fix(daemon): a cross-host handoff says when it overrode a claim (#2734 )

closeCrossHostHandoffSource takes the source's claim as the route read it before forwarding. If the row was claimed since, it still closes, records the overridden claim in the transition note, and stages one claim-overridden wake to the claimer in the close's transaction, naming the destination and the successor's plain id. That wake is current only while its row stays closed, owned by the claimer, and its latest close is the notice's own transition. Startup cleanup skips it, the recovery drain selects it on its closed row, and the prompt-clear retry finds such rows in one pass from the notices' id range on the outbox key, with ordinary wakes still pending-only. A full held-message quota records the wake as not retained without undoing the close. Refs #2631 (item 2); the opt-in --if-state pending is not included.
docs(plugin): record exactly what the person typed as their answer (#2733 )
2026-11-07 18:34:36 -10:00 · 2026-11-07 20:20:34 -08:00 · 2026-11-07 19:49:06 -08:00 · 2026-11-07 19:44:00 -08:00 · 2026-11-07 18:43:23 -08:00 · 2026-11-07 18:31:35 -08:00
513 changed files with 45369 additions and 1879 deletions
@@ -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."
+    }
+  ]
+}
@@ -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."
+    }
+  ]
+}
@@ -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:
@@ -8,6 +8,372 @@ deprecations, and behavioral changes. Breaking changes are called out explicitly

 ---

+## [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]

 Published 2026-11-05 as `@openrig/[email protected]`, built from the commit tagged
@@ -133,8 +499,13 @@ check timed out after 5 seconds, so no team work ran. Don't rely on this path in

 ### 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."
@@ -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 0.7.8
+## Released in 0.7.9

-[0.7.8 is published](https://github.com/mvschwarz/openrig/releases/tag/v0.7.8): it lets a person answer a runner seat's permission requests, with a conformance check and an example runner to build against. See its release notes for what shipped, known issues, upgrade instructions and contributor credits.
+[0.7.9 is published](https://github.com/mvschwarz/openrig/releases/tag/v0.7.9): it 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. See its release notes for what shipped, known issues, upgrade instructions and contributor credits.

 ## How to help

@@ -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`).
@@ -468,6 +468,16 @@ 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
@@ -363,6 +363,10 @@ over that token. Without it, the target's `POST /api/hosts/pair-request` answers
 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
@@ -414,9 +418,21 @@ 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 |
@@ -462,6 +478,11 @@ 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 |
@@ -470,16 +491,17 @@ 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>`--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>`--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` |
@@ -583,9 +605,10 @@ 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` |
@@ -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`,
@@ -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.
@@ -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 +288,8 @@ 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.

 **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 +372,10 @@ 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 +421,16 @@ 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 +439,55 @@ 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 +501,7 @@ 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 +543,7 @@ 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
@@ -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:
@@ -394,6 +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` and `codex` 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 (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 on missing/unexecutable wrapper, or for Codex non-POSIX pane/missing session identity, retrying long Pi commands via host runner with wrapper flags preserved; 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)) |

@@ -206,6 +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

@@ -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. |
@@ -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
@@ -65,14 +65,17 @@ or else from the file `$OPENRIG_HOME/terminal-token`, which the person writes; t

 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 token
-doesn't get past the browser boundary's 403 refusals, which [Browser access](browser-access.md) describes.
+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 three are server-sent events (`text/event-stream`):
+All of them are server-sent events (`text/event-stream`):

- **`GET /api/events`**: the daemon's event log. `?rigId=` limits it to one rig. Each event carries `id: <seq>`.
+- **`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.
@@ -80,6 +83,9 @@ All three are server-sent events (`text/event-stream`):
 - **`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`.
@@ -14,7 +14,7 @@ returns `503` with `error: "human_registry_unavailable"`. The result does not in
 ## Participants

 A participant is either:
- **a person**, registered with `rig gateway human add`, with or without a connector 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.

@@ -22,6 +22,9 @@ 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`,
@@ -0,0 +1,239 @@
+# 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`.
+- Each logical decision freezes its complete `relayEnvelope` onto the parent `DispatchBuffer` entry before the first network attempt, with no self-imposed outbound 64 KiB cap or multi-part splitting.
+- 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`. Bridge acceptance confirms custody by the bridge process only (and is projected as `accepted` / `"Accepted by bridge"`, whereas Slack posts use `posted` / `"Delivered"`). It does not indicate downstream platform delivery, posting, or human receipt.
+- 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.
+
+### 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
+
+```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
+```
+
+- `OPENRIG_RELAY_URL`: 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`: Non-empty Bearer token sent by OpenRig on outbound requests to the bridge.
+- `OPENRIG_RELAY_INBOUND_TOKEN`: Non-empty Bearer token required on inbound requests from the bridge. 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 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>`)
+
+```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`).
+
+### 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>` requires a full binding specification string. Supplying `unset` is not supported for binding slots. To reassign a participant, overwrite the slot with a full binding string or remove and re-add the human fragment.
+
+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
+Before removing a participant, 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`.
@@ -270,6 +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.
@@ -871,6 +882,7 @@ 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. |
@@ -1150,8 +1162,15 @@ 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
@@ -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
@@ -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
@@ -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.
@@ -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 +576,20 @@ 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"
+```
+
+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. Switching a seat between Pi
+and another runtime is not supported by this account-home feature.

 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 +769,7 @@ 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`. |
@@ -750,6 +779,7 @@ 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 built-in `pi` and `codex` members only. Literal argument prefix prepended before the managed launch command 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. |
@@ -887,14 +917,38 @@ 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.

+### Launch wrapper (`runtime: pi` and `codex`)
+
+Set `launch_wrapper` on a local built-in `pi` or `codex` member to prefix its managed launch command 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:** Supported only on local built-in `runtime: pi` and `runtime: codex` members (and legacy flat `pi`/`codex` nodes); other runtimes (`claude-code`, `omp`, `terminal`, `stub` and community harnesses) reject the field, including `[]`. There is no rig-level, pod-level or fork inheritance: a fork uses its target member's own `launch_wrapper`. Every element must be a string without NUL, CR or LF characters and is quoted literally with `shellQuote`, preserving argument order, spaces, quotes, shell metacharacters and empty non-executable arguments without shell or template expansion.
+- **Executable resolution and `exec` contract:** `launch_wrapper[0]` must be a non-empty bare command filename on the daemon's launch `PATH` or an absolute path (no relative paths with slashes such as `./x` or `../x`, no `/` in a bare name, and no shell aliases or functions). Preflight and launch resolve `launch_wrapper[0]` to an executable file using the daemon's launch `PATH` and daemon working directory (before the seat's `rig` bin directory is prepended). For `codex`, the prefix appears immediately before the resolved Codex entry (`'<abs node>' '<abs codex.js>' …` or `'<abs codex>' …`) inside the public seat environment envelope, preserving the session's TTY stdio. For `pi`, OpenRig's host runner stays host-side and unwrapped (`node <pi-runner.js> …`), preserving its RPC lifecycle, sidecars, and status reporting; the wrapper prefixes the resolved Pi child executable (`pi --mode rpc`) inside that runner, and the child's stdio pipes preserve RPC stdin, stdout, and stderr (not a TTY). The wrapper must prepare its environment and `exec` the target command in place so the child process preserves working directory, child stdio pipes (for Pi) or TTY stdio (for Codex), OpenRig session/identity environment, process lineage, exit status and signals; supervising wrappers that keep a parent process are unsupported.
+- **Fail-closed (never unwraps):** When a wrapper is selected, OpenRig applies it across fresh launch, relaunch, rebuild, native resume, fork, restore, handover successor and supported runtime switches, and never falls back to bare launch. For `codex`, wrapper resolution and launch environment assembly occur within the shell prefix: if the pane shell is non-POSIX (`nu`, `fish`, etc.), OpenRig session identity is missing or mismatched, `SeatLaunchEnvironment` is unavailable, or `launch_wrapper[0]` is missing or unexecutable, Codex launch fails closed (`OpenRig: launch_wrapper unavailable; no agent started.`). For `pi`, wrapper lookup resolves against the daemon launch `PATH` and daemon working directory before state creation or process dispatch, failing closed before pending launch state if `launch_wrapper[0]` cannot be resolved or executed; runtime child wrapper failures are subsequently reported via the host runner sidecars and markers. If Pi encounters a long launch command (`launch_path_too_long`), it retries execution using the raw host runner command; because that command carries the exact same child wrapper arguments, the Pi child executable never runs unwrapped. In both runtimes, malformed stored wrapper metadata fails closed during reads.
+- **Persistence, `[]` clear vs absent retain, and recreation:** On a new seat, omitting `launch_wrapper` or declaring `launch_wrapper: []` stores no wrapper (`NULL`). When an authored spec is applied to an existing seat, omitting `launch_wrapper` retains the seat's stored wrapper, `launch_wrapper: []` explicitly clears it, and a non-empty array replaces it. A rig that is recreated or imported, including a replacement import, creates new nodes from the incoming spec and uses only what that spec declares, so omitting the field gives the new seat no wrapper. Stored wrapper definitions remain strictly validated; corrupted wrapper rows fail closed during reads without silent unwrap. Restoring an older snapshot that predates `launch_wrapper` preserves the seat's current stored wrapper. First-start retry (`rig launch <rig> <member>`) compares the retained declaration with the node row and refuses to add, change or remove a wrapper on retry. Switching a wrapped seat to an unsupported runtime (`claude-code`, `omp`, etc.) is refused before side effects. There is no live `rig seat` edit verb or management route for `launch_wrapper`; to change or clear a wrapper, edit the `RigSpec` (use `launch_wrapper: []` to clear) and use the supported create/recreate/apply flow. Export (`rig export`) and cached node inventory / `rig ps --nodes --json` (including `--full` and `--fields launchWrapper`) include `launch_wrapper` / `launchWrapper` when configured and omit it when absent.
+- **Boundary and limits:** A `launch_wrapper` is only an argv prefix. Configuring or observing a prefix is **not** proof of sandbox isolation, tooling availability, memory wiring or governance enforcement; verify the wrapper's own guarantees and the 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.

@@ -1077,7 +1131,7 @@ 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. |

@@ -1302,6 +1356,7 @@ 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 only on `pi` and `codex` members (`[]` included); 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`.

