.blockedOn,
           wakeWatchdogId: opts.wakeWatchdog,
@@ -1298,9 +1356,9 @@ export function queueCommand(depsOverride?: QueueDeps): Command {
           summary: opts.summary,
           evidenceRef: opts.evidenceRef,
           transitionNote: opts.note,
-        });
-        await printQueueItemResult(client, qitemId, opts.json ?? false, res.data, res.status, writeReceipt(opts.full, { verb: "update" }));
-      });
+        }, target.options);
+        await printQueueItemResult(target.client, qitemId, opts.json ?? false, res.data, res.status, writeReceipt(opts.full, { verb: "update" }), target.remote);
+      }, remoteHostOf(opts.host) !== undefined, remoteHostOf(opts.host));
     });
 
   // OPR.0.4.4.19 FR-6 — the first-class park affordance (C5 leg 1). One verb,
@@ -1321,6 +1379,7 @@ export function queueCommand(depsOverride?: QueueDeps): Command {
     .option("--continuation <text>", "What resumes. Workspace deferred/not-imminent work belongs in a mission/slice")
     .option("--wake-watchdog <jobId>", "Attach an existing live watchdog id targeting the parked owner")
     .option("--wake-after <duration>", "Atomically arm a timer with the park (for example 90s, 15m, 2h)", wakeDurationSeconds)
+    .option("--host <id>", QUEUE_HOST_ROW_OPTION_HELP)
     .option("--full", "Print the whole response row(s) as today, even when output.compact is on (the receipt's waiting keeps only blocker and nextBackstop)")
     .option("--json", "JSON output for agents")
     .addHelpText("after", `
@@ -1342,6 +1401,7 @@ workspace home that is deferred/not-imminent belongs in its mission/slice.`)
       continuation?: string;
       wakeWatchdog?: string;
       wakeAfter?: number;
+      host?: string;
       json?: boolean;
       full?: boolean;
     }) => {
@@ -1349,8 +1409,10 @@ workspace home that is deferred/not-imminent belongs in its mission/slice.`)
       // ignored, no body actorSession. Verify the env, else the daemon returns 400 actor_required (no seat identity to record). Matches `resolve`.
       if (!requireSeatSession("actor")) return;
       const deps = getDeps();
-      await withClient(deps, async (client) => {
-        const res = await client.post<unknown>(`/api/queue/${encodeURIComponent(qitemId)}/update`, {
+      await withClient(deps, async (local) => {
+        const target = await queueHostTarget(deps, local, opts.host, true);
+        if ("error" in target) return emitHostResolutionError(target, opts.json ?? false);
+        const res = await target.client.post<unknown>(`/api/queue/${encodeURIComponent(qitemId)}/update`, {
           state: "blocked",
           blockedOn: opts.on,
           summary: opts.summary,
@@ -1358,9 +1420,9 @@ workspace home that is deferred/not-imminent belongs in its mission/slice.`)
           wakeWatchdogId: opts.wakeWatchdog,
           wakeAfterSeconds: opts.wakeAfter,
           transitionNote: opts.continuation ? `continuation: ${opts.continuation}` : (opts.note ?? `parked on ${opts.on}`),
-        });
-        await printQueueItemResult(client, qitemId, opts.json ?? false, res.data, res.status, writeReceipt(opts.full, { verb: "block" }));
-      });
+        }, target.options);
+        await printQueueItemResult(target.client, qitemId, opts.json ?? false, res.data, res.status, writeReceipt(opts.full, { verb: "block" }), target.remote);
+      }, remoteHostOf(opts.host) !== undefined, remoteHostOf(opts.host));
     });
 
   // OPR.0.4.4.19 FR-7 — the resolve verb's CLI wrapper: a THIN client of the
@@ -1655,10 +1717,14 @@ workspace home that is deferred/not-imminent belongs in its mission/slice.`)
     .description("Show one qitem and its derived waiting state (bounded preview; --full for complete body)")
     .option("--full", "Complete original record; may be large (use --full --json for lossless JSON)")
     .option("--json", "JSON preview with completeness, original byte size and exact full command")
-    .action(async (qitemId: string, opts: { full?: boolean; json?: boolean }) => {
+    .option("--host <id>", QUEUE_HOST_ROW_OPTION_HELP)
+    .action(async (qitemId: string, opts: { full?: boolean; json?: boolean; host?: string }) => {
       const deps = getDeps();
-      await withClient(deps, async (client) => {
-        const res = await client.get<unknown>(`/api/queue/${encodeURIComponent(qitemId)}`);
+      await withClient(deps, async (local) => {
+        const target = await queueHostTarget(deps, local, opts.host, false);
+        if ("error" in target) return emitHostResolutionError(target, opts.json ?? false);
+        const client = target.client;
+        const res = await client.get<unknown>(`/api/queue/${encodeURIComponent(qitemId)}`, target.options);
         const json = opts.json ?? false;
         const item = res.data;
         // --full is a pure passthrough of today's COMPLETE item shape (the
@@ -1666,7 +1732,7 @@ workspace home that is deferred/not-imminent belongs in its mission/slice.`)
         // passthrough on error responses / non-object payloads, where there is
         // no string body to preview.
         if (opts.full || res.status >= 400 || !isRecordWithStringBody(item)) {
-          await printQueueItemResult(client, qitemId, json, item, res.status);
+          await printQueueItemResult(client, qitemId, json, item, res.status, null, target.remote);
           return;
         }
         // The closing transition is for --full; the bounded preview prints the record as it did before.
@@ -1674,7 +1740,7 @@ workspace home that is deferred/not-imminent belongs in its mission/slice.`)
         const { preview, bodyBytes, bodyTruncated } = previewBody(item.body);
         // Append-only additions: keep `body` in place (now the preview) and add
         // the honest size + truncation flag. Object otherwise unchanged.
-        const fullCommand = `rig queue show ${shellQuote(qitemId)} --full --json`;
+        const fullCommand = `rig queue show ${shellQuote(qitemId)}${hostFlag(target)} --full --json`;
         // Everything held back from this preview is declared, so a reader that trusts readView runs --full.
         const omitted = [
           ...(bodyTruncated ? [omittedReadField("body (after preview)", item.body.slice(preview.length))] : []),
@@ -1693,6 +1759,47 @@ workspace home that is deferred/not-imminent belongs in its mission/slice.`)
           console.error(`… (bounded preview — complete body is ${bodyBytes} bytes; full record ${view.fullJsonBytes} JSON bytes: ${fullCommand})`);
         }
         if (compact && compact.emptyOmitted.length > 0) console.error(showEmptyNotice(compact.emptyOmitted, fullCommand));
+      }, remoteHostOf(opts.host) !== undefined, remoteHostOf(opts.host));
+    });
+
+  cmd
+    .command("artifact <qitemId>")
+    .description("Read a qitem's evidenceRef through the daemon's shared roots; a code-result manifest also comes back parsed as codeResult")
+    .option("--root <name>", "The shared root a relative evidenceRef is under")
+    .option("--json", "JSON output for agents")
+    .action(async (qitemId: string, opts: { root?: string; json?: boolean }) => {
+      const deps = getDeps();
+      await withClient(deps, async (client) => {
+        const json = opts.json ?? false;
+        const item = await client.get<unknown>(`/api/queue/${encodeURIComponent(qitemId)}`);
+        if (item.status >= 400) { await printQueueItemResult(client, qitemId, json, item.data, item.status); return; }
+        const ref = ((item.data ?? {}) as { evidenceRef?: unknown }).evidenceRef;
+        const rootsRes = await client.get<{ roots?: SharedRoot[] }>("/api/files/roots");
+        if (rootsRes.status >= 400) { printResult(json, rootsRes.data, rootsRes.status); return; }
+        const roots = rootsRes.data?.roots ?? [];
+        // The same rule as rig_artifact_read: only a file in a shared root is read, through the daemon.
+        const plan = planItemArtifact(ref, roots, opts.root);
+        if ("unavailable" in plan) {
+          if (json) console.log(JSON.stringify({ available: false, qitemId, reason: plan.unavailable, evidenceRef: typeof ref === "string" ? ref : null, roots: roots.map((entry) => entry.name) }));
+          else console.error(`${qitemId}: ${plan.unavailable}`);
+          process.exitCode = 1;
+          return;
+        }
+        const res = await client.get<FileRead>(`/api/files/read?${new URLSearchParams({ root: plan.read.root, path: plan.read.path }).toString()}`);
+        if (res.status >= 400) { printResult(json, res.data, res.status); return; }
+        const read = res.data ?? {};
+        const parsed = codeResultOf(read);
+        if (json) {
+          console.log(JSON.stringify({ available: true, qitemId, root: read.root, path: read.path, size: read.size, mtime: read.mtime, binary: read.binary === true,
+            ...(read.binary ? {} : { text: read.content ?? "" }), truncated: read.truncated === true, ...parsed }));
+          return;
+        }
+        if (read.binary) console.log(`A binary file (${read.root}:${read.path}): not shown as text.`);
+        else process.stdout.write(read.content ?? "");
+        // Notices go to stderr, so stdout stays the file.
+        if (parsed && "codeResult" in parsed) console.error(`code result: ${parsed.codeResult.repo} ${parsed.codeResult.base}..${parsed.codeResult.revision}${parsed.codeResult.paths.length ? ` (${parsed.codeResult.paths.length} paths)` : ""}`);
+        else if (parsed) console.error(parsed.codeResultIssue);
+        if (read.truncated) console.error(`… cut at ${read.truncatedAtBytes} of ${read.totalBytes} bytes, as the daemon reads it`);
       });
     });
 
@@ -1701,20 +1808,24 @@ workspace home that is deferred/not-imminent belongs in its mission/slice.`)
     .description("Show the append-only transition log for a qitem")
     .option("--json", "JSON output for agents")
     .option("--full", `The whole log, as today, even when output.compact is on (compact shows the latest transitions, ${TRANSITIONS_SHOWN} in all, always including the current disposition)`)
-    .action(async (qitemId: string, opts: { json?: boolean; full?: boolean }) => {
+    .option("--host <id>", QUEUE_HOST_ROW_OPTION_HELP)
+    .action(async (qitemId: string, opts: { json?: boolean; full?: boolean; host?: string }) => {
       const deps = getDeps();
-      await withClient(deps, async (client) => {
-        const res = await client.get<unknown>(`/api/queue/${encodeURIComponent(qitemId)}/transitions`);
+      await withClient(deps, async (local) => {
+        const target = await queueHostTarget(deps, local, opts.host, false);
+        if ("error" in target) return emitHostResolutionError(target, opts.json ?? false);
+        const client = target.client;
+        const res = await client.get<unknown>(`/api/queue/${encodeURIComponent(qitemId)}/transitions`, target.options);
         // Slice 15 P3: with output.compact on, the recent story; the notice goes to stderr so it stays
         // visible when the array is piped through jq.
         if (!opts.full && res.status < 400 && Array.isArray(res.data) && compactOutputEnabled()) {
           const { transitions, earlier } = compactTransitions(res.data as Record<string, unknown>[]);
           printResult(opts.json ?? false, transitions, res.status);
-          if (earlier > 0) console.error(transitionsNotice(earlier, res.data.length, `rig queue transitions ${shellQuote(qitemId)} --full${opts.json ? " --json" : ""}`));
+          if (earlier > 0) console.error(transitionsNotice(earlier, res.data.length, `rig queue transitions ${shellQuote(qitemId)}${hostFlag(target)} --full${opts.json ? " --json" : ""}`));
           return;
         }
-        await printQueueItemResult(client, qitemId, opts.json ?? false, res.data, res.status);
-      });
+        await printQueueItemResult(client, qitemId, opts.json ?? false, res.data, res.status, null, target.remote);
+      }, remoteHostOf(opts.host) !== undefined, remoteHostOf(opts.host));
     });
 
   cmd
@@ -1732,6 +1843,7 @@ workspace home that is deferred/not-imminent belongs in its mission/slice.`)
     .option("--target-repo <name>", "PL-007: filter qitems by target_repo (exact match)")
     .option("--limit <n>", "Result limit", positiveIntArg, 100)
     .option("--json", "JSON output (compact; use --full --json for complete fields)")
+    .option("--host <id>", `${QUEUE_HOST_ROW_OPTION_HELP} With --host, the rows of every rig on that host unless --destination or --source narrows them; --owned and --mine name this host's seat and are refused.`)
     .addHelpText("after", `
 Default: active items in your current rig, compact summary (like 'docker ps').
 Current rig is derived from OPENRIG_SESSION_NAME's @<rig> suffix.
@@ -1774,45 +1886,62 @@ Examples:
       targetRepo?: string;
       limit: string;
       json?: boolean;
+      host?: string;
     }) => {
       const deps = getDeps();
-      const params = new URLSearchParams();
-      const sessionName = readOpenRigEnv("OPENRIG_SESSION_NAME", "RIGGED_SESSION_NAME");
-      if (opts.owned && !sessionName) {
-        console.error("Cannot use --owned: checked OPENRIG_SESSION_NAME and RIGGED_SESSION_NAME, but neither caller seat identity is set. Set one to a canonical seat@rig address, or use --destination <session>.");
-        process.exitCode = 1;
-        return;
-      }
-      const hasExplicitScope = !!(opts.destination || opts.source || opts.owned);
-
-      if (opts.owned && sessionName) {
-        params.set("destinationSession", sessionName);
-      } else if (opts.mine && sessionName) {
-        params.set("as", sessionName);
-      } else if (!opts.allRigs && !hasExplicitScope) {
-        const rigName = sessionName ? extractRigName(sessionName) : undefined;
-        if (rigName) {
-          params.set("rig", rigName);
+      const remote = remoteHostOf(opts.host);
+      const useJson = opts.json || opts.o === "json";
+      // The seat env names a seat of THIS host. On another host's daemon (a --host that resolves remote) it scopes
+      // nothing that daemon holds: no rig is derived from it, and --owned/--mine are refused.
+      const seatScope = (onRemote: boolean): URLSearchParams | null => {
+        const params = new URLSearchParams();
+        const sessionName = onRemote ? undefined : readOpenRigEnv("OPENRIG_SESSION_NAME", "RIGGED_SESSION_NAME");
+        if (onRemote && (opts.owned || opts.mine)) {
+          emitHostResolutionError({ error: "host_scope_unsupported", message: `--host ${remote} lists another host's rows: --${opts.owned ? "owned" : "mine"} names this host's seat. Use --destination or --source with the address as that host stores it.` }, useJson);
+          return null;
         }
-      }
+        if (opts.owned && !sessionName) {
+          console.error("Cannot use --owned: checked OPENRIG_SESSION_NAME and RIGGED_SESSION_NAME, but neither caller seat identity is set. Set one to a canonical seat@rig address, or use --destination <session>.");
+          process.exitCode = 1;
+          return null;
+        }
+        const hasExplicitScope = !!(opts.destination || opts.source || opts.owned);
 
-      if (!opts.all) {
-        params.set("activeOnly", "1");
-      }
-      if (!opts.full) {
-        params.set("compact", "1");
-      }
-      if (opts.destination) params.set("destinationSession", opts.destination);
-      if (opts.source) params.set("sourceSession", opts.source);
+        if (opts.owned && sessionName) {
+          params.set("destinationSession", sessionName);
+        } else if (opts.mine && sessionName) {
+          params.set("as", sessionName);
+        } else if (!opts.allRigs && !hasExplicitScope) {
+          const rigName = sessionName ? extractRigName(sessionName) : undefined;
+          if (rigName) {
+            params.set("rig", rigName);
+          }
+        }
+
+        if (!opts.all) {
+          params.set("activeOnly", "1");
+        }
+        if (!opts.full) {
+          params.set("compact", "1");
+        }
+        if (opts.destination) params.set("destinationSession", opts.destination);
+        if (opts.source) params.set("sourceSession", opts.source);
+        if (opts.state) params.set("state", opts.state);
+        if (opts.targetRepo) params.set("targetRepo", opts.targetRepo);
+        if (opts.limit) params.set("limit", String(opts.limit));
+        return params;
+      };
+      // Without --host, the scope is settled before the daemon is contacted, as before.
+      if (!remote && !seatScope(false)) return;
       warnIfUnknownStates(opts.state);
-      if (opts.state) params.set("state", opts.state);
-      if (opts.targetRepo) params.set("targetRepo", opts.targetRepo);
-      if (opts.limit) params.set("limit", String(opts.limit));
-      await withClient(deps, async (client) => {
-        const res = await client.get<unknown>(`/api/queue/list?${params.toString()}`);
-        const useJson = opts.json || opts.o === "json";
+      await withClient(deps, async (local) => {
+        const target = await queueHostTarget(deps, local, opts.host, false);
+        if ("error" in target) return emitHostResolutionError(target, useJson);
+        const params = seatScope(target.remote !== undefined);
+        if (!params) return;
+        const res = await target.client.get<unknown>(`/api/queue/list?${params.toString()}`, target.options);
         printResult(useJson, res.data, res.status);
-      });
+      }, remote !== undefined, remote);
     });
 
   cmd
@@ -1870,6 +1999,100 @@ Examples:
       });
     });
 