 ---

@@ -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
@@ -173,6 +236,26 @@ 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
@@ -349,8 +349,7 @@ 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
@@ -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.
@@ -222,11 +222,26 @@ 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. Opening a tile can change the
+agent's terminal size; `ignore-size` does not guarantee isolation.

 ```bash
 rig terminal open qitem:example --follow --provider herdr --json
@@ -284,11 +299,17 @@ 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.
@@ -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.
@@ -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
@@ -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"
@@ -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",
@@ -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.
@@ -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`.
@@ -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);
+});
@@ -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"
+  ]
+}
@@ -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]"
+    }
+  }
+}
@@ -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>
@@ -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;
+    }
+  });
+}
@@ -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();
@@ -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(); }));
+  }
+}
@@ -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")) {
@@ -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));
@@ -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)")
@@ -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: ${advisory}`);
      if (data.valid) {
        const name = yamlScalar(yaml, "name") ?? "unknown";
        const version = yamlScalar(yaml, "version") ?? "unknown";
@@ -62,7 +62,7 @@ export function attachCommand(depsOverride?: StatusDeps): Command {
    .option("--member <name>", "Member name for pod attach mode")
    .option("--runtime <runtime>", "Runtime for pod attach mode; optional guard for node mode")
    .option("--cwd <path>", "Working directory to record", process.cwd())
-    .option("--display-name <name>", "Display label inside tmux; external session name outside tmux")
+    .option("--display-name <name>", "Display label inside tmux; external session address outside tmux (new bare names gain @rig)")
    .option("--print-env", "Print shell exports for OPENRIG_NODE_ID and OPENRIG_SESSION_NAME")
    .option("--json", "JSON output, including the activity env the attached app needs. Its token is a credential: don't print or paste it, and use rig whoami for identity")
    .action(async (opts: {
@@ -222,10 +222,13 @@ async function runCrossHostBroadcast(
 *  local route does, so per-target honesty prints identically. A partial
 *  fan-out sets a non-zero exit either way. */
 function renderBroadcastResults(data: Record<string, unknown>): void {
-  const results = (data["results"] as Array<{ sessionName: string; ok: boolean; error?: string; managedCompaction?: { notice?: unknown } }>) ?? [];
+  const results = (data["results"] as Array<{ sessionName: string; ok: boolean; error?: string; inputPhase?: string; sent?: boolean; managedCompaction?: { notice?: unknown } }>) ?? [];
  for (const r of results) {
    if (r.ok) {
      console.log(`${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"}`);
@@ -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.`);

@@ -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 +103,18 @@ 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)
  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 +134,13 @@ 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 +207,17 @@ 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 +233,22 @@ 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: ${envKeys.join(", ")}.`
+            : "Config reset to defaults.");
        }
      } catch (err) {
        console.error((err as Error).message);
@@ -0,0 +1,69 @@
+import { Command } from "commander";
+import { join, resolve } from "node:path";
+import { ConfigStore } from "../config-store.js";
+import { readOpenRigEnv } from "../openrig-compat.js";
+import { resolveWorkContextHome } from "../lib/work-install.js";
+import { ReconstructionError, readReconstruction, readStoreMappings, reconstructionOutput } from "../lib/reconstruction-reader.js";
+
+interface Options {
+  store?: string; storeMap?: string; project?: string; mission?: string; slice?: string; cwd?: string;
+  item: string[]; category: string[]; tag: string[]; checkSources?: boolean; maxBytes: string; json?: boolean;
+}
+const collect = (value: string, previous: string[]) => [...previous, value];
+
+/** A local config/filesystem consumer. This command never starts or contacts a
+ * daemon, invokes a helper, writes revisions, or takes over the selected work. */
+export function reconstructionCommand(): Command {
+  const cmd = new Command("reconstruction")
+    .description("Inspect or read an exact immutable reconstruction locally (read-only v0 consumer)");
+  for (const mode of ["inspect", "view"] as const) {
+    cmd.command(`${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;
+}
@@ -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"]);

@@ -252,6 +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),
@@ -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";
  };
 }

@@ -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 each part of the port's /healthz response may take (default 2000 ms). */
+  /** 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,
@@ -351,9 +353,16 @@ export function runDoctorChecks(deps: DoctorDeps, providerChecks: DoctorCheck[]
      } 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." };
          }
@@ -366,11 +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 ${version}, which isn't a tmux release known to be 3.7c or later.`,
+    reason: "Starting a control client can crash a tmux server older than 3.7c, so the daemon spawns a tmux client for every command instead. The running server's version decides this, not the tmux binary on PATH.",
+    fix: "Run the daemon's sessions on a tmux 3.7c or later server: rig context get reference/tmux-servers.md#moving-an-existing-team-onto-tmux-control-mode",
+  };
+}
+
 /**
 * Build B — does the rig spec still describe the rig that is running?
 *
@@ -512,18 +582,32 @@ class AmbiguousLiveRigError extends Error {
 export async function readLiveLogicalIds(base: string, rigName: string, ownerClient = new DaemonClient(base)): Promise<string[] | null> {
  try {
    await ownerClient.assertLocalOwner();
-    const rigsRes = await fetch(`${base}/api/rigs`);
+    let rigs: Array<{ id?: string; name?: string }> = [];
+    const rigsRes = await fetchWithTimeout(fetch, `${base}/api/rigs`, {}, {
+      timeoutMs: DOCTOR_HTTP_TIMEOUT_MS,
+      timeoutMessage: "Doctor's live rig read timed out.",
+      consumeResponse: async res => {
+        if (res.ok) rigs = (await res.json()) as typeof rigs;
+        else await res.body?.cancel();
+      },
+    });
    if (!rigsRes.ok) return null;
-    const rigs = (await rigsRes.json()) as Array<{ id?: string; name?: string }>;
    const matches = rigs.filter((r) => r.name === rigName);
    if (matches.length > 1) throw new AmbiguousLiveRigError(matches.map(r => r.id ?? "(missing id)"));
    if (matches.length === 0) return null;
    const rig = matches[0];
    if (!rig?.id) return null;
    await ownerClient.assertLocalOwner();
-    const nodesRes = await fetch(`${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 || error instanceof AmbiguousLiveRigError) throw error;
@@ -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) };
@@ -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;
+}
@@ -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
@@ -785,7 +785,7 @@ export function hostCommand(doctorDepsOverride?: DoctorDeps): Command {
    .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); 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 }) => {
@@ -857,11 +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) {
@@ -881,25 +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) {
@@ -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}: ${(err as Error).message}`);
@@ -603,6 +603,9 @@ API-key approval before the first turn. OpenRig does not enforce API budgets or
 Codex can select a saved API-key auth profile with --to-account just like a subscription profile. Its shared-home effects
 and native credential-store requirements are unchanged. No fresh account home or conversation is created.
 Do not combine --api-key-file with --account-home or --clear-account-home; select a home and switch credentials separately.
+Pi seats support --account-home with an existing readable auth.json, or --clear-account-home, through a fresh successor
+with a handover. Their current model and effort are retained unless overridden. Pi account changes work with or without
+a launch_wrapper; changing a seat between Pi and another runtime is not supported.
 --clear-account-home returns to the configured member/rig/daemon/native home through a fresh successor with a handover.
 Use it without --to-account, --stored-login, --account-home or --token-file. It removes an account-home token reference
 only after confirmation; it does not restore historical login bytes. Anthropic active/default profiles live outside
@@ -638,7 +641,7 @@ CLAUDE_CONFIG_DIR and are not isolated by an account home; check them before rel
        try {
          const status = await client.get<{ runtime?: string }>(`/api/seat/status/${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;
@@ -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;
@@ -217,6 +219,7 @@ const ALLOWED_NODE_FIELDS = new Set([
  "nodeKind",
  "runtime",
  "communityHarness",
+  "launchWrapper",
  "model",
  "modelObservation",
  "sessionStatus",
@@ -733,6 +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.
@@ -930,7 +934,7 @@ 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,
@@ -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" }));
@@ -1246,8 +1297,9 @@ export function queueCommand(depsOverride?: QueueDeps): Command {
    .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: {
@@ -1264,6 +1316,7 @@ export function queueCommand(depsOverride?: QueueDeps): Command {
      summary?: string;
      evidenceRef?: string;
      note?: string;
+      host?: string;
      json?: boolean;
      full?: boolean;
    }) => {
@@ -1287,8 +1340,10 @@ 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,
@@ -1301,9 +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,
@@ -1324,6 +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", `
@@ -1345,6 +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;
    }) => {
@@ -1352,8 +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,
@@ -1361,9 +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
@@ -1658,10 +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
@@ -1669,7 +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.
@@ -1677,7 +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))] : []),
@@ -1696,6 +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`);
      });
    });

@@ -1704,20 +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
@@ -1735,6 +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.
@@ -1777,45 +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
@@ -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.
@@ -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;
@@ -165,7 +169,7 @@ 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[];
@@ -208,6 +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}`);
@@ -243,6 +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}`);
@@ -357,7 +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}`);
@@ -376,8 +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: ${result.restorePacket.reason}.`);
+      console.log("Check the successor's screen before sending the context again; it may already be working on it.");
+    }
    console.log("No conversation continuity was carried: the successor started a fresh conversation.");
  } else if (result.source.mode === "fork") {
    console.log("No packet delivery: fork context rides the native conversation itself.");
@@ -550,7 +589,7 @@ Examples:
      const res = await client.get<SeatStatusResponse | SeatStatusError>(`/api/seat/status/${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;
      }
@@ -1399,7 +1438,9 @@ Examples:
    .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
+session is refused (use rig unclaim). 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.
@@ -242,7 +242,7 @@ 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)")
@@ -391,10 +391,10 @@ 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;
@@ -467,14 +467,10 @@ 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;
@@ -252,7 +252,7 @@ export function slackCommand(deps: SlackDeps = {}): Command {
        const connector = observation.connector as { configurationDigest?: string; inboundReady?: boolean; outboundReady?: boolean;
          inboundRouting?: { fallback: string; destination: string };
          settlement?: { pending: number; lastFailure: string | null };
-          deadLetterBacklog?: number | null; deadLetterBacklogState?: string; deadLetterBacklogReason?: string;
+          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;
@@ -291,6 +291,7 @@ export function slackCommand(deps: SlackDeps = {}): Command {
              + `${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") {
@@ -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>;
@@ -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);
      });
@@ -137,7 +137,7 @@ export function terminalCommand(depsOverride?: TerminalDeps): Command {
    .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("--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")
@@ -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)`);
+        }
      });
    });

@@ -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 +220,7 @@ export interface RiggedConfig {
    watchdogKeepPerJob: number;
    streamDays: number;
    streamMaxItems: number;
+    eventsDays: number;
    batchSize: number;
  };
  // S04 (OPR.0.5.5.4) — pickup-receipt stall threshold (twin of the daemon settings-store).
@@ -267,8 +272,8 @@ 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" },
@@ -387,6 +392,7 @@ const DEFAULTS = {
  transport: {
    verifyInputBox: false,
    localUserGuard: false,
+    draftDelivery: "deliver",
  },
  // Slice 22 — typed decisions OFF, no consumers, shipped templates only.
  decisions: {
@@ -394,6 +400,8 @@ 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",
@@ -406,6 +414,7 @@ const DEFAULTS = {
    watchdogKeepPerJob: 50,
    streamDays: 30,
    streamMaxItems: 20000,
+    eventsDays: 0,
    batchSize: 500,
  },
  // S04 — pickup-receipt stall threshold default (twin of daemon getDefaultValue).
@@ -425,6 +434,8 @@ 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-
@@ -470,6 +481,7 @@ 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",
@@ -535,10 +547,13 @@ 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",
@@ -548,6 +563,8 @@ 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",
  "stream.intake",
  "retention.batch_size",
  // S04 — pickup-receipt stall threshold; lockstep with the daemon settings-store twin.
@@ -579,6 +596,7 @@ 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" },
@@ -598,6 +616,7 @@ 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 /
@@ -655,9 +674,11 @@ 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" },
@@ -665,6 +686,7 @@ 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" },
  "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" },
@@ -688,6 +710,7 @@ 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"],
@@ -707,6 +730,7 @@ 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"],
@@ -756,15 +780,18 @@ 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"],
  "stream.intake": ["stream", "intake"],
  "retention.batch_size": ["retention", "batchSize"],
  "queue.pickup_stall_threshold_minutes": ["queue", "pickupStallThresholdMinutes"],
@@ -980,6 +1007,16 @@ 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+$/.test(raw.trim()) || typeof value !== "number" || !Number.isSafeInteger(value) || value < 1 || value > 1_000_000) {
      throw new Error("Invalid transcripts.lines: must be an integer in [1, 1000000]");
@@ -1135,6 +1172,12 @@ const KEY_CONSTRAINTS: Partial<Record<ValidKey, (raw: string, coerced: string |
      throw new Error(`Invalid value for stream.intake: must be "queue.worker", a canonical session (<seat>@<rig>), or "off", got "${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}"`);
+    }
+  },
  "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}"`);
@@ -1155,6 +1198,12 @@ 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 {
@@ -1251,6 +1300,7 @@ 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,
@@ -1329,7 +1379,7 @@ 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: {
@@ -1383,12 +1433,14 @@ 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,
@@ -1396,6 +1448,7 @@ 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,
        batchSize: v("retention.batch_size") as number,
      },
      queue: {
@@ -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;
@@ -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";
  }
@@ -679,6 +687,10 @@ 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 ${existing.port})`);
@@ -693,6 +705,10 @@ async function startOwnedDaemon(opts: StartOptions, deps: LifecycleDeps, lock: D
      await fetchDaemonProbe(deps, `http://${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 +988,16 @@ 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 ${check}; no signal sent, state preserved.`);
  }
@@ -1107,7 +1132,9 @@ export async function getDaemonStatus(
    const endpoint = (() => {
      try {
        const url = new URL(openrigUrl);
-        return { port: Number(url.port) || DEFAULT_PORT, host: url.hostname || DEFAULT_HOST };
+        const host = url.hostname || DEFAULT_HOST;
+        const loopback = host.toLowerCase() === "localhost" || host === "[::1]" || host === "::1" || /^127\.\d+\.\d+\.\d+$/.test(host);
+        return { port: Number(url.port) || DEFAULT_PORT, host, ...(!loopback ? { explicitRemote: true } : {}) };
      } catch {
        return {};
      }
@@ -72,12 +72,13 @@ export function daemonMayAnswer(status: DaemonStatus): boolean {

 /** Whether `rig up` or `rig start` should start a local daemon. Yes for one confirmed down (stopped, or a
 *  stale record). For an unverified status, as before this check existed, when another home's daemon is
- *  live (an independent home elsewhere, this one's own endpoint possibly free) or the probe was blocked;
+ *  live (an independent home elsewhere, this one's own endpoint possibly free);
 *  preflight and the start's own state and listener checks then decide. Not for an explicit HOME-MOVED
 *  marker, and not for a daemon of this home that is alive but too busy to answer (it gets the request). */
 export function shouldAutoStartDaemon(status: DaemonStatus): boolean {
  if (status.state === "stopped" || status.state === "stale") return true;
  if (status.state !== "unverified") return false;
+  if (isBlockedConnectionCode(status.probeErrorCode)) return !status.explicitRemote && !status.siblingHint?.moved;
  if (status.siblingHint) return !status.siblingHint.moved;
-  return isBlockedConnectionCode(status.probeErrorCode);
+  return false;
 }
@@ -34,6 +34,7 @@ import type { GatewayCommandDeps } from "./commands/gateway.js";
 import { parkedCommand } from "./commands/parked.js";
 import { mcpCommand } from "./commands/mcp.js";
 import { acpCommand } from "./commands/acp.js";
+import { execCommand, type ExecDeps } from "./commands/exec.js";
 import { agentCommand, type AgentDeps } from "./commands/agent.js";
 import { rigCommand, type RigDeps } from "./commands/rig.js";
 import { transcriptCommand } from "./commands/transcript.js";
@@ -120,6 +121,7 @@ export interface ProgramDeps {
  psDeps?: StatusDeps;
  mcpDeps?: StatusDeps;
  acpDeps?: StatusDeps;
+  execDeps?: ExecDeps;
  agentDeps?: AgentDeps;
  rigDeps?: RigDeps;
  transcriptDeps?: StatusDeps;
@@ -206,6 +208,7 @@ export function createProgram(depsOverride?: ProgramDeps): Command {
  program.addCommand(rosterCommand());
  program.addCommand(mcpCommand(depsOverride?.mcpDeps));
  program.addCommand(acpCommand(depsOverride?.acpDeps));
+  program.addCommand(execCommand(depsOverride?.execDeps));
  program.addCommand(agentCommand(depsOverride?.agentDeps));
  program.addCommand(rigCommand(depsOverride?.rigDeps));
  program.addCommand(transcriptCommand(depsOverride?.transcriptDeps));
@@ -42,6 +42,7 @@ export const COMMAND_LOADERS: ReadonlyMap<string, Load> = new Map<string, Load>(
  ["roster", async () => (await import("./commands/roster.js")).rosterCommand()],
  ["mcp", async () => (await import("./commands/mcp.js")).mcpCommand()],
  ["acp", async () => (await import("./commands/acp.js")).acpCommand()],
+  ["exec", async () => (await import("./commands/exec.js")).execCommand()],
  ["agent", async () => (await import("./commands/agent.js")).agentCommand()],
  ["spec", async () => (await import("./commands/rig.js")).rigCommand()],
  ["transcript", async () => (await import("./commands/transcript.js")).transcriptCommand()],
@@ -0,0 +1,382 @@
+import { createHash } from "node:crypto";
+import { constants, openSync, closeSync, fstatSync, lstatSync, readSync, existsSync } from "node:fs";
+import { isAbsolute, join, normalize } from "node:path";
+
+type Json = null | boolean | number | string | Json[] | { [key: string]: Json };
+type RecordValue = { [key: string]: Json };
+export const RECONSTRUCTION_SCHEMA = "openrig.reconstruction/v0";
+const MAX_DOCUMENT = 8 * 1024 * 1024;
+const MAX_CLOSURE = 32 * 1024 * 1024;
+const PRESERVED = ["authority", "constraint", "gap"];
+
+export class ReconstructionError extends Error {
+  readonly code: string;
+  constructor(code: string, message: string) { super(message); this.code = code; }
+}
+function requireValue(value: unknown, message: string): asserts value {
+  if (!value) throw new ReconstructionError("invalid_reconstruction", message);
+}
+function object(value: Json | undefined, label: string): RecordValue {
+  requireValue(value !== null && typeof value === "object" && !Array.isArray(value), `${label} must be an object`);
+  return value;
+}
+function list(value: Json | undefined, label: string): Json[] {
+  requireValue(Array.isArray(value), `${label} must be an array`); return value;
+}
+function text(value: Json | undefined, label: string): string {
+  requireValue(typeof value === "string" && value.trim(), `${label} must be a nonempty string`); return value;
+}
+function strings(value: Json | undefined, label: string): string[] {
+  const values = list(value === undefined ? [] : value, label);
+  requireValue(values.every(v => typeof v === "string"), `${label} must contain strings`); return values as string[];
+}
+function timestamp(value: Json | undefined, label: string, nullable = false): void {
+  if (value === null && nullable) return;
+  requireValue(typeof value === "string" && /(?:Z|[+-]\d\d:\d\d)$/.test(value) && Number.isFinite(Date.parse(value)), `${label} must have an ISO timestamp with timezone${nullable ? " or null" : ""}`);
+}
+function absolute(value: Json | undefined, label: string): string {
+  const name = text(value, label); requireValue(isAbsolute(name), `${label} must be an absolute local path`); return normalize(name);
+}
+function revision(value: Json | undefined): string {
+  const id = text(value, "revision"); requireValue(/^[a-f0-9]{64}$/.test(id), "revision must be a full lowercase SHA256 ID"); return id;
+}
+function validUnicode(value: string): void {
+  for (let i = 0; i < value.length; i++) {
+    const unit = value.charCodeAt(i);
+    if (unit >= 0xd800 && unit <= 0xdbff) {
+      const next = value.charCodeAt(++i);
+      requireValue(next >= 0xdc00 && next <= 0xdfff, "unpaired Unicode surrogate is unsupported by v0 UTF-8 hashing");
+    } else requireValue(unit < 0xdc00 || unit > 0xdfff, "unpaired Unicode surrogate is unsupported by v0 UTF-8 hashing");
+  }
+}
+
+/** v0 is Python sorted compact JSON, not JS property order. Reject numeric
+ * representations we cannot reproduce, rather than silently changing an ID. */
+export function parseReconstructionJson(input: string): Json {
+  let at = 0;
+  const space = () => { while (/[ \n\r\t]/.test(input[at] ?? "x")) at++; };
+  const quoted = (): string => {
+    const start = at++;
+    while (at < input.length) {
+      const char = input[at++];
+      if (char === "\\") at++;
+      else if (char === '"') {
+        const result: unknown = JSON.parse(input.slice(start, at));
+        requireValue(typeof result === "string", "invalid JSON string"); validUnicode(result); return result;
+      }
+    }
+    throw new ReconstructionError("invalid_reconstruction", "unterminated JSON string");
+  };
+  const value = (depth: number): Json => {
+    requireValue(depth <= 128, "JSON nesting exceeds 128"); space();
+    if (input[at] === '"') return quoted();
+    if (input[at] === "{") {
+      at++; space(); const result: RecordValue = Object.create(null) as RecordValue;
+      if (input[at] !== "}") while (true) {
+        requireValue(input[at] === '"', "expected JSON object key"); const key = quoted(); space();
+        requireValue(!Object.hasOwn(result, key), `duplicate JSON key: ${key}`);
+        requireValue(input[at++] === ":", "expected JSON colon"); result[key] = value(depth + 1); space();
+        if (input[at] !== ",") break; at++; space();
+      }
+      requireValue(input[at++] === "}", "expected JSON object end"); return result;
+    }
+    if (input[at] === "[") {
+      at++; space(); const result: Json[] = [];
+      if (input[at] !== "]") while (true) {
+        result.push(value(depth + 1)); space(); if (input[at] !== ",") break; at++;
+      }
+      requireValue(input[at++] === "]", "expected JSON array end"); return result;
+    }
+    for (const [token, result] of [["true", true], ["false", false], ["null", null]] as const) {
+      if (input.startsWith(token, at)) { at += token.length; return result; }
+    }
+    const number = /^-?(?:0|[1-9]\d*)(?:\.\d+)?(?:[eE][+-]?\d+)?/.exec(input.slice(at))?.[0];
+    requireValue(number !== undefined, "expected JSON value");
+    if (/[.eE]/.test(number) || !Number.isSafeInteger(Number(number))) {
+      throw new ReconstructionError("unsupported_v0_number", "v0 reader supports safe integer JSON numbers only; floating/exponent and unsafe integers require an explicit compatible reader, never an ID rewrite");
+    }
+    at += number.length; return Number(number);
+  };
+  try {
+    const parsed = value(0); space(); requireValue(at === input.length, "trailing JSON input"); return parsed;
+  } catch (error) {
+    if (error instanceof ReconstructionError) throw error;
+    throw new ReconstructionError("invalid_reconstruction", `invalid JSON: ${(error as Error).message}`);
+  }
+}
+
+function codepointOrder(a: string, b: string): number {
+  const aa = Array.from(a), bb = Array.from(b);
+  for (let i = 0; i < Math.min(aa.length, bb.length); i++) {
+    const difference = aa[i]!.codePointAt(0)! - bb[i]!.codePointAt(0)!; if (difference) return difference;
+  }
+  return aa.length - bb.length;
+}
+export function canonicalV0(value: Json): string {
+  if (value === null || typeof value !== "object") {
+    if (typeof value === "number") requireValue(Number.isSafeInteger(value), "unsupported v0 number");
+    if (typeof value === "string") validUnicode(value);
+    return JSON.stringify(value);
+  }
+  if (Array.isArray(value)) return `[${value.map(canonicalV0).join(",")}]`;
+  return `{${Object.keys(value).sort(codepointOrder).map(key => `${canonicalV0(key)}:${canonicalV0(value[key]!)}`).join(",")}}`;
+}
+const hash = (bytes: string | Buffer) => createHash("sha256").update(bytes).digest("hex");
+
+/** Regular-file, no-follow, bounded read; no source commands are executed. */
+function readRegular(file: string, limit: number, onRead?: (bytes: number) => void): Buffer {
+  const initial = lstatSync(file);
+  requireValue(initial.isFile() && !initial.isSymbolicLink(), `not a regular file: ${file}`);
+  const fd = openSync(file, constants.O_RDONLY | constants.O_NOFOLLOW | constants.O_NONBLOCK);
+  try {
+    const before = fstatSync(fd);
+    requireValue(before.isFile() && before.dev === initial.dev && before.ino === initial.ino && before.size <= limit, `file identity or byte bound failed: ${file}`);
+    const chunks: Buffer[] = []; let bytes = 0;
+    while (bytes < limit) {
+      const buffer = Buffer.alloc(Math.min(65536, limit - bytes));
+      const count = readSync(fd, buffer); if (!count) break;
+      bytes += count; onRead?.(count); chunks.push(buffer.subarray(0, count));
+    }
+    const after = fstatSync(fd), current = lstatSync(file);
+    requireValue(before.size === after.size && before.mtimeMs === after.mtimeMs && before.ctimeMs === after.ctimeMs
+      && after.dev === current.dev && after.ino === current.ino, `file changed during read: ${file}`);
+    return Buffer.concat(chunks);
+  } finally { closeSync(fd); }
+}
+function jsonFile(file: string, remainingBytes = MAX_DOCUMENT, onRead?: (bytes: number) => void): { value: Json; bytes: number } {
+  const bytes = readRegular(file, Math.min(MAX_DOCUMENT, remainingBytes), onRead);
+  return { value: parseReconstructionJson(new TextDecoder("utf-8", { fatal: true }).decode(bytes)), bytes: bytes.length };
+}
+
+export interface StoreMapping { sourceStore: string; localStore: string; sourceHost?: string }
+export function readStoreMappings(file: string): StoreMapping[] {
+  return list(jsonFile(file).value, "store mappings").map(value => {
+    const entry = object(value, "mapping");
+    return { sourceStore: absolute(entry.sourceStore, "sourceStore"), localStore: absolute(entry.localStore, "localStore"),
+      ...(entry.sourceHost === undefined ? {} : { sourceHost: text(entry.sourceHost, "sourceHost") }) };
+  });
+}
+interface Pointer { store: string; revision: string }
+function pointer(value: Json | undefined): Pointer {
+  const v = object(value, "pointer"); return { store: absolute(v.store, "pointer.store"), revision: revision(v.revision) };
+}
+function records(value: Json | undefined, kind: "items" | "sources"): RecordValue[] {
+  const seen = new Set<string>();
+  return list(value === undefined ? [] : value, kind).map(v => {
+    const r = object(v, kind), id = text(r.id, `${kind}.id`); requireValue(!seen.has(id), `duplicate ${kind} ID: ${id}`); seen.add(id);
+    timestamp(r.observed_at, `${kind}.observed_at`, true);
+    if (kind === "items") {
+      text(r.category, "item.category"); text(r.text, "item.text");
+      if (r.status !== undefined) text(r.status, "item.status");
+      for (const name of ["sources", "depends_on", "tags"]) strings(r[name], name);
+      requireValue(["source-belief", "assessment", "observed", "reported"].includes(String(r.perspective ?? "assessment")), "invalid item perspective");
+    } else {
+      requireValue(r.path !== undefined || r.address !== undefined, "source requires path or address");
+      if (r.path !== undefined) absolute(r.path, "source.path");
+      if (r.address !== undefined) text(r.address, "source.address");
+      if (r.sha256 !== undefined && r.sha256 !== null) revision(r.sha256);
+    }
+    return r;
+  });
+}
+function discovery(value: Json | undefined): Json[] {
+  const roots = list(value === undefined ? [] : value, "discovery"), ids = new Set<string>();
+  for (const value of roots) {
+    const root = object(value, "discovery entry"), id = text(root.id, "discovery.id");
+    requireValue(!ids.has(id), "duplicate discovery ID"); ids.add(id);
+    absolute(root.path, "discovery.path");
+    for (const name of strings(root.known_paths, "discovery.known_paths")) {
+      requireValue(!isAbsolute(name) && !name.split("/").includes(".."), "discovery known_paths must stay under root");
+    }
+    const entries = root.max_entries ?? 1000, depth = root.max_depth ?? 8;
+    requireValue(typeof entries === "number" && Number.isInteger(entries) && entries >= 1 && entries <= 10000, "discovery.max_entries must be 1..10000");
+    requireValue(typeof depth === "number" && Number.isInteger(depth) && depth >= 0 && depth <= 32, "discovery.max_depth must be 0..32");
+    timestamp(root.observed_at ?? null, "discovery.observed_at", true);
+  }
+  return roots;
+}
+function validate(doc: RecordValue): void {
+  if (doc.schema !== RECONSTRUCTION_SCHEMA) throw new ReconstructionError("unsupported_schema", `unsupported reconstruction schema: ${String(doc.schema)}`);
+  requireValue(["base", "delta", "composition"].includes(String(doc.kind)), "invalid reconstruction kind");
+  for (const key of ["subject", "purpose", "author"]) text(doc[key], key);
+  timestamp(doc.created_at, "created_at");
+  requireValue(["orientation", "working", "continuity", "full-domain"].includes(String(doc.depth)), "invalid reconstruction depth");
+  for (const key of ["items", "upsert_items"]) if (doc[key] !== undefined) records(doc[key], "items");
+  for (const key of ["sources", "upsert_sources"]) if (doc[key] !== undefined) records(doc[key], "sources");
+  if (doc.coverage !== undefined) {
+    const coverage = object(doc.coverage, "coverage"); strings(coverage.included, "coverage.included"); strings(coverage.excluded, "coverage.excluded");
+  }
+  strings(doc.unknowns, "unknowns"); discovery(doc.discovery);
+  for (const key of ["remove_items", "remove_sources"]) {
+    const ids = strings(doc[key], key); requireValue(new Set(ids).size === ids.length, `duplicate ${key}`);
+  }
+  if (doc.kind === "delta") {
+    pointer(doc.parent);
+    requireValue(!["items", "sources", "members"].some(key => key in doc), "delta uses upsert/remove fields; members remain pinned");
+  } else requireValue(!Object.keys(doc).some(key => key.startsWith("upsert_") || key.startsWith("remove_")), "base/composition uses items and sources, not delta fields");
+  if (doc.kind === "composition") {
+    const members = list(doc.members, "members"); requireValue(members.length > 0 && members.length <= 64, "composition requires 1..64 members"); members.forEach(pointer);
+  }
+}
+interface Node {
+  document: RecordValue; sourceStore: string; localStore: string; items: RecordValue[]; sources: RecordValue[];
+  coverage: Json; unknowns: string[]; discovery: Json[]; members: Node[]; lineage: Pointer[];
+}
+interface Issue { code: string; message: string; store?: string; revision?: string; subject?: string }
+export interface ReconstructionOptions {
+  store: string; revision: string; mappings?: StoreMapping[];
+  items?: string[]; categories?: string[]; tags?: string[]; checkSources?: boolean;
+}
+const flatten = (node: Node): Node[] => [node, ...node.members.flatMap(flatten)];
+
+export function readReconstruction(options: ReconstructionOptions) {
+  const rootStore = absolute(options.store, "store"), rootRevision = revision(options.revision);
+  const mappings = new Map<string, StoreMapping>();
+  for (const entry of options.mappings ?? []) {
+    const source = absolute(entry.sourceStore, "sourceStore");
+    requireValue(!mappings.has(source), "ambiguous duplicate source-store mapping; v0 pointers have no host discriminator");
+    mappings.set(source, { ...entry, sourceStore: source, localStore: absolute(entry.localStore, "localStore") });
+  }
+  const issues: Issue[] = [], loaded: Array<{ revision: string; sourceStore: string; localStore: string; fileBytes: number }> = [];
+  let totalBytes = 0, visited = 0;
+  let exhausted = false;
+  const materialize = (store: string, id: string, stack: string[]): Node | undefined => {
+    if (exhausted) return undefined;
+    const sourceStore = absolute(store, "pointer.store"), key = `${sourceStore}\0${id}`;
+    try {
+      if (stack.length >= 64 || ++visited > 64 || totalBytes >= MAX_CLOSURE) {
+        exhausted = true;
+        throw new ReconstructionError("closure_limit", "reconstruction closure exceeds 64 references or 32 MiB");
+      }
+      requireValue(!stack.includes(key), "reconstruction reference cycle");
+      const mapping = mappings.get(sourceStore);
+      if (!mapping && sourceStore !== rootStore && !existsSync(sourceStore)) throw new ReconstructionError("unmapped_external_store", `store ${sourceStore} is not available here; supply an explicit store mapping`);
+      const localStore = mapping?.localStore ?? sourceStore;
+      // Charge positive reads even when decoding, parsing or a late identity check fails.
+      const file = jsonFile(join(localStore, `${revision(id)}.json`), MAX_CLOSURE - totalBytes, count => { totalBytes += count; });
+      const doc = object(file.value, "revision"); requireValue(doc.id === id, "stored ID does not match requested revision");
+      const { id: ignored, ...body } = doc; void ignored;
+      requireValue(hash(canonicalV0(body)) === id, "revision content hash mismatch"); validate(doc);
+      loaded.push({ revision: id, sourceStore, localStore, fileBytes: file.bytes });
+      const trail = [...stack, key]; let result: Node;
+      if (doc.kind === "delta") {
+        const parent = pointer(doc.parent), inherited = materialize(parent.store, parent.revision, trail);
+        if (!inherited) return undefined;
+        requireValue(inherited.document.subject === doc.subject, "delta cannot change subject identity");
+        result = { ...inherited, members: [...inherited.members], lineage: [...inherited.lineage, { store: sourceStore, revision: id }] };
+        for (const collection of ["items", "sources"] as const) {
+          const values = new Map(inherited[collection].map(v => [String(v.id), v]));
+          const removed = strings(doc[`remove_${collection}`], `remove_${collection}`);
+          const updates = records(doc[`upsert_${collection}`], collection);
+          for (const remove of removed) {
+            requireValue(values.has(remove) && !updates.some(v => v.id === remove), `invalid removal/upsert: ${remove}`); values.delete(remove);
+          }
+          for (const v of updates) values.set(String(v.id), v);
+          result[collection] = [...values.values()];
+        }
+        if (doc.coverage !== undefined) result.coverage = doc.coverage;
+        if (doc.unknowns !== undefined) result.unknowns = strings(doc.unknowns, "unknowns");
+        if (doc.discovery !== undefined) result.discovery = list(doc.discovery, "discovery");
+      } else {
+        result = { document: doc, sourceStore, localStore, items: records(doc.items, "items"), sources: records(doc.sources, "sources"),
+          coverage: doc.coverage ?? { included: [], excluded: [] }, unknowns: strings(doc.unknowns, "unknowns"), discovery: discovery(doc.discovery),
+          members: [], lineage: [{ store: sourceStore, revision: id }] };
+        for (const member of list(doc.members === undefined ? [] : doc.members, "members")) {
+          const p = pointer(member), node = materialize(p.store, p.revision, trail); if (node) result.members.push(node);
+        }
+      }
+      return { ...result, document: doc, sourceStore, localStore };
+    } catch (error) {
+      issues.push({ code: error instanceof ReconstructionError ? error.code : "revision_unavailable", message: (error as Error).message, store: sourceStore, revision: id });
+      return undefined;
+    }
+  };
+  let root = materialize(rootStore, rootRevision, []);
+  const seenSubjects = new Set<string>();
+  if (root) for (const node of flatten(root)) {
+    const subject = String(node.document.subject);
+    if (seenSubjects.has(subject)) { issues.push({ code: "duplicate_subject", message: `composition repeats subject ${subject}` }); root = undefined; break; }
+    seenSubjects.add(subject);
+  }
+  const nodes = root ? flatten(root) : [], allIds = new Set(nodes.flatMap(node => node.items.map(v => String(v.id))));
+  const filtered = Boolean(options.items?.length || options.categories?.length || options.tags?.length);
+  for (const id of options.items ?? []) if (!allIds.has(id)) issues.push({ code: "missing_item", message: `selected item is missing: ${id}` });
+  const selected = nodes.map(node => {
+    const byId = new Map(node.items.map(v => [String(v.id), v]));
+    const ids = new Set(node.items.filter(v => !filtered || options.items?.includes(String(v.id)) || options.categories?.includes(String(v.category))
+      || strings(v.tags, "tags").some(tag => options.tags?.includes(tag)) || PRESERVED.includes(String(v.category))).map(v => String(v.id)));
+    const pending = [...ids];
+    while (pending.length) for (const dependency of strings(byId.get(pending.pop()!)!.depends_on, "depends_on")) {
+      if (!byId.has(dependency)) issues.push({ code: "missing_dependency", message: `missing dependency: ${dependency}`, subject: String(node.document.subject) });
+      else if (!ids.has(dependency)) { ids.add(dependency); pending.push(dependency); }
+    }
+    const items = node.items.filter(v => ids.has(String(v.id))), cited = new Set(items.flatMap(v => strings(v.sources, "sources")));
+    const sources = new Map(node.sources.map(v => [String(v.id), v]));
+    for (const id of cited) if (!sources.has(id)) issues.push({ code: "missing_citation", message: `missing source: ${id}`, subject: String(node.document.subject) });
+    return { node, items, cited, ids };
+  });
+  let sourceBytes = 0, sourceCount = 0;
+  const sourceChecks = options.checkSources ? nodes.flatMap(node => node.sources.map(source => {
+    const base = { subject: node.document.subject, id: source.id, observedAt: source.observed_at };
+    if (source.path === undefined) return { ...base, status: "unchecked", reason: "opaque/nonlocal route; not fetched" };
+    try {
+      requireValue(++sourceCount <= 128 && sourceBytes < 64 * 1024 * 1024, "source checking budget exhausted (128 files / 64 MiB total)");
+      const bytes = readRegular(String(source.path), 64 * 1024 * 1024 - sourceBytes, count => { sourceBytes += count; });
+      const actual = hash(bytes);
+      return { ...base, sha256: actual, status: source.sha256 ? (actual === source.sha256 ? "unchanged" : "changed") : "unchecked", reason: "byte comparison only; observation dates and semantic freshness are not refreshed" };
+    } catch (error) { return { ...base, status: "unavailable", reason: (error as Error).message }; }
+  })) : undefined;
+  const materializedBytes = Buffer.byteLength(JSON.stringify(root ?? null));
+  return {
+    presentationSchema: "openrig.reconstruction-reader/v1", storedSchema: RECONSTRUCTION_SCHEMA,
+    requestedStore: rootStore, requestedRevision: rootRevision, complete: issues.length === 0, issues,
+    completenessScope: "record closure and selected item/citation references; not source freshness",
+    caveat: "Context is evidence for understanding, not authority, competence, acceptance or current runtime state. Authors are claimed attribution. No discovery cursor or implicit remote reads.",
+    selectors: { mode: filtered ? "filtered" : "full", items: options.items ?? [], categories: options.categories ?? [], tags: options.tags ?? [], combination: "union", alwaysIncludeCategories: PRESERVED, includeTransitiveDependencies: true },
+    mappings: [...mappings.values()], records: loaded, closureBytesRead: totalBytes, atomicGroupCapture: false,
+    size: { materializedJsonUtf8Bytes: materializedBytes, approximateTokens: Math.ceil(materializedBytes / 4),
+      estimateMethod: "materialized reader JSON UTF-8 bytes / 4; rough, not a model tokenizer" },
+    nodes: selected.map(({ node, items, cited, ids }) => ({
+      subject: node.document.subject, revision: node.document.id, kind: node.document.kind, author: node.document.author,
+      purpose: node.document.purpose, depth: node.document.depth, createdAt: node.document.created_at,
+      sourceStore: node.sourceStore, localStore: node.localStore, lineage: node.lineage,
+      declaredCoverage: node.coverage, unknowns: node.unknowns, discovery: node.discovery,
+      sourceManifest: node.sources, itemObservations: node.items.map(v => ({ id: v.id, observedAt: v.observed_at })),
+      items, viewSources: node.sources.filter(v => !filtered || cited.has(String(v.id))),
+      viewCoverage: { totalItems: node.items.length, includedIds: items.map(v => v.id), omittedIds: node.items.filter(v => !ids.has(String(v.id))).map(v => v.id),
+        registeredSources: node.sources.length, citedSourceIds: [...cited], omittedRouteIds: filtered ? node.sources.filter(v => !cited.has(String(v.id))).map(v => v.id) : [] },
+    })), ...(sourceChecks ? { sourceChecks } : {}),
+  };
+}
+
+export type ReconstructionRead = ReturnType<typeof readReconstruction>;
+export function reconstructionOutput(result: ReconstructionRead, mode: "inspect" | "view", json: boolean, maxBytes = 256 * 1024): string {
+  requireValue(Number.isInteger(maxBytes) && maxBytes >= 1024 && maxBytes <= MAX_DOCUMENT, "max-bytes must be 1024..8388608");
+  // Inspect never delivers item bodies; view carries its actual omissions and
+  // complete-manifest route rather than presenting full coverage as selected.
+  const data = { ...result, nodes: result.nodes.map(node => mode === "inspect" ? { ...node, items: undefined, viewSources: undefined } : { ...node, sourceManifest: undefined }) };
+  let output: string;
+  if (json || mode === "inspect") output = JSON.stringify(data, null, 2);
+  else {
+    const lines = ["# Reconstruction view", "", result.caveat, `Complete: ${result.complete} (${result.completenessScope})`, `Applied selectors: ${JSON.stringify(result.selectors)}`];
+    for (const issue of result.issues) lines.push(`UNAVAILABLE: ${issue.code}: ${issue.message}`);
+    for (const node of result.nodes) {
+      lines.push("", `## ${String(node.subject)}`, `Revision: ${String(node.revision)}`, `Author (claimed): ${String(node.author)}`, `Purpose: ${String(node.purpose)}`, `Depth: ${String(node.depth)}`,
+        `Source store: ${node.sourceStore}; local read store: ${node.localStore}`,
+        `Created: ${String(node.createdAt)} (not inherited-fact verification time)`, `Author-declared coverage: ${JSON.stringify(node.declaredCoverage)}`, `Selected view coverage: ${JSON.stringify(node.viewCoverage)}`,
+        "Composition members retain independent observation dates; no atomic capture is claimed.");
+      for (const unknown of node.unknowns) lines.push(`Unknown: ${unknown}`);
+      for (const item of node.items) lines.push("", `### ${String(item.id)} [${String(item.category)}]`, String(item.text),
+        `Perspective: ${String(item.perspective ?? "assessment")}; observed: ${String(item.observed_at)}; status: ${String(item.status ?? "unspecified")}`,
+        `Sources: ${JSON.stringify(item.sources ?? [])}; dependencies: ${JSON.stringify(item.depends_on ?? [])}`);
+      lines.push("", "Evidence routes:");
+      for (const source of node.viewSources) lines.push(JSON.stringify(source));
+      if (node.viewCoverage.omittedRouteIds.length) lines.push("Uncited routes omitted; inspect this exact revision for its complete source manifest.");
+    }
+    if (result.sourceChecks) lines.push("Source byte checks (not semantic reverification):", JSON.stringify(result.sourceChecks));
+    output = lines.join("\n");
+  }
+  const bytes = Buffer.byteLength(output);
+  if (bytes > maxBytes) throw new ReconstructionError("output_limit", `output is ${bytes} UTF-8 bytes; max-bytes is ${maxBytes}. Select a narrower view or explicitly raise the bound; no complete output was delivered.`);
+  return output;
+}
@@ -215,45 +215,22 @@ function resolveExplicitSlice(
  return matches[0]!;
 }

-export function resolveWorkPosition(opts: {
+export interface WorkProjectOptions {
  workspaceRoot: string;
  catalogPath?: string;
-  contextRoot: string;
-  systemWorldSelection: string;
-  systemWorldSource: SystemWorldSource;
  project?: string;
-  mission?: string;
-  slice?: string;
-  /** Working directory for selection step 4 (the CLI passes --cwd or the process cwd). */
  cwd?: string;
-  /** Calling seat's canonical session, for selection step 3; absent in a plain shell. */
  sessionName?: string;
-}): WorkInstallResult {
+}
+
+/** Catalog/home resolution shared by read-only work consumers and work-install.
+ * No System World, skill selection or projection is performed here. */
+export function resolveWorkProject(opts: WorkProjectOptions, warnings: string[]) {
  if (opts.project !== undefined && !SEGMENT.test(opts.project)) {
    return failure("invalid_project", "project must be a single bounded segment");
  }
-  if (opts.mission !== undefined && !SEGMENT.test(opts.mission)) {
-    return failure("invalid_mission", "mission must be a single bounded segment");
-  }
-  if (opts.slice !== undefined && !SEGMENT.test(opts.slice)) {
-    return failure("invalid_slice", "slice must be a single bounded segment");
-  }
-  if (opts.slice !== undefined && opts.mission === undefined) {
-    return failure("mission_required", "slice requires an exact mission selection");
-  }
-
  const workspaceRoot = canonicalExisting(opts.workspaceRoot);
-  if (!workspaceRoot) {
-    return failure("workspace_root_missing", `workspace root does not exist: ${resolve(opts.workspaceRoot)}`);
-  }
-
-  const warnings: string[] = [];
-  const systemWorld = resolveSystemWorld({
-    contextRoot: opts.contextRoot,
-    selection: opts.systemWorldSelection,
-    source: opts.systemWorldSource,
-  });
-  if (!systemWorld.ok) return failure(systemWorld.error.code, systemWorld.error.message);
+  if (!workspaceRoot) return failure("workspace_root_missing", `workspace root does not exist: ${resolve(opts.workspaceRoot)}`);
  const catalogPath = resolve(opts.catalogPath ?? join(workspaceRoot, "workspace.yaml"));
  let projectId: string | null = null;
  let projectRoot = workspaceRoot;
@@ -303,6 +280,85 @@ export function resolveWorkPosition(opts: {
    return failure("missions_root_escape", "project.yaml missions.root resolves outside the selected project root");
  }

+  return { workspaceRoot, projectId, projectRoot, selectedBy, projectManifest, missionsRoot };
+}
+
+/** A context home is a location, not task ownership or runtime authority. */
+export function resolveWorkContextHome(opts: WorkProjectOptions & { mission?: string; slice?: string }) {
+  const warnings: string[] = [];
+  if (opts.mission !== undefined && !SEGMENT.test(opts.mission)) return failure("invalid_mission", "mission must be a single bounded segment");
+  if (opts.slice !== undefined && !SEGMENT.test(opts.slice)) return failure("invalid_slice", "slice must be a single bounded segment");
+  if (opts.slice !== undefined && opts.mission === undefined) return failure("mission_required", "slice requires an exact mission selection");
+  const project = resolveWorkProject(opts, warnings);
+  if ("error" in project) return project;
+  let home = project.projectRoot;
+  if (opts.mission !== undefined) {
+    const nominal = join(project.missionsRoot, opts.mission);
+    const mission = canonicalExisting(nominal);
+    if (!mission) return failure("mission_not_found", `mission '${opts.mission}' does not exist in the selected project`);
+    if (!inside(project.projectRoot, mission)) return failure("mission_root_escape", "mission resolves outside the selected project root");
+    home = mission;
+    if (opts.slice !== undefined) {
+      let declared: Set<string> | undefined;
+      const manifestPath = join(mission, "mission.yaml");
+      if (existsSync(manifestPath)) {
+        const parsed = readYaml(manifestPath), composition = parsed.value?.["composition"];
+        if (parsed.value && isRecord(composition) && "slices" in composition) {
+          try { declared = new Set(validateMissionComposition(parsed.value, manifestPath).map(member => dirname(member.path))); }
+          catch (error) { return failure("mission_composition_invalid", (error as Error).message); }
+        } else if (!parsed.value) warnings.push(`${parsed.error}; kept conventional slice addresses`);
+      }
+      const slice = resolveExplicitSlice(mission, opts.slice, declared);
+      if ("error" in slice) return slice;
+      home = slice.root;
+    }
+  }
+  return { projectId: project.projectId, projectRoot: project.projectRoot, selectedBy: project.selectedBy, home, warnings };
+}
+
+export function resolveWorkPosition(opts: {
+  workspaceRoot: string;
+  catalogPath?: string;
+  contextRoot: string;
+  systemWorldSelection: string;
+  systemWorldSource: SystemWorldSource;
+  project?: string;
+  mission?: string;
+  slice?: string;
+  /** Working directory for selection step 4 (the CLI passes --cwd or the process cwd). */
+  cwd?: string;
+  /** Calling seat's canonical session, for selection step 3; absent in a plain shell. */
+  sessionName?: string;
+}): WorkInstallResult {
+  if (opts.project !== undefined && !SEGMENT.test(opts.project)) {
+    return failure("invalid_project", "project must be a single bounded segment");
+  }
+  if (opts.mission !== undefined && !SEGMENT.test(opts.mission)) {
+    return failure("invalid_mission", "mission must be a single bounded segment");
+  }
+  if (opts.slice !== undefined && !SEGMENT.test(opts.slice)) {
+    return failure("invalid_slice", "slice must be a single bounded segment");
+  }
+  if (opts.slice !== undefined && opts.mission === undefined) {
+    return failure("mission_required", "slice requires an exact mission selection");
+  }
+
+  const workspaceRoot = canonicalExisting(opts.workspaceRoot);
+  if (!workspaceRoot) {
+    return failure("workspace_root_missing", `workspace root does not exist: ${resolve(opts.workspaceRoot)}`);
+  }
+
+  const warnings: string[] = [];
+  const systemWorld = resolveSystemWorld({
+    contextRoot: opts.contextRoot,
+    selection: opts.systemWorldSelection,
+    source: opts.systemWorldSource,
+  });
+  if (!systemWorld.ok) return failure(systemWorld.error.code, systemWorld.error.message);
+  const project = resolveWorkProject({ ...opts, workspaceRoot }, warnings);
+  if ("error" in project) return project;
+  const { projectId, projectRoot, selectedBy, projectManifest, missionsRoot } = project;
+
  let projectIntent = "SPEC.md";
  let projectIntentSource: WorkInstallSource = "default";
  let projectContext: string[] = [];
@@ -19,7 +19,7 @@ import { createHash, randomBytes, timingSafeEqual } from "node:crypto";
 import { createMcpHandler, type AuthInfo } from "@modelcontextprotocol/server";
 import { toNodeHandler } from "@modelcontextprotocol/node";
 import { DaemonClient } from "./client.js";
-import { createMcpServer, type McpAccess } from "./mcp-server.js";
+import { createMcpServer, mintUnknownSender, type McpAccess, type UnfilledSeat } from "./mcp-server.js";
 import { getOpenRigHome } from "./openrig-compat.js";
 import { MCP_PRESENCE_FILE, presenceRecorder } from "./mcp-presence.js";
 import { resolveRigHandle } from "./rig-handle.js";
@@ -310,6 +310,12 @@ export type McpHttpOptions = {
  clientFactory?: (daemonUrl: string, sessionName?: string, principal?: McpPrincipal) => DaemonClient;
  /** Grants come from an external grant service instead of the token file, which is then never read. */
  grantService?: GrantServiceOptions;
+  /** Operator-configured receiver for external callers omitting a task destination; never an authority grant. */
+  taskDestination?: string;
+  /** rig mcp serve --front-door: a caller that isn't external gets the front-door instructions, not the member workflow. */
+  frontDoor?: boolean;
+  /** The seat value rig mcp serve dropped as empty or unfilled, which rig_whoami names to a caller with no seat. */
+  unfilledSeat?: UnfilledSeat;
  /** With grantService: OAuth discovery for clients that sign in with OAuth. The public resource this server is reached
   *  as and the authorization server that issues its keys, both served exactly as given. */
  oauth?: { resource: string; issuer: string };
@@ -338,6 +344,12 @@ function readBody(req: http.IncomingMessage): Promise<string> {
 export async function startMcpHttpServer(opts: McpHttpOptions): Promise<{ url: string; authenticated: boolean; close(): Promise<void> }> {
  const loopback = isLoopbackHost(opts.host);
  const service = opts.grantService;
+  if (opts.taskDestination !== undefined) {
+    if (!service) throw new McpHttpConfigError("--task-destination needs --grant-service");
+    if (!/^[^\s@]+@[^\s@]+$/.test(opts.taskDestination) || opts.taskDestination.endsWith("@external")) {
+      throw new McpHttpConfigError("--task-destination takes one local team seat, member@rig");
+    }
+  }
  const serviceUrl = service ? grantServiceUrl(service.url) : undefined;
  // The metadata describes the keys a grant service checks, so it's served only with one.
  if (opts.oauth && !service) throw new McpHttpConfigError("--oauth-resource and --oauth-issuer need --grant-service");
@@ -377,12 +389,16 @@ export async function startMcpHttpServer(opts: McpHttpOptions): Promise<{ url: s
  };
  // One handler for both protocol revisions: a request carrying 2026-07-28's per-request envelope is served as that
  // revision, anything else (initialize and the rest of 2025) statelessly as before. Each request gets a fresh server for
-  // the caller decided below, handed over as the request's authInfo.
+  // the caller decided below, handed over as the request's authInfo. A caller with no seat writes as this process's one
+  // unknown sender, so its later requests carry the label rig_whoami named and a same-id create retry matches.
+  const unknownSender = mintUnknownSender();
  const mcpHandler = createMcpHandler(({ authInfo }) => {
    const caller = (authInfo?.extra as { caller?: RequestCaller } | undefined)?.caller;
    // Never a default seat: a request reaches the handler only with its caller decided.
    if (!caller) throw new Error("a request reached the MCP handler without its caller");
-    return createMcpServer(makeClient(opts.daemonUrl, caller.seat, caller.principal), { seat: caller.seat, access: caller.access, presenceFile: () => presenceFile });
+    return createMcpServer(makeClient(opts.daemonUrl, caller.seat, caller.principal), { seat: caller.seat, access: caller.access,
+      taskDestination: opts.taskDestination, presenceFile: () => presenceFile, frontDoor: opts.frontDoor === true,
+      unknownSender, ...(opts.unfilledSeat ? { unfilledSeat: opts.unfilledSeat } : {}) });
  });
  const serveMcp = toNodeHandler(mcpHandler);
  const bearer = (req: http.IncomingMessage): string | undefined => /^Bearer\s+(\S+)\s*$/i.exec(req.headers.authorization ?? "")?.[1];
@@ -12,6 +12,7 @@ import { CLI_VERSION } from "./version.js";
 import { ConfigStore } from "./config-store.js";
 import { readRosters } from "./commands/roster.js";
 import { mcpPresenceFilePath, readPresence } from "./mcp-presence.js";
+import { codeResultOf, planItemArtifact, type FileRead, type SharedRoot } from "./queue-artifact.js";

 const LONG_RUNNING_UP_TIMEOUT_MS = 120_000;
 const WAIT_FOR_IDLE_REQUEST_OVERHEAD_MS = 5_000;
@@ -396,20 +397,6 @@ function outOfScope(what: string): TextResult {
  return errorResult(JSON.stringify({ ok: false, error: "out_of_scope", message }), "out_of_scope", message);
 }

-type SharedRoot = { name: string; path: string };
-type FileRead = { root?: string; path?: string; content?: string; binary?: boolean; size?: number; mtime?: string; truncated?: boolean; truncatedAtBytes?: number | null; totalBytes?: number };
-
-/** The shared root an absolute path falls under as written, and its path inside that root. The reference names a file
- *  where the daemon runs, which may be another machine, so it's never resolved through this machine's filesystem; the
- *  daemon still checks containment itself when the file is read. */
-function underSharedRoot(roots: SharedRoot[], absolute: string): { root: string; path: string } | null {
-  for (const root of roots) {
-    const inside = nodePath.relative(root.path, absolute);
-    if (inside && inside !== ".." && !inside.startsWith(`..${nodePath.sep}`) && !nodePath.isAbsolute(inside)) return { root: root.name, path: inside };
-  }
-  return null;
-}
-
 /** A file the daemon read through a shared root, as text from offset: at most CONTEXT_TEXT_LIMIT characters, with
 *  nextOffset when more of what the daemon read remains, and its byte cut when the file was longer than the daemon reads. */
 function artifactView(source: string, read: FileRead, offset: number): Record<string, unknown> {
@@ -418,7 +405,7 @@ function artifactView(source: string, read: FileRead, offset: number): Record<st
  const text = read.content ?? "";
  const end = wholeCharEnd(text, offset + CONTEXT_TEXT_LIMIT);
  return {
-    ...base, binary: false, offset, text: text.slice(offset, end), totalChars: text.length,
+    ...base, binary: false, offset, text: text.slice(offset, end), totalChars: text.length, ...codeResultOf(read),
    truncated: text.length > end || read.truncated === true,
    ...(text.length > end ? { nextOffset: end } : {}),
    ...(read.truncated ? { truncatedAtBytes: read.truncatedAtBytes, totalBytes: read.totalBytes } : {}),
@@ -572,18 +559,54 @@ const SHARED_INSTRUCTIONS = [
  "A question reserved for a person is theirs to answer: don't answer it on their behalf. rig_queue_decide records a decision only inside the decision room of the mode they have on, beside their answers and never as one.",
 ];
 const INSTRUCTIONS = [...SHARED_INSTRUCTIONS, MEMBER_WORKFLOW].join("\n");
-const EXTERNAL_INSTRUCTIONS = SHARED_INSTRUCTIONS.join("\n");
+
+/** rig mcp serve --front-door: the seat is a person's agent in their own app, the front door to their team, not a member