+  cmd
+    .command("erase")
+    .description("Erase queue rows in one exact scope: an account's rows, a list of ids, or the closed-work retention windows; a dry run unless --apply")
+    .option("--members <file>", "Account deletion: the account's recorded session names (exact, departed seats included), one per line or a JSON array")
+    .option("--rig <id>", "With --members: the account rig's id. Also erases the rig's own records that outlive it (events, snapshots, node and seat records); refused unless every member is a session of it")
+    .option("--ids <file>", "These qitem ids, one per line or a JSON array; a report's qitems entries, or an erasure ledger's JSON lines, keep each one's action")
+    .option("--action <action>", "With --ids: delete (default) removes each row and keeps only its id; scrub removes its content and keeps the row", enumArg(["delete", "scrub"]))
+    .option("--routine", "Run this daemon's closed-work windows now (retention.closed_content_days, retention.closed_metadata_days)")
+    .option("--apply", "Erase. Without it the command reports what would be erased and changes nothing")
+    .option("--actor <session>", "Who is erasing, when OPENRIG_SESSION_NAME isn't set (recorded as a claimed identity)")
+    .option("--json", "JSON output for agents")
+    .addHelpText("after", [
+      "",
+      "Name exactly one scope. What each one erases:",
+      "  --members  every row whose source and destination are both members, in any state, deleted with its transitions,",
+      "             outbox and inbox copies, stream actions, events and audit rows; then member-to-member outbox and inbox",
+      "             entries no row owns, and the members' stream items. A row with one member endpoint is reported under",
+      "             mixed and left untouched. With --rig, also the rig's own events, snapshots and node and seat records.",
+      "  --ids      the listed rows, deleted or scrubbed (a report's qitems entries keep their action); a deleted id",
+      "             that's already gone keeps (or regains) its id marker.",
+      "  --routine  closed rows (done, canceled, denied, handed-off) past the windows, aged from when they closed.",
+      "A deleted row keeps only its id, and a create that reuses that id is refused (qitem_erased). A row a workflow still",
+      "references is reported under gaps and left. Output is JSON: clean (true only when nothing was left), complete,",
+      "qitems, mixed, gaps, alreadyAbsent, linked, unlinked and rig. Refusals: erase_scope_required, erase_scope_conflict,",
+      "erase_members_invalid, erase_rig_invalid, erase_rig_mismatch, erase_ids_invalid, erase_action_invalid,",
+      "erase_routine_off. A run that stops on an error returns what it committed, with error.stage, and exits 2.",
+      "Reference: docs/reference/queue-erasure.md",
+    ].join("\n"))
+    .action(async (opts: { members?: string; rig?: string; ids?: string; action?: string; routine?: boolean; apply?: boolean; actor?: string; json?: boolean }) => {
+      const list = (flag: string, path: string): unknown[] | undefined => {
+        let text: string;
+        try {
+          text = fs.readFileSync(path, "utf8");
+        } catch (err) {
+          console.error(`--${flag} ${path}: can't be read (${err instanceof Error ? err.message : String(err)}). Nothing was erased.`);
+          process.exitCode = 1;
+          return undefined;
+        }
+        if (text.trimStart().startsWith("[")) {
+          try {
+            const parsed: unknown = JSON.parse(text);
+            if (Array.isArray(parsed)) return parsed;
+          } catch { /* reported below */ }
+          console.error(`--${flag} ${path}: not a JSON array. Nothing was erased.`);
+          process.exitCode = 1;
+          return undefined;
+        }
+        // One entry per line: a plain id or session, or (an erasure ledger's line) a JSON object.
+        const entries: unknown[] = [];
+        for (const line of text.split(/\r?\n/).map((entry) => entry.trim()).filter((entry) => entry.length > 0 && !entry.startsWith("#"))) {
+          if (!line.startsWith("{")) {
+            entries.push(line);
+            continue;
+          }
+          try {
+            entries.push(JSON.parse(line));
+          } catch {
+            console.error(`--${flag} ${path}: a line isn't valid JSON (${line.slice(0, 60)}). Nothing was erased.`);
+            process.exitCode = 1;
+            return undefined;
+          }
+        }
+        return entries;
+      };
+      const body: Record<string, unknown> = { apply: opts.apply === true };
+      if (opts.members !== undefined) {
+        const members = list("members", opts.members);
+        if (!members) return;
+        body.members = members;
+      }
+      // Sent as given: the daemon refuses a rig without members (erase_rig_invalid), never drops it.
+      if (opts.rig !== undefined) body.rig = opts.rig;
+      if (opts.ids !== undefined) {
+        const ids = list("ids", opts.ids);
+        if (!ids) return;
+        body.ids = ids;
+        if (opts.action) body.action = opts.action;
+      }
+      if (opts.routine) body.routine = true;
+      if (!readOpenRigEnv("OPENRIG_SESSION_NAME", "RIGGED_SESSION_NAME")) {
+        if (!opts.actor) {
+          console.error("OPENRIG_SESSION_NAME is not set and --actor wasn't given, so there's no one to record as erasing. Nothing was erased.");
+          process.exitCode = 1;
+          return;
+        }
+        body.actorSession = opts.actor;
+      }
+      const deps = getDeps();
+      await withClient(deps, async (client) => {
+        const res = await client.post<unknown>("/api/queue/erase", body);
+        printResult(opts.json ?? false, res.data, res.status);
+      });
+    });
+
   // ---- Inbox subcommands ----
 
   cmd

packages/cli/src/commands/roster.ts

@@ -60,7 +60,19 @@ export function enrichRosterMember(member: RosterMember, inventory: RosterInvent
   const node = identities.size === 1 ? matches[0] : undefined;
   const sessionStatus = node ? text(node.sessionStatus) : null;
   const current = sessionStatus === "running";
-  const activity = current && object(node?.agentActivity) ? node.agentActivity : null;
+  const activities = matches.map(n => {
+    const activity = current && object(n.agentActivity) ? n.agentActivity : null;
+    return { state: text(activity?.state) ?? "unknown", reason: text(activity?.reason), sampledAt: text(activity?.sampledAt) };
+  });
+  const agreed = node && new Set(activities.map(a => JSON.stringify([a.state, a.reason]))).size === 1;
+  const times = activities.map(a => a.sampledAt);
+  const sameTime = new Set(times).size === 1;
+  const comparable = times.every((t): t is string => t !== null && Number.isFinite(Date.parse(t)));
+  // A later timestamp is meaningful only when every route supplied a comparable time.
+  const sampledAt = sameTime ? times[0] ?? null : comparable
+    ? [...times].sort((a, b) => Date.parse(b!) - Date.parse(a!) || b!.localeCompare(a!))[0] ?? null : null;
+  const activity = agreed ? { ...activities[0]!, sampledAt }
+    : { state: "unknown", reason: matches.length > 1 ? "Conflicting activity observations" : null, sampledAt: null };
   const observations = matches.map(n => modelObservation(n.modelObservation, current));
   const observation = node && new Set(observations.map(o => JSON.stringify(o))).size === 1 ? observations[0]!
     : unknownModelObservation(matches.length ? "conflicting-observations" : "host-or-seat-unavailable");
@@ -75,7 +87,7 @@ export function enrichRosterMember(member: RosterMember, inventory: RosterInvent
     observedModel: observation.model,
     modelObservation: observation,
     sessionStatus,
-    agentActivity: { state: text(activity?.state) ?? "unknown", reason: text(activity?.reason), sampledAt: text(activity?.sampledAt) },
+    agentActivity: activity,
     // A hosted member's last call to this host's MCP server, for a roster row on this host only: another host's member,
     // or one whose host can't be told, is unknown (null), never this host's record for a same-named seat. Its last ring
     // is in rig_team.

packages/cli/src/commands/scope.ts

@@ -1236,6 +1236,8 @@ function buildAuditCommand(): Command {
         for (const sr of sliceResults) {
           const shadow = shadowedNodeFileFinding(path.join(slicesDir, sr.name), "slice");
           if (shadow) sr.result.findings.push(shadow);
+          const shared = sharedSliceNameFinding(missionsRoot, missionName, sr.name);
+          if (shared) sr.result.findings.push(shared);
         }
 
         const allFindings = [
@@ -2156,6 +2158,45 @@ function shadowedNodeFileFinding(dir: string, level: "mission" | "slice"): {
   };
 }
 
+/**
+ * Advisory-only: another mission has a slice folder of the same name. The slice list and the routes that
+ * find a slice by name keep one slice per name, so the others are hidden there with no sign they exist.
+ * Twin of the daemon's finding of the same name.
+ */
+function sharedSliceNameFinding(missionsRoot: string, missionName: string, sliceName: string): {
+  kind: "slice_name_shared"; severity: "low"; path: string; message: string; remediation: string;
+} | null {
+  let others: string[];
+  try {
+    others = fs.readdirSync(missionsRoot, { withFileTypes: true })
+      .filter((e) => e.isDirectory() && !e.name.startsWith(".") && e.name !== missionName
+        && fs.statSync(path.join(missionsRoot, e.name, "slices", sliceName), { throwIfNoEntry: false })?.isDirectory() === true)
+      .map((e) => e.name)
+      .sort();
+  } catch {
+    return null;
+  }
+  if (others.length === 0) return null;
+  return {
+    kind: "slice_name_shared",
+    severity: "low",
+    path: path.join(missionsRoot, missionName, "slices", sliceName),
+    message: `Mission ${others.join(", ")} also ${others.length === 1 ? "has" : "have"} a slice named "${sliceName}". The slice list (GET /api/slices) and the routes that find a slice by name keep one slice per name, so all but one of these slices are hidden there.`,
+    remediation: "Rename one of the slice folders so each slice name is unique across missions. Advisory only — nothing is blocked.",
+  };
+}
+
+/** Show the inherited workspace route because Commander lists a parent option only in the parent's help. */
+export function scopeWorkspaceHelp(tier: "slice" | "mission", verb?: string): string {
+  return `
+Workspace: the parent option --workspace selects the missions root this command reads.
+  rig scope --workspace <project root> ${tier} ${verb ?? "<verb>"} ...
+A project root uses its project.yaml missions.root, else its missions/ folder; a
+folder named missions is used as is. Use it for a catalog project, for example
+$(rig config get workspace.projects_root)/<project>. Without --workspace,
+OPENRIG_WORK_ROOT applies, then the configured workspace.slices_root.`;
+}
+
 export function scopeCommand(): Command {
   const cmd = new Command("scope")
     .description("Scope tree primitive: missions, slices, sub-slices (per conventions/scope-and-versioning)")
@@ -2186,6 +2227,10 @@ export function scopeCommand(): Command {
   mission.addCommand(buildMissionVerifiedCommand());
   mission.addCommand(buildApproveCommand("mission"));
   cmd.addCommand(mission);
+  for (const [tier, group] of [["slice", slice], ["mission", mission]] as const) {
+    group.addHelpText("after", scopeWorkspaceHelp(tier));
+    for (const verb of group.commands) verb.addHelpText("after", scopeWorkspaceHelp(tier, verb.name()));
+  }
   cmd.addCommand(buildResolveNotesCommand());
   cmd.addCommand(buildAuditCommand());
 

packages/cli/src/commands/seat.ts

@@ -21,6 +21,10 @@ async function defaultReadStdin(): Promise<string> {
 }
 
 interface SeatStatusResponse {
+  permission_stop?: { state: "unavailable"; command: "unavailable"; origin: "unknown" }
+    | { state: "request_observed" | "completed"; command: string; origin: "harness";
+        source: "PermissionRequest"; observed_at: string; actor: "main" | "subagent"; pending: "unknown" | false;
+        suggested_rules?: { state: "missing" | "present" | "unavailable"; entries: unknown[]; unsupported: boolean; size_cut: boolean } };
   seat_ref: string;
   rig_id: string;
   rig_name: string;
@@ -165,8 +169,10 @@ interface SeatHandoverMutationResult {
   sourceOutcome?:
     | { mode: "fork"; forkedFrom: string; forkId?: "captured" | "pending"; limit?: string }
     | { mode: "rebuild"; primedArtifacts: Array<{ address: string; label: string }>; gaps: string[]; emptyChainReason?: string; permissions?: { state: string; detail: string } };
-  /** The restore packet was pasted and Enter pressed, but its submission couldn't be confirmed. */
+  /** The restore or rebuild priming packet was pasted and Enter pressed, but its submission couldn't be confirmed. */
   restorePacket?: { status: "unconfirmed"; reason: string };
+  /** #2456: the required startup files the successor started without, where a start would have failed. */
+  startedWithoutRequired?: string[];
   sideEffects: {
     departingSessionKilled: false;
     /** Absent from an older daemon, whose receipt reported both side effects as a fixed false. */
@@ -206,6 +212,13 @@ const LEVEL_WORDS: Record<string, string> = {
   "sends-only": "Deliberate sends from people and agents type; OpenRig's own notices are held.",
 };
 