+ *  that works its inbox. The team's questions to it are the person's to answer. */
+export const FRONT_DOOR_WORKFLOW_VERSION = "front-door/v2";
+export const FRONT_DOOR_WORKFLOW = [
+  `Front door (${FRONT_DOOR_WORKFLOW_VERSION}). This seat is a person's agent in their own app: you hand the team their work and bring back what it produced, and you aren't one of its members.`,
+  "Items addressed to your seat (rig_queue_list with owned set) are usually the team's questions about work you handed it. They're the person's to answer, not yours: show each to the person with the team's draft, and record only what the person tells you.",
+  "To record an answer, claim the item (rig_queue_claim), then close it done with closure reason no-follow-on and exactly what the person typed in resolution, with no label, quotation marks or words of your own (rig_queue_update). If the person isn't there, leave the item open and tell them it's waiting. Never choose an answer for them, and don't claim an item you can't close with their words.",
+  "Nothing wakes this conversation when the team finishes: check your work and your items when the person asks or comes back.",
+].join(" ");
+const FRONT_DOOR_INSTRUCTIONS = [...SHARED_INSTRUCTIONS, FRONT_DOOR_WORKFLOW].join("\n");
+const EXTERNAL_INSTRUCTIONS = [
+  "This connection sends requests to an OpenRig team as the caller named by its grant. The daemon attributes requests to that caller. Use only the tools this connection offers.",
+  "To ask the connected team to do work, use rig_queue_create with the task in body. You may omit destination to use the receiver configured by the server operator. If no receiver is available, ask the operator to configure it; don't guess an address. An explicit destination still must be within this connection's grant.",
+  "Keep the returned qitemId and use rig_queue_show to read the result. Acceptance is not completed work. If a create times out, check that same qitemId before retrying and reuse it; never retry with a new id. Use revises for a requested revision of earlier work, preserving its lineage. rig_queue_list shows your own work, including an empty list before your first request.",
+  "Read returned files with rig_artifact_read, within the shared roots and items this connection may reach. A question reserved for a person is theirs to answer; don't answer on their behalf.",
+].join("\n");
+
 /** For an external grant that names claim and update: how it works the items addressed to it, messages included. */
 const EXTERNAL_INBOX_INSTRUCTIONS = [...SHARED_INSTRUCTIONS,
  "Items addressed to your seat are yours: rig_queue_list with owned set lists them. For each, claim it (rig_queue_claim), do it, then close it done with closure reason no-follow-on and the answer in resolution (rig_queue_update); the sender reads that answer on the same item. A message is an item too: read it, close it done, and reply, when it needs one, with a new item to its sender. Nothing promises to wake you, so check your items again.",
 ].join("\n");

-function instructionsFor(access: McpAccess | undefined): string {
-  if (!access?.external) return INSTRUCTIONS;
+function instructionsFor(access: McpAccess | undefined, frontDoor = false): string {
+  if (!access?.external) return frontDoor ? FRONT_DOOR_INSTRUCTIONS : INSTRUCTIONS;
  const tools = externalTools(access.external.objects);
  return tools.has("rig_queue_claim") && tools.has("rig_queue_update") ? EXTERNAL_INBOX_INSTRUCTIONS : EXTERNAL_INSTRUCTIONS;
 }

+/** A configured default must identify exactly one existing local node. No topology relationship, name convention
+ *  for a lead, caller identity or task history supplies a missing receiver. Resolve again on each omitted destination. */
+async function taskDestinationExists(client: DaemonClient, destination: string): Promise<boolean> {
+  const rigName = destination.split("@")[1]!;
+  const rig = await resolveRigHandle(client, rigName, { strict: true });
+  if (rig.kind !== "resolved") return false;
+  const nodes = await client.get(`/api/rigs/${encodeURIComponent(rig.id)}/nodes`);
+  if (nodes.status !== 200 || !Array.isArray(nodes.data)) return false;
+  let matches = 0;
+  for (const entry of nodes.data as unknown[]) {
+    const { logicalId, rigName: name } = (entry ?? {}) as Record<string, unknown>;
+    if (typeof logicalId !== "string" || name !== rigName) return false;
+    const dot = logicalId.indexOf(".");
+    if (dot <= 0 || dot === logicalId.length - 1) return false;
+    if (`${logicalId.slice(0, dot)}-${logicalId.slice(dot + 1)}@${name}` === destination) matches++;
+  }
+  return matches === 1;
+}
+
 /**
 * Maps a DaemonResponse to MCP tool result.
 * Three error levels: HTTP status, in-body errors, and structural failure fields.
@@ -626,21 +649,73 @@ function mapResult(
  return opts.structured ? { content: [{ type: "text", text }], structuredContent: structured(res.data) } : { content: [{ type: "text", text }] };
 }

+/** A seat value an app passed on without filling it in: empty, or still a placeholder. */
+export interface UnfilledSeat { variable: "OPENRIG_SESSION_NAME" | "RIGGED_SESSION_NAME"; value: string }
+
+// An app that doesn't substitute its settings passes the template through as written: ${user_config.seat}, ${input:seat},
+// $SEAT. No seat address can contain these.
+const SEAT_PLACEHOLDER = /\$\{[^}]*\}|^\s*\$[A-Za-z_]/;
+
+/** Treat an empty or unfilled seat as no seat, for the whole process: drops both names from env, so neither requests
+ *  nor rig_whoami act as the placeholder, or as a seat inherited under the legacy name behind an empty one. A set
+ *  OPENRIG_SESSION_NAME decides; RIGGED_SESSION_NAME counts only when it's absent. Returns what was dropped. */
+export function dropUnfilledSeat(env: NodeJS.ProcessEnv = process.env): UnfilledSeat | undefined {
+  const variable = env.OPENRIG_SESSION_NAME !== undefined ? "OPENRIG_SESSION_NAME" : env.RIGGED_SESSION_NAME !== undefined ? "RIGGED_SESSION_NAME" : undefined;
+  const value = variable ? env[variable]! : undefined;
+  if (!variable || value === undefined || (value.trim() && !SEAT_PLACEHOLDER.test(value))) return undefined;
+  delete env.OPENRIG_SESSION_NAME;
+  delete env.RIGGED_SESSION_NAME;
+  return { variable, value };
+}
+
+/** How an MCP client named itself in its handshake, cut to one short line; it's the client's claim, never an identity. */
+function clientSaid(client: { name?: unknown; version?: unknown } | undefined): string | null {
+  const part = (value: unknown) => typeof value === "string" ? value.replace(/[\u0000-\u001f\u007f]+/g, " ").trim().slice(0, 60) : "";
+  const name = part(client?.name);
+  if (!name) return null;
+  const version = part(client?.version);
+  return JSON.stringify(version ? `${name} ${version}` : name);
+}
+
+export function unfilledSeatText(seat: UnfilledSeat): string {
+  return seat.value.trim()
+    ? `${seat.variable} is ${JSON.stringify(seat.value)}, a placeholder the app didn't fill in`
+    : `${seat.variable} is empty`;
+}
+
+/** A label for a caller with no seat: no seat's address, and random, so each `rig mcp serve` process has its own. */
+export function mintUnknownSender(): string {
+  return `unknown-mcp-${randomBytes(4).toString("hex")}`;
+}
+
 /**
 * Creates an MCP server wrapping the daemon HTTP API.
 * @param client - DaemonClient connected to the daemon
 * @returns McpServer instance (not yet connected to a transport)
 */
-export function createMcpServer(client: DaemonClient, options: { seat?: string; access?: McpAccess; rostersFolder?: () => string; presenceFile?: () => string } = {}): McpServer {
+export function createMcpServer(client: DaemonClient, options: { seat?: string; access?: McpAccess; taskDestination?: string; rostersFolder?: () => string; presenceFile?: () => string; frontDoor?: boolean; unfilledSeat?: UnfilledSeat; unknownSender?: string } = {}): McpServer {
  const server = new McpServer({
    name: "openrig",
    version: CLI_VERSION,
-  }, { instructions: instructionsFor(options.access) });
+  }, { instructions: instructionsFor(options.access, options.frontDoor === true) });
  const { access } = options;
  const rostersFolder = options.rostersFolder ?? (() => nodePath.join(new ConfigStore().get("workspace.root") as string, "rosters"));
  const presenceFile = options.presenceFile ?? (() => mcpPresenceFilePath());
  // The seat this server acts as: the authenticated caller's, else OPENRIG_SESSION_NAME where it was configured.
  const seatOf = (): string | undefined => options.seat ?? readOpenRigEnv("OPENRIG_SESSION_NAME", "RIGGED_SESSION_NAME")?.trim();
+  // With no seat (none configured, or one an app left unfilled), messages, new tasks and chat still go through, as an
+  // unknown sender. Its label is this server's own and no seat's, so two unknown callers never count as one author (a
+  // delegated decision compares authors). Its note says only what is known: why there's no seat, and how the client
+  // describes itself in its handshake, which is the client's claim, not an identity. Every other write keeps its rules.
+  // HTTP builds a server per request and passes its process's label, so a caller keeps the one rig_whoami names.
+  const unknownSender = options.unknownSender ?? mintUnknownSender();
+  const unknownSenderNote = (kind: "task" | "message"): string | null => {
+    if (seatOf()) return null;
+    const why = options.unfilledSeat ? unfilledSeatText(options.unfilledSeat) : "OPENRIG_SESSION_NAME isn't set where it was configured";
+    const said = clientSaid(server.server.getClientVersion());
+    return `[Unknown sender ${unknownSender}: this app has no OpenRig seat (${why}).${said ? ` Its MCP client says it is ${said}.` : ""}`
+      + (kind === "task" ? " There's no return address: it follows this by its task id.]" : " There's no return address.]");
+  };
  const accessClass = new Map(MCP_OPERATIONS.map((op) => [op.tool, op.access]));
  // A local caller (no access) is offered every tool; an authenticated one, its class's; an external caller, its
  // grant's pairs.
@@ -1035,9 +1110,11 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
        }
        if (scope && !scope.session(session)) return outOfScope(`Session ${session}`);
        const waitForIdleMs = waitForIdleSeconds === undefined ? undefined : Math.ceil(waitForIdleSeconds * 1000);
+        // Unsigned, as before (the transport names it "<unknown sender>"); the note tells the recipient what is known.
+        const note = unknownSenderNote("message");
        const res = await client.post("/api/transport/send", {
          session,
-          text,
+          text: note ? `${note}\n${text}` : text,
          verify,
          force,
          waitForIdleMs,
@@ -1094,7 +1171,7 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
  if (offered("rig_chatroom_send")) server.registerTool(
    "rig_chatroom_send",
    {
-      description: "Send a message to a rig's chatroom as this server's seat",
+      description: "Send a message to a rig's chatroom as this server's seat, or, with no seat, as an unknown sender with a note saying so",
      inputSchema: z.object({
        rigName: z.string().describe("Rig name or identifier"),
        body: z.string().describe("Message body"),
@@ -1108,10 +1185,10 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
        const rig = await chatRigId(client, rigName);
        if ("content" in rig) return rig;
        // P21: no body sender — the daemon derives it from the transport header (X-OpenRig-Session,
-        // stamped by the MCP server's DaemonClient from its seat env).
-        const res = await client.post(`/api/rigs/${encodeURIComponent(rig.id)}/chat/send`, {
-          body,
-        });
+        // stamped by the MCP server's DaemonClient from its seat env). With no seat, the unknown sender is declared.
+        const note = unknownSenderNote("message");
+        const res = await client.post(`/api/rigs/${encodeURIComponent(rig.id)}/chat/send`, note
+          ? { body: `${note}\n${body}`, sender: unknownSender } : { body });
        return mapResult(res);
      } catch (err) {
        return caught(err);
@@ -1199,7 +1276,9 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
    {
      description: "Hand a task to another seat as a durable queue item. Returns its qitemId: read the answer later with rig_queue_show. If the call times out, check rig_queue_show with the same qitemId before retrying; never retry with a new id. To ask for a revision of earlier work, name it in revises: the new item opens with 'Revision of <id>' and records it in its chainOfRecord.",
      inputSchema: z.object({
-        destination: z.string().describe("Destination seat (e.g. dev-impl@my-rig)"),
+        destination: access?.external
+          ? z.string().optional().describe("Destination seat; omit to use the connected team's configured receiver")
+          : z.string().describe("Destination seat (e.g. dev-impl@my-rig)"),
        body: z.string().describe("The task: what to do and what to return"),
        summary: z.string().optional().describe("One-line summary"),
        tags: z.array(z.string()).optional().describe("Tags, e.g. mission:<name>"),
@@ -1212,6 +1291,18 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
      annotations: ADDITIVE,
    },
    async ({ destination, body, summary, tags, priority, tier, qitemId, revises, alert }) => {
+      if (destination === undefined && access?.external) {
+        destination = options.taskDestination;
+        const unavailable = () => errorResult("The connected team's task receiver is unavailable. Ask the server operator to configure it.",
+          "task_destination_unavailable", "The connected team's task receiver is unavailable. Ask the server operator to configure it.");
+        if (!destination) return unavailable();
+        // Check scope before looking up the configured receiver: configuration never expands a grant or reveals
+        // another rig's destination. An explicit destination continues through the existing path below.
+        if (!scope?.destination(destination)) return outOfScope("Configured task destination");
+        try {
+          if (!await taskDestinationExists(client, destination)) return unavailable();
+        } catch { return unavailable(); }
+      }
      if (scope && !scope.destination(destination)) return outOfScope(`Destination ${destination}`);
      // A revision carries its original's lineage as a handoff does: the original's chain, then the original.
      let lineage: { chainOfRecord: string[]; body: string } | undefined;
@@ -1229,9 +1320,12 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
        }
      }
      const id = qitemId ?? `qitem-${new Date().toISOString().replace(/\D/g, "").slice(0, 14)}-${randomBytes(8).toString("hex")}`;
+      // With no seat, the task goes to its destination as from the unknown sender, with the note at the top of its body.
+      const note = unknownSenderNote("task");
      try {
        const res = await client.post("/api/queue/create", {
-          qitemId: id, destinationSession: destination, body: lineage?.body ?? body, summary, priority, tier,
+          qitemId: id, destinationSession: destination, body: note ? `${note}\n\n${lineage?.body ?? body}` : lineage?.body ?? body,
+          ...(note ? { sourceSession: unknownSender } : {}), summary, priority, tier,
          tags: access?.external ? tags?.filter((tag) => !OPENRIG_PROVENANCE_TAG.test(tag)) : tags,
          ...(alert ? { humanAlert: true } : {}),
          ...(lineage ? { chainOfRecord: lineage.chainOfRecord } : {}),
@@ -1245,7 +1339,10 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
        }
        // A create's return, a retry's stored row included, goes through the same external projection as show and list.
        const view = externalView(scope);
-        return mapResult(view && res.status < 400 ? { ...res, data: projectForExternal(res.data, view) } : res, undefined, { structured: true });
+        // An unknown sender has no return address, so it says how to follow the task.
+        const sent = note && res.status < 400 && res.data && typeof res.data === "object" ? { ...res, data: { ...res.data,
+          unknownSender: { label: unknownSender, followUp: `No return address: read this task with rig_queue_show and qitemId ${id}.` } } } : res;
+        return mapResult(view && sent.status < 400 ? { ...sent, data: projectForExternal(sent.data, view) } : sent, undefined, { structured: true });
      } catch (err) {
        const failure = caught(err);
        const text = `${(err as Error).message} The outcome is unknown: check rig_queue_show with qitemId ${id} before retrying, and retry only with that same qitemId.`;
@@ -1312,7 +1409,7 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
        // An external caller lists its own items only, so the daemon's limit counts those and not the team's.
        const mine = mineAsked || (scope?.ownListsOnly === true && !owned);
        if ((owned || mine) && !sessionName) {
-          const message = "owned and mine need this server's seat: set OPENRIG_SESSION_NAME in its env, or filter by destination or source";
+          const message = "owned and mine need this server's seat (rig_whoami says how to give it one), or filter by destination or source";
          return errorResult(JSON.stringify({ ok: false, error: message }), "seat_required", message);
        }
        if (scope) {
@@ -1388,7 +1485,7 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
        resolution: z.string().optional().describe("The outcome the item keeps, stored only by the update that closes an open item (as rig queue update --resolution)"),
        reopen: z.boolean().optional().describe("Reopen a closed item: give an active state (pending, in-progress or blocked) and a note (as rig queue update --reopen)"),
        summary: z.string().optional().describe("Replace the one-line summary"),
-        evidenceRef: z.string().optional().describe("Pointer to the durable artifact, stored on the item when closing with state done or parking on a human seat; other updates refuse it"),
+        evidenceRef: z.string().optional().describe("Pointer to the durable artifact, stored on the item when closing with state done or parking on a human seat; other updates refuse it. For a code result, a JSON file of kind openrig.code-result/v1 with repo, base and revision (full commit ids) and paths lets a reader diff it exactly"),
        blockedOn: z.string().optional().describe("What it waits on, when blocking: another qitemId (this item returns to pending when that one closes), or the address of the person this item came from, not a shared channel (they're asked with this item's summary, and their answer returns it to you in progress)"),
        wakeAfterSeconds: z.number().int().positive().optional().describe("Wake this seat after this many seconds, when blocking. Delivered only to a seat with a terminal: an attached app can't receive it today, so the item just waits until it is checked."),
      }),
@@ -1487,7 +1584,18 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
      try {
        const seat = seatOf();
        if (!seat) {
-          const none = { ok: true, seat: null, caller, message: "This server has no seat: OPENRIG_SESSION_NAME isn't set where it was configured, so its writes aren't attributed to a seat." };
+          const why = options.unfilledSeat ? unfilledSeatText(options.unfilledSeat) : "OPENRIG_SESSION_NAME isn't set where it was configured";
+          const message = `This server has no seat: ${why}, so it acts as no seat and never as another seat. Reads work. `
+            + `Messages (rig_send), new tasks (rig_queue_create) and chat go through as the unknown sender ${unknownSender}, each `
+            + "with a note saying so: there's no return address, so follow a task by its qitemId with rig_queue_show. "
+            + "Claiming, updating, handing off or deciding a task, answering a question asked of a seat, and listing your own "
+            + "items (owned or mine) need this app's seat; reply to the team's questions as a message or a new task naming the "
+            + "question's qitemId. To act as this app's address on the team, such as cursor@my-team, the server needs "
+            + "OPENRIG_SESSION_NAME: in Claude Code, set the plugin's OpenRig seat; in Codex, pass it with --env on codex mcp add. "
+            + "VS Code, Copilot CLI and Cursor can't give a plugin's server a value: add your own rig mcp serve --front-door "
+            + "server with OPENRIG_SESSION_NAME in its env and turn the plugin's openrig server off (in VS Code, give yours "
+            + "another name, such as openrig-seat). Starting the app from a shell with it set lasts only for that launch.";
+          const none = { ok: true, seat: null, caller, unknownSender, ...(options.unfilledSeat ? { unfilledSeat: options.unfilledSeat } : {}), message };
          return { content: [{ type: "text" as const, text: JSON.stringify(none) }], structuredContent: none };
        }
        const params = new URLSearchParams({ sessionName: seat, compact: "1" });
@@ -1502,6 +1610,17 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
            message: "This seat has no session here: it's the seat this connection's grant names, such as a hosted member's. rig_team shows its rig and hosted members, and rig_queue_list with owned set lists its work." };
          return { content: [{ type: "text" as const, text: JSON.stringify(member) }], structuredContent: member };
        }
+        // A front door (--front-door) usually has no session either: its seat is the address the person picked for their
+        // app, declared in OPENRIG_SESSION_NAME. Answer with it when its rig exists here, still marked declared.
+        if (res.status === 404 && !access && options.frontDoor === true) {
+          const at = seat.indexOf("@");
+          const rigName = at > 0 ? seat.slice(at + 1) : "";
+          if (rigName && (await resolveRigHandle(client, rigName, { strict: true })).kind === "resolved") {
+            const frontDoor = { ok: true, seat, rig: rigName, session: null, caller,
+              message: "This seat has no session here: it's the front-door address this server was configured with (OPENRIG_SESSION_NAME), declared, not verified. rig_team shows its rig and lead, and rig_queue_list with owned set lists the team's questions to it." };
+            return { content: [{ type: "text" as const, text: JSON.stringify(frontDoor) }], structuredContent: frontDoor };
+          }
+        }
        if (res.status >= 400) return mapResult(res);
        return mapResult({ ...res, data: { ...(res.data as Record<string, unknown>), caller } }, undefined, { structured: true });
      } catch (err) {
@@ -1661,7 +1780,7 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
  if (offered("rig_artifact_read")) server.registerTool(
    "rig_artifact_read",
    {
-      description: `Read a finished item's artifact (its evidenceRef), or a named file, through the daemon's shared roots (OPENRIG_FILES_ALLOWLIST), as text: at most ${CONTEXT_TEXT_LIMIT} characters from offset; nextOffset reads on. An artifact outside every shared root, a link, or an item with no evidenceRef comes back as available: false with the reason and the shared roots' names; nothing outside those roots is read.`,
+      description: `Read a finished item's artifact (its evidenceRef), or a named file, through the daemon's shared roots (OPENRIG_FILES_ALLOWLIST), as text: at most ${CONTEXT_TEXT_LIMIT} characters from offset; nextOffset reads on. An artifact outside every shared root, a link, or an item with no evidenceRef comes back as available: false with the reason and the shared roots' names; nothing outside those roots is read. A code-result manifest (kind openrig.code-result/v1) also comes back parsed as codeResult: its repo, the exact base and revision commits, and paths; a branch or fetch ref in it is a hint, never the revision.`,
      inputSchema: z.object({
        qitemId: z.string().optional().describe("Read this item's evidenceRef"),
        root: z.string().optional().describe("A shared root's name: with path, the file to read; with qitemId, the root a relative evidenceRef is under"),
@@ -1694,15 +1813,9 @@ export function createMcpServer(client: DaemonClient, options: { seat?: string;
        const unavailable = (reason: string) => mapResult({ status: 200, data: {
          available: false, qitemId, reason, evidenceRef: typeof ref === "string" ? ref : null, roots: roots.map((entry) => entry.name),
        } }, undefined, { structured: true });
-        if (typeof ref !== "string" || !ref.trim()) return unavailable("This item names no artifact: it has no evidenceRef.");
-        if (/^[a-z][a-z0-9+.-]*:\/\//i.test(ref)) return unavailable("The evidenceRef is a link, not a file in a shared root: open it where it lives.");
-        if (roots.length === 0) return unavailable("This daemon shares no roots (OPENRIG_FILES_ALLOWLIST is empty), so no artifact can be read through it.");
-        if (!nodePath.isAbsolute(ref)) {
-          return root ? await read("evidenceRef", root, ref) : unavailable("The evidenceRef is a relative path: name the shared root it's under (root).");
-        }
-        const shared = underSharedRoot(roots, ref);
-        if (!shared) return unavailable("The evidenceRef is outside every shared root as the daemon lists them, so it isn't read. If it reaches a root through a link, read the file with root and path.");
-        return await read("evidenceRef", shared.root, shared.path);
+        const plan = planItemArtifact(ref, roots, root);
+        if ("unavailable" in plan) return unavailable(plan.unavailable);
+        return await read("evidenceRef", plan.read.root, plan.read.path);
      } catch (err) {
        return caught(err);
      }
@@ -0,0 +1,75 @@
+import nodePath from "node:path";
+
+/** A shared root as GET /api/files/roots lists it, and a file as GET /api/files/read returns it. */
+export type SharedRoot = { name: string; path: string };
+export type FileRead = { root?: string; path?: string; content?: string; binary?: boolean; size?: number; mtime?: string; truncated?: boolean; truncatedAtBytes?: number | null; totalBytes?: number };
+
+/** The shared root an absolute path falls under as written, and its path inside that root. The reference names a file
+ *  where the daemon runs, which may be another machine, so it's never resolved through this machine's filesystem; the
+ *  daemon still checks containment itself when the file is read. */
+export function underSharedRoot(roots: SharedRoot[], absolute: string): { root: string; path: string } | null {
+  for (const root of roots) {
+    const inside = nodePath.relative(root.path, absolute);
+    if (inside && inside !== ".." && !inside.startsWith(`..${nodePath.sep}`) && !nodePath.isAbsolute(inside)) return { root: root.name, path: inside };
+  }
+  return null;
+}
+
+/** Which file an item's evidenceRef names under the daemon's shared roots, or why none is read. `root` names the shared
+ *  root a relative evidenceRef is under. Shared by rig_artifact_read and rig queue artifact. */
+export function planItemArtifact(ref: unknown, roots: SharedRoot[], root?: string): { read: { root: string; path: string } } | { unavailable: string } {
+  if (typeof ref !== "string" || !ref.trim()) return { unavailable: "This item names no artifact: it has no evidenceRef." };
+  if (/^[a-z][a-z0-9+.-]*:\/\//i.test(ref)) return { unavailable: "The evidenceRef is a link, not a file in a shared root: open it where it lives." };
+  if (roots.length === 0) return { unavailable: "This daemon shares no roots (OPENRIG_FILES_ALLOWLIST is empty), so no artifact can be read through it." };
+  if (!nodePath.isAbsolute(ref)) {
+    return root ? { read: { root, path: ref } } : { unavailable: "The evidenceRef is a relative path: name the shared root it's under (root)." };
+  }
+  const shared = underSharedRoot(roots, ref);
+  return shared ? { read: shared }
+    : { unavailable: "The evidenceRef is outside every shared root as the daemon lists them, so it isn't read. If it reaches a root through a link, read the file with root and path." };
+}
+
+export const CODE_RESULT_KIND = "openrig.code-result/v1";
+
+/** A code result's exact pair: the commit the work started from and the one it produced, in one repository. */
+export interface CodeResult {
+  repo: string;
+  base: string;
+  revision: string;
+  paths: string[];
+  /** Informational: a branch moves, so it's never the revision. */
+  branch?: string;
+  /** Where the revision can be fetched from when a local clone lacks it. */
+  fetch?: { remote: string; ref: string };
+}
+
+// A full commit id, SHA-1 or SHA-256. A short id or a branch isn't exact.
+const COMMIT = /^(?:[0-9a-f]{40}|[0-9a-f]{64})$/;
+const isRecord = (value: unknown): value is Record<string, unknown> => typeof value === "object" && value !== null && !Array.isArray(value);
+
+/** The code result in an artifact an assignee wrote as its evidenceRef: only from a whole text file of this kind whose
+ *  base and revision are full commit ids. A file of the kind that doesn't qualify says why; any other file is no code
+ *  result, and reads as before. */
+export function codeResultOf(read: FileRead): { codeResult: CodeResult } | { codeResultIssue: string } | null {
+  if (read.binary || read.truncated || typeof read.content !== "string") return null;
+  let doc: unknown;
+  try { doc = JSON.parse(read.content); } catch { return null; }
+  if (!isRecord(doc) || doc.kind !== CODE_RESULT_KIND) return null;
+  const issue = (why: string) => ({ codeResultIssue: `${CODE_RESULT_KIND}: ${why}` });
+  if (typeof doc.repo !== "string" || !doc.repo.trim()) return issue("repo must name the repository, as an absolute path or a URL");
+  for (const key of ["base", "revision"] as const) {
+    const value = doc[key];
+    if (typeof value !== "string" || !COMMIT.test(value)) return issue(`${key} must be a full commit id (40 or 64 lowercase hex), not a branch or a short id`);
+  }
+  const { paths, branch, fetch } = doc;
+  if (paths !== undefined && (!Array.isArray(paths) || !paths.every((p) => typeof p === "string" && p.length > 0))) return issue("paths must be a list of repository paths");
+  if (branch !== undefined && typeof branch !== "string") return issue("branch must be a string");
+  if (fetch !== undefined && !(isRecord(fetch) && typeof fetch.remote === "string" && fetch.remote.trim() && typeof fetch.ref === "string" && fetch.ref.trim())) {
+    return issue("fetch must name a remote and a ref");
+  }
+  return { codeResult: {
+    repo: doc.repo, base: doc.base as string, revision: doc.revision as string, paths: (paths as string[] | undefined) ?? [],