+// JSON escapes C0; also escape C1, format/separator and default-ignorable characters.
+// Escaping the serialized representation leaves JSON.parse's rule strings unchanged.
+function terminalJson(value: unknown, space?: number): string {
+  return JSON.stringify(value, null, space).replace(/[\u007f-\u009f\p{Cf}\p{Zl}\p{Zp}\p{Default_Ignorable_Code_Point}]/gu,
+    character => character.split("").map(unit => `\\u${unit.charCodeAt(0).toString(16).padStart(4, "0")}`).join(""));
+}
+
 /** Show configured seat state, including selected account references, without claiming the running login is verified. */
 function printHuman(status: SeatStatusResponse): void {
   console.log(`Seat ${status.seat_ref}`);
@@ -241,6 +254,24 @@ function printHuman(status: SeatStatusResponse): void {
     else console.log("Hold pauses all automatic input, including writing lifecycle. Changing mode does not replay held messages.");
   }
   console.log(`Session: ${display(status.session_status, "unknown")}`);
+  const stop = status.permission_stop;
+  if (stop && stop.state !== "unavailable") {
+    console.log(`Latest permission observation at ${stop.observed_at}: ${stop.command}; harness permission request (${stop.source}, ${stop.actor})`);
+    console.log(stop.state === "completed" ? "  Displayed actor's tool batch completed; this observation is no longer pending."
+      : "  Displayed actor's pending status: unknown (an observation does not establish a waiting prompt).");
+    console.log("  Other actors' pending state: unknown (only the latest observation is shown).");
+    console.log("  Claude's suggested rules: read-only evidence; not applied or approved, effectiveness unknown.");
+    console.log("  These are neither the exact command nor a complete list of dialog options.");
+    const rules = stop.suggested_rules;
+    if (!rules || rules.state === "unavailable") console.log("  Suggested-rule evidence: unavailable.");
+    else if (rules.state === "missing") console.log("  Suggested-rule field: missing from the observed hook.");
+    else {
+      if (!rules.entries.length) console.log("  Supported suggested-rule entries: empty.");
+      for (const entry of rules.entries) console.log(`  ${terminalJson(entry)}`);
+    }
+    if (rules?.unsupported) console.log("  Unsupported or malformed suggested-rule evidence was withheld.");
+    if (rules?.size_cut) console.log("  Suggested-rule evidence is incomplete because of size bounds.");
+  }
   if (status.compaction) console.log(`Managed compaction (${status.compaction.stage}): ${status.compaction.detail}`);
   const reason = sessionReason(status.session_reason);
   if (reason) console.log(`Reason: ${reason}`);
@@ -325,7 +356,11 @@ function printHumanHandoverPlan(plan: SeatHandoverPlan): void {
 }
 
 function printHumanHandoverResult(result: SeatHandoverMutationResult): void {
-  console.log(`Seat handover complete: ${result.seat.ref}`);
+  // #2456: a successor that started without a required startup file isn't a plain success; a start would have failed.
+  const withoutRequired = result.startedWithoutRequired ?? [];
+  console.log(`Seat handover complete${withoutRequired.length > 0 ? ", degraded" : ""}: ${result.seat.ref}`);
+  for (const file of withoutRequired) console.log(`The successor started without required file ${file}.`);
+  if (withoutRequired.length > 0) console.log("rig ps --nodes shows the seat as needing attention until it's next launched.");
   console.log(`Rig: ${result.seat.rigName}`);
   console.log(`Logical ID: ${result.seat.logicalId}`);
   console.log(`Source: ${result.source.mode}${result.source.ref ? `:${result.source.ref}` : ""}`);
@@ -351,7 +386,8 @@ function printHumanHandoverResult(result: SeatHandoverMutationResult): void {
     const outcome = result.sourceOutcome;
     if (outcome.permissions) console.log(`Codex rebuild permissions: ${outcome.permissions.state}. ${outcome.permissions.detail}`);
     if (outcome.primedArtifacts.length > 0) {
-      console.log(`Rebuild primed from ${outcome.primedArtifacts.length} durable artifact${outcome.primedArtifacts.length === 1 ? "" : "s"}:`);
+      const heading = result.restorePacket?.status === "unconfirmed" ? "Rebuild packet includes" : "Rebuild primed from";
+      console.log(`${heading} ${outcome.primedArtifacts.length} durable artifact${outcome.primedArtifacts.length === 1 ? "" : "s"}:`);
       for (const artifact of outcome.primedArtifacts) console.log(`  - ${artifact.address} — ${artifact.label}`);
     }
     for (const gap of outcome.gaps) console.log(`  ! declared but missing on disk: ${gap}`);
@@ -370,8 +406,17 @@ function printHumanHandoverResult(result: SeatHandoverMutationResult): void {
     console.log("No conversation continuity was carried: the successor started a fresh conversation.");
   } else if (result.restorePacket?.status === "unconfirmed") {
     // The packet was pasted and Enter pressed: the successor may already be working on it, so don't send it again.
-    console.log(`The captured startup context (restore packet) was sent to the successor, but its submission is unconfirmed: ${result.restorePacket.reason}.`);
-    console.log("Check the successor's screen before sending the context again; it may already be working on it.");
+    if (result.source.mode === "rebuild") {
+      console.log(`The rebuild priming packet was sent to the successor, but its submission is unconfirmed: ${result.restorePacket.reason}.`);
+      if (result.restorePacket.reason === "the prompt still looked staged in the input box") {
+        console.log("Check the successor's current screen. Only if that exact packet is still staged, submit the existing packet once. Do not resend blindly or rerun the handover.");
+      } else {
+        console.log("Check the successor's screen before sending the priming packet again; it may already be working on it.");
+      }
+    } else {
+      console.log(`The captured startup context (restore packet) was sent to the successor, but its submission is unconfirmed: ${result.restorePacket.reason}.`);
+      console.log("Check the successor's screen before sending the context again; it may already be working on it.");
+    }
     console.log("No conversation continuity was carried: the successor started a fresh conversation.");
   } else if (result.source.mode === "fork") {
     console.log("No packet delivery: fork context rides the native conversation itself.");
@@ -544,7 +589,7 @@ Examples:
       const res = await client.get<SeatStatusResponse | SeatStatusError>(`/api/seat/status/${encodeURIComponent(seat)}`);
 
       if (opts.json) {
-        console.log(JSON.stringify(res.data, null, 2));
+        console.log(terminalJson(res.data, 2));
         if (res.status >= 400) process.exitCode = res.status >= 500 ? 2 : 1;
         return;
       }
@@ -990,13 +1035,13 @@ resume/successor launch reads it at call time. Examples:
   cmd
     .command("set-runtime")
     .argument("<seat>", "Canonical session name or logical seat ref")
-    .option("--runtime <runtime>", "codex or claude-code")
+    .option("--runtime <runtime>", "codex, claude-code or pi")
     .option("--reason <text>", "Why the seat moves; recorded with the switch")
     .option("--model <id>", "Target model; otherwise the seat's agent spec for the target, else the target's default")
     .option("--effort <level>", "Target effort, the target runtime's own level (not translated)")
     .option("--codex-config-profile <name>", "Codex configuration profile for a Codex successor")
     .option("--clear-account-home", "Return to member/rig/daemon/native login selection through a fresh successor")
-    .option("--account-home <path>", "Isolated target CODEX_HOME or CLAUDE_CONFIG_DIR; also permits a fresh successor on the same runtime")
+    .option("--account-home <path>", "Target Codex/Claude home or Pi account directory containing auth.json; also permits a same-runtime successor")
     .option("--token-file <path>", "With a Claude account home, its required subscription token reference")
     .option("--account-label <label>", "With --token-file, the selected account's label")
     .option("--permission-mode <mode>", "Target native permission choice; default: inherit the seat's member and rig policy")
@@ -1006,7 +1051,7 @@ resume/successor launch reads it at call time. Examples:
     .option("--note <text>", "With --confirm: what you'll do first")
     .option("--recover", "Finish an interrupted switch from its record")
     .option("--json", "JSON output for agents")
-    .description("Move a live Codex or Claude Code seat to the other runtime, successor first")
+    .description("Move a live Codex, Claude Code or Pi seat to another runtime, successor first")
     .addHelpText("after", `
 The seat keeps its address, connections, queue rows and claims. Its successor
 starts on the target runtime beside the current occupant, reads a handover, and
@@ -1020,7 +1065,7 @@ later converge reports a deferred change_runtime); the source runtime's model,
 effort, Codex profile and native permission choice are cleared, not mapped; two
 processes run until cutover; messages sent to the seat during the switch are
 held, not typed, and need sending again afterwards.
-Codex and Claude Code, single-pane seats only. --dry-run lists the details.
+Codex, Claude Code and Pi, single-pane seats only. --dry-run lists the details.
 A seat that has never started has no pane to switch: declare the other runtime for it in its rig spec and bring the team
 up again from that spec. That's a new generation of the seat, without this seat's edges, settings or queued work.
 --clear-account-home returns to configured member/rig/daemon/native selection; it does not restore historical credentials.
@@ -1392,10 +1437,16 @@ Examples:
     .option("--json", "JSON output for agents")
     .description("Stop exactly one LIVE managed seat (kills only that seat's tmux session; audited)")
     .addHelpText("after", `
-Siblings are untouched. A dead seat is refused (use rig seat clean); an adopted
-session is refused (use rig unclaim). Stopping ends that agent's session and any
-work in progress: check it first (rig capture <seat>) and stop only a seat you
-own or were asked to stop. rig seat launch <seat> --fresh --reason <text> starts
+Siblings are untouched. A dead seat is refused (use rig seat clean). An adopted
+session is refused: to end only it, run rig unclaim <session> (it releases the
+session and leaves its process running), then end that process in its own
+terminal. rig down <rig> ends every seat in its rig, so it is the route only
+for a one-seat rig.
+A seat handed over (fresh, fork or rebuild) before OpenRig 0.7.9 still reads
+adopted and is refused too, until its next launched handover or managed launch;
+the refusal says so. Stopping ends that agent's session and any work in progress:
+check it first (rig capture <seat>) and stop only a seat you own or were asked to
+stop. rig seat launch <seat> --fresh --reason <text> starts
 a fresh occupant there; it does not resume the stopped conversation.
 A community harness stop closes its managed pane; its native processes or compute
 may outlive it. Inspect their own lifecycle receipts.

packages/cli/src/commands/send.ts

@@ -1,3 +1,4 @@
+import { isHumanSeatSessionRef } from "../session-name.js";
 import { Command } from "commander";
 import { resolveEffectiveHost } from "../host-selection.js";
 import { DaemonClient, DaemonConnectionError, terminalAuthHeaders } from "../client.js";
@@ -90,6 +91,9 @@ export function wrapSendBody(
   // renders EXACTLY AS RECEIVED — bare member@rig locally (copy-paste reply hint), and a
   // cross-host caller constructs the origin triple at the forwarding boundary before wrapping.
   const senderLabel = sender && sender.trim().length > 0 ? sender : SENDER_FALLBACK;
+  // Only reply classification uses the bare address; From and the destination retain the origin host.
+  const senderParts = senderLabel.split("@");
+  const personAddress = senderParts.length === 3 && senderParts[2] ? senderParts.slice(0, 2).join("@") : senderLabel;
   const header = [`From: ${senderLabel}`, renderToLine(recipient, meta?.scope)];
   if (meta?.stampISO) {
     // GHOST-STAGE (g): the sender's occupant generation rides the Sent: line as a short suffix
@@ -102,7 +106,9 @@ export function wrapSendBody(
   }
   const reply = senderLabel.endsWith("@external")
     ? `↩ Reply if needed: rig queue create --destination ${senderLabel}${meta?.replyToQitem ? ` --human-intent update --reply-to ${meta.replyToQitem}` : ""} --summary "<one line>" --body "..." --verify`
-    : `↩ Reply: rig send ${senderLabel} "..."`;
+    : isHumanSeatSessionRef(personAddress)
+      ? `↩ Reply if needed: rig queue create --destination ${senderLabel} --human-intent update --summary "<one line>" --evidence-ref "<path>" --body "..." --verify`
+      : `↩ Reply: rig send ${senderLabel} "..."`;
   return [...header, "---", body, "---", reply].join("\n");
 }
 
@@ -236,11 +242,11 @@ export function sendCommand(depsOverride?: SendDeps): Command {
     .option("--to <sessions>", "Multi-recipient: comma-list or repeated (--to a,b or --to a --to b)", collectSessions, [] as string[])
     .option("--pod <name>", "Send to every seat in a pod (fan-out, per-recipient results)")
     .option("--rig <name>", "Send to every seat in a rig (fan-out, per-recipient results)")
-    .option("--verify", "Verify pane only delivery by checking content after send")
+    .option("--verify", "Verify pane only delivery by checking content after send. Exit 1 can mean unconfirmed, not failed: the message may have arrived, so inspect with rig capture before any resend")
     .option("--force", "Back-compat no-op: a mid-task/busy pane already sends-with-advisory by default; --force never bypasses the interactive-prompt/permission guard")
     .option("--wait-for-idle <seconds>", "Wait until the target is explicitly idle before sending. If it's still busy, or its activity is unknown, when the wait ends, nothing is sent (exit 1) and nothing is kept to send later; put separate work in a queue row (rig queue create)")
     .option("--raw", "Send exact text/keystrokes without the From/To messaging envelope (still guarded against interactive prompts)")
-    .option("--dangerously-interact", "DANGEROUS: type an answer as keystrokes into an interactive prompt (implies --raw; requires --reason). Letters can select menu shortcuts; CR/ESC act as keys. Enter is added only for the complete visibly staged answer.")
+    .option("--dangerously-interact", "DANGEROUS: type an answer as keystrokes into an interactive prompt (implies --raw; requires --reason). Letters can select menu shortcuts; CR/ESC act as keys. Enter is added for the complete visibly staged answer, or away from a recognized prompt when the input box can't be found or read.")
     .option("--type-without-checking", "With --dangerously-interact: type without reading or clearing existing input; may join or submit that text")
     .option("--reason <text>", "Why the prompt is being driven (required with --dangerously-interact; recorded in the audit log)")
     .option("--urgent", "Mark this message urgent: it types at seats whose interruption level holds other messages. Anyone may mark one; the recipient sees who did")
@@ -316,8 +322,10 @@ would run the envelope lines; the output says so. Use
 --dangerously-interact --reason "<why>" to DELIBERATELY drive a prompt (select an
 option or approve a permission) — the only override of the prompt guard; it
 implies --raw and is audit-logged. The answer is typed as keystrokes: letters may
-be menu shortcuts, and CR/ESC act as keys. An Enter is added only when the complete
-answer is visibly staged in a recognized text input. Combining --verify with
+be menu shortcuts, and CR/ESC act as keys. An Enter is added when the complete
+answer is visibly staged in a recognized text input, and, away from a recognized
+prompt or menu, when the input box can't be found or read (the result says so).
+Combining --verify with
 --dangerously-interact never adds an automatic submit retry, even when the daemon
 did not detect a prompt or did not report its disposition. Unknown consumption
 is not a send failure. On a managed seat with the delivery guard attached, a
@@ -326,7 +334,10 @@ pasted text stops the send and is shown only to you, separately from recognized
 dim suggestions. Collapsed paste contents and unrecognized inputs are not known.
 Inspect the native terminal, deliberately clear its prompt with native/tmux
 controls, then retry. Manual clearing runs outside OpenRig's lease and observation
-checks and is not recorded. --type-without-checking explicitly skips this read;
+checks and is not recorded. An input that can't be read doesn't stop the send:
+the answer is typed, nothing is cleared, the result says the input couldn't be
+read, and the answer may join or submit what is there.
+--type-without-checking explicitly skips this read;
 it does not clear anything and may join or submit existing text. Human typing
 can still race a capture and paste; the check is not atomic.
 
@@ -385,10 +396,10 @@ agent@rig@host is sugar for --host when the suffix is a REGISTERED host id
       }
       const isFanOut = fanModes.length === 1;
 
-      // Slice-03 Atom 6b: --context is single-seat LOCAL in v1 (fan-out and
-      // cross-host --context are follow-ons). Reject loudly rather than silently
-      // dropping the requested context.
-      if (opts.context && (isFanOut || opts.host)) {
+      // --context is single-seat LOCAL in v1. Reject fan-out here; the
+      // cross-host check follows target resolution so a self-qualified target
+      // can override a persisted remote selection.
+      if (opts.context && isFanOut) {
         console.error("--context is supported on a single-seat LOCAL send in v1 (not with --to/--pod/--rig or --host).");
         process.exitCode = 1;
         return;
@@ -461,14 +472,10 @@ agent@rig@host is sugar for --host when the suffix is a REGISTERED host id
       session = targetResolution.target;
       const crossHostHint = targetResolution.hint;
       if (targetResolution.warning) console.error(targetResolution.warning);
-      opts.host = explicitHost ?? targetResolution.sugarHost ?? opts.host;
+      opts.host = targetResolution.localHost ? undefined : explicitHost ?? targetResolution.sugarHost ?? opts.host;
 
-      // Atom 6b QA fix (root cause): re-reject --context on the cross-host path
-      // AFTER the agent@rig@host sugar host folds into opts.host. The early guard
-      // runs before resolveCrossHostTarget, so it cannot see the sugar host — the
-      // hole that let a sugar-form --context reach the remote argv (shipping a
-      // literal null with no message, or silently dropping the context with one).
-      // --context is single-seat LOCAL in v1; it is never handed to a remote send.
+      // Reject --context after resolving the target, before a remote send can
+      // silently drop the context or forward an empty message.
       if (opts.context && opts.host) {
         console.error("--context is supported on a LOCAL send in v1 (not with --host or an agent@rig@host cross-host target).");
         process.exitCode = 1;
@@ -565,7 +572,7 @@ agent@rig@host is sugar for --host when the suffix is a REGISTERED host id
         const envelope = !effect
           ? res.data
           : effectUnresolved(effect)
-            ? { ...res.data, verified: false, outcome: "staged-not-consumed", effectCheck: effect }
+            ? { ...res.data, verified: false, outcome: effectSubmissionUnconfirmed(effect) ? "submission-unconfirmed" : "staged-not-consumed", effectCheck: effect }
             : { ...res.data, effectCheck: effect };
         console.log(JSON.stringify(envelope));
         if (res.status >= 400) process.exitCode = res.status >= 500 ? 2 : 1;
@@ -632,7 +639,7 @@ agent@rig@host is sugar for --host when the suffix is a REGISTERED host id
  * recheck, then stops. Pure classification: no printing here.
  */
 type EffectCheck =
-  | { checked: true; state: "staged"; remedy: "submitted-cleared" | "submitted-still-staged" | "submit-refused"; detail?: string }
+  | { checked: true; state: "staged"; remedy: "submitted-cleared" | "submitted-still-staged" | "submit-refused" | "submit-unconfirmed"; detail?: string }
   | { checked: true; state: "no-staged-residual" }
   | { checked: false; why: string };
 
@@ -683,16 +690,26 @@ async function classifyDeliveryEffect(
   const probe = await detectStagedAtPrompt(client, session, identity);
   if (probe.state === "unchecked") return { checked: false, why: probe.why };
   if (probe.state === "not-staged") return { checked: true, state: "no-staged-residual" };
-  const enter = await client.post<Record<string, unknown>>("/api/transport/send", {
-    session,
-    submitOnly: true,
-    expectedStagedText: identity.expectedStagedText,
-    expectedStagedLineCount: identity.expectedLines,
-  }, transportRequestOptions(waitForIdleMs));
+  let enter: { status: number; data: Record<string, unknown> };
+  try {
+    enter = await client.post<Record<string, unknown>>("/api/transport/send", {
+      session,
+      submitOnly: true,
+      expectedStagedText: identity.expectedStagedText,
+      expectedStagedLineCount: identity.expectedLines,
+    }, transportRequestOptions(waitForIdleMs));
+  } catch (err) {
+    return { checked: true, state: "staged", remedy: "submit-unconfirmed", detail: (err as Error).message };
+  }
   if (enter.status >= 400) {
-    return { checked: true, state: "staged", remedy: "submit-refused", detail: (enter.data?.["error"] as string | undefined) ?? `HTTP ${enter.status}` };
+    const phase = enter.data?.["inputPhase"];
+    const reason = enter.data?.["reason"];
+    const unconfirmed = phase === "attempted" || phase === "written"
+      || (phase !== "not_reached" && enter.data?.["sent"] !== false && (reason === "submit_failed" || enter.status >= 500));
+    return { checked: true, state: "staged", remedy: unconfirmed ? "submit-unconfirmed" : "submit-refused", detail: (enter.data?.["error"] as string | undefined) ?? `HTTP ${enter.status}` };
   }
   const recheck = await detectStagedAtPrompt(client, session, identity);
+  if (recheck.state === "unchecked") return { checked: true, state: "staged", remedy: "submit-unconfirmed", detail: recheck.why };
   return recheck.state === "staged"
     ? { checked: true, state: "staged", remedy: "submitted-still-staged" }
     : { checked: true, state: "staged", remedy: "submitted-cleared" };
@@ -705,13 +722,17 @@ function effectHumanLines(effect: EffectCheck, session: string): string[] {
     return [`Note: the pane-effect check could not run (${effect.why}); the verdict above is transport-level only.`];
   }
   if (effect.state === "no-staged-residual") return [];
+  if (effect.remedy === "submit-unconfirmed") return [
+    `Delivery: submission unconfirmed — the text was staged before one guarded Enter attempt, but the result could not be confirmed (${effect.detail}). Inspect with: rig capture ${session}`,
+  ];
+  if (effect.remedy === "submit-refused") return [
+    `Delivery: unconfirmed — the guarded Enter was refused (${effect.detail}); the earlier staged observation may have changed. Inspect with: rig capture ${session}`,
+  ];
   const out = ["Delivery: staged, not consumed (checked: post-send pane capture; observed: the sent text is still at the prompt — typed, never submitted)"];
   if (effect.remedy === "submitted-cleared") {
     out.push("Remedy: one guarded Enter submitted — the staged text left the prompt.");
   } else if (effect.remedy === "submitted-still-staged") {
     out.push(`Remedy: one guarded Enter submitted, but the text is STILL at the prompt — not consumed; stopping here (one submit is the contract). Inspect with: rig capture ${session}`);
-  } else {
-    out.push(`Remedy: the single guarded Enter (submit path) was refused (${effect.detail}); stopping here — one submit is the contract. Inspect with: rig capture ${session}`);
   }
   return out;
 }
@@ -721,6 +742,10 @@ function effectUnresolved(effect: EffectCheck | undefined): boolean {
   return !!effect && effect.checked && effect.state === "staged" && effect.remedy !== "submitted-cleared";
 }
 
+function effectSubmissionUnconfirmed(effect: EffectCheck | undefined): boolean {
+  return !!effect && effect.checked && effect.state === "staged" && (effect.remedy === "submit-unconfirmed" || effect.remedy === "submit-refused");
+}
+
 /**
  * S3 (OPR.0.5.4.6) — the staged-at-prompt detector, walk's staged-evidence
  * primitive generalized to plain sends. POSITIVE evidence only: text rendered
@@ -754,7 +779,7 @@ async function detectStagedAtPrompt(
   for (let i = lines.length - 1; i >= 0; i--) {
     if (PROMPT_LINE_PATTERN.test(lines[i]!)) { lastPrompt = i; break; }
   }
-  if (lastPrompt === -1) return { state: "not-staged" };
+  if (lastPrompt === -1) return { state: "unchecked", why: "no input prompt was visible in the capture" };
   const inputRegionRaw = lines.slice(lastPrompt).join("\n").replace(PROMPT_LINE_PATTERN, "");
   // IDENTITY FIRST (round-2 F2, desk-binding): the literal-residual match uses
   // the daemon submit-precheck's own normalization (strip ALL whitespace,
@@ -1083,6 +1108,7 @@ async function runFanOutSend(params: {
   }
   const effectBySeat = new Map((effects ?? []).map((e) => [e.sessionName, e.effect]));
   const unresolvedCount = (effects ?? []).filter((e) => effectUnresolved(e.effect)).length;
+  const unconfirmedCount = (effects ?? []).filter((e) => effectSubmissionUnconfirmed(e.effect)).length;
 
   if (opts.json) {
     // Round-2 F1: ONE verdict per recipient — a staged-unresolved recipient's
@@ -1090,7 +1116,7 @@ async function runFanOutSend(params: {
     const rawResults = (res.data["results"] as Array<{ sessionName: string; ok: boolean; outcome?: string }> | undefined) ?? [];
     const encodedResults = rawResults.map((r) =>
       effectUnresolved(effectBySeat.get(r.sessionName))
-        ? { ...r, ok: false, verified: false, outcome: "staged-not-consumed", error: "staged, not consumed (pane effect); the one guarded submit did not clear it" }
+        ? { ...r, ok: false, verified: false, outcome: effectSubmissionUnconfirmed(effectBySeat.get(r.sessionName)) ? "submission-unconfirmed" : "staged-not-consumed", error: effectSubmissionUnconfirmed(effectBySeat.get(r.sessionName)) ? "delivery unconfirmed; inspect the recipient after the guarded submit attempt" : "staged, not consumed (pane effect); the one guarded submit did not clear it" }
         : r
     );
     // Round-3 (r2 row 00fb3a68): the machine-readable AGGREGATE derives from
@@ -1113,9 +1139,11 @@ async function runFanOutSend(params: {
   }
 
   const data = res.data;
-  const results = (data["results"] as Array<{ sessionName: string; ok: boolean; error?: string; outcome?: string; envelopeOmitted?: boolean; managedCompaction?: { notice?: unknown }; existingInput?: unknown }>) ?? [];
-  // Context beside a recipient's own result: it was in a managed compaction or restore when the send ran.
+  const results = (data["results"] as Array<{ sessionName: string; ok: boolean; error?: string; outcome?: string; envelopeOmitted?: boolean; managedCompaction?: { notice?: unknown }; existingInput?: unknown; warning?: string }>) ?? [];
+  // Context beside a recipient's own result: its own advisory (such as unreadable input), and whether
+  // it was in a managed compaction or restore when the send ran.
   const printNotice = (r: (typeof results)[number]) => {
+    if (r.warning) console.log(`${r.sessionName}: Advisory: ${r.warning}`);
     if (typeof r.managedCompaction?.notice === "string") console.log(`Notice: ${r.managedCompaction.notice}`);
   };
   // Round-2 F1 (desk-binding): ONE verdict line per recipient. A
@@ -1133,6 +1161,8 @@ async function runFanOutSend(params: {
     if (ec && ec.checked && ec.state === "staged") {
       if (ec.remedy === "submitted-cleared") {
         console.log(`${r.sessionName}: sent — staged at the prompt, cleared by one guarded Enter (consumed)`);
+      } else if (effectSubmissionUnconfirmed(ec)) {
+        console.log(`${r.sessionName}: ${effectHumanLines(ec, r.sessionName)[0]}`);
       } else {
         console.log(`${r.sessionName}: staged, not consumed — ${effectHumanLines(ec, r.sessionName).slice(-1)[0]}`);
       }
@@ -1145,7 +1175,7 @@ async function runFanOutSend(params: {
   }
   // The delivered count never includes a staged-unresolved recipient.
   const deliveredCount = Math.max(0, ((data["sent"] as number) ?? 0) - unresolvedCount);
-  console.log(`${deliveredCount}/${data["total"]} delivered${unresolvedCount > 0 ? `, ${unresolvedCount} staged-not-consumed` : ""}`);
+  console.log(`${deliveredCount}/${data["total"]} delivered${unresolvedCount - unconfirmedCount > 0 ? `, ${unresolvedCount - unconfirmedCount} staged-not-consumed` : ""}${unconfirmedCount > 0 ? `, ${unconfirmedCount} submission-unconfirmed` : ""}`);
   // S2 (OPR.0.5.4.3): surface additive advisories (e.g. the unknown-sender
   // sign-it notice) — an env-less operator must see it, not "sent" lines alone.
   const fanoutAdvisory = data["warning"] as string | undefined;

packages/cli/src/commands/slack.ts

@@ -122,7 +122,7 @@ const CHANNEL_MAP_APPLY =
 const POST_OPEN_ASKS = "`rig slack post-open-asks`";
 
 type OpenAskOutcome = "posted" | "already-posted" | "queued" | "held" | "below-floor" | "no-longer-active" | "superseded"
-  | "recipient-changed" | "state-unavailable" | "transport-failed" | "uncertain";
+  | "recipient-changed" | "no-connector" | "state-unavailable" | "transport-failed" | "uncertain";
 interface OpenAsk { qitemId: string; recipient: string; summary: string | null; outcome: OpenAskOutcome; detail?: string }
 interface PostOpenAsksResponse { ok: boolean; asks: OpenAsk[] }
 
@@ -135,6 +135,7 @@ const OUTCOME_LABEL: Record<OpenAskOutcome, string> = {
   "no-longer-active": "no longer active",
   superseded: "superseded",
   "recipient-changed": "recipient changed",
+  "no-connector": "no connector (Steering and the Feed only)",
   "state-unavailable": "state unavailable",
   "transport-failed": "transport failed",
   uncertain: "uncertain",
@@ -251,16 +252,21 @@ export function slackCommand(deps: SlackDeps = {}): Command {
         log(`Daemon: ${observation.state}${observation.reason ? ` (${observation.reason})` : ""}`);
         const connector = observation.connector as { configurationDigest?: string; inboundReady?: boolean; outboundReady?: boolean;
           inboundRouting?: { fallback: string; destination: string };
-          deadLetterBacklog?: number | null; deadLetterBacklogState?: string; deadLetterBacklogReason?: string;
+          settlement?: { pending: number; lastFailure: string | null };
+          deadLetterBacklog?: number | null; deadLetterBacklogState?: string; deadLetterBacklogReason?: string; relayAttemptsInFlight?: number;
           inbound?: InboundDeliveryStatus & { state?: string; generation?: number; lastEventAt?: string };
           recovery?: { state?: string; reason?: string; lastScanAt?: string; acceptedThisProcess?: number; deadLetteredThisProcess?: number;
             coverage?: { coverageStart: string; coveredThrough: string; pending?: { upper: string; nextLatest: string }; nextRetryAt?: number } | null;
             channels?: Array<{ channel?: string | null; state?: string; reason?: string; lastScanAt?: string;
               coverage?: { coverageStart: string; coveredThrough: string; nextRetryAt?: number } | null }>;
-            threads?: { state?: string; reason?: string; tracked?: number; scanned?: number; pending?: number; oldestPendingSeconds?: number | null;
+            threads?: { state?: string; reason?: string; tracked?: number; scanned?: number; pending?: number; settled?: number; settledRescanHours?: number; oldestPendingSeconds?: number | null;
               backoff?: Record<string, number>; incomplete?: Record<string, number> };
             limits?: string[] } } | undefined;
         if (connector) {
+          if (connector.settlement) {
+            log(`  Ask edits: ${connector.settlement.pending} queued for reconciliation`);
+            if (connector.settlement.lastFailure) log(`  Last ask-edit failure: ${connector.settlement.lastFailure}`);
+          }
           if (connector.inboundRouting?.fallback === "local") log(`  Configured routing fallback: local to ${connector.inboundRouting.destination} (after forwarding, thread and channel routes)`);
           log(`  Socket: ${connector.inbound?.state ?? "unknown"}; generation ${connector.inbound?.generation ?? "unknown"}; last event ${connector.inbound?.lastEventAt ?? "unknown"}`);
           for (const line of socketDeliveryLines(connector.inbound)) log(`  ${line}`);
@@ -283,9 +289,11 @@ export function slackCommand(deps: SlackDeps = {}): Command {
             const behind = threads.oldestPendingSeconds != null ? ` (oldest ${Math.round(threads.oldestPendingSeconds / 60)} min behind)` : "";
             log(`  Thread recovery: ${threads.state ?? "unknown"}${threads.reason ? ` (${threads.reason})` : ""}; ${threads.tracked ?? 0} tracked (open, or closed within seven days); `
               + `${threads.scanned ?? 0} scanned, ${threads.pending ?? 0} waiting their turn${behind}`
+              + `${threads.settled ? `, ${threads.settled} settled (nobody waiting; rescanned ${!threads.settledRescanHours || threads.settledRescanHours === 24 ? "daily" : `every ${threads.settledRescanHours} hours`})` : ""}`
               + `${reasons(threads.backoff) ? `; backing off: ${reasons(threads.backoff)}` : ""}${reasons(threads.incomplete) ? `; incomplete: ${reasons(threads.incomplete)}` : ""}`);
           }
           log(`  Recovery counts since connector start: accepted ${recovery?.acceptedThisProcess ?? "unknown"}; dead-lettered ${recovery?.deadLetteredThisProcess ?? "unknown"} (custody, not delivery)`);
+          if (connector.relayAttemptsInFlight !== undefined) log(`  Relay attempts awaiting response: ${connector.relayAttemptsInFlight} (this process, point-in-time; zero does not prove global quiescence)`);
           if (connector.deadLetterBacklogState !== undefined) {
             const scope = "inbound messages, reactions and click answers";
             if (connector.deadLetterBacklogState === "unknown") {

packages/cli/src/commands/stream.ts

@@ -1,6 +1,8 @@
 import { Command } from "commander";
-import { DaemonClient } from "../client.js";
+import { DaemonClient, DaemonResponseError, DaemonTimeoutError } from "../client.js";
 import { getDaemonStatus, getDaemonUrl , daemonStatusGuard} from "../daemon-lifecycle.js";
+import { readOpenRigEnv } from "../openrig-compat.js";
+import { shellQuote } from "../cross-host-executor.js";
 import { realDeps } from "./daemon.js";
 import type { StatusDeps } from "./status.js";
 import { streamOutcome } from "./stream-outcome.js";
@@ -42,6 +44,21 @@ function printResult(json: boolean, body: unknown, status: number): void {
   if (status >= 400) process.exitCode = status >= 500 ? 2 : 1;
 }
 