+    ...(typeof branch === "string" ? { branch } : {}),
+    ...(isRecord(fetch) ? { fetch: { remote: fetch.remote as string, ref: fetch.ref as string } } : {}),
+  } };
+}
@@ -67,8 +67,9 @@ const str = (v: unknown): string | null => (typeof v === "string" && v !== "" ?
 const isRow = (v: unknown): v is Row => typeof v === "object" && v !== null && !Array.isArray(v);

 function showCommand(id: string, remoteHost?: string): string {
-  const show = `rig queue show ${shellQuote(id)} --full --json`;
-  return remoteHost ? `OPENRIG_URL='<daemon-url of host ${remoteHost}>' ${show}` : show;
+  return remoteHost
+    ? `rig queue show ${shellQuote(id)} --host ${shellQuote(remoteHost)} --full --json`
+    : `rig queue show ${shellQuote(id)} --full --json`;
 }

 interface WakeEvidence {
@@ -95,7 +96,7 @@ function wakeEvidence(row: Row, ctx: ReceiptContext): WakeEvidence {
  }
  const maybeExisting = ctx.idGiven
    ? "; if --id named an existing row, it was returned with no new wake"
-    : ctx.remoteHost && ctx.verb !== "create"
+    : ctx.remoteHost && (ctx.verb === "handoff" || ctx.verb === "handoff-and-complete")
      ? "; a re-driven cross-host handoff returns the existing successor with no new wake"
      : "";
  const earlier = maybeExisting ? { maybeEarlier: true as const } : {};
@@ -6,6 +6,7 @@ import type { ConfigStore, RiggedConfig } from "./config-store.js";
 import type { DaemonStatus } from "./daemon-lifecycle.js";
 import { buildTmuxControlFailure, probeTmuxControlAsync } from "./tmux-health.js";
 import { classifyNodeVersion } from "./node-support.js";
+import { isBlockedConnectionCode } from "./daemon-reachability.js";

 export interface PreflightCheck {
  name: string;
@@ -49,22 +50,23 @@ interface WritableHomeCheckDeps {
  checkWritable?: (path: string) => void;
 }

-function checkPort(host: string, port: number): Promise<boolean> {
-  if (port <= 0) return Promise.resolve(true); // port 0 is always available
+function checkPort(host: string, port: number): Promise<"available" | "occupied" | { blocked: string }> {
+  if (port <= 0) return Promise.resolve("available"); // port 0 is always available
  return new Promise((resolve) => {
    const socket = new net.Socket();
    socket.setTimeout(1000);
    socket.on("connect", () => {
      socket.destroy();
-      resolve(false); // port is in use
+      resolve("occupied"); // port is in use
    });
-    socket.on("error", () => {
+    socket.on("error", (error) => {
      socket.destroy();
-      resolve(true); // port is available
+      const code = (error as NodeJS.ErrnoException).code;
+      resolve(isBlockedConnectionCode(code) ? { blocked: code } : "available");
    });
    socket.on("timeout", () => {
      socket.destroy();
-      resolve(true); // timeout = nothing listening
+      resolve("available"); // timeout = nothing listening
    });
    socket.connect(port, host);
  });
@@ -176,12 +178,16 @@ export class SystemPreflight {
      (status.host ?? "127.0.0.1") === effectiveHost &&
      status.port === effectivePort;

-    if (daemonOnSameEndpoint) {
+    if (daemonOnSameEndpoint && isBlockedConnectionCode(status.probeErrorCode)) {
+      checks.push({ name: "port_available", ok: true, warning: `Port availability could not be checked: the connection was blocked (${status.probeErrorCode}). A daemon may still be running.` });
+    } else if (daemonOnSameEndpoint) {
      // Our daemon is already running on this exact endpoint — skip
      checks.push({ name: "port_available", ok: true });
    } else {
      const available = await checkPort(effectiveHost, effectivePort);
-      if (available) {
+      if (typeof available === "object") {
+        checks.push({ name: "port_available", ok: true, warning: `Port availability could not be checked: the connection was blocked (${available.blocked}). A daemon may still be running.` });
+      } else if (available === "available") {
        checks.push({ name: "port_available", ok: true });
      } else {
        checks.push({
@@ -220,6 +220,40 @@ describe("rig acp run", () => {
    expect(readFileSync(run.logPath, "utf8")).not.toContain("session/new");
  });

+  it("reads a load error as missing only by the harness's own rule, and never starts a fresh session", async () => {
+    vi.stubEnv("FAKE_ACP_LOAD", "1");
+    vi.stubEnv("FAKE_ACP_MISSING", "claude");
+    const claudeRule = (error: { code: number; data?: unknown }, id: string) => error.code === -32002 && (error.data as { sessionId?: string })?.sessionId === id;
+    const q = memoryQueue("scenario:reply PONG");
+    const run = options({ loadSession: "sess-gone", missingSession: claudeRule });
+    expect(await runAcpAssignment(q.queue, run)).toMatchObject({ kind: "missing" });
+    expect(outcomeNote(q)).toContain('outcome missing: session sess-gone doesn\'t exist for this agent (session/load returned -32002 Resource not found {"sessionId":"sess-gone"}); no new session was started');
+    expect(readFileSync(run.logPath, "utf8")).not.toContain("session/new");
+
+    // The same agent error with no rule, or a rule it doesn't match, reads failed with the agent's own code and text.
+    vi.stubEnv("FAKE_ACP_MISSING", "codex");
+    for (const extra of [{}, { missingSession: claudeRule }]) {
+      const other = memoryQueue("scenario:reply PONG");
+      const outcome = await runAcpAssignment(other.queue, options({ loadSession: "sess-gone", ...extra }));
+      expect(outcome.kind).toBe("failed");
+      expect(outcome.summary).toBe('session/load returned error -32603 Internal error {"details":"no rollout found for thread id sess-gone"}');
+    }
+  });
+
+  it("puts the agent's reported usage and cost on the outcome with where each came from, and unknown when absent", async () => {
+    vi.stubEnv("FAKE_ACP_COST", "1");
+    vi.stubEnv("FAKE_ACP_PROMPT_USAGE", "1");
+    const q = memoryQueue("scenario:reply PONG");
+    const outcome = await runAcpAssignment(q.queue, options());
+    expect(outcome.usage).toEqual({ tokens: { total: 1500, input: 1200, output: 300, thought: undefined, cachedRead: undefined, cachedWrite: undefined }, cost: { amount: 0.00199, currency: "USD" } });
+    expect(outcomeNote(q)).toContain("Usage as the agent reported it: 1500 tokens (1200 in, 300 out), from the agent's prompt response; cost 0.00199 USD, the agent's own cumulative figure for the session (usage_update).");
+
+    vi.unstubAllEnvs();
+    const none = memoryQueue("scenario:reply PONG");
+    expect((await runAcpAssignment(none.queue, options())).usage).toEqual({});
+    expect(outcomeNote(none)).toContain("Usage as the agent reported it: tokens unknown (the agent's prompt response reported none); cost unknown (the agent reported none).");
+  });
+
  it("reports a process the agent launched that outlives the connection, without stopping it", async () => {
    const q = memoryQueue("scenario:reply PONG");
    const run = options();
@@ -0,0 +1,146 @@
+// @openrig/acp: the one-executable launcher the ACP registry starts with npx. It must stay in lockstep with the CLI it
+// pins (a release that bumps one bumps all of them), pass its listing's checks, and hand the app's stdio, arguments,
+// environment, working directory, exit code and signals straight to `rig acp serve`.
+import { afterEach, describe, expect, it } from "vitest";
+import { spawn } from "node:child_process";
+import { copyFileSync, mkdirSync, mkdtempSync, readFileSync, realpathSync, rmSync, statSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { ConfigStore } from "../src/config-store.js";
+import { acpReceiver } from "../src/commands/acp.js";
+
+const repo = fileURLToPath(new URL("../../../", import.meta.url));
+const json = (path: string) => JSON.parse(readFileSync(join(repo, path), "utf8"));
+const cli = json("packages/cli/package.json");
+const launcher = json("packages/acp/package.json");
+const listing = json("packages/acp/registry/agent.json");
+
+describe("@openrig/acp stays in lockstep with @openrig/cli", () => {
+  it("has the CLI's version, pins exactly that CLI, and has one executable", () => {
+    expect(launcher.name).toBe("@openrig/acp");
+    expect(launcher.version).toBe(cli.version);
+    expect(launcher.dependencies).toEqual({ "@openrig/cli": cli.version });
+    expect(launcher.bin).toEqual({ "openrig-acp": "bin/openrig-acp.js" });
+    expect(statSync(join(repo, "packages/acp/bin/openrig-acp.js")).mode & 0o111).not.toBe(0);
+  });
+
+  it("the registry listing names this version of the launcher, with no arguments", () => {
+    expect(listing).toMatchObject({ id: "openrig", name: "OpenRig", version: cli.version, license: "Apache-2.0" });
+    expect(listing.distribution).toEqual({ npx: { package: `@openrig/acp@${cli.version}` } });
+  });
+
+  it("the icon passes the registry's checks: 16x16, currentColor only, no comments", () => {
+    const svg = readFileSync(join(repo, "packages/acp/registry/icon.svg"), "utf8");
+    expect(svg).not.toContain("<!--");
+    const root = svg.match(/^<svg\b[^>]*>/)![0];
+    expect(root).toMatch(/\swidth="16"/);
+    expect(root).toMatch(/\sheight="16"/);
+    const paints = [...svg.matchAll(/\s(?:fill|stroke)="([^"]*)"/g)].map((m) => m[1]!.toLowerCase());
+    expect(paints).toContain("currentcolor");
+    expect(paints.every((v) => ["currentcolor", "none", "inherit"].includes(v))).toBe(true);
+    expect(svg).not.toMatch(/style=|<style/);
+  });
+});
+
+describe("acp.to: the receiver a registry launch uses", () => {
+  it("is empty by default, takes member@rig, and refuses anything else", () => {
+    const dir = mkdtempSync(join(tmpdir(), "acp-to-"));
+    try {
+      const store = new ConfigStore(join(dir, "config.json"));
+      expect(store.get("acp.to")).toBe("");
+      for (const bad of ["lead", "lead@", "@team", "lead @team", "a@b@c"]) expect(() => store.set("acp.to", bad)).toThrow(/Invalid acp\.to/);
+      store.set("acp.to", "lead@my-team");
+      expect(store.resolve().acp).toEqual({ to: "lead@my-team" });
+      store.set("acp.to", "");
+      expect(store.get("acp.to")).toBe("");
+    } finally { rmSync(dir, { recursive: true, force: true }); }
+  });
+
+  it("--to wins; a settings file that can't be read is the reason the app's sign-in step shows", () => {
+    const dir = mkdtempSync(join(tmpdir(), "acp-to-"));
+    try {
+      const store = new ConfigStore(join(dir, "config.json"));
+      expect(acpReceiver(undefined, store)).toEqual({ receiver: undefined });
+      store.set("acp.to", "lead@my-team");
+      expect(acpReceiver(undefined, store)).toEqual({ receiver: "lead@my-team" });
+      writeFileSync(join(dir, "config.json"), "{not json");
+      expect(acpReceiver(" other@team ", store)).toEqual({ receiver: "other@team" });
+      const unreadable = acpReceiver(undefined, store);
+      expect("error" in unreadable && unreadable.error).toMatch(/malformed[\s\S]*adding --to <member@rig> to the agent's command skips it\.$/);
+    } finally { rmSync(dir, { recursive: true, force: true }); }
+  });
+});
+
+describe("openrig-acp hands everything to rig acp serve", () => {
+  const dirs: string[] = [];
+  afterEach(() => { for (const dir of dirs.splice(0)) rmSync(dir, { recursive: true, force: true }); });
+
+  /** The launcher installed beside a stand-in @openrig/cli (the version it pins, unless given) whose `rig` reports what it was given. */
+  function install(withCli = true, cliVersion: string = cli.version): string {
+    const root = mkdtempSync(join(tmpdir(), "openrig-acp-"));
+    dirs.push(root);
+    const acp = join(root, "node_modules/@openrig/acp");
+    mkdirSync(join(acp, "bin"), { recursive: true });
+    copyFileSync(join(repo, "packages/acp/package.json"), join(acp, "package.json"));
+    copyFileSync(join(repo, "packages/acp/bin/openrig-acp.js"), join(acp, "bin/openrig-acp.js"));
+    if (withCli) {
+      const fake = join(root, "node_modules/@openrig/cli");
+      mkdirSync(join(fake, "dist"), { recursive: true });
+      writeFileSync(join(fake, "package.json"), JSON.stringify({ name: "@openrig/cli", version: cliVersion, type: "module", bin: { rig: "dist/rig.js", "openrig-tui": "dist/tui.js" } }));
+      writeFileSync(join(fake, "dist/rig.js"), [
+        `process.stdout.write(JSON.stringify({ argv: process.argv.slice(2), cwd: process.cwd(), mark: process.env.FAKE_MARK }) + "\\n");`,
+        `process.stdin.on("data", (d) => process.stdout.write(d));`,
+        `process.stdin.on("end", () => process.exit(Number(process.env.FAKE_EXIT ?? 0)));`,
+      ].join("\n"));
+    }
+    mkdirSync(join(root, "work"));
+    return root;
+  }
+
+  function run(root: string, args: string[], env: Record<string, string> = {}) {
+    const child = spawn(process.execPath, [join(root, "node_modules/@openrig/acp/bin/openrig-acp.js"), ...args],
+      { cwd: join(root, "work"), env: { ...process.env, ...env }, stdio: ["pipe", "pipe", "pipe"] });
+    let out = "", err = "";
+    child.stdout.on("data", (d: Buffer) => { out += d.toString(); });
+    child.stderr.on("data", (d: Buffer) => { err += d.toString(); });
+    const done = new Promise<{ code: number | null; signal: NodeJS.Signals | null }>((resolve) => child.on("exit", (code, signal) => resolve({ code, signal })));
+    return { child, done, out: () => out, err: () => err };
+  }
+
+  it("passes arguments, stdin, stdout, environment, working directory and the exit code through", async () => {
+    const root = install();
+    const app = run(root, ["--to", "lead@team"], { FAKE_MARK: "from-the-app", FAKE_EXIT: "3" });
+    app.child.stdin.end('{"jsonrpc":"2.0","id":1,"method":"initialize"}\n');
+    expect(await app.done).toEqual({ code: 3, signal: null });
+    const [first, echoed] = app.out().trim().split("\n");
+    expect(JSON.parse(first!)).toEqual({ argv: ["acp", "serve", "--to", "lead@team"], cwd: realpathSync(join(root, "work")), mark: "from-the-app" });
+    expect(echoed).toBe('{"jsonrpc":"2.0","id":1,"method":"initialize"}');
+    expect(app.err()).toBe("");
+  });
+
+  it("a signal from the app reaches the server, and the launcher ends the way the server did", async () => {
+    const app = run(install(), []);
+    await new Promise<void>((resolve) => app.child.stdout.once("data", () => resolve()));
+    app.child.kill("SIGTERM");
+    expect(await app.done).toEqual({ code: null, signal: "SIGTERM" });
+  });
+
+  it("without its @openrig/cli it says so on stderr, writes nothing to stdout, and exits 1", async () => {
+    const app = run(install(false), []);
+    app.child.stdin.end();
+    expect(await app.done).toEqual({ code: 1, signal: null });
+    expect(app.out()).toBe("");
+    expect(app.err()).toMatch(/^openrig-acp: @openrig\/cli isn't installed beside this package/);
+  });
+
+  it("an @openrig/cli of another version is named on stderr and run anyway, with the arguments and exit code passed", async () => {
+    const root = install(true, "0.0.0");
+    const app = run(root, ["--to", "lead@team"], { FAKE_EXIT: "3" });
+    app.child.stdin.end();
+    expect(await app.done).toEqual({ code: 3, signal: null });
+    expect(JSON.parse(app.out().trim().split("\n")[0]!).argv).toEqual(["acp", "serve", "--to", "lead@team"]);
+    expect(app.err()).toMatch(/^openrig-acp: found @openrig\/cli 0\.0\.0 at /);
+    expect(app.err()).toContain(`not the ${cli.version} this package pins; running it anyway. Reinstall @openrig/acp to get ${cli.version}.`);
+  });
+});
@@ -0,0 +1,739 @@
+// rig acp serve against the real queue: the daemon's queue and answer routes over a database with every migration, on
+// loopback, driven by a scripted ACP client over a stream pair. The team's moves go through the same HTTP routes a seat's
+// CLI uses, each as its own seat.
+import { afterAll, afterEach, beforeAll, describe, expect, it } from "vitest";
+import { serve } from "@hono/node-server";
+import { Hono } from "hono";
+import { PassThrough } from "node:stream";
+import { mkdtempSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import type Database from "better-sqlite3";
+import { createDb } from "../../daemon/src/db/connection.js";
+import { migrate } from "../../daemon/src/db/migrate.js";
+import { ALL_MIGRATIONS } from "../../daemon/src/db/all-migrations.js";
+import { EventBus } from "../../daemon/src/domain/event-bus.js";
+import { QueueRepository } from "../../daemon/src/domain/queue-repository.js";
+import { queueRoutes } from "../../daemon/src/routes/queue.js";
+import { missionControlRoutes } from "../../daemon/src/routes/mission-control.js";
+import { MissionControlWriteContract } from "../../daemon/src/domain/mission-control/mission-control-write-contract.js";
+import { MissionControlActionLog } from "../../daemon/src/domain/mission-control/mission-control-action-log.js";
+import { DaemonClient } from "../src/client.js";
+import { daemonTeamQueue, unregisteredPersonNotice } from "../src/commands/acp.js";
+import { addHumanFragment } from "../../daemon/src/domain/gateway/human-registry.js";
+import { AUTH_METHOD_ID, serveAcpAgent, type AcpAgentOptions, type TeamConnection, type TeamQueue } from "../src/acp/agent-server.js";
+import { JsonRpcRequestError, StreamJsonRpcPeer } from "../src/acp/json-rpc-stdio.js";
+
+const CALLER = "human-ada@external";
+const LEAD = "lead@team";
+const IMPL = "impl@team";
+let db: Database.Database;
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+let server: any;
+let port = 0;
+const url = () => `http://127.0.0.1:${port}`;
+const ambient = ["OPENRIG_URL", "RIGGED_URL", "OPENRIG_PORT", "OPENRIG_SESSION_NAME", "RIGGED_SESSION_NAME"].map((key) => [key, process.env[key]] as const);
+
+beforeAll(async () => {
+  for (const [key] of ambient) delete process.env[key];
+  db = createDb();
+  migrate(db, ALL_MIGRATIONS);
+  const bus = new EventBus(db);
+  // The caller is a registered person, so the team's asks reach them and their answers record as theirs.
+  const registry = { ok: true as const, entities: [{ entityId: "human-ada", class: "human" as const, displayName: "Ada", address: CALLER, connectorBindings: [],
+    prefs: { deliveryClass: "B" as const, availability: "available" as const } }] };
+  const repo = new QueueRepository(db, bus, { validateRig: (session) => !session.endsWith("@gone-rig"), loadHumanRegistry: () => registry as never });
+  const writeContract = new MissionControlWriteContract({ db, eventBus: bus, queueRepo: repo, actionLog: new MissionControlActionLog(db) });
+  const app = new Hono();
+  app.use("*", async (c, next) => {
+    c.set("eventBus" as never, bus); c.set("queueRepo" as never, repo); c.set("missionControlWriteContract" as never, writeContract as never);
+    await next();
+  });
+  app.route("/api/queue", queueRoutes());
+  app.route("/api/mission-control", missionControlRoutes());
+  await new Promise<void>((resolve) => { server = serve({ fetch: app.fetch, hostname: "127.0.0.1", port: 0 }, (info: { port: number }) => { port = info.port; resolve(); }); });
+});
+
+afterAll(async () => {
+  if (server) await new Promise<void>((resolve) => server.close(() => resolve()));
+  db?.close();
+  for (const [key, value] of ambient) if (value !== undefined) process.env[key] = value;
+});
+
+const ends: PassThrough[] = [];
+afterEach(() => { for (const end of ends.splice(0)) end.end(); });
+
+/** A scripted ACP client wired to an in-process rig acp serve. */
+function connect(overrides: Partial<AcpAgentOptions> = {}, receiver: string | null = LEAD, wrap: (queue: TeamQueue) => TeamQueue = (q) => q) {
+  const toAgent = new PassThrough();
+  const fromAgent = new PassThrough();
+  ends.push(toAgent);
+  const out: string[] = [];
+  fromAgent.on("data", (chunk: Buffer) => out.push(chunk.toString()));
+  const team: TeamConnection = { queue: wrap(daemonTeamQueue(new DaemonClient(url(), { sessionName: CALLER }))), caller: CALLER, receiver: receiver ?? undefined };
+  const agent = new StreamJsonRpcPeer(toAgent, fromAgent);
+  serveAcpAgent(agent, { connect: async () => team, version: "9.9.9-test", pollMs: 10, closed: agent.closed, ...overrides });
+  const inbound: Array<Record<string, any>> = [];
+  const client = new StreamJsonRpcPeer(fromAgent, toAgent, (direction, packet) => { if (direction === "in") inbound.push(packet as Record<string, any>); });
+  const updates: Array<Record<string, unknown>> = [];
+  client.onNotification("session/update", (params) => updates.push((params as { update: Record<string, unknown> }).update));
+  return { client, updates, out, inbound };
+}
+
+async function open(client: StreamJsonRpcPeer, clientCapabilities: Record<string, unknown> = {}): Promise<string> {
+  await client.request("initialize", { protocolVersion: 1, clientCapabilities, clientInfo: { name: "zed", title: "Zed" } });
+  await client.request("authenticate", { methodId: AUTH_METHOD_ID });
+  return ((await client.request("session/new", { cwd: "/tmp", mcpServers: [] })) as { sessionId: string }).sessionId;
+}
+
+const prompt = (sessionId: string, text: string) => ({ sessionId, prompt: [{ type: "text", text }] });
+const seat = (session: string) => new DaemonClient(url(), { sessionName: session });
+/** A team seat's move through the queue routes; a refused move fails the test with the daemon's answer. */
+async function move(session: string, path: string, body: Record<string, unknown>): Promise<void> {
+  const response = await seat(session).post(path, body);
+  if (response.status >= 400) throw new Error(`${session} ${path}: HTTP ${response.status} ${JSON.stringify(response.data)}`);
+}
+const rowsTo = (destination: string) => db.prepare("SELECT qitem_id, source_session, state, body, tags FROM queue_items WHERE destination_session = ? ORDER BY rowid").all(destination) as Array<{ qitem_id: string; source_session: string; state: string; body: string; tags: string }>;
+const messages = (updates: Array<Record<string, unknown>>) => updates.filter((u) => u.sessionUpdate === "agent_message_chunk").map((u) => (u.content as { text: string }).text);
+
+async function until<T>(read: () => T | undefined, what: string): Promise<T> {
+  for (let i = 0; i < 300; i++) {
+    const value = read();
+    if (value !== undefined) return value;
+    await new Promise((resolve) => setTimeout(resolve, 10));
+  }
+  throw new Error(`timed out waiting for ${what}`);
+}
+
+describe("rig acp serve", () => {
+  it("answers initialize without reaching OpenRig, and offers the one real sign-in method", async () => {
+    let connected = 0;
+    const { client, out } = connect({ connect: async () => { connected += 1; return { error: "unused" }; } });
+    const init = (await client.request("initialize", { protocolVersion: 1, clientCapabilities: { terminal: true } })) as Record<string, any>;
+    expect(init.protocolVersion).toBe(1);
+    expect(init.authMethods).toEqual([expect.objectContaining({ id: AUTH_METHOD_ID, name: "OpenRig on this machine" })]);
+    expect(init.authMethods[0].type).toBeUndefined();
+    expect(init.agentInfo).toEqual({ name: "openrig", title: "OpenRig", version: "9.9.9-test" });
+    expect(init.agentCapabilities.loadSession).toBe(true);
+    expect(init.agentCapabilities.sessionCapabilities).toEqual({ list: {} });
+    expect(connected).toBe(0);
+    for (const line of out.join("").split("\n").filter(Boolean)) expect(JSON.parse(line).jsonrpc).toBe("2.0");
+  });
+
+  it("sign-in and a new session fail with auth_required and the guidance when OpenRig can't be reached", async () => {
+    const { client } = connect({ connect: async () => ({ error: "OpenRig daemon is not running. Start it with: rig daemon start" }) });
+    await client.request("initialize", { protocolVersion: 1, clientCapabilities: {} });
+    for (const [method, params] of [["authenticate", { methodId: AUTH_METHOD_ID }], ["session/new", { cwd: "/tmp", mcpServers: [] }]] as const) {
+      const error = await client.request(method, params).then(() => undefined, (e: unknown) => e);
+      expect(error).toBeInstanceOf(JsonRpcRequestError);
+      expect((error as JsonRpcRequestError).error).toEqual({ code: -32000, message: "OpenRig daemon is not running. Start it with: rig daemon start" });
+    }
+    const unknown = await client.request("authenticate", { methodId: "api-key" }).then(() => undefined, (e: unknown) => e);
+    expect((unknown as JsonRpcRequestError).error.code).toBe(-32602);
+  });
+
+  it("a prompt becomes the caller's row to the receiver; handoffs and notes come back attributed; the resolution is the reply", async () => {
+    const { client, updates } = connect();
+    const sessionId = await open(client);
+    const turn = client.request("session/prompt", prompt(sessionId, "Please fix the login bug\nIt fails on Safari."));
+
+    const task = await until(() => rowsTo(LEAD)[0], "the row to the lead");
+    expect(task.source_session).toBe(CALLER);
+    expect(task.state).toBe("pending");
+    expect(task.body).toMatch(/^Please fix the login bug\nIt fails on Safari\.\n\n---\nSent by human-ada@external from Zed through rig acp serve \(session acps-/);
+    expect(task.body).toContain("close it with a resolution written for human-ada@external");
+    expect(JSON.parse(task.tags)).toEqual(["acp-serve", `acp-session:${sessionId}`, "acp-cwd:/tmp"]);
+
+    await move(LEAD, `/api/queue/${task.qitem_id}/claim`, {});
+    await move(LEAD, `/api/queue/${task.qitem_id}/update`, { state: "in-progress", transitionNote: "reproduced on Safari 19" });
+    await move(LEAD, `/api/queue/${task.qitem_id}/handoff`, { toSession: IMPL, transitionNote: "impl owns the auth code", nudge: false });
+    const next = await until(() => rowsTo(IMPL)[0], "the impl row");
+    await move(IMPL, `/api/queue/${next.qitem_id}/claim`, {});
+    await move(IMPL, `/api/queue/${next.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Fixed in PR 123: the cookie lacked SameSite=None." });
+
+    expect(await turn).toEqual({ stopReason: "end_turn" });
+    const calls = updates.filter((u) => u.sessionUpdate === "tool_call");
+    expect(calls.map((u) => [u.toolCallId, u.title])).toEqual([
+      [task.qitem_id, `Sent to ${LEAD} (${task.qitem_id})`],
+      [next.qitem_id, `Handed to ${IMPL} (${next.qitem_id})`],
+    ]);
+    const lastOf = (id: string) => updates.filter((u) => u.sessionUpdate === "tool_call_update" && u.toolCallId === id).at(-1)!;
+    expect(lastOf(task.qitem_id).status).toBe("completed");
+    const leadLines = ((lastOf(task.qitem_id).content as Array<{ content: { text: string } }>)[0]!.content.text);
+    expect(leadLines).toMatch(/Z lead@team · in-progress: reproduced on Safari 19/);
+    // An app that shows the step collapsed (Zed) still shows its title, which leads with the row's latest note.
+    const latest = db.prepare("SELECT actor_session, transition_note FROM queue_transitions WHERE qitem_id = ? AND transition_note IS NOT NULL ORDER BY transition_id DESC LIMIT 1")
+      .get(task.qitem_id) as { actor_session: string; transition_note: string };
+    const titled = updates.filter((u) => u.sessionUpdate === "tool_call_update" && u.toolCallId === task.qitem_id && u.title).at(-1)!;
+    expect(titled.title).toBe(`${latest.actor_session}: ${latest.transition_note} · Sent to ${LEAD} (${task.qitem_id})`);
+    expect(lastOf(next.qitem_id).status).toBe("completed");
+    expect(messages(updates).at(-1)).toBe(`${IMPL} closed ${next.qitem_id} done.\n\nFixed in PR 123: the cookie lacked SameSite=None.`);
+  });
+
+  it("a handoff-and-complete passes the work on: the reply is the successor's result, not the first row's done", async () => {
+    const { client, updates } = connect({}, "front@team");
+    const sessionId = await open(client);
+    const turn = client.request("session/prompt", prompt(sessionId, "Upgrade the parser"));
+    const task = await until(() => rowsTo("front@team")[0], "the front row");
+    await move("front@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("front@team", `/api/queue/${task.qitem_id}/handoff-and-complete`, { toSession: "back@team", transitionNote: "back owns the parser", nudge: false });
+    expect(db.prepare("SELECT state, closure_reason FROM queue_items WHERE qitem_id = ?").get(task.qitem_id)).toEqual({ state: "done", closure_reason: "handed_off_to" });
+    const next = await until(() => rowsTo("back@team")[0], "the back row");
+    await move("back@team", `/api/queue/${next.qitem_id}/claim`, {});
+    await move("back@team", `/api/queue/${next.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Parser upgraded to 3.2." });
+    expect(await turn).toEqual({ stopReason: "end_turn" });
+    expect(updates.filter((u) => u.sessionUpdate === "tool_call").map((u) => u.toolCallId)).toEqual([task.qitem_id, next.qitem_id]);
+    expect(messages(updates).at(-1)).toBe(`back@team closed ${next.qitem_id} done.\n\nParser upgraded to 3.2.`);
+  });
+
+  it("the reply names the seat that closed the row: not its owner by default, nor one that wrote a note after the close", async () => {
+    // A slow poll, so the close and the later note both land before the agent reads the row again.
+    const { client, updates } = connect({ pollMs: 500 }, "closer@team");
+    const sessionId = await open(client);
+    const turn = client.request("session/prompt", prompt(sessionId, "Ship the hotfix"));
+    const task = await until(() => rowsTo("closer@team")[0], "the row");
+    // The lead closes its report's row: the closer is neither the row's owner nor the newest writer.
+    await move("closer@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("lead@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Hotfix shipped." });
+    await move("watcher@team", `/api/queue/${task.qitem_id}/update`, { transitionNote: "seen, thanks" });
+    const notes = db.prepare("SELECT actor_session FROM queue_transitions WHERE qitem_id = ? ORDER BY rowid").all(task.qitem_id) as Array<{ actor_session: string }>;
+    expect(notes.at(-1)!.actor_session).toBe("watcher@team");
+    expect(await turn).toEqual({ stopReason: "end_turn" });
+    expect(messages(updates).at(-1)).toBe(`lead@team closed ${task.qitem_id} done.\n\nHotfix shipped.`);
+  });
+
+  it("a create whose answer is lost but whose row landed carries on with that row", async () => {
+    const lostAck = (q: TeamQueue): TeamQueue => ({ ...q, create: async (row) => { await q.create(row); throw new Error("socket hang up"); } });
+    const { client, updates } = connect({}, "acked@team", lostAck);
+    const sessionId = await open(client);
+    const turn = client.request("session/prompt", prompt(sessionId, "Rotate the logs"));
+    const task = await until(() => rowsTo("acked@team")[0], "the row");
+    await move("acked@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("acked@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Rotated." });
+    expect(await turn).toEqual({ stopReason: "end_turn" });
+    expect(rowsTo("acked@team")).toHaveLength(1);
+    expect(messages(updates).at(-1)).toBe(`acked@team closed ${task.qitem_id} done.\n\nRotated.`);
+  });
+
+  it("a create that can't be confirmed either way is named as unknown, and the next prompt names it", async () => {
+    let readbacks = 0;
+    const blind = (q: TeamQueue): TeamQueue => ({
+      ...q,
+      create: async (row) => { await q.create(row); throw new Error("socket hang up"); },
+      find: async (id) => { readbacks += 1; if (readbacks === 1) throw new Error("connect ECONNREFUSED"); return q.find(id); },
+    });
+    const { client, updates } = connect({}, "unsure@team", blind);
+    const sessionId = await open(client);
+    const error = await client.request("session/prompt", prompt(sessionId, "Archive March")).then(() => undefined, (e: unknown) => e);
+    const first = rowsTo("unsure@team")[0]!;
+    expect((error as JsonRpcRequestError).error).toMatchObject({ code: -32603, message: `Couldn't confirm ${first.qitem_id}: socket hang up` });
+    expect(messages(updates).at(-1)).toBe(`Sending ${first.qitem_id} to unsure@team got no answer (socket hang up), and 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 ${first.qitem_id}\` shows it if it landed.`);
+
+    const second = client.request("session/prompt", prompt(sessionId, "Archive March, please"));
+    const followUp = await until(() => rowsTo("unsure@team")[1], "the follow-up row");
+    expect(followUp.body).toContain(`Follows up on ${first.qitem_id} (pending).`);
+    expect(messages(updates)).toContain(`${first.qitem_id} is still pending with unsure@team. Sending this as a follow-up.\n\n`);
+    client.notify("session/cancel", { sessionId });
+    expect(await second).toEqual({ stopReason: "cancelled" });
+  });
+
+  it("an unanswered create that isn't found on the immediate readback stays unconfirmed, and is named when it lands late", async () => {
+    // The request is still in flight when the agent reads back: it lands only after the readback found nothing.
+    let late: (() => Promise<void>) | undefined;
+    const slow = (q: TeamQueue): TeamQueue => ({ ...q, create: async (row) => {
+      if (late) return q.create(row);
+      late = () => q.create(row);
+      throw new Error("socket hang up");
+    } });
+    const { client, updates } = connect({}, "late@team", slow);
+    const sessionId = await open(client);
+    const error = await client.request("session/prompt", prompt(sessionId, "Tag the release")).then(() => undefined, (e: unknown) => e);
+    expect(rowsTo("late@team")).toHaveLength(0);
+    const id = /Couldn't confirm (qitem-\S+):/.exec((error as JsonRpcRequestError).error.message)![1]!;
+    expect(messages(updates).at(-1)).toBe(`Sending ${id} to late@team got no answer (socket hang up), and it wasn't found on the immediate readback: delivery is unconfirmed, so it may still reach the team. A new prompt in this session names it, and \`rig queue show ${id}\` shows it if it landed.`);
+
+    await late!();
+    expect(rowsTo("late@team")[0]!.qitem_id).toBe(id);
+    const second = client.request("session/prompt", prompt(sessionId, "Tag the release, please"));
+    const followUp = await until(() => rowsTo("late@team")[1], "the follow-up row");
+    expect(followUp.body).toContain(`Follows up on ${id} (pending).`);
+    client.notify("session/cancel", { sessionId });
+    expect(await second).toEqual({ stopReason: "cancelled" });
+  });
+
+  it("a create the daemon answers with a refusal is said plainly, and nothing is kept as sent", async () => {
+    // A seat whose rig has left the topology can't be sent work: the queue answers with a refusal.
+    const { client, updates } = connect({}, "lead@gone-rig");
+    const sessionId = await open(client);
+    const error = await client.request("session/prompt", prompt(sessionId, "Anything")).then(() => undefined, (e: unknown) => e);
+    expect((error as JsonRpcRequestError).error.code).toBe(-32603);
+    expect((error as JsonRpcRequestError).error.message).toMatch(/^The team's queue refused the task: HTTP 4\d\d: /);
+    expect(rowsTo("lead@gone-rig")).toHaveLength(0);
+    expect(messages(updates)).toEqual([]);
+  });
+
+  it("cancel stops waiting, never the work; the next prompt is a follow-up that names the open item", async () => {
+    const { client, updates } = connect({}, "solo@team");
+    const sessionId = await open(client);
+    const first = client.request("session/prompt", prompt(sessionId, "Draft the release notes"));
+    const row = await until(() => rowsTo("solo@team")[0], "the first row");
+    client.notify("session/cancel", { sessionId });
+    expect(await first).toEqual({ stopReason: "cancelled" });
+    expect(messages(updates).at(-1)).toBe(`Stopped waiting. ${row.qitem_id} is still pending with solo@team: 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.`);
+    expect(rowsTo("solo@team")[0]!.state).toBe("pending");
+
+    const second = client.request("session/prompt", prompt(sessionId, "Also mention the Safari fix"));
+    const followUp = await until(() => rowsTo("solo@team")[1], "the follow-up row");
+    expect(followUp.body).toContain(`Follows up on ${row.qitem_id} (pending).`);
+    expect(messages(updates)).toContain(`${row.qitem_id} is still pending with solo@team. Sending this as a follow-up.\n\n`);
+    await move("solo@team", `/api/queue/${followUp.qitem_id}/claim`, {});
+    await move("solo@team", `/api/queue/${followUp.qitem_id}/update`, { state: "done", closureReason: "canceled", resolution: "Needs the first item first." });
+    expect(await second).toEqual({ stopReason: "end_turn" });
+    expect(messages(updates).at(-1)).toBe(`solo@team closed ${followUp.qitem_id} done (canceled).\n\nNeeds the first item first.`);
+  });
+
+  it("with no receiver configured, a prompt sends nothing and says how to set one", async () => {
+    const { client, updates } = connect({}, null);
+    const sessionId = await open(client);
+    const before = (db.prepare("SELECT count(*) AS n FROM queue_items").get() as { n: number }).n;
+    expect(await client.request("session/prompt", prompt(sessionId, "anything"))).toEqual({ stopReason: "end_turn" });
+    expect((db.prepare("SELECT count(*) AS n FROM queue_items").get() as { n: number }).n).toBe(before);
+    expect(messages(updates).at(-1)).toContain("`rig config set acp.to <member@rig>`, or add `--to <member@rig>`");
+  });
+
+  it("a question tagged for this chat comes as a permission request and the choice is recorded as the caller's answer; an untagged one is only named", async () => {
+    const { client, updates } = connect({}, "asker@team");
+    const asked: Array<Record<string, any>> = [];
+    client.onRequest("session/request_permission", async (params) => {
+      asked.push(params as Record<string, any>);
+      return { outcome: { outcome: "selected", optionId: "b" } };
+    });
+    const sessionId = await open(client);
+    // Asked before this turn by a seat the work never reaches: not this turn's question.
+    await move("bystander@team", "/api/queue/create", { qitemId: "qitem-unrelated-ask", destinationSession: CALLER, body: "unrelated", summary: "Unrelated", nudge: false,
+      humanQuestions: [{ id: "q1", question: "Unrelated?", options: [{ id: "y", label: "Yes" }, { id: "n", label: "No" }] }] });
+    const turn = client.request("session/prompt", prompt(sessionId, "Pick a database for the cache"));
+    const task = await until(() => rowsTo("asker@team")[0], "the work row");
+    await move("asker@team", `/api/queue/${task.qitem_id}/claim`, {});
+    // Untagged, from the seat working on it: it could be another chat's, so it's named here, never asked.
+    await move("asker@team", "/api/queue/create", { qitemId: "qitem-ask-untagged", destinationSession: CALLER, body: "Which region?", summary: "Region?", nudge: false,
+      humanQuestions: [{ id: "q1", question: "Which region?", options: [{ id: "eu", label: "EU" }, { id: "us", label: "US" }] }] });
+    await move("asker@team", "/api/queue/create", { qitemId: "qitem-ask-db", destinationSession: CALLER, body: "Redis is faster; SQLite is already here.", summary: "Which store?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which store?", options: [{ id: "a", label: "Redis" }, { id: "b", label: "SQLite", recommended: true }] }] });
+    const answered = await until(() => {
+      const row = db.prepare("SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-ask-db'").get() as { human_answers: string | null };
+      return row.human_answers ? JSON.parse(row.human_answers) : undefined;
+    }, "the recorded answer");
+    expect(Object.keys(answered)).toEqual(["q1"]);
+    expect(JSON.stringify(answered)).toContain("b");
+    await move("asker@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "SQLite, as you chose." });
+    expect(await turn).toEqual({ stopReason: "end_turn" });
+
+    expect(asked).toHaveLength(1);
+    expect(asked[0]!.toolCall).toMatchObject({ toolCallId: "qitem-ask-db#q1", title: "asker@team asks: Which store?" });
+    expect(asked[0]!.options).toEqual([{ optionId: "a", name: "Redis", kind: "allow_once" }, { optionId: "b", name: "SQLite (recommended)", kind: "allow_once" }]);
+    const result = updates.filter((u) => u.toolCallId === "qitem-ask-db#q1").at(-1)!;
+    expect(result.status).toBe("completed");
+    expect((result.content as Array<{ content: { text: string } }>)[0]!.content.text).toBe("You answered: SQLite. asker@team gets it as your answer.");
+    // "asker@team gets it as your answer": the answer path's reply row, from the caller to the seat that asked.
+    const replies = db.prepare("SELECT body FROM queue_items WHERE source_session = ? AND destination_session = 'asker@team' AND reply_to = 'qitem-ask-db'").all(CALLER) as Array<{ body: string }>;
+    expect(replies).toHaveLength(1);
+    expect(replies[0]!.body).toContain("SQLite");
+    const unrelated = db.prepare("SELECT state, human_answers FROM queue_items WHERE qitem_id = 'qitem-unrelated-ask'").get() as { state: string; human_answers: string | null };
+    expect(unrelated).toEqual({ state: "pending", human_answers: null });
+    expect(messages(updates)).toContain("\n\nasker@team, which this work reached, has a question for you that isn't tagged for this chat (qitem-ask-untagged: Region?). Answer it in the OpenRig TUI or Slack.\n\n");
+    expect(db.prepare("SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-ask-untagged'").get()).toEqual({ human_answers: null });
+    expect(task.body).toContain(`--tags acp-session:${sessionId}`);
+  });
+
+  it("an app that asks forms gets a question's choices as a form, each by its label, and the choice is recorded", async () => {
+    const { client, updates } = connect({}, "former@team");
+    const asked: Array<Record<string, any>> = [];
+    client.onRequest("elicitation/create", async (params) => { asked.push(params as Record<string, any>); return { action: "accept", content: { q1: "SQLite (recommended)" } }; });
+    client.onRequest("session/request_permission", async () => { throw new Error("no permission request was expected"); });
+    const sessionId = await open(client, { elicitation: { form: {} } });
+    const turn = client.request("session/prompt", prompt(sessionId, "Pick a store for the cache"));
+    const task = await until(() => rowsTo("former@team")[0], "the work row");
+    await move("former@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("former@team", "/api/queue/create", { qitemId: "qitem-form-db", destinationSession: CALLER, body: "Redis is faster; SQLite is already here.", summary: "Which store?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which store?", options: [{ id: "a", label: "Redis" }, { id: "b", label: "SQLite", recommended: true }] }] });
+    const answered = await until(() => {
+      const row = db.prepare("SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-form-db'").get() as { human_answers: string | null };
+      return row.human_answers ? JSON.parse(row.human_answers) : undefined;
+    }, "the recorded answer");
+    expect(JSON.stringify(answered)).toContain("b");
+    await move("former@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "SQLite." });
+    expect(await turn).toEqual({ stopReason: "end_turn" });
+    expect(asked).toEqual([{
+      sessionId, toolCallId: "qitem-form-db#q1", mode: "form", message: "former@team asks: Which store?",
+      requestedSchema: { type: "object", required: ["q1"], properties: { q1: { type: "string", title: "Which store?", enum: ["Redis", "SQLite (recommended)"] } } },
+    }]);
+    const result = updates.filter((u) => u.toolCallId === "qitem-form-db#q1").at(-1)!;
+    expect((result.content as Array<{ content: { text: string } }>)[0]!.content.text).toBe("You answered: SQLite. former@team gets it as your answer.");
+  });
+
+  it("#2708 choices whose labels repeat carry their ids, so the answer names one choice", async () => {
+    const { client } = connect({}, "twins@team");
+    const asked: Array<Record<string, any>> = [];
+    client.onRequest("elicitation/create", async (params) => { asked.push(params as Record<string, any>); return { action: "accept", content: { q1: "Same (b)" } }; });
+    const sessionId = await open(client, { elicitation: { form: {} } });
+    const turn = client.request("session/prompt", prompt(sessionId, "Pick one"));
+    const task = await until(() => rowsTo("twins@team")[0], "the work row");
+    await move("twins@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("twins@team", "/api/queue/create", { qitemId: "qitem-form-twins", destinationSession: CALLER, body: "Twins", summary: "Which?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which?", options: [{ id: "a", label: "Same" }, { id: "b", label: "Same" }] }] });
+    const answered = await until(() => {
+      const row = db.prepare("SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-form-twins'").get() as { human_answers: string | null };
+      return row.human_answers ? JSON.parse(row.human_answers) : undefined;
+    }, "the recorded answer");
+    expect(asked[0]!.requestedSchema.properties.q1.enum).toEqual(["Same (a)", "Same (b)"]);
+    expect(JSON.stringify(answered)).toContain('"b"');
+    client.notify("session/cancel", { sessionId });
+    expect(await turn).toEqual({ stopReason: "cancelled" });
+  });
+
+  it("a form declined, cancelled, or answered with no option of the question's records nothing; form: null gets a permission request", async () => {
+    for (const [i, reply] of [{ action: "decline" }, { action: "cancel" }, { action: "accept", content: { q1: "zzz" } }].entries()) {
+      const { client, updates } = connect({}, `former${i}@team`);
+      client.onRequest("elicitation/create", async () => reply);
+      const sessionId = await open(client, { elicitation: { form: {} } });
+      const turn = client.request("session/prompt", prompt(sessionId, "Pick one"));
+      const task = await until(() => rowsTo(`former${i}@team`)[0], "the work row");
+      await move(`former${i}@team`, `/api/queue/${task.qitem_id}/claim`, {});
+      await move(`former${i}@team`, "/api/queue/create", { qitemId: `qitem-form-${i}`, destinationSession: CALLER, body: "Two ways", summary: "Way?", nudge: false,
+        tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which way?", options: [{ id: "a", label: "Left" }, { id: "b", label: "Right" }] }] });
+      const said = await until(() => updates.find((u) => u.toolCallId === `qitem-form-${i}#q1` && u.sessionUpdate === "tool_call_update"), "the outcome");
+      expect((said.content as Array<{ content: { text: string } }>)[0]!.content.text, JSON.stringify(reply)).toBe("Not answered here. The question is still open in the OpenRig TUI and Slack.");
+      expect(messages(updates), JSON.stringify(reply)).toContain(`\n\nNot answered here. former${i}@team's question "Which way?" (qitem-form-${i}) is still open in the OpenRig TUI and Slack.\n\n`);
+      expect(db.prepare(`SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-form-${i}'`).get()).toEqual({ human_answers: null });
+      client.notify("session/cancel", { sessionId });
+      expect(await turn).toEqual({ stopReason: "cancelled" });
+    }
+    const { client } = connect({}, "nullform@team");
+    let permissionAsked = false;
+    client.onRequest("elicitation/create", async () => { throw new Error("form: null advertises no form mode"); });
+    client.onRequest("session/request_permission", async () => { permissionAsked = true; return { outcome: { outcome: "cancelled" } }; });
+    const sessionId = await open(client, { elicitation: { form: null } });
+    const turn = client.request("session/prompt", prompt(sessionId, "Pick one"));
+    const task = await until(() => rowsTo("nullform@team")[0], "the work row");
+    await move("nullform@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("nullform@team", "/api/queue/create", { qitemId: "qitem-form-null", destinationSession: CALLER, body: "Two ways", summary: "Way?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which way?", options: [{ id: "a", label: "Left" }, { id: "b", label: "Right" }] }] });
+    await until(() => (permissionAsked ? true : undefined), "the permission request");
+    client.notify("session/cancel", { sessionId });
+    expect(await turn).toEqual({ stopReason: "cancelled" });
+  });
+
+  it("a question tagged for one chat is asked in that chat only, though another chat's work reached the same seat", async () => {
+    const chatA = connect({}, "shared@team");
+    const chatB = connect({}, "shared@team");
+    const askedIn: Record<string, string[]> = { a: [], b: [] };
+    chatA.client.onRequest("session/request_permission", async (params) => { askedIn.a.push((params as any).toolCall.toolCallId); return { outcome: { outcome: "cancelled" } }; });
+    chatB.client.onRequest("session/request_permission", async (params) => { askedIn.b.push((params as any).toolCall.toolCallId); return { outcome: { outcome: "selected", optionId: "y" } }; });
+    const a = await open(chatA.client);
+    const b = await open(chatB.client);
+    const turnA = chatA.client.request("session/prompt", prompt(a, "Chat A's task"));
+    const turnB = chatB.client.request("session/prompt", prompt(b, "Chat B's task"));
+    const [rowA, rowB] = await until(() => { const rows = rowsTo("shared@team"); return rows.length === 2 ? rows : undefined; }, "both rows");
+    await move("shared@team", "/api/queue/create", { qitemId: "qitem-ask-for-b", destinationSession: CALLER, body: "For B only", summary: "B?", nudge: false,
+      tags: [`acp-session:${b}`], humanQuestions: [{ id: "q1", question: "Ship B?", options: [{ id: "y", label: "Yes" }, { id: "n", label: "No" }] }] });
+    await until(() => (askedIn.b.length > 0 ? true : undefined), "B's question");
+    for (const row of [rowA!, rowB!]) {
+      await move("shared@team", `/api/queue/${row.qitem_id}/claim`, {});
+      await move("shared@team", `/api/queue/${row.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Done." });
+    }
+    await Promise.all([turnA, turnB]);
+    expect(askedIn).toEqual({ a: [], b: ["qitem-ask-for-b#q1"] });
+    expect(chatA.updates.some((u) => String(u.toolCallId ?? "").startsWith("qitem-ask-for-b"))).toBe(false);
+    expect(messages(chatA.updates).some((m) => m.includes("qitem-ask-for-b"))).toBe(false);
+  });
+
+  it("cancel waits for the app to settle the question it's showing, and nothing about it follows the cancelled answer", async () => {
+    const { client, updates, inbound } = connect({}, "asker3@team");
+    let answerCancelled!: () => void;
+    client.onRequest("session/request_permission", () => new Promise((resolve) => {
+      answerCancelled = () => setTimeout(() => resolve({ outcome: { outcome: "cancelled" } }), 50);
+    }));
+    const sessionId = await open(client);
+    const turn = client.request("session/prompt", prompt(sessionId, "Choose a theme"));
+    await until(() => rowsTo("asker3@team")[0], "the work row");
+    await move("asker3@team", "/api/queue/create", { qitemId: "qitem-ask-theme", destinationSession: CALLER, body: "Light or dark", summary: "Theme?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which theme?", options: [{ id: "l", label: "Light" }, { id: "d", label: "Dark" }] }] });
+    await until(() => (answerCancelled ? true : undefined), "the question on screen");
+    // As ACP has it: the app sends the cancel, then answers its open question as cancelled.
+    client.notify("session/cancel", { sessionId });
+    answerCancelled();
+    expect(await turn).toEqual({ stopReason: "cancelled" });
+    await new Promise((resolve) => setTimeout(resolve, 100));
+    const end = inbound.findIndex((p) => p.result?.stopReason === "cancelled");
+    const about = inbound.map((p, i) => ({ p, i })).filter(({ p }) => p.params?.update?.toolCallId === "qitem-ask-theme#q1");
+    expect(about.at(-1)!.i).toBeLessThan(end);
+    expect(about.at(-1)!.p.params.update.content[0].content.text).toBe("Not answered here. The question is still open in the OpenRig TUI and Slack.");
+    const said = inbound.map((p, i) => ({ p, i })).filter(({ p }) => p.params?.update?.sessionUpdate === "agent_message_chunk"
+      && p.params.update.content.text.includes(`asker3@team's question "Which theme?" (qitem-ask-theme) is still open`));
+    expect(said).toHaveLength(1);
+    expect(said[0]!.i).toBeLessThan(end);
+    expect(updates.filter((u) => u.toolCallId === "qitem-ask-theme#q1")).toHaveLength(2);
+    expect(db.prepare("SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-ask-theme'").get()).toEqual({ human_answers: null });
+  });
+
+  it("after a cancel no new question goes to the app, even one already waiting when the cancel wakes the turn", async () => {
+    // The poll never fires on its own here, so the only pass after the first is the one the cancel's wake runs.
+    let scans = 0;
+    let shown = false;
+    const hiddenUntilCancel = (q: TeamQueue): TeamQueue => ({ ...q, openAsks: async (person) => { scans += 1; return shown ? q.openAsks(person) : []; } });
+    const { client, updates } = connect({ pollMs: 60_000, sleep: (ms) => new Promise<void>((resolve) => { if (ms !== 60_000) setTimeout(resolve, ms); }) },
+      "asker5@team", hiddenUntilCancel);
+    const asked: string[] = [];
+    client.onRequest("session/request_permission", async (params) => { asked.push((params as any).toolCall.toolCallId); return { outcome: { outcome: "cancelled" } }; });
+    const sessionId = await open(client);
+    const turn = client.request("session/prompt", prompt(sessionId, "Pick a font"));
+    await until(() => (scans > 0 ? true : undefined), "the first read of open questions");
+    await move("asker5@team", "/api/queue/create", { qitemId: "qitem-ask-after-cancel", destinationSession: CALLER, body: "Serif or sans", summary: "Font?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which font?", options: [{ id: "s", label: "Serif" }, { id: "n", label: "Sans" }] }] });
+    shown = true;
+    client.notify("session/cancel", { sessionId });
+    expect(await turn).toEqual({ stopReason: "cancelled" });
+    expect(asked).toEqual([]);
+    expect(updates.some((u) => String(u.toolCallId ?? "").startsWith("qitem-ask-after-cancel"))).toBe(false);
+    expect(db.prepare("SELECT state, human_answers FROM queue_items WHERE qitem_id = 'qitem-ask-after-cancel'").get()).toEqual({ state: "pending", human_answers: null });
+  });
+
+  it("a cancel while an answer is being recorded keeps that answer and opens no further question from the ask", async () => {
+    let release!: () => void;
+    const held = new Promise<void>((resolve) => { release = resolve; });
+    const slowAnswer = (q: TeamQueue): TeamQueue => ({ ...q, answer: async (...args) => { await held; return q.answer(...args); } });
+    // The answer write is held until the cancel's settle wait (2 s, the only sleep that long here) has begun.
+    const settleThenRelease = (ms: number) => { if (ms === 2000) release(); return new Promise<void>((resolve) => setTimeout(resolve, ms)); };
+    const { client, updates } = connect({ sleep: settleThenRelease }, "asker6@team", slowAnswer);
+    const asked: string[] = [];
+    client.onRequest("session/request_permission", async (params) => {
+      const id = (params as any).toolCall.toolCallId as string;
+      asked.push(id);
+      return id.endsWith("#q1") ? { outcome: { outcome: "selected", optionId: "y" } } : { outcome: { outcome: "cancelled" } };
+    });
+    const sessionId = await open(client);
+    const turn = client.request("session/prompt", prompt(sessionId, "Plan the launch"));
+    await until(() => rowsTo("asker6@team")[0], "the work row");
+    await move("asker6@team", "/api/queue/create", { qitemId: "qitem-ask-two", destinationSession: CALLER, body: "Two calls to make", summary: "Launch?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [
+        { id: "q1", question: "Ship this week?", options: [{ id: "y", label: "Yes" }, { id: "n", label: "No" }] },
+        { id: "q2", question: "Announce it?", options: [{ id: "y", label: "Yes" }, { id: "n", label: "No" }] }] });
+    await until(() => (asked.length > 0 ? true : undefined), "the first question");
+    client.notify("session/cancel", { sessionId });
+    expect(await turn).toEqual({ stopReason: "cancelled" });
+    expect(asked).toEqual(["qitem-ask-two#q1"]);
+    expect(updates.some((u) => u.toolCallId === "qitem-ask-two#q2")).toBe(false);
+    const recorded = db.prepare("SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-ask-two'").get() as { human_answers: string | null };
+    expect(JSON.parse(recorded.human_answers ?? "{}")).toEqual({ q1: "y" });
+  });
+
+  it("when the work closes with a question still on screen, the app is told before the turn ends, and a later reply answers nothing", async () => {
+    const { client, inbound } = connect({}, "asker4@team");
+    let answerLate!: () => void;
+    client.onRequest("session/request_permission", () => new Promise((resolve) => { answerLate = () => resolve({ outcome: { outcome: "selected", optionId: "x" } }); }));
+    const sessionId = await open(client);
+    const turn = client.request("session/prompt", prompt(sessionId, "Plan the offsite"));
+    const task = await until(() => rowsTo("asker4@team")[0], "the work row");
+    await move("asker4@team", "/api/queue/create", { qitemId: "qitem-ask-venue", destinationSession: CALLER, body: "Two venues", summary: "Venue?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which venue?", options: [{ id: "x", label: "Lisbon" }, { id: "z", label: "Oslo" }] }] });
+    await until(() => (answerLate ? true : undefined), "the question on screen");
+    await move("asker4@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("asker4@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Booked without it." });
+    expect(await turn).toEqual({ stopReason: "end_turn" });
+    answerLate();
+    await new Promise((resolve) => setTimeout(resolve, 150));
+    const end = inbound.findIndex((p) => p.result?.stopReason === "end_turn");
+    const about = inbound.map((p, i) => ({ p, i })).filter(({ p }) => p.params?.update?.toolCallId === "qitem-ask-venue#q1");
+    expect(about.at(-1)!.i).toBeLessThan(end);
+    expect(about.at(-1)!.p.params.update.content[0].content.text).toBe("The work closed before this was answered here. The question is still open in the OpenRig TUI and Slack.");
+    const said = inbound.findIndex((p) => p.params?.update?.sessionUpdate === "agent_message_chunk"
+      && p.params.update.content.text === `\n\nThe work closed before this was answered here. asker4@team's question "Which venue?" (qitem-ask-venue) is still open in the OpenRig TUI and Slack.\n\n`);
+    expect(said).toBeGreaterThan(-1);
+    expect(said).toBeLessThan(end);
+    expect(db.prepare("SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-ask-venue'").get()).toEqual({ human_answers: null });
+  });
+
+  it("a question the caller dismisses stays open and unanswered", async () => {
+    const { client, updates } = connect({}, "asker2@team");
+    client.onRequest("session/request_permission", async () => ({ outcome: { outcome: "cancelled" } }));
+    const sessionId = await open(client);
+    const turn = client.request("session/prompt", prompt(sessionId, "Rename the service"));
+    const task = await until(() => rowsTo("asker2@team")[0], "the work row");
+    await move("asker2@team", "/api/queue/create", { qitemId: "qitem-ask-name", destinationSession: CALLER, body: "Two names", summary: "Name?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which name?", options: [{ id: "a", label: "Atlas" }, { id: "b", label: "Beacon" }] }] });
+    await until(() => updates.find((u) => u.toolCallId === "qitem-ask-name#q1" && u.sessionUpdate === "tool_call_update"), "the dismissal");
+    await move("asker2@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("asker2@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Kept the old name." });
+    expect(await turn).toEqual({ stopReason: "end_turn" });
+    const dismissed = updates.filter((u) => u.toolCallId === "qitem-ask-name#q1").at(-1)!;
+    expect(dismissed.status).toBeUndefined();
+    expect((dismissed.content as Array<{ content: { text: string } }>)[0]!.content.text).toBe("Not answered here. The question is still open in the OpenRig TUI and Slack.");
+    expect(messages(updates)).toContain(`\n\nNot answered here. asker2@team's question "Which name?" (qitem-ask-name) is still open in the OpenRig TUI and Slack.\n\n`);
+    expect(db.prepare("SELECT state, human_answers FROM queue_items WHERE qitem_id = 'qitem-ask-name'").get()).toEqual({ state: "pending", human_answers: null });
+  });
+
+  it("a question the app can't show is named in the chat as still open, with why", async () => {
+    const { client, updates } = connect({}, "unshown@team");
+    client.onRequest("elicitation/create", async () => { throw new Error("form rejected"); });
+    const sessionId = await open(client, { elicitation: { form: {} } });
+    const turn = client.request("session/prompt", prompt(sessionId, "Pick a color"));
+    const task = await until(() => rowsTo("unshown@team")[0], "the work row");
+    await move("unshown@team", "/api/queue/create", { qitemId: "qitem-ask-color", destinationSession: CALLER, body: "Two colors", summary: "Color?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which color?", options: [{ id: "r", label: "Red" }, { id: "g", label: "Green" }] }] });
+    const card = await until(() => updates.find((u) => u.toolCallId === "qitem-ask-color#q1" && u.sessionUpdate === "tool_call_update"), "the failure");
+    expect((card.content as Array<{ content: { text: string } }>)[0]!.content.text).toMatch(/^This app couldn't show the question \(.*form rejected.*\)\. The question is still open in the OpenRig TUI and Slack\.$/);
+    expect(messages(updates).some((m) => /^\n\nThis app couldn't show the question \(.*form rejected.*\)\. unshown@team's question "Which color\?" \(qitem-ask-color\) is still open in the OpenRig TUI and Slack\.\n\n$/.test(m))).toBe(true);