+/**
+ * Added to an emit whose reply came too late or unreadable: the item may still have landed, and the
+ * daemon keeps one item per stream item ID, so only a retry with the same ID can't add a second one.
+ */
+function emitRetryGuidance(id: string | undefined, source: string, sentAt: number): string {
+  if (id !== undefined) {
+    return `The item may still land. To retry safely, repeat the same emit with --id ${shellQuote(id)}; the daemon keeps one item per ID. To check first: rig stream show ${shellQuote(id)}.`;
+  }
+  // The daemon records the seat the request came from (its seat env), not --source, when the two differ.
+  const recorded = readOpenRigEnv("OPENRIG_SESSION_NAME", "RIGGED_SESSION_NAME")?.trim() || source;
+  // Five minutes before the send, for a daemon whose clock runs behind this one.
+  const since = new Date(sentAt - 5 * 60_000).toISOString();
+  return `The item may still land, and this emit had no --id, so emitting again can add a second item. Check first: rig stream list --source ${shellQuote(recorded)} --since ${since}. Pass a stable --id on emits so a later retry is safe.`;
+}
+
 function isWatchedStreamItem(value: unknown): value is WatchedStreamItem {
   if (!value || typeof value !== "object") return false;
   const item = value as Record<string, unknown>;
@@ -104,7 +121,7 @@ export function streamCommand(depsOverride?: StreamDeps): Command {
     .description("Append a stream item")
     .requiredOption("--source <session>", "Source session (e.g. dev-build@starter)")
     .requiredOption("--body <text>", "Stream item body")
-    .option("--hint-destination <session>", "Hint at intended destination seat")
+    .option("--hint-destination <session>", "Hint at intended destination seat (the queue worker routes by this hint)")
     .option("--hint-type <type>", "Hint type (e.g. review, handoff, idea)")
     .option("--hint-urgency <urgency>", "Hint urgency (routine, urgent, critical)")
     .option("--hint-tags <tags>", "Comma-separated hint tags")
@@ -112,6 +129,11 @@ export function streamCommand(depsOverride?: StreamDeps): Command {
     .option("--interrupt", "Mark item as interrupting")
     .option("--id <streamItemId>", "Idempotent stream_item_id (skip if not provided)")
     .option("--json", "JSON output for agents")
+    .addHelpText("after", `
+Stream intake is on by default: the seat named by stream.intake (queue.worker unless set)
+gets an intake row for each item, which wakes it, or a note on its intake row still pending.
+--hint-urgency critical or --interrupt makes that row urgent. The reply names only the stream item.
+To keep emits passive: rig config set stream.intake off`)
     .action(async (opts: {
       source: string;
       body: string;
@@ -127,6 +149,7 @@ export function streamCommand(depsOverride?: StreamDeps): Command {
       const deps = getDeps();
       const tags = opts.hintTags ? opts.hintTags.split(",").map((s) => s.trim()).filter(Boolean) : null;
       await withClient(deps, async (client) => {
+        const sentAt = Date.now();
         const res = await client.post<Record<string, unknown>>("/api/stream/emit", {
           streamItemId: opts.id,
           sourceSession: opts.source,
@@ -137,6 +160,12 @@ export function streamCommand(depsOverride?: StreamDeps): Command {
           hintUrgency: opts.hintUrgency ?? null,
           hintTags: tags,
           interrupt: opts.interrupt ?? false,
+        }).catch((error: unknown): never => {
+          // The shared renderer reads both as an UNKNOWN outcome; name the retry that can't duplicate.
+          if (error instanceof DaemonTimeoutError || error instanceof DaemonResponseError) {
+            error.message += ` ${emitRetryGuidance(opts.id, opts.source, sentAt)}`;
+          }
+          throw error;
         });
         printResult(opts.json ?? false, res.data, res.status);
       });

packages/cli/src/commands/terminal.ts

@@ -24,10 +24,39 @@ import { realDeps } from "./daemon.js";
 import type { StatusDeps } from "./status.js";
 import { openTerminalWindow, type WindowDeps } from "../terminal-window.js";
 import { assertLocalTmux, attachTmux, isLocalTmuxHost, readTmuxServerArgs, type TmuxAttachment } from "../tmux-attachment.js";
+import { existsSync } from "node:fs";
+import { join, resolve } from "node:path";
+import { homedir } from "node:os";
+
+/** herdr-app's home, resolved the way herdr-app resolves it (as the TUI does). */
+function herdrAppHome(env: NodeJS.ProcessEnv): string {
+  return env["HERDR_APP_HOME"] || join(env["XDG_CONFIG_HOME"] || join(env["HOME"] || homedir(), ".config"), "herdr-app");
+}
 
 export interface TerminalDeps extends StatusDeps {
   windowDeps?: WindowDeps;
   attachTmux?: (target: TmuxAttachment) => Promise<number>;
+  /** This process's environment (herdr-app context); injectable for tests. */
+  env?: NodeJS.ProcessEnv;
+  socketExists?: (path: string) => boolean;
+}
+
+/**
+ * The herdr-app client a command addresses, or null when it doesn't use herdr-app: `--provider herdr-app` opens the
+ * view in the herdr-app window you run it in. Inside a herdr-app terminal, plain `rig terminal open <view>` does the
+ * same when that window is answering (its socket file is a hint; the window is asked before it is chosen). `--window`
+ * always asks for a new desktop window.
+ */
+export function herdrAppClient(opts: { provider?: string; window?: boolean; appHome?: string }, deps: TerminalDeps):
+  { appHome: string; ambient: boolean } | null {
+  if (opts.window) return null;
+  const env = deps.env ?? process.env;
+  const named = opts.provider?.trim();
+  const exists = deps.socketExists ?? existsSync;
+  const ambient = !named
+    && !!env["HC_TERMINAL_ID"] && !!env["HC_TERMINAL_SOCKET"] && exists(join(herdrAppHome(env), "herdr-app.sock"));
+  if (named !== "herdr-app" && !ambient) return null;
+  return { appHome: opts.appHome ? resolve(opts.appHome) : herdrAppHome(env), ambient };
 }
 export interface FollowControls {
   enabled: boolean; maxTiles?: number; graceMs?: number; focusRetentionMs?: number;
@@ -121,7 +150,7 @@ function printOpen(json: boolean, r: OpenViewResult, status: number): void {
 export function terminalCommand(depsOverride?: TerminalDeps): Command {
   const cmd = new Command("terminal").description(
     "Open the OpenRig view and agent conversations in terminal tiles",
-  ).addHelpText("after", "\nThe OpenRig TUI and operator (after install, and whenever the person wants their agents):\n  rig terminal open saved:kernel --window\n  Opens the window itself on the daemon's desktop; an agent can run it from its shell.\n  Inside Rex it opens a Rex session and shows it in the Rex app; otherwise it uses herdr when installed, else the same layout in plain tmux.\n  Herdr needs a terminal the person can see; TUI navigation does not open one.\n  Over headless SSH, give the exact connection/attach command for a new terminal/tab.\n  Only if the window cannot open: rig tui --shared is the dashboard-only fallback.\n  On a desktop, the agent opens the view; it does not finish by printing a command to copy.\n");
+  ).addHelpText("after", "\nThe OpenRig TUI and operator (after install, and whenever the person wants their agents):\n  rig terminal open saved:kernel --window\n  Opens the window itself on the daemon's desktop; an agent can run it from its shell.\n  Inside a herdr-app terminal, omit --window to open in that window.\n  Inside Rex it opens a Rex session and shows it in the Rex app; otherwise it uses herdr when installed, else the same layout in plain tmux.\n  Herdr needs a terminal the person can see; TUI navigation does not open one.\n  Over headless SSH, give the exact connection/attach command for a new terminal/tab.\n  Only if the window cannot open: rig tui --shared is the dashboard-only fallback.\n  On a desktop, the agent opens the view; it does not finish by printing a command to copy.\n");
 
   const getDeps = (): TerminalDeps =>
     depsOverride ?? {
@@ -133,11 +162,12 @@ export function terminalCommand(depsOverride?: TerminalDeps): Command {
     .command("open")
     .argument("<view>", "a rig name, pod:<rig>/<pod>, seat:<rig>/<logical-id>, mission:<id>, slice:<id>, qitem:<id> (current local obligation owners), or a saved-view id")
     .description("Open a desktop terminal showing the view's live agents as interactive tiles")
-    .option("--provider <name>", "herdr, cmux or rex: open in the running provider without a desktop window; each open adds a new workspace (inside Herdr, --window reuses a matching one); tmux requires --window. Inside Rex, the default opens in Rex")
+    .option("--provider <name>", "herdr, herdr-app, cmux or rex: open in the running provider without a desktop window; each open adds a new workspace (inside Herdr, --window reuses a matching one), and herdr-app opens a Rig view in the herdr-app window this runs in, which must be answering; tmux requires --window. Inside Rex, the default opens in Rex")
     .option("--session <name>", "Herdr session on the daemon's host; --session default selects Herdr's default session")
-    .option("--window", "Open a new desktop terminal tab/window (the default when --provider is omitted)")
+    .option("--window", "Open a new desktop terminal tab/window (the default when --provider is omitted outside herdr-app); inside a herdr-app terminal --window still asks for a desktop window, and plain open reuses herdr-app")
+    .option("--app-home <dir>", "herdr-app: the client to open in (default: herdr-app's own home, from HERDR_APP_HOME or ~/.config/herdr-app)")
     .option("--expected-plan <id>", "Open only if the view still matches this preview")
-    .option("--follow", "Keep one qitem:<id>, slice:<id-or-name> or mission:<id> workspace following current obligation owners (Linux Herdr 0.9.5/protocol 22)")
+    .option("--follow", "Keep one qitem:<id>, slice:<id-or-name> or mission:<id> workspace following current obligation owners (Linux/macOS Herdr protocol 22)")
     .option("--no-follow", "Stop following and keep the existing workspace and tiles as a fixed layout")
     .option("--max-tiles <count>", "Following tile limit, 1–256; explicitly raise it to show more members", Number)
     .option("--pin <seats...>", "Retain these following tiles after they leave the task")
@@ -148,7 +178,7 @@ export function terminalCommand(depsOverride?: TerminalDeps): Command {
     .option("--grace-ms <ms>", "Departure grace, default 4000, at most 60000", Number)
     .option("--focus-retention-ms <ms>", "Extra focused-reading retention, default 12000, at most 300000", Number)
     .option("--json", "JSON output for agents")
-    .action(async (view: string, opts: { provider?: string; session?: string; json?: boolean; window?: boolean; expectedPlan?: string;
+    .action(async (view: string, opts: { provider?: string; session?: string; json?: boolean; window?: boolean; expectedPlan?: string; appHome?: string;
       follow?: boolean; maxTiles?: number; pin?: string[]; unpin?: string[]; suppress?: string[]; unsuppress?: string[];
       suppressionMs?: number; graceMs?: number; focusRetentionMs?: number }) => {
       const deps = getDeps();
@@ -157,12 +187,51 @@ export function terminalCommand(depsOverride?: TerminalDeps): Command {
       const hasControls = [maxTiles, pin, unpin, suppress, unsuppress, suppressionMs, graceMs, focusRetentionMs].some(v => v !== undefined);
       const following: FollowControls | undefined = opts.follow !== undefined || hasControls
         ? { enabled: opts.follow !== false, maxTiles, pin, unpin, suppress, unsuppress, suppressionMs, graceMs, focusRetentionMs } : undefined;
+      const herdrApp = herdrAppClient(opts, deps);
       await withClient(deps, async (client) => {
+        // An ambient herdr-app terminal does not claim --follow; only an explicit --provider herdr-app errors with follow_unsupported.
+        let selectedHerdrApp = herdrApp?.ambient && following ? null : herdrApp;
+        let ambientFallbackNote: string | undefined;
+        if (selectedHerdrApp?.ambient) {
+          const probe = await client.get<{
+            available?: boolean;
+            launch?: { socketPath?: string };
+            providers?: Array<{ status?: { available?: boolean; launch?: { socketPath?: string } } }>;
+          }>(
+            `/api/terminal/status?provider=herdr-app&appHome=${encodeURIComponent(selectedHerdrApp.appHome)}`,
+          );
+          const report = probe.data?.providers?.[0]?.status ?? probe.data;
+          if (probe.status >= 400 || typeof report?.available !== "boolean") {
+            const socketPath = report?.launch?.socketPath ?? join(selectedHerdrApp.appHome, "herdr-app.sock");
+            ambientFallbackNote = `herdr-app's socket exists at ${socketPath} but the daemon could not confirm whether it is answering (HTTP ${probe.status}); opening a desktop window instead. Use --provider herdr-app to require it.`;
+            console.error(ambientFallbackNote);
+            selectedHerdrApp = null;
+          } else if (!report.available) {
+            const socketPath = report.launch?.socketPath ?? join(selectedHerdrApp.appHome, "herdr-app.sock");
+            ambientFallbackNote = `herdr-app's socket exists at ${socketPath} but it isn't answering; opening a desktop window instead. Use --provider herdr-app to require it.`;
+            console.error(ambientFallbackNote);
+            selectedHerdrApp = null;
+          }
+        }
+        if (selectedHerdrApp) {
+          // herdr-app: a fresh Rig view in that client's own window, through the daemon. No desktop window.
+          if (following) { printResult(opts.json ?? false, { ok: false, provider: "herdr-app", code: "follow_unsupported", error: "Following isn't available in herdr-app yet; open the view without --follow." }, 400); return; }
+          const { appHome } = selectedHerdrApp;
+          const res = await client.post<OpenViewResult>("/api/terminal/open", { view, provider: "herdr-app",
+            herdrApp: { appHome },
+            ...(opts.expectedPlan !== undefined ? { expectedPlan: opts.expectedPlan } : {}) }, { timeoutMs: TERMINAL_OPEN_TIMEOUT_MS });
+          if (!Array.isArray(res.data?.opened)) { printResult(opts.json ?? false, res.data, res.status); return; }
+          printOpen(opts.json ?? false, res.data, res.status);
+          return;
+        }
         // Herdr's control socket can answer with no desktop client attached.
         // The default requests a desktop; an explicit provider reuses its workspace.
+        // --window never selects herdr-app
         if (following?.enabled !== false && (opts.window || !opts.provider)) {
-          const result = await openTerminalWindow(client, view, opts.provider, deps.windowDeps, opts.expectedPlan, session, following);
-          printOpen(opts.json ?? false, result, 200);
+          const windowProvider = opts.provider?.trim() === "herdr-app" ? undefined : opts.provider;
+          const result = await openTerminalWindow(client, view, windowProvider, deps.windowDeps, opts.expectedPlan, session, following);
+          const out = ambientFallbackNote ? { ...result, notes: [...(result.notes ?? []), ambientFallbackNote] } : result;
+          printOpen(opts.json ?? false, out, 200);
           return;
         }
         let expectedPlan = opts.expectedPlan;
@@ -203,18 +272,38 @@ export function terminalCommand(depsOverride?: TerminalDeps): Command {
       });
     });
 