+    await move("unshown@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("unshown@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Picked red." });
+    expect(await turn).toEqual({ stopReason: "end_turn" });
+    expect(db.prepare("SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-ask-color'").get()).toEqual({ human_answers: null });
+  });
+
+  it("a session is listed and reopened from its rows: the prompt, its work and the reply replay, and a follow-up continues it", async () => {
+    const first = connect({}, "keeper@team");
+    const sessionId = await open(first.client);
+    const turn = first.client.request("session/prompt", prompt(sessionId, "Summarize yesterday's incidents"));
+    const task = await until(() => rowsTo("keeper@team")[0], "the row");
+    await move("keeper@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("keeper@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Two incidents, both closed." });
+    await turn;
+
+    // A new process, as after the app restarts.
+    const second = connect({}, "keeper@team");
+    await second.client.request("initialize", { protocolVersion: 1, clientCapabilities: {} });
+    const listed = (await second.client.request("session/list", {})) as { sessions: Array<Record<string, unknown>> };
+    expect(listed.sessions.find((s) => s.sessionId === sessionId)).toMatchObject({ cwd: "/tmp", title: "Summarize yesterday's incidents" });
+    expect(((await second.client.request("session/list", { cwd: "/elsewhere" })) as { sessions: unknown[] }).sessions.find((s) => (s as { sessionId: string }).sessionId === sessionId)).toBeUndefined();
+    expect(await second.client.request("session/load", { sessionId, cwd: "/tmp", mcpServers: [] })).toEqual({});
+    expect(second.updates.map((u) => u.sessionUpdate)).toEqual(["user_message_chunk", "tool_call", "tool_call_update", "agent_message_chunk"]);
+    expect((second.updates[0]!.content as { text: string }).text).toBe("Summarize yesterday's incidents");
+    expect(messages(second.updates).at(-1)).toBe(`keeper@team closed ${task.qitem_id} done.\n\nTwo incidents, both closed.`);
+
+    const followUp = second.client.request("session/prompt", prompt(sessionId, "And the week before?"));
+    const next = await until(() => rowsTo("keeper@team")[1], "the follow-up row");
+    expect(next.body).toContain(`Follows up on ${task.qitem_id} (done).`);
+    expect(JSON.parse(next.tags)).toContain(`acp-session:${sessionId}`);
+    second.client.notify("session/cancel", { sessionId });
+    expect(await followUp).toEqual({ stopReason: "cancelled" });
+    const missing = await second.client.request("session/load", { sessionId: "acps-nope", cwd: "/tmp", mcpServers: [] }).then(() => undefined, (e: unknown) => e);
+    expect((missing as JsonRpcRequestError).error.code).toBe(-32602);
+  });
+
+  it("an app that wraps the person's words in its own instructions: they're the row's summary and the session's title", async () => {
+    const { client } = connect({}, "wrapped@team");
+    const sessionId = await open(client);
+    const sent = "<t3_code_instructions>\nUse the preview_* tools.\n</t3_code_instructions>\n<runtime_info>\nmacOS\n</runtime_info>\n"
+      + "<user_request>\nFix the flaky login test\n</user_request>";
+    const turn = client.request("session/prompt", prompt(sessionId, sent));
+    const task = await until(() => rowsTo("wrapped@team")[0], "the row");
+    expect(db.prepare("SELECT summary FROM queue_items WHERE qitem_id = ?").get(task.qitem_id)).toEqual({ summary: "ACP: Fix the flaky login test" });
+    expect(task.body.startsWith(sent)).toBe(true);
+    const listed = (await client.request("session/list", {})) as { sessions: Array<Record<string, unknown>> };
+    expect(listed.sessions.find((s) => s.sessionId === sessionId)).toMatchObject({ title: "Fix the flaky login test" });
+    client.notify("session/cancel", { sessionId });
+    expect(await turn).toEqual({ stopReason: "cancelled" });
+  });
+
+  it("#2708 a prompt whose request marks aren't one pair keeps its own first line: an example is never taken for the request", async () => {
+    for (const [i, sent] of [
+      "<t3_code_instructions>\nFor example: <user_request>Delete the cache</user_request>\n</t3_code_instructions>\n<user_request>\nRepair the login test\n</user_request>",
+      "<t3_code_instructions>\nThe request follows in <user_request>.\n</t3_code_instructions>\n<user_request>\nRepair the login test\n</user_request>",
+    ].entries()) {
+      const { client } = connect({}, `marked${i}@team`);
+      const sessionId = await open(client);
+      const turn = client.request("session/prompt", prompt(sessionId, sent));
+      const task = await until(() => rowsTo(`marked${i}@team`)[0], "the row");
+      expect(db.prepare("SELECT summary FROM queue_items WHERE qitem_id = ?").get(task.qitem_id), sent).toEqual({ summary: "ACP: <t3_code_instructions>" });
+      const listed = (await client.request("session/list", {})) as { sessions: Array<Record<string, unknown>> };
+      expect(listed.sessions.find((s) => s.sessionId === sessionId), sent).toMatchObject({ title: "<t3_code_instructions>" });
+      client.notify("session/cancel", { sessionId });
+      expect(await turn).toEqual({ stopReason: "cancelled" });
+    }
+  });
+
+  it("a reopened session replays the questions asked during it with what's recorded, and asks nothing again", async () => {
+    const first = connect({}, "asker7@team");
+    first.client.onRequest("session/request_permission", async () => ({ outcome: { outcome: "selected", optionId: "b" } }));
+    const sessionId = await open(first.client);
+    const turn = first.client.request("session/prompt", prompt(sessionId, "Pick a store"));
+    const task = await until(() => rowsTo("asker7@team")[0], "the work row");
+    await move("asker7@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("asker7@team", "/api/queue/create", { qitemId: "qitem-ask-replayed", destinationSession: CALLER, body: "Redis or SQLite", summary: "Store?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Which store?", options: [{ id: "a", label: "Redis" }, { id: "b", label: "SQLite" }] }] });
+    await until(() => ((db.prepare("SELECT human_answers FROM queue_items WHERE qitem_id = 'qitem-ask-replayed'").get() as { human_answers: string | null }).human_answers ? true : undefined), "the answer");
+    await move("asker7@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "SQLite it is." });
+    await turn;
+
+    const second = connect({}, "asker7@team");
+    const asked: unknown[] = [];
+    second.client.onRequest("session/request_permission", async (params) => { asked.push(params); return { outcome: { outcome: "cancelled" } }; });
+    await second.client.request("initialize", { protocolVersion: 1, clientCapabilities: {} });
+    expect(await second.client.request("session/load", { sessionId, cwd: "/tmp", mcpServers: [] })).toEqual({});
+    expect(second.updates.map((u) => u.sessionUpdate)).toEqual(["user_message_chunk", "tool_call", "tool_call_update", "tool_call", "agent_message_chunk"]);
+    expect(second.updates[3]).toMatchObject({ toolCallId: "qitem-ask-replayed#q1", status: "completed", title: "asker7@team asks: Which store?" });
+    expect((second.updates[3]!.content as Array<{ content: { text: string } }>).map((c) => c.content.text)).toEqual(["Redis or SQLite", "You answered: SQLite."]);
+    expect(messages(second.updates).at(-1)).toBe(`asker7@team closed ${task.qitem_id} done.\n\nSQLite it is.`);
+    expect(asked).toEqual([]);
+  });
+
+  it("a reopened session shows, in the order asked, a question closed unanswered as closed and one still open as pending", async () => {
+    const first = connect({}, "asker8@team");
+    const sessionId = await open(first.client);
+    const turn = first.client.request("session/prompt", prompt(sessionId, "Plan the release"));
+    const task = await until(() => rowsTo("asker8@team")[0], "the work row");
+    await move("asker8@team", `/api/queue/${task.qitem_id}/claim`, {});
+    await move("asker8@team", `/api/queue/${task.qitem_id}/update`, { state: "done", closureReason: "no-follow-on", resolution: "Planned." });
+    await turn;
+    // Asked after the turn ended, so no live turn puts them to the app.
+    const choices = [{ id: "y", label: "Yes" }, { id: "n", label: "No" }];
+    await move("asker8@team", "/api/queue/create", { qitemId: "qitem-ask-closed-unanswered", destinationSession: CALLER, body: "Freeze on Friday?", summary: "Freeze?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Freeze on Friday?", options: choices }] });
+    await move(CALLER, "/api/queue/qitem-ask-closed-unanswered/claim", {});
+    await move(CALLER, "/api/queue/qitem-ask-closed-unanswered/update", { state: "done", closureReason: "no-follow-on", resolution: "Moot now." });
+    await move("asker8@team", "/api/queue/create", { qitemId: "qitem-ask-still-open", destinationSession: CALLER, body: "Tag it 1.0?", summary: "Tag?", nudge: false,
+      tags: [`acp-session:${sessionId}`], humanQuestions: [{ id: "q1", question: "Tag it 1.0?", options: choices }] });
+
+    const second = connect({}, "asker8@team");
+    const asked: unknown[] = [];
+    second.client.onRequest("session/request_permission", async (params) => { asked.push(params); return { outcome: { outcome: "cancelled" } }; });
+    await second.client.request("initialize", { protocolVersion: 1, clientCapabilities: {} });
+    expect(await second.client.request("session/load", { sessionId, cwd: "/tmp", mcpServers: [] })).toEqual({});
+    const questions = second.updates.filter((u) => u.sessionUpdate === "tool_call" && String(u.toolCallId).includes("#"));
+    expect(questions.map((u) => [u.toolCallId, u.status, (u.content as Array<{ content: { text: string } }>).at(-1)!.content.text])).toEqual([
+      ["qitem-ask-closed-unanswered#q1", "completed", "Not answered: qitem-ask-closed-unanswered was closed (done)."],
+      ["qitem-ask-still-open#q1", "pending", "Not answered yet. It's still open in the OpenRig TUI and Slack."],
+    ]);
+    expect(asked).toEqual([]);
+  });
+
+  it("a connection's notice is said once in each session, before its first prompt's work", async () => {
+    const team: TeamConnection = { queue: daemonTeamQueue(new DaemonClient(url(), { sessionName: CALLER })), caller: CALLER, receiver: "noticed@team", notice: "Your team can't ask you questions here yet." };
+    const { client, updates } = connect({ connect: async () => team });
+    const sessionId = await open(client);
+    for (const [i, text] of ["One", "Two"].entries()) {
+      const turn = client.request("session/prompt", prompt(sessionId, text));
+      await until(() => rowsTo("noticed@team")[i], `row ${i + 1}`);
+      client.notify("session/cancel", { sessionId });
+      expect(await turn).toEqual({ stopReason: "cancelled" });
+    }
+    expect(messages(updates).filter((m) => m.includes("can't ask you questions"))).toEqual(["Your team can't ask you questions here yet.\n\n"]);
+    expect(messages(updates)[0]).toBe("Your team can't ask you questions here yet.\n\n");
+  });
+
+  it("names an unregistered person and how to register; says nothing for a registered person or a seat", async () => {
+    const home = mkdtempSync(join(tmpdir(), "acp-people-"));
+    try {
+      expect(await unregisteredPersonNotice("dev@team", home)).toBeUndefined();
+      expect(await unregisteredPersonNotice("cq-person@external", home)).toBe("Your team can't ask you questions here yet: cq-person@external 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 cq-person --display-name \"<your name>\" --delivery-class <A|B|C|D>`.");
+      expect(addHumanFragment({ entityId: "ada", class: "human", displayName: "Ada", address: "ada@external", connectorBindings: [], prefs: { deliveryClass: "B" } }, home).ok).toBe(true);
+      expect(await unregisteredPersonNotice("ada@external", home)).toBeUndefined();
+    } finally { rmSync(home, { recursive: true, force: true }); }
+  });
+
+  it("an unknown method gets JSON-RPC's method-not-found", async () => {
+    const { client } = connect();
+    const error = await client.request("session/fork", { sessionId: "x" }).then(() => undefined, (e: unknown) => e);
+    expect((error as JsonRpcRequestError).error).toEqual({ code: -32601, message: "Method not found: session/fork" });
+  });
+});
@@ -70,7 +70,11 @@ function createMockDaemon() {
      let body = "";
      req.on("data", (c: Buffer) => { body += c.toString(); });
      req.on("end", () => {
-        if (body.includes("INVALID")) {
+        if (body.includes("ADVISED")) {
+          res.writeHead(200, { "Content-Type": "application/json" });
+          const invalid = body.includes("ADVISED-INVALID");
+          res.end(JSON.stringify({ valid: !invalid, errors: invalid ? ["name is required"] : [], advisories: ['defaults.effort: non-string value "3" ignored; effort must be a text value'] }));
+        } else if (body.includes("INVALID")) {
          res.writeHead(400, { "Content-Type": "application/json" });
          res.end(JSON.stringify({ valid: false, errors: ["name is required", "missing runtime"] }));
        } else {
@@ -128,6 +132,16 @@ describe("rig agent", () => {
    expect(exitCode).toBeUndefined();
  });

+  it("agent validate prints each advisory, valid or not, and an advisory never fails it", async () => {
+    for (const [content, exit] of [["name: my-agent\nversion: 1.2.0\n# ADVISED\n", undefined], ["# ADVISED-INVALID\n", 1]] as const) {
+      const program = new Command();
+      program.addCommand(agentCommand(agentDeps(content)));
+      const { logs, exitCode } = await captureLogs(() => program.parseAsync(["node", "rig", "agent", "validate", "agent.yaml"]));
+      expect(logs).toContain('⚠ agent spec advisory: defaults.effort: non-string value "3" ignored; effort must be a text value');
+      expect(exitCode).toBe(exit);
+    }
+  });
+
  // T2: rig agent validate invalid spec -> exit 1, output contains error
  it("agent validate invalid spec: prints errors, exitCode 1", async () => {
    const deps = agentDeps("INVALID");
@@ -1,6 +1,9 @@
 import { afterEach, describe, expect, it } from "vitest";
+import { spawnSync } from "node:child_process";
+import ts from "typescript";
 import fs from "node:fs";
 import os from "node:os";
+import module from "node:module";
 import path from "node:path";
 import { fileURLToPath, pathToFileURL } from "node:url";
 import { resolveBinEntry, resolveNodeReexecBinary, run } from "../src/bin-wrapper.js";
@@ -15,6 +18,40 @@ describe("bin-wrapper", () => {
    }
  });

+  it.skipIf(!module.getCompileCacheDir || !module.enableCompileCache).each([
+    { args: [], interactive: false, explicitCache: false, expected: false },
+    { args: ["--version"], interactive: false, explicitCache: false, expected: true },
+    { args: [], interactive: false, explicitCache: true, expected: true },
+    { args: [], interactive: true, explicitCache: false, expected: true },
+  ])("compile cache for args=$args, interactive=$interactive, explicitCache=$explicitCache", ({ args, interactive, explicitCache, expected }) => {
+    tempDir = fs.mkdtempSync(path.join(os.tmpdir(), "openrig-bin-cache-"));
+    const dist = path.join(tempDir, "dist");
+    fs.mkdirSync(dist);
+    fs.writeFileSync(path.join(tempDir, "package.json"), '{ "type": "module" }');
+    const wrapper = path.join(dist, "bin-wrapper.js");
+    const source = fs.readFileSync(new URL("../src/bin-wrapper.ts", import.meta.url), "utf8");
+    const { outputText } = ts.transpileModule(source, {
+      compilerOptions: { target: ts.ScriptTarget.ES2022, module: ts.ModuleKind.ESNext },
+    });
+    fs.writeFileSync(wrapper, outputText);
+    const report = `import { getCompileCacheDir } from "node:module";
+export async function runProgram() { console.log(JSON.stringify({ enabled: !!getCompileCacheDir() })); }
+`;
+    fs.writeFileSync(path.join(dist, "index.js"), report + "export function createProgram() { return {}; }\n");
+    fs.writeFileSync(path.join(dist, "lazy-program.js"), report + "export async function createLazyProgram() { return {}; }\n");
+    const env = { ...process.env, TMPDIR: tempDir };
+    delete env.NODE_DISABLE_COMPILE_CACHE;
+    delete env.NODE_COMPILE_CACHE;
+    if (explicitCache) env.NODE_COMPILE_CACHE = path.join(tempDir, "explicit-cache");
+    // Set stream flags before the wrapper loads; front-door.test.ts covers a real PTY launch.
+    const nodeArgs = interactive
+      ? ["--import", "data:text/javascript,process.stdin.isTTY=true;process.stdout.isTTY=true"]
+      : [];
+    const result = spawnSync(process.execPath, [...nodeArgs, wrapper, ...args], { env, encoding: "utf8", timeout: 5000 });
+    expect(result.status, result.stderr).toBe(0);
+    expect(JSON.parse(result.stdout)).toEqual({ enabled: expected });
+  });
+
  it("resolves the real dist entry from a symlinked wrapper path", () => {
    tempDir = fs.mkdtempSync(path.join(os.tmpdir(), "openrig-bin-wrapper-"));
    const wrapperSource = fileURLToPath(new URL("../src/bin-wrapper.ts", import.meta.url));
@@ -70,6 +70,19 @@ describe("Broadcast CLI", () => {
                { ok: false, sessionName: "c@notice-rig", error: "send failed", managedCompaction: { stage: "compacting", observed: "before_send", notice: "c@notice-rig was being compacted: /compact was submitted." } },
              ],
            }));
+          } else if (parsed.rig === "phase-rig") {
+            // #2507: a recipient whose input was attempted or written is recorded indeterminate, not failed.
+            res.writeHead(200, { "Content-Type": "application/json" });
+            res.end(JSON.stringify({
+              total: 5, sent: 1, failed: 4,
+              results: [
+                { ok: true, sessionName: "a@phase-rig", inputPhase: "written", sent: true },
+                { ok: false, sessionName: "b@phase-rig", inputPhase: "not_reached", sent: false, error: "target_needs_input" },
+                { ok: false, sessionName: "c@phase-rig", inputPhase: "attempted", sent: false, error: "paste lost" },
+                { ok: false, sessionName: "d@phase-rig", inputPhase: "written", sent: false, error: "enter unconfirmed" },
+                { ok: false, sessionName: "e@phase-rig", sent: true, error: "receipt lost" },
+              ],
+            }));
          } else if (parsed.rig === "warn-rig") {
            // S2 (OPR.0.5.4.3): an unattributed broadcast's response carries the
            // sign-it notice as an additive warning; the renderer must surface it.
@@ -156,6 +169,22 @@ describe("Broadcast CLI", () => {
    expect(exitCode).toBe(1);
  });

+  it("broadcast prints UNCONFIRMED, not FAILED, for a recipient whose input was attempted or written; exit code unchanged", async () => {
+    const { stdout, exitCode } = await captureLogs(async () => {
+      await makeCmd().parseAsync(["node", "rig", "broadcast", "--rig", "phase-rig", "hello"]);
+    });
+    const check = "the text may have reached it, so check with rig capture before resending";
+    expect(stdout).toEqual([
+      "a@phase-rig: sent",
+      "b@phase-rig: FAILED — target_needs_input",
+      `c@phase-rig: UNCONFIRMED — paste lost; ${check}`,
+      `d@phase-rig: UNCONFIRMED — enter unconfirmed; ${check}`,
+      `e@phase-rig: UNCONFIRMED — receipt lost; ${check}`,
+      "1/5 delivered",
+    ]);
+    expect(exitCode).toBe(1);
+  });
+
  it("broadcast --rig prints per-target summary", async () => {
    const { logs } = await captureLogs(async () => {
      await makeCmd().parseAsync(["node", "rig", "broadcast", "--rig", "my-rig", "hello"]);
@@ -64,6 +64,24 @@ describe("auto-start only for a daemon confirmed down", () => {
      expect(c.requests.length).toBeGreaterThan(0);
    });

+    it.each(["EPERM", "EACCES"])(`${name}: a blocked health probe reaches the shared start path (%s)`, async (code) => {
+      lifecycle.status = { state: "unverified", host: "127.0.0.1", port: 7433, probeErrorCode: code };
+      const c = client();
+      const cmd = (command as (d: unknown) => import("commander").Command)({ lifecycleDeps: {}, clientFactory: c.factory, preflightExec: async () => "" });
+      cmd.exitOverride();
+      await cmd.parseAsync(argv as unknown as string[]).catch(() => {});
+      expect(lifecycle.startDaemon).toHaveBeenCalledOnce();
+    });
+
+    it.each(["EPERM", "EACCES"])(`${name}: a blocked explicit remote target never starts a local daemon (%s)`, async (code) => {
+      lifecycle.status = { state: "unverified", host: "remote.example", port: 7433, probeErrorCode: code, explicitRemote: true };
+      const c = client();
+      const cmd = (command as (d: unknown) => import("commander").Command)({ lifecycleDeps: {}, clientFactory: c.factory, preflightExec: async () => "" });
+      cmd.exitOverride();
+      await cmd.parseAsync(argv as unknown as string[]).catch(() => {});
+      expect(lifecycle.startDaemon).not.toHaveBeenCalled();
+    });
+
  }

  it("rig up --plan starts nothing: a stopped daemon is daemon_not_running, a blocked probe keeps its own guidance", async () => {
@@ -0,0 +1,21 @@
+import { it, expect, vi } from "vitest";
+import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { ConfigStore } from "../src/config-store.js";
+import { SettingsStore } from "../../daemon/src/domain/user-settings/settings-store.js";
+it("launch.codex_login_shell defaults off in both stores, with file and env precedence", () => {
+  const home=mkdtempSync(join(tmpdir(),"codex-login-shell-")), file=join(home,"config.json");
+  const cli=new ConfigStore(file), daemon=new SettingsStore(file);
+  try {
+    vi.stubEnv("OPENRIG_LAUNCH_CODEX_LOGIN_SHELL",undefined);
+    expect(cli.resolveWithSource("launch.codex_login_shell")).toMatchObject({value:false,source:"default"});
+    expect(daemon.resolveOne("launch.codex_login_shell")).toMatchObject({value:false,source:"default"});
+    writeFileSync(file, JSON.stringify({launch:{codexLoginShell:true}}));
+    expect(cli.resolveWithSource("launch.codex_login_shell")).toMatchObject({value:true,source:"file"});
+    expect(daemon.resolveOne("launch.codex_login_shell")).toMatchObject({value:true,source:"file"});
+    vi.stubEnv("OPENRIG_LAUNCH_CODEX_LOGIN_SHELL","false");
+    expect(cli.resolveWithSource("launch.codex_login_shell")).toMatchObject({value:false,source:"env"});
+    expect(daemon.resolveOne("launch.codex_login_shell")).toMatchObject({value:false,source:"env"});
+  } finally { vi.unstubAllEnvs(); rmSync(home,{recursive:true,force:true}); }
+});
@@ -5,10 +5,10 @@ import * as fs from "node:fs";
 import * as os from "node:os";
 import * as path from "node:path";
 import { configCommand } from "../src/commands/config.js";
+import { ENV_MAP } from "../src/config-store.js";

 let dir: string;
 let configPath: string;
-let savedPort: string | undefined;

 async function run(args: string[]): Promise<string> {
  const logSpy = vi.spyOn(console, "log").mockImplementation(() => {});
@@ -25,13 +25,14 @@ describe("rig config get --with-source", () => {
  beforeEach(() => {
    dir = fs.mkdtempSync(path.join(os.tmpdir(), "config-get-with-source-"));
    configPath = path.join(dir, "config.json");
-    savedPort = process.env["OPENRIG_PORT"];
-    delete process.env["OPENRIG_PORT"];
+    for (const { primary, legacy } of Object.values(ENV_MAP)) {
+      vi.stubEnv(primary, undefined);
+      if (legacy) vi.stubEnv(legacy, undefined);
+    }
  });

  afterEach(() => {
-    if (savedPort === undefined) delete process.env["OPENRIG_PORT"];
-    else process.env["OPENRIG_PORT"] = savedPort;
+    vi.unstubAllEnvs();
    fs.rmSync(dir, { recursive: true, force: true });
  });

@@ -72,4 +73,55 @@ describe("rig config get --with-source", () => {
    const json = JSON.parse(await run(["get", "daemon.port", "--json"]));
    expect(json).toMatchObject({ value: 7555, source: "file" });
  });
+
+  it("reports the active environment override when setting or resetting a file value", async () => {
+    process.env["OPENRIG_PORT"] = "7666";
+    expect(await run(["set", "daemon.port", "7555"])).toContain("7666 (source: env; file override saved but masked)");
+    expect(JSON.parse(fs.readFileSync(configPath, "utf8")).daemon.port).toBe(7555);
+    expect(await run(["reset", "daemon.port"])).toContain("effective value 7666 (source: env; default: 7433)");
+    expect(await run(["reset"])).toContain("environment overrides them: daemon.port.");
+  });
+
+  it("retains the default reset message when no environment override remains", async () => {
+    expect(await run(["reset"])).toBe("Config reset to defaults.");
+  });
+
+  it("reports a successful reset when reading effective configuration fails", async () => {
+    fs.writeFileSync(configPath, JSON.stringify({ daemon: { port: 7555 } }));
+    vi.stubEnv("OPENRIG_CONTEXT_PACKS_ROOT", "/old-context");
+    const previousExitCode = process.exitCode;
+    process.exitCode = 0;
+    const errorSpy = vi.spyOn(console, "error").mockImplementation(() => {});
+    try {
+      expect(await run(["reset"])).toBe("Config reset to defaults.");
+      expect(fs.existsSync(configPath)).toBe(false);
+      expect(errorSpy.mock.calls.flat().join(" ")).toContain("OPENRIG_CONTEXT_PACKS_ROOT was removed");
+      expect(process.exitCode).toBe(0);
+    } finally {
+      process.exitCode = previousExitCode;
+      errorSpy.mockRestore();
+    }
+  });
+
+  it("does not promise a masked boot setting will take effect on restart", async () => {
+    process.env["OPENRIG_PORT"] = "7666";
+    const errorSpy = vi.spyOn(console, "error").mockImplementation(() => {});
+    try {
+      await configCommand(configPath, { probeDaemonRunning: async () => true })
+        .parseAsync(["node", "rig", "set", "daemon.port", "7555"]);
+      expect(errorSpy.mock.calls.flat().join(" ")).toContain("remains masked after restart");
+    } finally { errorSpy.mockRestore(); }
+  });
+
+  it("distinguishes the saved threshold from the default used after an invalid environment pair", async () => {
+    vi.stubEnv("OPENRIG_HEALTH_CONTEXT_PRESSURE_WARNING_PERCENT", "98");
+    vi.stubEnv("OPENRIG_HEALTH_CONTEXT_PRESSURE_CRITICAL_PERCENT", undefined);
+    const stderr = vi.spyOn(process.stderr, "write").mockImplementation(() => true);
+    try {
+      expect(await run(["set", "health.context_pressure.critical_percent", "97"]))
+        .toContain("file override saved; effective value 99 (source: default)");
+      expect(JSON.parse(fs.readFileSync(configPath, "utf8")).health.contextPressure.criticalPercent).toBe(97);
+      expect(stderr).toHaveBeenCalled();
+    } finally { stderr.mockRestore(); }
+  });
 });
@@ -121,7 +121,7 @@ describe("ConfigStore — extended namespaces (User Settings v0)", () => {

  it("VALID_KEYS includes legacy, user-settings, ui.preview, recovery, agents, feed, and transcript-rotation keys", () => {
    const expected = [
-      "daemon.port", "daemon.host", "daemon.db_conflict",
+      "daemon.port", "daemon.host", "daemon.db_conflict", "daemon.seat_poll_interval_ms",
      // OPR.0.4.6.MH1 FR-1/FR-4 — host-selection pointer + own-host name.
      "host.selected",
      "host.name",
@@ -143,6 +143,7 @@ describe("ConfigStore — extended namespaces (User Settings v0)", () => {
      "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", "progress.scan_roots",
@@ -201,10 +202,12 @@ describe("ConfigStore — extended namespaces (User Settings v0)", () => {
      "usage.pi_openrouter_charge",
      "transport.verify_input_box",
      "transport.local_user_guard",
+      "transport.draft_delivery",
      // Slice 22 (OPR.0.7.2.22) — typed decisions, off by default.
      "decisions.enabled",
      "decisions.consumers",
      "decisions.template_dir",
+      "acp.to",
      // OPR.0.4.6.FS-1 W2 — queue-retention maintenance knobs.
      "retention.enabled",
      "retention.transitions_days",
@@ -212,6 +215,7 @@ describe("ConfigStore — extended namespaces (User Settings v0)", () => {
      "retention.watchdog_keep_per_job",
      "retention.stream_days", // slice 22 outcome 9: stream retention
      "retention.stream_max_items",
+      "retention.events_days", // the history-only event types of events
      "stream.intake",
      "retention.batch_size",
      "queue.pickup_stall_threshold_minutes",
@@ -328,11 +332,11 @@ describe("ConfigStore — extended namespaces (User Settings v0)", () => {
  it("transport.verify_input_box defaults off, resolves its env twin, and appears in the resolved config", () => {
    const store = new ConfigStore(configPath);
    expect(store.resolveWithSource("transport.verify_input_box")).toMatchObject({ value: false, source: "default", defaultValue: false });
-    expect(store.resolve().transport).toEqual({ verifyInputBox: false, localUserGuard: false });
+    expect(store.resolve().transport).toEqual({ verifyInputBox: false, localUserGuard: false, draftDelivery: "deliver" });
    try {
      process.env.OPENRIG_TRANSPORT_VERIFY_INPUT_BOX = "1";
      expect(store.resolveWithSource("transport.verify_input_box")).toMatchObject({ value: true, source: "env" });
-      expect(store.resolve().transport).toEqual({ verifyInputBox: true, localUserGuard: false });
+      expect(store.resolve().transport).toEqual({ verifyInputBox: true, localUserGuard: false, draftDelivery: "deliver" });
    } finally {
      delete process.env.OPENRIG_TRANSPORT_VERIFY_INPUT_BOX;
    }
@@ -402,6 +406,29 @@ describe("ConfigStore — extended namespaces (User Settings v0)", () => {
    expect(readFileSync(configPath, "utf8")).toBe(before);
  });

+  it("rig config set retention.events_days 3 succeeds and rig config get returns 3", async () => {
+    const log = vi.spyOn(console, "log").mockImplementation(() => {});
+    const error = vi.spyOn(console, "error").mockImplementation(() => {});
+    const oldExitCode = process.exitCode;
+    try {
+      process.exitCode = undefined;
+      await configCommand(configPath, { probeDaemonRunning: async () => false })
+        .parseAsync(["node", "rig", "set", "retention.events_days", "3"]);
+      expect(process.exitCode).toBeUndefined();
+      expect(error).not.toHaveBeenCalled();
+      log.mockClear();
+      await configCommand(configPath, { probeDaemonRunning: async () => false })
+        .parseAsync(["node", "rig", "get", "retention.events_days"]);
+      expect(process.exitCode).toBeUndefined();
+      expect(log.mock.calls.flat().join(" ")).toMatch(/\b3\b/);
+      expect(new ConfigStore(configPath).get("retention.events_days")).toBe(3);
+    } finally {
+      log.mockRestore();
+      error.mockRestore();
+      process.exitCode = oldExitCode;
+    }
+  });
+
  it("shows the file-pair explanation through the config command", async () => {
    process.env.OPENRIG_HEALTH_CONTEXT_PRESSURE_WARNING_PERCENT = "50";
    const error = vi.spyOn(console, "error").mockImplementation(() => {});