+  cmd
+    .command("close <view>")
+    .description("Close a herdr-app Rig view (rv-…): its cells leave the layout, then only its own viewers end; seats keep running")
+    .option("--app-home <dir>", "herdr-app home the view is in (default: herdr-app's own)")
+    .option("--json", "JSON output for agents")
+    .action(async (view: string, opts: { appHome?: string; json?: boolean }) => {
+      const deps = getDeps();
+      const env = deps.env ?? process.env;
+      await withClient(deps, async (client) => {
+        const res = await client.post<OpenViewResult>("/api/terminal/close", { view, provider: "herdr-app",
+          herdrApp: { appHome: opts.appHome ? resolve(opts.appHome) : herdrAppHome(env) } }, { timeoutMs: TERMINAL_OPEN_TIMEOUT_MS });
+        if (opts.json) { printResult(true, res.data, res.status); return; }
+        console.log(res.data?.ok ? `Closed Rig view ${view}.` : `Not closed: ${res.data?.error ?? `HTTP ${res.status}`}`);
+        for (const note of res.data?.notes ?? []) console.log(`  ${note}`);
+        if (res.status >= 400) process.exitCode = 1;
+      });
+    });
+
   cmd
     .command("status")
     .description("Show terminal provider availability + liveness (doctor)")
-    .option("--provider <name>", "restrict to one provider (herdr / cmux / rex)")
+    .option("--provider <name>", "restrict to one provider (herdr / herdr-app / cmux / rex)")
     .option("--session <name>", "Inspect a named Herdr session on the daemon's host")
+    .option("--app-home <dir>", "herdr-app: the client to inspect (default: herdr-app's own home)")
     .option("--json", "JSON output for agents")
-    .action(async (opts: { provider?: string; session?: string; json?: boolean }) => {
+    .action(async (opts: { provider?: string; session?: string; appHome?: string; json?: boolean }) => {
       const deps = getDeps();
       await withClient(deps, async (client) => {
         let path = opts.provider
           ? `/api/terminal/status?provider=${encodeURIComponent(opts.provider)}`
           : "/api/terminal/status";
+        if (opts.provider === "herdr-app") path += `&appHome=${encodeURIComponent(opts.appHome ? resolve(opts.appHome) : herdrAppHome(deps.env ?? process.env))}`;
         const session = opts.session?.trim();
         if (session !== undefined) path += `${opts.provider ? "&" : "?"}session=${encodeURIComponent(session)}`;
         const res = await client.get<{ session?: string; error?: string }>(path);

packages/cli/src/commands/up.ts

@@ -11,6 +11,7 @@ import { getDaemonStatus, getDaemonUrl, startDaemon, type LifecycleDeps, daemonS
 import { shouldAutoStartDaemon } from "../daemon-reachability.js";
 import { prepareDaemonAutoStart } from "../daemon-auto-start.js";
 import { freshSeatCommand, launchFreshSeat } from "../fresh-seat-command.js";
+import { launchStatusRunning } from "./launch.js";
 import type { StartOptions } from "../daemon-lifecycle.js";
 import { realDeps } from "./daemon.js";
 import type { StatusDeps } from "./status.js";
@@ -120,6 +121,17 @@ snapshot, resuming seats where their harness allows; it refuses while any of its
 whose rig name already exists starts a new team in place of the stopped one, which
 is archived, and refuses if the old one may still be running. To bring a stopped
 team back with its conversations, run rig up <rig-name> --existing, not the spec.
+
+After a restore, rig up exits 0 only when the rig came back whole (rigResult
+fully_restored). It exits 1 when it didn't (partially_restored, failed or
+not_attempted), including when every seat is running but some started fresh
+without their conversation, or came back with a guidance or restore gap. With
+--json, read rigResult and each node's status, noConversation, guidanceGaps and
+restoreGaps to tell these apart: a seat with a gap keeps its own status, such as
+resumed, and a fresh seat marked noConversation (a terminal seat) lost nothing.
+A refused request exits 1 for a conflict (HTTP 409, such as a rig still running
+or a restore not attempted) and 2 for any other refusal (400, 404 or 500, such
+as no rig by that name or a failed boot).
 `);
   const getDepsF = () => depsOverride ?? { lifecycleDeps: realDeps(), clientFactory: (url: string) => new DaemonClient(url) };
 
@@ -706,7 +718,7 @@ team back with its conversations, run rig up <rig-name> --existing, not the spec
         }
         console.log(`Rig "${rigName ?? rigId}" restored (ID: ${rigId})`);
         if (rigResult) console.log(`Result: ${rigResult}`);
-        const nodes = (res.data["nodes"] as Array<{ logicalId: string; status: string; error?: string }>) ?? [];
+        const nodes = (res.data["nodes"] as Array<{ logicalId: string; status: string; error?: string; noConversation?: true; guidanceGaps?: string[]; restoreGaps?: string[] }>) ?? [];
         for (const n of nodes) {
           // OPR.0.3.4.2 — the five-term vocabulary renders distinctly; the
           // awaiting-decision reason is part of the line (not a dead end), and so is a failed seat's (such as a seat held
@@ -717,6 +729,16 @@ team back with its conversations, run rig up <rig-name> --existing, not the spec
             console.log(`  ${n.logicalId}: ${n.status}`);
           }
         }
+        // Every seat running still exits 1 when the restore isn't whole; say why, so nobody retries a start that worked.
+        if (rigResult === "partially_restored" && nodes.length > 0 && nodes.every((n) => launchStatusRunning(n.status))) {
+          const fresh = nodes.filter((n) => (n.status === "fresh" || n.status === "fresh-primed") && !n.noConversation).length;
+          const gapped = nodes.filter((n) => (n.guidanceGaps?.length ?? 0) > 0 || (n.restoreGaps?.length ?? 0) > 0).map((n) => n.logicalId);
+          const reasons = [
+            ...(fresh > 0 ? [`${fresh} started fresh without their conversation`] : []),
+            ...(gapped.length > 0 ? [`${gapped.join(", ")} came back with a gap named in a warning`] : []),
+          ];
+          console.log(`${nodes.length === 1 ? "The one seat is" : `All ${nodes.length} seats are`} running${reasons.length > 0 ? `; ${reasons.join("; ")}` : ""}. The restore isn't whole, so rig up exits 1.`);
+        }
         const warnings = (res.data["warnings"] as string[]) ?? [];
         for (const w of warnings) {
           console.error(`  warning: ${w}`);

packages/cli/src/commands/walk.ts

@@ -273,7 +273,7 @@ initial record is explicitly reported as unverified delivery.`)
 
             let verdict = await pollConsumed();
             if (verdict === "generation-rolled") {
-              failPiece(i, piece.label, "the seat's generation rolled mid-walk (a re-prime); the walk cannot continue into a different generation.");
+              failPiece(i, piece.label, "the seat's generation rolled mid-walk (a re-prime); delivery is unconfirmed. Inspect the recipient before any resend; the walk cannot continue into a different generation.");
               return;
             }
             if (verdict === "timeout") {
@@ -289,23 +289,39 @@ initial record is explicitly reported as unverified delivery.`)
                 /\[Pasted text #\d+ \+\d+ lines\]/.test(pane)
               );
               if (stagedEvidence) {
-                const enter = await client.post<Record<string, unknown>>("/api/transport/send", {
-                  session: seat,
-                  submitOnly: true,
-                  expectedStagedText: piece.content, // FULL bytes — the transport checks the rendered literal residual for contiguous containment
-                  expectedStagedLineCount: piece.content.split("\n").length,
-                }, { headers: terminalAuthHeaders() });
+                let enter: { status: number; data: Record<string, unknown> };
+                try {
+                  enter = await client.post<Record<string, unknown>>("/api/transport/send", {
+                    session: seat,
+                    submitOnly: true,
+                    expectedStagedText: piece.content, // FULL bytes — the transport checks the rendered literal residual for contiguous containment
+                    expectedStagedLineCount: piece.content.split("\n").length,
+                  }, { headers: terminalAuthHeaders() });
+                } catch (error) {
+                  failPiece(i, piece.label, `the guarded submit result is unconfirmed (${(error as Error).message}); inspect the recipient before any resend.`);
+                  return;
+                }
                 if (enter.status >= 400) {
-                  failPiece(i, piece.label, `typed but not consumed; the single submit retry was refused (${(enter.data?.["error"] as string | undefined) ?? `HTTP ${enter.status}`}).`);
+                  const phase = enter.data?.["inputPhase"], reason = enter.data?.["reason"];
+                  const unconfirmed = phase === "attempted" || phase === "written"
+                    || (phase !== "not_reached" && enter.data?.["sent"] !== false && (reason === "submit_failed" || enter.status >= 500));
+                  const detail = (enter.data?.["error"] as string | undefined) ?? `HTTP ${enter.status}`;
+                  failPiece(i, piece.label, unconfirmed
+                    ? `the guarded submit result is unconfirmed (${detail}); inspect the recipient before any resend.`
+                    : `the single guarded submit was refused (${detail}); the earlier staged observation may have changed. Inspect the recipient before any resend.`);
                   return;
                 }
                 verdict = await pollConsumed();
+                if (verdict === "generation-rolled") {
+                  failPiece(i, piece.label, "the seat's generation rolled after the guarded submit; delivery is unconfirmed. Inspect the recipient before any resend.");
+                  return;
+                }
                 if (verdict !== "consumed") {
-                  failPiece(i, piece.label, `typed and staged, but not consumed even after the single submit retry — the piece never entered the seat's conversation record. Not re-sending (one retry is the contract).`);
+                  failPiece(i, piece.label, "no matching conversation record was observed within the verification window after the single submit retry; delivery is unconfirmed. Inspect the recipient before any resend.");
                   return;
                 }
               } else {
-                failPiece(i, piece.label, `send reported ${sendOutcome === "ok" ? "success" : "a transport error"} but the piece is neither consumed in the generation record nor staged in the pane — delivery lost; consumption verification fails this walk closed.`);
+                failPiece(i, piece.label, `send reported ${sendOutcome === "ok" ? "success" : "a transport error"}, but no matching record or staged input was observed within the verification window. Delivery and consumption are unconfirmed; inspect the recipient before any resend.`);
                 return;
               }
             }

packages/cli/src/commands/workflow-errors.ts

@@ -72,7 +72,7 @@ export function describeDaemonRejection(body: unknown, instanceHint?: string): T
       return {
         fact: "The failed step no longer exists in the cached workflow spec.",
         consequence: "The spec drifted after this instance failed; rebinding would route a step the workflow no longer declares.",
-        action: "Restore the step in the spec (re-validate to refresh the cache) or instantiate a fresh run.",
+        action: "Restore the step in the spec (then re-register it through the workflows-folder scan or an instantiate by path to refresh the cache) or instantiate a fresh run.",
       };
     case "instance_version_conflict":
       return {

packages/cli/src/commands/workflow-render.ts

@@ -30,6 +30,8 @@ export interface RenderInstance {
   hopCount?: number;
   /** OPR.0.4.6.FAC1: the instance's bound rig (API-carried; null/absent = unbound, renders nothing). */
   boundRig?: string | null;
+  /** #2730: what the run is for, as its `--root-objective` gave it; null for an instance from before it was kept. */
+  rootObjective?: string | null;
   /**
    * WF-1 FR-2's API-carried classification (present since the
    * completion fixback at 384e60f1). RAIL 1 (arch ruling): the CLI
@@ -181,6 +183,7 @@ export function renderTraceTree(
   const lines: string[] = [];
   const name = instance.workflowName ? `${instance.workflowName}${instance.workflowVersion ? ` v${instance.workflowVersion}` : ""}` : "";
   lines.push(`${statusGlyph(instance.status)} ${instance.instanceId}  ${name}  status=${instance.status}${instance.hopCount !== undefined ? `  hops=${instance.hopCount}` : ""}${instance.boundRig ? `  rig=${instance.boundRig}` : ""}`);
+  if (instance.rootObjective) lines.push(`  objective: ${instance.rootObjective}`);
   if (instance.createdAt) {
     lines.push(`  created ${instance.createdAt}${instance.createdBySession ? ` by ${instance.createdBySession}` : ""}  age ${humanAge(instance.createdAt, nowIso)}`);
   }
@@ -254,6 +257,8 @@ function isStuck(instance: RenderInstance): boolean {
 export interface AttentionRow {
   instanceId: string;
   workflowName: string;
+  /** #2730: what the run is for; null when the instance doesn't record it. */
+  rootObjective: string | null;
   /** All classes for this instance — exactly-once row, combined. */
   classes: string[];
   reasons: string[];
@@ -320,6 +325,7 @@ export function composeAttentionRollup(instances: RenderInstance[]): AttentionRo
     attention.push({
       instanceId: inst.instanceId,
       workflowName: inst.workflowName ?? "",
+      rootObjective: inst.rootObjective ?? null,
       classes,
       reasons,
       affordance,
@@ -345,6 +351,7 @@ export function renderStatus(rollup: AttentionRollup): string[] {
     lines.push(
       `${fitCell("▲", 2)}${fitCell(row.instanceId, 30)}${fitCell(row.workflowName, 20)}${fitCell(row.classes.join("+"), 16)}${row.reasons.join("; ")}`,
     );
+    if (row.rootObjective) lines.push(`  ${fitCell("", 2)}  objective: ${row.rootObjective}`);
     lines.push(`  ${fitCell("", 2)}└ ${row.affordance}`);
   }
   return lines;
@@ -374,6 +381,7 @@ export function renderInstanceShow(
   const lines: string[] = [];
   lines.push(`${statusGlyph(instance.status)} ${instance.instanceId}  status=${instance.status}`);
   if (instance.workflowName) lines.push(`  workflow: ${instance.workflowName}${instance.workflowVersion ? ` v${instance.workflowVersion}` : ""}`);
+  if (instance.rootObjective) lines.push(`  objective: ${instance.rootObjective}`);
   // OPR.0.4.6.FAC1: the bound rig renders when present (unbound rows unchanged).
   if (instance.boundRig) lines.push(`  rig:      ${instance.boundRig}`);
   if (instance.createdAt) lines.push(`  created:  ${instance.createdAt}${instance.createdBySession ? ` by ${instance.createdBySession}` : ""}  (age ${humanAge(instance.createdAt, nowIso)})`);

packages/cli/src/commands/workflow.ts

@@ -172,7 +172,7 @@ export function workflowCommand(depsOverride?: WorkflowDeps): Command {
 
   cmd
     .command("validate <specPath>")
-    .description("Validate a workflow spec file (returns structured ok/error report)")
+    .description("Validate a workflow spec file in memory, without registering it (returns structured ok/error report)")
     .option("--json", "JSON output for agents")
     .addHelpText("after", `
 Examples:
@@ -637,19 +637,86 @@ exception routes only. Existing obligations keep their owners and evidence.
     });
 
   cmd
-    .command("trace <instanceId>")
-    .description("Show one workflow instance + its append-only step trail (audit-only verdict)")
+    .command("unregister [spec]")
+    .description("Take a workflow spec out of the registry: <name> or <name>@<version>, or --prune-missing for versions whose file is gone")
+    .option("--prune-missing", "Remove every registered version whose source file no longer exists (scoped to [spec] when given)")
     .option("--json", "JSON output for agents")
     .addHelpText("after", `
+A version that unfinished work (an active, waiting or resumable failed instance) is pinned to
+is never removed; with --prune-missing it is retained, readable by that work but hidden from
+the library, and a later prune removes it. A file still inside the workflows folder or the
+built-in starters is kept, because the next scan registers it again: move or delete the file.
+A file whose presence can't be checked (an access error, not "no such file") is kept and named.
+An empty version (name@) is a usage error.
+
+Examples:
+  $ rig workflow unregister qa-probe
+  $ rig workflow unregister pedido-do-board@1
+  $ rig workflow unregister --prune-missing
+  $ rig workflow unregister --prune-missing --json | jq '.removed[] | "\\(.name)@\\(.version) \\(.sourcePath)"'
+`)
+    .action(async (spec: string | undefined, opts: { pruneMissing?: boolean; json?: boolean }) => {
+      if (!spec && !opts.pruneMissing) {
+        console.error("Name a workflow spec (<name> or <name>@<version>) or pass --prune-missing.");
+        process.exitCode = 1;
+        return;
+      }
+      let name: string | undefined;
+      let version: string | undefined;
+      if (spec) {
+        const at = spec.lastIndexOf("@");
+        name = at > 0 ? spec.slice(0, at) : spec;
+        version = at > 0 ? spec.slice(at + 1) : undefined;
+      }
+      if (version !== undefined && version.trim() === "") {
+        console.error(`The version after @ is empty in "${spec}": use <name>@<version>, or <name> for every version.`);
+        process.exitCode = 1;
+        return;
+      }
+      const deps = getDeps();
+      await withClient(deps, async (client) => {
+        const res = await client.post<{
+          removed?: Array<{ name: string; version: string; sourcePath: string }>;
+          kept?: Array<{ name: string; version: string; sourcePath: string; reason: "pinned" | "scanned" | "unchecked" }>;
+        }>("/api/workflow/specs/unregister", { name, version, pruneMissing: opts.pruneMissing === true });
+        if (opts.json || res.status >= 400) {
+          printResult(opts.json ?? false, res.data, res.status);
+          return;
+        }
+        const removed = res.data.removed ?? [];
+        const kept = res.data.kept ?? [];
+        if (removed.length === 0 && kept.length === 0) console.log("Nothing to remove.");
+        for (const row of removed) console.log(`removed  ${row.name}@${row.version}  ${row.sourcePath}`);
+        for (const row of kept) {
+          const why = row.reason === "pinned"
+            ? "unfinished work is pinned to this version"
+            : row.reason === "unchecked"
+              ? "its file couldn't be checked (not ENOENT), so it isn't treated as missing"
+              : "the file is still in a scanned folder; move or delete it";
+          console.log(`kept     ${row.name}@${row.version}  ${row.sourcePath}  (${why})`);
+        }
+      });
+    });
+
+  cmd
+    .command("trace <instanceId>")
+    .description("Show one workflow instance + its append-only step trail (audit-only verdict)")
+    .option("--full", "Return the complete step trail instead of the 50 most recent closures (trailCompleteness: complete)")
+    .option("--json", "JSON output for agents")
+    .addHelpText("after", `
+The default trail holds the 50 most recent closures; trailCompleteness says
+"recent" when older ones were left out. --full asks the daemon for all of them.
+
 Examples:
   $ rig workflow trace WF01ABC
   $ rig workflow trace WF01ABC --json | jq '.trail[] | {step, actor, exit}'
+  $ rig workflow trace WF01ABC --full --json | jq '.trailCompleteness, (.trail | length)'
 `)
-    .action(async (instanceId: string, opts: { json?: boolean }) => {
+    .action(async (instanceId: string, opts: { json?: boolean; full?: boolean }) => {
       const deps = getDeps();
       await withClient(deps, async (client) => {
-        const res = await client.get<{ instance?: unknown; reconciliation?: unknown; trail?: unknown[]; frontier?: unknown[]; failures?: unknown[]; boundaryObligations?: unknown[]; unknowns?: string[] }>(
-          `/api/workflow/${encodeURIComponent(instanceId)}/trace`,
+        const res = await client.get<{ instance?: unknown; reconciliation?: unknown; trail?: unknown[]; trailCompleteness?: string; frontier?: unknown[]; failures?: unknown[]; boundaryObligations?: unknown[]; unknowns?: string[] }>(
+          `/api/workflow/${encodeURIComponent(instanceId)}/trace${opts.full ? "?full=true" : ""}`,
         );
         // WF3 FR-2: human mode renders the per-step tree (mini-req 2's
         // one-screen bar); --json byte-identical (BR-2).
@@ -667,6 +734,9 @@ Examples:
         } as Parameters<typeof renderTraceTree>[0];
         const trail = (res.data.trail ?? []) as Parameters<typeof renderTraceTree>[1];
         for (const line of renderTraceTree(instance, trail)) console.log(line);
+        if (res.data.trailCompleteness === "recent") {
+          console.log(`  (the ${trail.length} most recent closures; older ones left out - all of them: rig workflow trace ${instanceId} --full)`);
+        }
       });
     });
 

packages/cli/src/config-store.ts

@@ -14,8 +14,8 @@ import {
 
 export interface RiggedConfig {
   specs: { builtinsVisible: boolean };
-  launch: { nonInterruptive: boolean };
-  daemon: { port: number; host: string; dbConflict?: "warn" | "stop" };
+  launch: { nonInterruptive: boolean; codexLoginShell: boolean };
+  daemon: { port: number; host: string; dbConflict?: "warn" | "stop"; seatPollIntervalMs?: number };
   // OPR.0.4.6.MH1 FR-1 — the persisted host-selection pointer.
   // OPR.0.4.6.MH1 FR-4 — the own-host display name (default "localhost").
   host: { selected: string; name: string };
@@ -199,6 +199,8 @@ export interface RiggedConfig {
   transport: {
     verifyInputBox: boolean;
     localUserGuard: boolean;
+    // Whether an ordinary send or work notice reaches a seat whose prompt holds a person's draft (default deliver; refuse opts out).
+    draftDelivery: "refuse" | "deliver";
   };
   // Slice 22 (OPR.0.7.2.22) — typed decisions (OpenRouter + Jev), off until enabled; read by the
   // invoking rig process. templateDir holds optional <name>.yaml templates. Lockstep with the daemon twin.
@@ -207,6 +209,8 @@ export interface RiggedConfig {
     consumers: string;
     templateDir: string;
   };
+  // rig acp serve's receiver when --to isn't given (an ACP registry launch has no args). Lockstep with the daemon twin.
+  acp: { to: string };
   // OPR.0.4.6.FS-1 W2 — queue-retention maintenance knobs (twin of the daemon
   // settings-store): enabled + the four bounded numeric tunables.
   retention: {
@@ -216,6 +220,10 @@ export interface RiggedConfig {
     watchdogKeepPerJob: number;
     streamDays: number;
     streamMaxItems: number;
+    eventsDays: number;
+    closedContentDays: number;
+    closedMetadataDays: number;
+    erasureLedger: string;
     batchSize: number;
   };
   // S04 (OPR.0.5.5.4) — pickup-receipt stall threshold (twin of the daemon settings-store).
@@ -230,6 +238,10 @@ export interface RiggedConfig {
     wakeSwapGraceSeconds: number;
     wakeHumanRung: string;
   };
+  // Issue #1856 — stream intake setting
+  stream: {
+    intake: string;
+  };
 }
 
 const DEFAULT_WORKSPACE_ROOT = getDefaultOpenRigPath("workspace");
@@ -263,8 +275,8 @@ const DEFAULT_CLAUDE_COMPACTION_EXTRA_INSTRUCTION_FILE_PATH = getDefaultOpenRigP
 );
 
 const DEFAULTS = {
-  launch: { nonInterruptive: false },
-  daemon: { port: 7433, host: "127.0.0.1", dbConflict: "warn" },
+  launch: { nonInterruptive: false, codexLoginShell: false },
+  daemon: { port: 7433, host: "127.0.0.1", dbConflict: "warn", seatPollIntervalMs: 1000 },
   // OPR.0.4.6.MH1 FR-1 — "local" ≡ no remote selection (LOCAL_HOST_ID).
   // FR-4 — own-host display name; default "localhost" (PRD-named).
   host: { selected: "local", name: "localhost" },
@@ -383,6 +395,7 @@ const DEFAULTS = {
   transport: {
     verifyInputBox: false,
     localUserGuard: false,
+    draftDelivery: "deliver",
   },
   // Slice 22 — typed decisions OFF, no consumers, shipped templates only.
   decisions: {
@@ -390,6 +403,12 @@ const DEFAULTS = {
     consumers: "",
     templateDir: "",
   },
+  // No ACP receiver: rig acp serve needs --to until one is set.
+  acp: { to: "" },
+  // Issue #1856 — stream intake default queue.worker; read live by the stream intake coordinator.
+  stream: {
+    intake: "queue.worker",
+  },
   // OPR.0.4.6.FS-1 W2 — retention defaults (twin of daemon getDefaultValue).
   retention: {
     enabled: true,
@@ -398,6 +417,10 @@ const DEFAULTS = {
     watchdogKeepPerJob: 50,
     streamDays: 30,
     streamMaxItems: 20000,
+    eventsDays: 0,
+    closedContentDays: 0,
+    closedMetadataDays: 0,
+    erasureLedger: "",
     batchSize: 500,
   },
   // S04 — pickup-receipt stall threshold default (twin of daemon getDefaultValue).
@@ -417,6 +440,8 @@ export const VALID_KEYS = [
   "daemon.port",
   "daemon.host",
   "daemon.db_conflict",
+  // #2687: the seat pollers' cadence; startup-only (read once when the daemon starts). Twin of the daemon key.
+  "daemon.seat_poll_interval_ms",
   // OPR.0.4.6.MH1 FR-1 — the persisted host-selection pointer (the
   // kubectl current-context shape). ONE static key, lockstep with the
   // daemon settings-store twin (parity test pins both). Registry-
@@ -462,6 +487,7 @@ export const VALID_KEYS = [
   "skills.router_line.enabled",
   "onboarding.default_pack.enabled",
   "launch.non_interruptive",
+  "launch.codex_login_shell",
   "health.context_pressure.warning_percent",
   "health.context_pressure.critical_percent",
   "files.allowlist",
@@ -527,10 +553,13 @@ export const VALID_KEYS = [
   "usage.pi_openrouter_charge",
   "transport.verify_input_box",
   "transport.local_user_guard",
+  "transport.draft_delivery",
   // Slice 22 — typed decisions (OpenRouter + Jev), off by default; lockstep with the daemon twin.
   "decisions.enabled",
   "decisions.consumers",
   "decisions.template_dir",
+  // rig acp serve's receiver when --to isn't given; lockstep with the daemon twin.
+  "acp.to",
   // OPR.0.4.6.FS-1 W2 — queue-retention maintenance knobs; CLI-settable twin,
   // lockstep with the daemon settings-store SETTINGS_VALID_KEYS.
   "retention.enabled",
@@ -540,6 +569,15 @@ export const VALID_KEYS = [
   // Slice 22 outcome 9 — stream retention: unreferenced items older than N days or beyond the newest N.
   "retention.stream_days",
   "retention.stream_max_items",
+  // The history-only event types of `events` older than N days, 0 (default) keeps them; lockstep with the daemon twin.
+  "retention.events_days",
+  // Closed queue work: content erased N days after it closed, then the row (only its id kept); 0 (default) keeps them;
+  // lockstep with the daemon twin.
+  "retention.closed_content_days",
+  "retention.closed_metadata_days",
+  // Queue erasure's ledger file (JSON lines of erased ids and actions); empty keeps none; lockstep with the daemon twin.
+  "retention.erasure_ledger",
+  "stream.intake",
   "retention.batch_size",
   // S04 — pickup-receipt stall threshold; lockstep with the daemon settings-store twin.
   "queue.pickup_stall_threshold_minutes",
@@ -570,6 +608,7 @@ export const ENV_MAP: Record<ValidKey, { primary: string; legacy?: string }> = {
   "workflow.exception_routing": { primary: "OPENRIG_WORKFLOW_EXCEPTION_ROUTING" },
   "daemon.host": { primary: "OPENRIG_HOST", legacy: "RIGGED_HOST" },
   "daemon.db_conflict": { primary: "OPENRIG_DAEMON_DB_CONFLICT" },
+  "daemon.seat_poll_interval_ms": { primary: "OPENRIG_SEAT_POLL_INTERVAL_MS" },
   "db.path": { primary: "OPENRIG_DB", legacy: "RIGGED_DB" },
   "transcripts.enabled": { primary: "OPENRIG_TRANSCRIPTS_ENABLED", legacy: "RIGGED_TRANSCRIPTS_ENABLED" },
   "transcripts.path": { primary: "OPENRIG_TRANSCRIPTS_PATH", legacy: "RIGGED_TRANSCRIPTS_PATH" },
@@ -589,6 +628,7 @@ export const ENV_MAP: Record<ValidKey, { primary: string; legacy?: string }> = {
   "skills.router_line.enabled": { primary: "OPENRIG_SKILLS_ROUTER_LINE_ENABLED" },
   "onboarding.default_pack.enabled": { primary: "OPENRIG_ONBOARDING_DEFAULT_PACK_ENABLED" },
   "launch.non_interruptive": { primary: "OPENRIG_LAUNCH_NON_INTERRUPTIVE" },
+  "launch.codex_login_shell": { primary: "OPENRIG_LAUNCH_CODEX_LOGIN_SHELL" },
   "health.context_pressure.warning_percent": { primary: "OPENRIG_HEALTH_CONTEXT_PRESSURE_WARNING_PERCENT" },
   "health.context_pressure.critical_percent": { primary: "OPENRIG_HEALTH_CONTEXT_PRESSURE_CRITICAL_PERCENT" },
   // UEP env-var graduation: existing OPENRIG_FILES_ALLOWLIST /
@@ -646,9 +686,11 @@ export const ENV_MAP: Record<ValidKey, { primary: string; legacy?: string }> = {
   "usage.pi_openrouter_charge": { primary: "OPENRIG_USAGE_PI_OPENROUTER_CHARGE" },
   "transport.local_user_guard": { primary: "OPENRIG_TRANSPORT_LOCAL_USER_GUARD" },
   "transport.verify_input_box": { primary: "OPENRIG_TRANSPORT_VERIFY_INPUT_BOX" },
+  "transport.draft_delivery": { primary: "OPENRIG_TRANSPORT_DRAFT_DELIVERY" },
   "decisions.enabled": { primary: "OPENRIG_DECISIONS_ENABLED" },
   "decisions.consumers": { primary: "OPENRIG_DECISIONS_CONSUMERS" },
   "decisions.template_dir": { primary: "OPENRIG_DECISIONS_TEMPLATE_DIR" },
+  "acp.to": { primary: "OPENRIG_ACP_TO" },
   // OPR.0.4.6.FS-1 W2 — retention knobs; net-new keys, OPENRIG_* primary only.
   "retention.enabled": { primary: "OPENRIG_RETENTION_ENABLED" },
   "retention.transitions_days": { primary: "OPENRIG_RETENTION_TRANSITIONS_DAYS" },
@@ -656,6 +698,11 @@ export const ENV_MAP: Record<ValidKey, { primary: string; legacy?: string }> = {
   "retention.watchdog_keep_per_job": { primary: "OPENRIG_RETENTION_WATCHDOG_KEEP_PER_JOB" },
   "retention.stream_days": { primary: "OPENRIG_RETENTION_STREAM_DAYS" },
   "retention.stream_max_items": { primary: "OPENRIG_RETENTION_STREAM_MAX_ITEMS" },
+  "retention.events_days": { primary: "OPENRIG_RETENTION_EVENTS_DAYS" },
+  "retention.closed_content_days": { primary: "OPENRIG_RETENTION_CLOSED_CONTENT_DAYS" },
+  "retention.closed_metadata_days": { primary: "OPENRIG_RETENTION_CLOSED_METADATA_DAYS" },
+  "retention.erasure_ledger": { primary: "OPENRIG_RETENTION_ERASURE_LEDGER" },
+  "stream.intake": { primary: "OPENRIG_STREAM_INTAKE" },
   "retention.batch_size": { primary: "OPENRIG_RETENTION_BATCH_SIZE" },
   "queue.pickup_stall_threshold_minutes": { primary: "OPENRIG_QUEUE_PICKUP_STALL_THRESHOLD_MINUTES" },
   "queue.stuck_sweep_interval_seconds": { primary: "OPENRIG_QUEUE_STUCK_SWEEP_INTERVAL_SECONDS" },
@@ -678,6 +725,7 @@ const KEY_TO_PATH: Record<ValidKey, string[]> = {
   "workflow.exception_routing": ["workflow", "exceptionRouting"],
   "daemon.host": ["daemon", "host"],
   "daemon.db_conflict": ["daemon", "dbConflict"],
+  "daemon.seat_poll_interval_ms": ["daemon", "seatPollIntervalMs"],
   "db.path": ["db", "path"],
   "transcripts.enabled": ["transcripts", "enabled"],
   "transcripts.path": ["transcripts", "path"],
@@ -697,6 +745,7 @@ const KEY_TO_PATH: Record<ValidKey, string[]> = {
   "skills.router_line.enabled": ["skills", "routerLine", "enabled"],
   "onboarding.default_pack.enabled": ["onboarding", "defaultPack", "enabled"],
   "launch.non_interruptive": ["launch", "nonInterruptive"],
+  "launch.codex_login_shell": ["launch", "codexLoginShell"],
   "health.context_pressure.warning_percent": ["health", "contextPressure", "warningPercent"],
   "health.context_pressure.critical_percent": ["health", "contextPressure", "criticalPercent"],
   "files.allowlist": ["files", "allowlist"],
@@ -746,15 +795,22 @@ const KEY_TO_PATH: Record<ValidKey, string[]> = {
   "usage.pi_openrouter_charge": ["usage", "piOpenRouterCharge"],
   "transport.local_user_guard": ["transport", "localUserGuard"],
   "transport.verify_input_box": ["transport", "verifyInputBox"],
+  "transport.draft_delivery": ["transport", "draftDelivery"],
   "decisions.enabled": ["decisions", "enabled"],
   "decisions.consumers": ["decisions", "consumers"],
   "decisions.template_dir": ["decisions", "templateDir"],
+  "acp.to": ["acp", "to"],
   "retention.enabled": ["retention", "enabled"],
   "retention.transitions_days": ["retention", "transitionsDays"],
   "retention.watchdog_days": ["retention", "watchdogDays"],
   "retention.watchdog_keep_per_job": ["retention", "watchdogKeepPerJob"],
   "retention.stream_days": ["retention", "streamDays"],
   "retention.stream_max_items": ["retention", "streamMaxItems"],
+  "retention.events_days": ["retention", "eventsDays"],
+  "retention.closed_content_days": ["retention", "closedContentDays"],
+  "retention.closed_metadata_days": ["retention", "closedMetadataDays"],
+  "retention.erasure_ledger": ["retention", "erasureLedger"],
+  "stream.intake": ["stream", "intake"],
   "retention.batch_size": ["retention", "batchSize"],
   "queue.pickup_stall_threshold_minutes": ["queue", "pickupStallThresholdMinutes"],
   "queue.stuck_sweep_interval_seconds": ["queue", "stuckSweepIntervalSeconds"],
@@ -969,6 +1025,16 @@ const KEY_CONSTRAINTS: Partial<Record<ValidKey, (raw: string, coerced: string |
   "daemon.db_conflict": (raw) => {
     if (raw !== "warn" && raw !== "stop") throw new Error("daemon.db_conflict must be warn or stop");
   },
+  "daemon.seat_poll_interval_ms": (raw, value) => {
+    // Startup-only: the seat pollers read it once when the daemon starts. Only slower than the 1 s default, up to the
+    // ceiling the structural observation's expiry allows (seat-activity-service.ts MAX_SEAT_POLL_INTERVAL_MS).
+    if (!/^\d+$/.test(raw.trim()) || typeof value !== "number" || !Number.isSafeInteger(value) || value < 1000 || value > 2000) {
+      throw new Error("Invalid daemon.seat_poll_interval_ms: must be a whole number of milliseconds in [1000, 2000]");
+    }
+  },
+  "transport.draft_delivery": (raw) => {
+    if (raw !== "refuse" && raw !== "deliver") throw new Error(`Invalid value for transport.draft_delivery: must be "refuse" or "deliver", got "${raw}"`);
+  },
   "transcripts.lines": (raw, value) => {
     if (!/^\d+$/.test(raw.trim()) || typeof value !== "number" || !Number.isSafeInteger(value) || value < 1 || value > 1_000_000) {
       throw new Error("Invalid transcripts.lines: must be an integer in [1, 1000000]");
@@ -1118,6 +1184,36 @@ const KEY_CONSTRAINTS: Partial<Record<ValidKey, (raw: string, coerced: string |
       throw new Error(`Invalid value for retention.stream_max_items: must be an integer >= 1, got "${raw}"`);
     }
   },
+  "stream.intake": (raw) => {
+    const val = (raw ?? "").trim();
+    if (val !== "off" && val !== "queue.worker" && !/^[a-zA-Z0-9\-_.]+@[a-zA-Z0-9\-_.]+$/.test(val)) {
+      throw new Error(`Invalid value for stream.intake: must be "queue.worker", a canonical session (<seat>@<rig>), or "off", got "${raw}"`);
+    }
+  },
+  // 0 (the default) keeps these events; a positive day count turns their pruning on.
+  "retention.events_days": (raw, coerced) => {
+    if (!/^-?\d+$/.test((raw ?? "").trim()) || typeof coerced !== "number" || !Number.isInteger(coerced) || coerced < 0) {
+      throw new Error(`Invalid value for retention.events_days: must be an integer >= 0 (0 keeps the events), got "${raw}"`);
+    }
+  },
+  // 0 (the default) keeps closed work; a positive day count turns its erasure on (queue-erasure).
+  "retention.closed_content_days": (raw, coerced) => {
+    if (!/^-?\d+$/.test((raw ?? "").trim()) || typeof coerced !== "number" || !Number.isInteger(coerced) || coerced < 0) {
+      throw new Error(`Invalid value for retention.closed_content_days: must be an integer >= 0 (0 keeps the content), got "${raw}"`);
+    }
+  },
+  "retention.closed_metadata_days": (raw, coerced) => {
+    if (!/^-?\d+$/.test((raw ?? "").trim()) || typeof coerced !== "number" || !Number.isInteger(coerced) || coerced < 0) {
+      throw new Error(`Invalid value for retention.closed_metadata_days: must be an integer >= 0 (0 keeps the rows), got "${raw}"`);
+    }
+  },
+  // Empty (the default) keeps no erasure ledger; otherwise an absolute file path, kept outside the backed-up data.
+  "retention.erasure_ledger": (raw) => {
+    const value = (raw ?? "").trim();
+    if (value !== "" && !value.startsWith("/")) {
+      throw new Error(`Invalid value for retention.erasure_ledger: must be empty or an absolute file path, got "${raw}"`);
+    }
+  },
   "retention.batch_size": (raw, coerced) => {
     if (!/^-?\d+$/.test((raw ?? "").trim()) || typeof coerced !== "number" || !Number.isInteger(coerced) || coerced < 1) {
       throw new Error(`Invalid value for retention.batch_size: must be an integer >= 1, got "${raw}"`);
@@ -1138,6 +1234,12 @@ const KEY_CONSTRAINTS: Partial<Record<ValidKey, (raw: string, coerced: string |
       throw new Error("Invalid decisions.template_dir: use an absolute path, or an empty string for the shipped templates only");
     }
   },
+  // rig acp serve's receiver when --to isn't given: a team seat, member@rig, or empty for none.
+  "acp.to": (_raw, value) => {
+    if (typeof value !== "string" || (value !== "" && !/^[^@\s]+@[^@\s]+$/.test(value))) {
+      throw new Error("Invalid acp.to: use the seat that receives your ACP tasks as member@rig, or an empty string for none");
+    }
+  },
 };
 
 function validateKeyConstraints(key: ValidKey, raw: string, coerced: string | number | boolean): void {
@@ -1234,6 +1336,7 @@ export class ConfigStore {
         port: v("daemon.port") as number,
         host: v("daemon.host") as string,
         dbConflict: v("daemon.db_conflict") as "warn" | "stop",
+        seatPollIntervalMs: v("daemon.seat_poll_interval_ms") as number,
       },
       host: {
         selected: v("host.selected") as string,
@@ -1312,7 +1415,7 @@ export class ConfigStore {
           auditLog: v("feed.subscriptions.audit_log") as boolean,
         },
       },
-      launch: { nonInterruptive: v("launch.non_interruptive") as boolean },
+      launch: { nonInterruptive: v("launch.non_interruptive") as boolean, codexLoginShell: v("launch.codex_login_shell") as boolean },
       runtime: {
         readinessTimeoutSeconds: v("runtime.readiness_timeout_seconds") as number,
         codex: {
@@ -1366,12 +1469,14 @@ export class ConfigStore {
       transport: {
         localUserGuard: v("transport.local_user_guard") as boolean,
         verifyInputBox: v("transport.verify_input_box") as boolean,
+        draftDelivery: v("transport.draft_delivery") as "refuse" | "deliver",
       },
       decisions: {
         enabled: v("decisions.enabled") as boolean,
         consumers: v("decisions.consumers") as string,
         templateDir: v("decisions.template_dir") as string,
       },
+      acp: { to: v("acp.to") as string },
       retention: {
         enabled: v("retention.enabled") as boolean,
         transitionsDays: v("retention.transitions_days") as number,
@@ -1379,6 +1484,10 @@ export class ConfigStore {
         watchdogKeepPerJob: v("retention.watchdog_keep_per_job") as number,
         streamDays: v("retention.stream_days") as number,
         streamMaxItems: v("retention.stream_max_items") as number,
+        eventsDays: v("retention.events_days") as number,
+        closedContentDays: v("retention.closed_content_days") as number,
+        closedMetadataDays: v("retention.closed_metadata_days") as number,
+        erasureLedger: v("retention.erasure_ledger") as string,
         batchSize: v("retention.batch_size") as number,
       },
       queue: {
@@ -1391,6 +1500,9 @@ export class ConfigStore {
         wakeSwapGraceSeconds: v("queue.wake_swap_grace_seconds") as number,
         wakeHumanRung: v("queue.wake_human_rung") as string,
       },
+      stream: {
+        intake: v("stream.intake") as string,
+      },
     };
   }
 
@@ -1531,7 +1643,7 @@ export class ConfigStore {
       const critical = getNestedValue(fileConfig, KEY_TO_PATH["health.context_pressure.critical_percent"])
         ?? getDefaultValue("health.context_pressure.critical_percent", workspaceRoot);
       if ((warning as number) >= (critical as number)) {
-        throw new Error(`Invalid context-pressure policy: warning (${warning}) must be less than critical (${critical})`);
+        throw new Error(`Invalid context-pressure policy for the config-file pair: warning (${warning}) must be less than critical (${critical}). Unset file values use defaults; environment overrides apply when each process resolves configuration and are not part of this check.`);
       }
     }
     mkdirSync(dirname(this.configPath), { recursive: true });
@@ -1583,7 +1695,7 @@ export class ConfigStore {
         const critical = getNestedValue(fileConfig, KEY_TO_PATH["health.context_pressure.critical_percent"])
           ?? getDefaultValue("health.context_pressure.critical_percent", workspaceRoot);
         if ((warning as number) >= (critical as number)) {
-          throw new Error(`Invalid context-pressure policy: warning (${warning}) must be less than critical (${critical})`);
+          throw new Error(`Invalid context-pressure policy for the config-file pair: warning (${warning}) must be less than critical (${critical}). Unset file values use defaults; environment overrides apply when each process resolves configuration and are not part of this check.`);
         }
       }
       pruneEmptyAncestors(fileConfig, parentParts);

packages/cli/src/cross-host-target.ts

@@ -36,6 +36,8 @@ export interface CrossHostTargetResolution {
   target: string;
   /** The sugar-derived host id (registered suffix), if any. */
   sugarHost: string | undefined;
+  /** The qualifier explicitly names this host, overriding a persisted remote selection. */
+  localHost?: boolean;
   /**
    * Loud-failure hint when the target was 3-part-SHAPED but the suffix
    * matched no registered host. Callers append it to failure surfaces for
@@ -102,7 +104,7 @@ export function resolveCrossHostTarget(
       };
     }
     // Routes home: strip the self-suffix, no cross-host sugarHost (local send).
-    return { ok: true, target: base, sugarHost: undefined, hint: undefined };
+    return { ok: true, target: base, sugarHost: undefined, localHost: true, hint: undefined };
   }
 
   const loader = registryLoader ?? loadHostRegistry;

packages/cli/src/daemon-lifecycle.ts

@@ -39,6 +39,8 @@ export interface DaemonEventLoopEvidence {
 }
 
 export interface DaemonStatus {
+  /** The selected OPENRIG_URL names a non-loopback endpoint. */
+  explicitRemote?: boolean;
   databaseWarnings?: string[];
   /** RULING 1ae863d2 — C3 3-state semantics (canonical: daemon crash-cart-detect.ts, kept in
    *  lockstep): "stopped" requires POSITIVE evidence (connection refused); a probe TIMEOUT or
@@ -432,9 +434,13 @@ function resolveLifecycleFile(deps: LifecycleDeps, filename: "daemon.json" | "da
  *  - "openrig" — healthz responded (ok or not) → this is our daemon
  *  - "not_openrig" — connection refused → PID is alive but not listening on our port
  *  - "unresponsive" — pid is alive but healthz probe timed out
- *  - "dead" — PID not alive */
-async function checkPid(state: DaemonState, deps: LifecycleDeps): Promise<"openrig" | "not_openrig" | "unresponsive" | "dead"> {
-  if (!deps.isProcessAlive(state.pid)) return "dead";
+ *  - "dead" — PID not alive
+ *  - { blocked } — the OS refused the probe; identity is unknown */
+async function checkPid(state: DaemonState, deps: LifecycleDeps, knownLiveness?: ProcessLiveness): Promise<"openrig" | "not_openrig" | "unresponsive" | "dead" | { blocked: string }> {
+  const liveness = knownLiveness ?? (deps.processLiveness
+    ? deps.processLiveness(state.pid)
+    : deps.isProcessAlive(state.pid) ? "alive" : "dead");
+  if (liveness === "dead") return "dead";
   const host = state.host ?? DEFAULT_HOST;
   try {
     await fetchDaemonProbe(deps, `http://${formatDaemonHostForUrl(host)}:${state.port}/healthz`, HEALTHZ_PROBE_TIMEOUT_MS);
@@ -442,6 +448,8 @@ async function checkPid(state: DaemonState, deps: LifecycleDeps): Promise<"openr
     return "openrig";
   } catch (err) {
     if (err instanceof HealthProbeTimeoutError) return "unresponsive";
+    const code = connectionErrorCode(err);
+    if (isBlockedConnectionCode(code)) return { blocked: code };
     // Connection refused → PID alive but not our daemon
     return "not_openrig";
   }
@@ -487,6 +495,13 @@ const ENV_SCRUB_EXACT = new Set([
   "TERM_PROGRAM",
   "TERM_PROGRAM_VERSION",
   "TERMINFO",
+  // The herdr-core terminal this command runs in (HERDR_CORE_TERMINAL_ENV_KEYS in the daemon's
+  // parent-session-env.ts). Core ends every process started with that terminal's HC_SCOPE when the
+  // terminal closes, and it reads the environment the process was started with, so the daemon must
+  // never be started with it: closing the terminal would end the daemon, its tmux server and every seat.
+  "HC_SCOPE",
+  "HC_TERMINAL_ID",
+  "HC_TERMINAL_SOCKET",
 ]);
 
 export function buildDaemonEnv(
@@ -679,6 +694,10 @@ async function startOwnedDaemon(opts: StartOptions, deps: LifecycleDeps, lock: D
   const existing = readState(deps);
   if (existing) {
     const pidState = await checkPid(existing, deps);
+    if (typeof pidState === "object") {
+      const guidance = blockedConnectionGuidance(pidState.blocked);
+      throw new Error(`${guidance.consequence} ${guidance.action}`);
+    }
     if (pidState === "openrig") {
       // Our daemon is running (possibly unhealthy, but alive on our port)
       throw new Error(`Daemon already running (pid ${existing.pid} on port ${existing.port})`);
@@ -693,6 +712,10 @@ async function startOwnedDaemon(opts: StartOptions, deps: LifecycleDeps, lock: D
       await fetchDaemonProbe(deps, `http://${formatDaemonHostForUrl(probeHost)}:${port}/healthz`, HEALTHZ_PROBE_TIMEOUT_MS);
       recoveredRunning = true;
     } catch (error) {
+      const code = connectionErrorCode(error);
+      if (isBlockedConnectionCode(code)) {
+        console.warn(`Warning: daemon health probe was blocked (${code}); the port could not be checked and a daemon may already be running.`);
+      }
       // WSL2 can silently drop probes to an unused port. A slow live daemon can
       // also time out: do not run a second child's initialization against its DB.
       if (error instanceof HealthProbeTimeoutError && await deps.isPortOccupied?.(probeHost, port)) {
@@ -972,7 +995,16 @@ export async function stopDaemon(deps: LifecycleDeps): Promise<"stopped" | "no-t
   }
   const stateFile = resolveLifecycleFile(deps, "daemon.json");
 
-  const pidState = await checkPid(state, deps);
+  const liveness = deps.processLiveness
+    ? deps.processLiveness(state.pid)
+    : deps.isProcessAlive(state.pid) ? "alive" : "dead";
+  if (liveness === "unknown") {
+    throw new Error(`Cannot stop safely: PID ${state.pid} liveness is unknown; no signal sent, state preserved.`);
+  }
+  const pidState = await checkPid(state, deps, liveness);
+  if (typeof pidState === "object") {
+    throw new Error(`Cannot stop safely: the identity probe was blocked (${pidState.blocked}); no signal sent, state preserved.`);
+  }
   if (pidState === "not_openrig") {
     throw new Error(`Cannot stop safely: PID ${state.pid} is present but identity is not confirmed at ${check}; no signal sent, state preserved.`);
   }
@@ -1107,7 +1139,9 @@ export async function getDaemonStatus(
     const endpoint = (() => {
       try {
         const url = new URL(openrigUrl);
-        return { port: Number(url.port) || DEFAULT_PORT, host: url.hostname || DEFAULT_HOST };
+        const host = url.hostname || DEFAULT_HOST;
+        const loopback = host.toLowerCase() === "localhost" || host === "[::1]" || host === "::1" || /^127\.\d+\.\d+\.\d+$/.test(host);
+        return { port: Number(url.port) || DEFAULT_PORT, host, ...(!loopback ? { explicitRemote: true } : {}) };
       } catch {
         return {};
       }