@@ -501,6 +528,21 @@ describe("ConfigStore — extended namespaces (User Settings v0)", () => {
    }
  });

+  it("retention.events_days: set 3 then get 3, default 0 (keep), 0 accepted, negative and non-integer refused", () => {
+    const store = new ConfigStore(configPath);
+    expect(store.get("retention.events_days")).toBe(0);
+    expect(store.resolve().retention.eventsDays).toBe(0);
+    store.set("retention.events_days", "3");
+    expect(store.get("retention.events_days")).toBe(3);
+    expect(store.resolve().retention.eventsDays).toBe(3);
+    expect(JSON.parse(readFileSync(configPath, "utf-8")).retention.eventsDays).toBe(3);
+    store.set("retention.events_days", "0");
+    expect(store.get("retention.events_days")).toBe(0);
+    for (const raw of ["-1", "1.5"]) {
+      expect(() => store.set("retention.events_days", raw)).toThrow();
+    }
+  });
+
  // OPR.0.3.4.9 — CLI config resolve shape + malformed write rejection.
  it("resolve().snapshots.periodic returns defaults", () => {
    const store = new ConfigStore(configPath);
@@ -24,7 +24,9 @@ const VIEWER_DESCRIPTION_FRAGMENT = "context-usage across running agents";
 // situation-composed delivery verb. Both are delivery-FREE library verbs
 // (get/profile SERVE bytes to the caller's stdout; nothing sends to a seat),
 // so the delivery-seam exclusion below still holds.
-const LIBRARY_SUBCOMMANDS = ["work-install", "trace", "compose", "list", "show", "preview", "get", "profile", "recap-write", "sync", "add", "source", "rm"];
+// Reconstruction inspect/view reads local immutable records; it neither
+// delivers to a seat nor starts or contacts a daemon.
+const LIBRARY_SUBCOMMANDS = ["reconstruction", "work-install", "trace", "compose", "list", "show", "preview", "get", "profile", "recap-write", "sync", "add", "source", "rm"];

 function topLevelNames(program: Command): string[] {
  return program.commands.map((c) => c.name());
@@ -0,0 +1,99 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+import { mkdirSync, mkdtempSync, readFileSync, realpathSync, readdirSync, rmSync, symlinkSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join, relative } from "node:path";
+import { Command } from "commander";
+import { contextCommand } from "../src/commands/context.js";
+import { resolveWorkContextHome } from "../src/lib/work-install.js";
+
+const fixture = JSON.parse(readFileSync(new URL("./fixtures/reconstruction-v0.json", import.meta.url), "utf8")) as {
+  base: { id: string }; sourceStore: string;
+};
+
+describe("rig context reconstruction", () => {
+  let root: string, catalog: string, alpha: string, beta: string, home: string;
+  beforeEach(() => {
+    root = realpathSync(mkdtempSync(join(tmpdir(), "context-reconstruction-")));
+    catalog = join(root, "catalog"); alpha = join(root, "alpha"); beta = join(root, "beta"); home = join(root, "instance");
+    for (const path of [catalog, alpha, beta, home]) mkdirSync(path);
+    for (const [id, path] of [["alpha", alpha], ["beta", beta]]) {
+      const mission = join(path!, "workplans", "same-task");
+      mkdirSync(join(mission, "slices", "01-example", "reconstructions"), { recursive: true });
+      writeFileSync(join(path!, "project.yaml"), `schema: openrig.project/v0alpha1\nkind: project\nid: ${id}\nmissions: {root: workplans}\n`);
+      writeFileSync(join(mission, "slices", "01-example", "SPEC.md"), "---\nid: OPR.99.0.1\n---\nExample slice\n");
+      writeFileSync(join(mission, "slices", "01-example", "reconstructions", `${fixture.base.id}.json`), JSON.stringify(fixture.base));
+    }
+    writeFileSync(join(catalog, "workspace.yaml"), `schema: openrig.workspace/v0alpha1\nprojects:\n  - id: alpha\n    root: ${relative(catalog, alpha)}\n    rigs: [example-rig]\n  - id: beta\n    root: ${relative(catalog, beta)}\n    rigs: [other-rig]\n`);
+    vi.stubEnv("OPENRIG_HOME", home);
+    vi.stubEnv("OPENRIG_WORKSPACE_ROOT", catalog);
+    vi.stubEnv("OPENRIG_WORKSPACE_CATALOG_PATH", join(catalog, "workspace.yaml"));
+    vi.stubEnv("OPENRIG_SESSION_NAME", ""); vi.stubEnv("RIGGED_SESSION_NAME", "");
+    // If this local command accidentally contacts the daemon, fail in the test.
+    vi.spyOn(globalThis, "fetch").mockImplementation(() => { throw new Error("local reconstruction must not fetch"); });
+  });
+  afterEach(() => { vi.restoreAllMocks(); vi.unstubAllEnvs(); rmSync(root, { recursive: true, force: true }); });
+
+  async function run(args: string[]) {
+    const log = vi.spyOn(console, "log").mockImplementation(() => {});
+    const error = vi.spyOn(console, "error").mockImplementation(() => {});
+    const saved = process.exitCode; process.exitCode = undefined;
+    try {
+      const command = new Command().exitOverride(); command.addCommand(contextCommand());
+      await command.parseAsync(["node", "rig", "context", "reconstruction", ...args]);
+      return { output: log.mock.calls.map(call => call.join(" ")).join("\n"), errors: error.mock.calls, exitCode: process.exitCode };
+    } finally { process.exitCode = saved; log.mockRestore(); error.mockRestore(); }
+  }
+
+  it("resolves a catalog project, custom missions root and slice ID without writing or installing context", async () => {
+    const result = await run(["inspect", fixture.base.id, "--project", "beta", "--mission", "same-task", "--slice", "OPR.99.0.1"]);
+    const data = JSON.parse(result.output);
+    expect(result.exitCode).toBeUndefined();
+    expect(data.complete).toBe(true);
+    expect(data.resolution.projectId).toBe("beta");
+    expect(data.resolution.home).toBe(join(beta, "workplans", "same-task", "slices", "01-example"));
+    expect(data.requestedStore).toBe(join(data.resolution.home, "reconstructions"));
+    expect(data.nodes[0].sourceManifest).toHaveLength(2);
+    expect(data.nodes[0].items).toBeUndefined();
+    expect(readdirSync(home)).toEqual([]);
+    expect(globalThis.fetch).not.toHaveBeenCalled();
+  });
+
+  it("uses shared cwd/rig inference rather than choosing another project's same mission", () => {
+    const common = { workspaceRoot: catalog, catalogPath: join(catalog, "workspace.yaml"), mission: "same-task" };
+    expect(resolveWorkContextHome({ ...common, cwd: join(beta, "workplans") })).toMatchObject({ projectId: "beta", selectedBy: "cwd" });
+    expect(resolveWorkContextHome({ ...common, sessionName: "agent@example-rig", cwd: beta })).toMatchObject({ projectId: "alpha", selectedBy: "rig" });
+    expect(resolveWorkContextHome({ ...common, cwd: root })).toMatchObject({ error: { code: "project_required" } });
+  });
+
+  it("reads an explicit imported store even when the workspace configuration is unavailable", async () => {
+    vi.stubEnv("OPENRIG_WORKSPACE_ROOT", join(root, "missing"));
+    const store = join(alpha, "workplans", "same-task", "slices", "01-example", "reconstructions");
+    const result = await run(["view", fixture.base.id, "--store", store, "--item", "decision", "--json"]);
+    const data = JSON.parse(result.output);
+    expect(data.complete).toBe(true);
+    expect(data.resolution.mode).toBe("explicit-store");
+    expect(data.nodes[0].items.map((v: { id: string }) => v.id)).toEqual(["decision", "dependency", "constraint", "gap", "authority"]);
+    expect(data.nodes[0].viewCoverage.omittedIds).toEqual(["remove", "tagged", "omitted"]);
+    expect(readdirSync(home)).toEqual([]);
+    expect(globalThis.fetch).not.toHaveBeenCalled();
+  });
+
+  it("returns structured failures and nonzero status for missing context, invalid bounds or ambiguous projects", async () => {
+    const ambiguous = await run(["inspect", fixture.base.id, "--cwd", root]);
+    expect(ambiguous.exitCode).toBe(1); expect(JSON.parse(ambiguous.output).error.code).toBe("project_required");
+    const unavailable = await run(["inspect", "b".repeat(64), "--store", root]);
+    expect(unavailable.exitCode).toBe(1); expect(JSON.parse(unavailable.output).complete).toBe(false);
+    const bound = await run(["view", fixture.base.id, "--store", root, "--max-bytes", "NaN", "--json"]);
+    expect(bound.exitCode).toBe(1); expect(JSON.parse(bound.output).error.message).toContain("max-bytes");
+  });
+
+  it("rejects paths escaping the selected mission and invalid mission composition", () => {
+    const opts = { workspaceRoot: catalog, project: "alpha", mission: "same-task" };
+    expect(resolveWorkContextHome({ ...opts, mission: "../beta" })).toMatchObject({ error: { code: "invalid_mission" } });
+    expect(resolveWorkContextHome({ ...opts, slice: "../other" })).toMatchObject({ error: { code: "invalid_slice" } });
+    symlinkSync(beta, join(alpha, "workplans", "same-task", "slices", "outside"));
+    expect(resolveWorkContextHome({ ...opts, slice: "outside" })).toMatchObject({ error: { code: "slice_root_escape" } });
+    writeFileSync(join(alpha, "workplans", "same-task", "mission.yaml"), "composition: {slices: not-a-list}\n");
+    expect(resolveWorkContextHome({ ...opts, slice: "01-example" })).toMatchObject({ error: { code: "mission_composition_invalid" } });
+  });
+});
@@ -158,6 +158,23 @@ describe("transport render: blocked is not down", () => {
 });

 describe("status probe keeps the cause; the precheck guard agrees with the transport render", () => {
+  it.each([
+    ["http://remote.example:7433", true],
+    ["http://127.0.0.1:7433", false],
+    ["http://localhost:7433", false],
+    ["http://[::1]:7433", false],
+  ])("marks blocked explicit remote targets for %s", async (url, remote) => {
+    vi.stubEnv("OPENRIG_URL", url);
+    try {
+      const status = await getDaemonStatus(lifecycleDeps({ fetch: probeFailing("EPERM") }));
+      expect(status.state).toBe("unverified");
+      expect(status.probeErrorCode).toBe("EPERM");
+      expect(status.explicitRemote === true).toBe(remote);
+    } finally {
+      vi.unstubAllEnvs();
+    }
+  });
+
  it("records EPERM on every probe path, and never calls a blocked probe 'stopped'", async () => {
    process.env.OPENRIG_URL = URL;
    try {
@@ -260,7 +277,7 @@ describe("sandboxed liveness: EPERM on kill(pid, 0) is not a dead daemon", () =>
    expect(live("sent", "Z")).toBe("dead");
    expect(live("sent", null)).toBe("unknown");
    expect(live("missing", "S")).toBe("dead");
-    expect(live("not-permitted", "S")).toBe("alive");
+    expect(live("not-permitted", "S")).toBe("unknown");
    expect(live("not-permitted", "Z")).toBe("dead");
    expect(live("not-permitted", null)).toBe("unknown");
  });
@@ -3,7 +3,7 @@ import path from "node:path";
 import fs from "node:fs";
 import os from "node:os";
 import { createServer } from "node:net";
-import { realDeps } from "../src/commands/daemon.js";
+import { createProcessLiveness, realDeps } from "../src/commands/daemon.js";
 import { EventEmitter } from "node:events";
 import type { ChildProcess } from "node:child_process";
 import {
@@ -81,6 +81,57 @@ function writtenState(deps: LifecycleDeps): DaemonState {
 }

 describe("Daemon Lifecycle", () => {
+  it.each(["EPERM", "EACCES"])("warns and starts when state is absent and the initial probe is blocked (%s)", async (code) => {
+    let probes = 0;
+    const deps = startableDeps({ fetch: vi.fn(async () => { if (probes++ === 0) throw Object.assign(new Error("probe blocked"), { code }); return startupHealth(); }) });
+    const warn = vi.spyOn(console, "warn").mockImplementation(() => {});
+    try { await startDaemon({ port: 7433 }, deps); expect(deps.spawn).toHaveBeenCalledOnce(); expect(warn).toHaveBeenCalled(); } finally { warn.mockRestore(); }
+  });
+
+  it("does not spawn when recorded process liveness is unknown and its probe is blocked", async () => {
+    const state: DaemonState = { pid: 999, port: 7433, db: "fixture.db", startedAt: "2026-01-01T00:00:00Z" };
+    const deps = mockDeps({ exists: vi.fn((p: string) => p === STATE_FILE), readFile: vi.fn((p: string) => p === STATE_FILE ? JSON.stringify(state) : null), isProcessAlive: vi.fn(() => false), processLiveness: vi.fn(() => "unknown"), fetch: vi.fn(async () => { throw Object.assign(new Error("blocked"), { code: "EPERM" }); }) });
+    await expect(startDaemon({ port: 7433 }, deps)).rejects.toThrow(/blocked the connection/i);
+    expect(deps.spawn).not.toHaveBeenCalled();
+  });
+
+  it("continues startup when the recorded process is definitely dead", async () => {
+    const state: DaemonState = { pid: 999, port: 7433, db: "fixture.db", startedAt: "2026-01-01T00:00:00Z" };
+    const deps = startableDeps({ exists: vi.fn((p: string) => p === STATE_FILE), readFile: vi.fn((p: string) => p === STATE_FILE ? JSON.stringify(state) : null), processLiveness: vi.fn(() => "dead") });
+    await startDaemon({ port: 7433 }, deps);
+    expect(deps.spawn).toHaveBeenCalledOnce();
+  });
+
+  it("does not signal or remove state when recorded process liveness is unknown", async () => {
+    const state: DaemonState = { pid: 999, port: 7433, db: "fixture.db", startedAt: "2026-01-01T00:00:00Z" };
+    const deps = mockDeps({
+      exists: vi.fn((p: string) => p === STATE_FILE),
+      readFile: vi.fn((p: string) => p === STATE_FILE ? JSON.stringify(state) : null),
+      isProcessAlive: vi.fn(() => false),
+      processLiveness: createProcessLiveness({ signal: () => "not-permitted", readProcessState: () => "S" }),
+      fetch: vi.fn(neverFetch),
+    });
+    await expect(stopDaemon(deps)).rejects.toThrow(/liveness is unknown.*no signal sent.*state preserved/i);
+    expect(deps.fetch).not.toHaveBeenCalled();
+    expect(deps.kill).not.toHaveBeenCalled();
+    expect(deps.removeFile).not.toHaveBeenCalled();
+  });
+
+  it.each(["EPERM", "EACCES"])("preserves a live recorded PID when its identity probe is blocked (%s)", async (code) => {
+    const state: DaemonState = { pid: 999, port: 7433, db: "fixture.db", startedAt: "2026-01-01T00:00:00Z" };
+    const deps = mockDeps({
+      exists: vi.fn((p: string) => p === STATE_FILE),
+      readFile: vi.fn((p: string) => p === STATE_FILE ? JSON.stringify(state) : null),
+      isProcessAlive: vi.fn(() => true),
+      fetch: vi.fn(async () => { throw Object.assign(new Error("probe blocked"), { code }); }),
+    });
+    await expect(startDaemon({ port: 7433 }, deps)).rejects.toThrow(/blocked the connection/i);
+    expect(deps.spawn).not.toHaveBeenCalled();
+    await expect(stopDaemon(deps)).rejects.toThrow(/blocked.*no signal sent/i);
+    expect(deps.kill).not.toHaveBeenCalled();
+    expect(deps.removeFile).not.toHaveBeenCalled();
+  });
+
  // A started child whose own endpoint names its PID and reports `hosts`; `probe` scripts each later
  // request: "ok" answers, any other value throws with that error code.
  function listenerStart(hosts: string[], probe: (url: string) => string) {
@@ -1,7 +1,8 @@
 import { describe, it, expect, vi } from "vitest";
 import { createProgram } from "../src/index.js";
 import { Command } from "commander";
-import { doctorCommand, runDoctorChecks, type DoctorDeps } from "../src/commands/doctor.js";
+import { doctorCommand, runDoctorChecks, readLiveLogicalIds, type DoctorDeps } from "../src/commands/doctor.js";
+import { DaemonClient } from "../src/client.js";
 import { resolveCmuxSettingsPath } from "../src/cmux-config.js";

 const CMUX_SETTINGS_PATH = resolveCmuxSettingsPath();
@@ -567,8 +568,9 @@ describe("runDoctorChecks", () => {
  it("settles a /healthz body that never finishes as owner unknown within its bound, and releases the request", async () => {
    const signals: AbortSignal[] = [];
    const deps = makeDeps({ checkPort: async () => false, callerUid: 1002, ownerReportTimeoutMs: 50,
-      fetch: async (_url: string, init?: { signal?: AbortSignal }) => {
-        if (init?.signal) signals.push(init.signal);
+      fetch: async (url: string, init?: { signal?: AbortSignal }) => {
+        // This owner-body cell measures the two health probes; status has its own deadline cells below.
+        if (init?.signal && url.endsWith("/healthz")) signals.push(init.signal);
        return { ok: true, json: () => new Promise<unknown>(() => {}) };
      } });
    const started = Date.now();
@@ -796,3 +798,173 @@ describe("rig doctor", () => {
    expect(portArgs).toContain(7433);
  });
 });
+
+
+describe("doctor JSON HTTP deadlines", () => {
+  it.each(["cmux headers", "cmux body", "healthy"] as const)("returns an honest diagnostic when a real listener holds %s", async (held) => {
+    const { createServer } = await import("node:http");
+    const requests: string[] = [];
+    const server = createServer((req, res) => {
+      requests.push(req.url!);
+      if (req.url === "/healthz") {
+        res.writeHead(200, { "content-type": "application/json" });
+        res.end(JSON.stringify({ ok: true, pid: 4242, euid: 1002 }));
+      } else if (held === "healthy") {
+        res.writeHead(200, { "content-type": "application/json" });
+        res.end(JSON.stringify({ available: true }));
+      } else if (held === "cmux body") {
+        res.writeHead(200, { "content-type": "application/json" });
+        res.write('{"available":');
+      }
+    });
+    await new Promise<void>(resolve => server.listen(0, "127.0.0.1", resolve));
+    const port = (server.address() as { port: number }).port;
+    const log = vi.spyOn(console, "log").mockImplementation(() => {});
+    const priorExitCode = process.exitCode;
+    const program = new Command();
+    program.addCommand(doctorCommand(makeDeps({ checkPort: async () => false, callerUid: 1002,
+      ownerReportTimeoutMs: 100, fetch: (url, init) => globalThis.fetch(url, init),
+      configStore: { resolve: () => ({ ...defaultConfig, daemon: { host: "127.0.0.1", port } }) },
+    })));
+    const task = program.parseAsync(["node", "rig", "doctor", "--json"]);
+    let observationTimer: ReturnType<typeof setTimeout> | undefined;
+    try {
+      const outcome = await Promise.race([task.then(() => "settled"), new Promise<string>(resolve => {
+        observationTimer = setTimeout(() => resolve("still-waiting"), 1500);
+      })]);
+      expect(["darwin", "linux"]).toContain(process.platform);
+      expect(requests).toContain("/healthz");
+      expect(requests).toContain("/api/adapters/cmux/status");
+      console.info("[doctor HTTP deadline]", JSON.stringify({ platform: process.platform, held, requests, outcome }));
+      expect(outcome).toBe("settled");
+      const result = JSON.parse(String(log.mock.calls.at(-1)![0]));
+      const portCheck = result.checks.find((c: { name: string }) => c.name === "port");
+      const cmux = result.checks.find((c: { name: string }) => c.name === "cmux_daemon");
+      expect(portCheck.status).toBe("pass");
+      expect(cmux.status).toBe(held === "healthy" ? "pass" : "warn");
+    } finally {
+      clearTimeout(observationTimer);
+      server.closeAllConnections();
+      await new Promise<void>(resolve => server.close(() => resolve()));
+      await Promise.allSettled([task]);
+      log.mockRestore(); process.exitCode = priorExitCode;
+      expect(server.listening).toBe(false);
+    }
+  });
+});
+
+
+describe("doctor --spec topology HTTP deadlines", () => {
+  it.each(["rig headers", "rig body", "nodes headers", "nodes body", "healthy"] as const)(
+    "returns unknown or measured conformance when the topology listener is %s", async held => {
+    const { createServer } = await import("node:http");
+    const requests: string[] = [];
+    const heldPath = held.startsWith("rig") ? "/api/rigs" : "/api/rigs/rig-id/nodes";
+    const server = createServer((req, res) => {
+      requests.push(req.url!);
+      if (held !== "healthy" && req.url === heldPath) {
+        if (held.endsWith("body")) { res.writeHead(200, { "content-type": "application/json" }); res.write("["); }
+        return;
+      }
+      res.writeHead(200, { "content-type": "application/json" });
+      res.end(JSON.stringify(req.url === "/healthz" ? { ok: true, pid: 4242, euid: 1002 }
+        : req.url === "/api/rigs" ? [{ id: "rig-id", name: "fixture" }]
+        : req.url === "/api/rigs/rig-id/nodes" ? [{ logicalId: "dev.worker" }] : { available: true }));
+    });
+    await new Promise<void>(resolve => server.listen(0, "127.0.0.1", resolve));
+    const port = (server.address() as { port: number }).port;
+    const base = `http://127.0.0.1:${port}`;
+    vi.stubEnv("OPENRIG_TRANSPORT_LOCAL_USER_GUARD", "true");
+    vi.stubEnv("OPENRIG_URL", ""); vi.stubEnv("RIGGED_URL", "");
+    const ownerProbe = vi.fn(async () => ({ state: "unknown" as const, reason: "fixture-unconfirmed" }));
+    const ownerClient = new DaemonClient(base, { ownerProbe });
+    const specPath = "/fixture/rig.yaml";
+    const spec = 'version: "0.2"\nname: fixture\npods:\n  - id: dev\n    label: Development\n    members:\n      - id: worker\n';
+    const log = vi.spyOn(console, "log").mockImplementation(() => {});
+    const priorExitCode = process.exitCode;
+    const program = new Command().addCommand(doctorCommand(makeDeps({ callerUid: 1002,
+      checkPort: async () => false, specPath, readFile: path => path === specPath ? spec : null,
+      configStore: { resolve: () => ({ ...defaultConfig, daemon: { host: "127.0.0.1", port } }) },
+      fetch: (url, init) => globalThis.fetch(url, init),
+      fetchLiveLogicalIds: name => readLiveLogicalIds(base, name, ownerClient),
+    })));
+    const task = program.parseAsync(["node", "rig", "doctor", "--json", "--spec", specPath]);
+    let timer: ReturnType<typeof setTimeout> | undefined;
+    try {
+      const outcome = await Promise.race([task.then(() => "settled"), new Promise<string>(resolve => {
+        timer = setTimeout(() => resolve("still-waiting"), 4500);
+      })]);
+      expect(["darwin", "linux"]).toContain(process.platform);
+      expect(requests).toContain(held === "healthy" ? "/api/rigs/rig-id/nodes" : heldPath);
+      expect(ownerProbe).toHaveBeenCalled();
+      console.info("[doctor spec HTTP deadline]", JSON.stringify({ platform: process.platform, held, requests, outcome }));
+      expect(outcome).toBe("settled");
+      const result = JSON.parse(String(log.mock.calls.at(-1)![0]));
+      const check = result.checks.find((c: { name: string }) => c.name === "spec_live_conformance");
+      expect(check.status).toBe(held === "healthy" ? "pass" : "skipped");
+      if (held === "healthy") expect(ownerProbe).toHaveBeenCalledTimes(2);
+    } finally {
+      clearTimeout(timer); server.closeAllConnections();
+      await new Promise<void>(resolve => server.close(() => resolve()));
+      await Promise.allSettled([task]);
+      log.mockRestore(); process.exitCode = priorExitCode; vi.unstubAllEnvs();
+      expect(server.listening).toBe(false);
+    }
+  });
+});
+
+describe("#161 tmux control mode", () => {
+  const daemonSays = (body: unknown, env: NodeJS.ProcessEnv = {}) => makeDeps({
+    env,
+    fetch: async (url: string) => url.endsWith("/api/adapters/tmux/control")
+      ? ({ ok: true, json: async () => body } as Response) : ({ ok: false } as Response),
+  });
+  const controlCheck = async (deps: DoctorDeps) =>
+    (await Promise.all(runDoctorChecks(deps).asyncChecks)).find((c) => c.name === "tmux_control_mode");
+
+  it("warns when it was asked for and the daemon's tmux server is older than 3.7c: the version, why, and what to do", async () => {
+    const check = await controlCheck(daemonSays({ requested: true, state: "off", serverVersion: "3.4" }));
+    expect(check?.status).toBe("warn");
+    expect(check?.message).toContain("tmux 3.4");
+    expect(check?.message).toContain("3.7c or later");
+    expect(check?.reason).toContain("not the tmux binary");
+    expect(check?.fix).toContain("rig context get reference/tmux-servers.md#moving-an-existing-team-onto-tmux-control-mode");
+  });
+
+  it("passes when control mode is on, naming the server's version", async () => {
+    const check = await controlCheck(daemonSays({ requested: true, state: "on", serverVersion: "3.8" }));
+    expect(check).toMatchObject({ status: "pass", message: expect.stringContaining("tmux 3.8") });
+  });
+
+  it("#2707 warns when the server allows it but no control client is connected: commands are spawned meanwhile", async () => {
+    const check = await controlCheck(daemonSays({ requested: true, state: "allowed", serverVersion: "3.8" }));
+    expect(check?.status).toBe("warn");
+    expect(check?.message).toContain("tmux 3.8 allows it");
+    expect(check?.message).toContain("no control client is connected right now");
+  });
+
+  it("#2707 names a version it doesn't recognize as unrecognized, and a missing client wrapper by its causes", async () => {
+    const unknown = await controlCheck(daemonSays({ requested: true, state: "off", serverVersion: "next-3.8" }));
+    expect(unknown?.message).toContain("tmux next-3.8, which isn't a tmux release known to be 3.7c or later");
+    const wrapperless = await controlCheck(daemonSays({ requested: true, state: "off" }));
+    expect(wrapperless?.message).toContain("Windows");
+    expect(wrapperless?.message).toContain("OPENRIG_TMUX_SHELL=1");
+  });
+
+  it("says it isn't decided while the daemon hasn't read its server's version yet", async () => {
+    const check = await controlCheck(daemonSays({ requested: true, state: "undecided" }));
+    expect(check).toMatchObject({ status: "skipped", message: expect.stringContaining("hasn't read its tmux server's version yet") });
+  });
+
+  it("warns when this shell sets it but the daemon wasn't started with it", async () => {
+    const check = await controlCheck(daemonSays({ requested: false, state: "off" }, { OPENRIG_TMUX_CONTROL: "1" }));
+    expect(check).toMatchObject({ status: "warn", message: expect.stringContaining("wasn't started with it") });
+    expect(check?.fix).toContain("daemon's own environment");
+  });
+
+  it("skips when it isn't asked for, or when the daemon doesn't say", async () => {
+    expect(await controlCheck(daemonSays({ requested: false, state: "off" }))).toMatchObject({ status: "skipped" });
+    const older = makeDeps({ env: { OPENRIG_TMUX_CONTROL: "1" }, fetch: async () => ({ ok: false } as Response) });
+    expect(await controlCheck(older)).toMatchObject({ status: "skipped", message: expect.stringContaining("didn't say") });
+  });
+});
@@ -0,0 +1,405 @@
+import { afterEach, describe, expect, it } from "vitest";
+import { spawn, type ChildProcess } from "node:child_process";
+import { chmodSync, existsSync, mkdirSync, mkdtempSync, readdirSync, readFileSync, rmSync, symlinkSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { delimiter, dirname, join } from "node:path";
+import { fileURLToPath, pathToFileURL } from "node:url";
+import { processStart } from "../src/commands/acp.js";
+import { claimRow, HARNESSES, resolveHarness, rowRun, type ExecRun } from "../src/commands/exec.js";
+import { readRows, writeRows } from "./fixtures/exec-file-queue.js";
+
+// `rig exec` end to end: each `rig exec …` is its own process that exits as the CLI does, the connector it starts is
+// real and detached, and the adapters are fake installed packages whose program is the fake ACP agent.
+
+const FAKE_AGENT = fileURLToPath(new URL("./fixtures/fake-acp-agent.mjs", import.meta.url));
+const RUNNER = fileURLToPath(new URL("./fixtures/exec-cli-runner.ts", import.meta.url));
+const CONNECTOR_ENTRY = fileURLToPath(new URL("./fixtures/exec-connector-entry.ts", import.meta.url));
+const SEAT = "builder@studio";
+const cleanup: Array<() => void> = [];
+afterEach(() => { for (const fn of cleanup.splice(0)) fn(); });
+
+function install(root: string, pkg: string, version: string, bin?: string, extra: Record<string, unknown> = {}) {
+  const dir = join(root, "node_modules", pkg);
+  mkdirSync(join(dir, "dist"), { recursive: true });
+  writeFileSync(join(dir, "package.json"), JSON.stringify({ name: pkg, version, ...extra }));
+  if (!bin) return;
+  const program = join(dir, "dist", "index.mjs");
+  writeFileSync(program, `#!/usr/bin/env node\nawait import(${JSON.stringify(pathToFileURL(FAKE_AGENT).href)});\n`);
+  chmodSync(program, 0o755);
+  mkdirSync(join(root, "node_modules", ".bin"), { recursive: true });
+  symlinkSync(program, join(root, "node_modules", ".bin", bin));
+}
+
+/** An OpenRig home with both adapters installed and one row owned by this seat, whose body picks the agent's scenario. */
+function setup(body = "scenario:reply PONG", env: Record<string, string> = {}) {
+  const home = mkdtempSync(join(tmpdir(), "rig-exec-"));
+  cleanup.push(() => rmSync(home, { recursive: true, force: true }));
+  install(home, "@agentclientprotocol/claude-agent-acp", "0.101.0", "claude-agent-acp");
+  install(home, "@anthropic-ai/claude-agent-sdk", "0.3.318", undefined, { claudeCodeVersion: "2.1.318" });
+  install(home, "@agentclientprotocol/codex-acp", "2.5.0", "codex-acp");
+  install(home, "@openai/codex", "0.165.0");
+  const queueFile = join(home, "queue.json");
+  writeRows(queueFile, { "qitem-task": { qitemId: "qitem-task", state: "pending", destinationSession: SEAT, body, resolution: null, tags: ["mission:demo"], notes: [] } });
+  const runEnv: NodeJS.ProcessEnv = {
+    ...process.env, OPENRIG_HOME: home, OPENRIG_SESSION_NAME: SEAT, EXEC_QUEUE_FILE: queueFile,
+    PATH: [join(home, "node_modules", ".bin"), dirname(process.execPath), process.env.PATH].join(delimiter), ...env,
+  };
+  for (const name of ["FAKE_ACP_LOAD", "FAKE_ACP_MISSING", "FAKE_ACP_COMPANION_PID_FILE", "FAKE_ACP_COST", "FAKE_ACP_PROMPT_USAGE", "FAKE_ACP_IGNORE_EOF", "EXEC_TEST_CONNECTOR", "EXEC_TEST_WINNER", "EXEC_TEST_HOOK", "EXEC_TEST_STALL_MS", "OPENRIG_EXEC_CONNECTOR"]) if (!(name in env)) delete runEnv[name];
+  return { home, env: runEnv, queueFile };
+}
+type Home = ReturnType<typeof setup>;
+
+function start(h: Home, args: string[], options: { group?: boolean } = {}) {
+  const child: ChildProcess = spawn(process.execPath, ["--import", "tsx", RUNNER, ...args], { env: h.env, stdio: ["ignore", "pipe", "pipe"], detached: options.group });
+  let out = "";
+  child.stdout!.on("data", (data: Buffer) => { out += data.toString(); });
+  child.stderr!.on("data", (data: Buffer) => { out += data.toString(); });
+  const exited = new Promise<number | null>((resolve) => child.once("exit", (code) => resolve(code)));
+  cleanup.push(() => { if (child.exitCode === null) child.kill("SIGKILL"); });
+  return { child, exited, out: () => out };
+}
+async function rig(h: Home, args: string[]) {
+  const run = start(h, args);
+  const code = await run.exited;
+  return { code, out: run.out() };
+}
+
+async function until<T>(check: () => T | undefined, timeoutMs = 20_000): Promise<T> {
+  for (const deadline = Date.now() + timeoutMs; Date.now() < deadline; await new Promise((r) => setTimeout(r, 50))) {
+    const value = check();
+    if (value !== undefined) return value;
+  }
+  throw new Error("timed out");
+}
+const runFile = (h: Home, runId: string) => JSON.parse(readFileSync(join(h.home, "acp", "runs", `${runId}.json`), "utf8")) as ExecRun;
+const ended = (h: Home, runId: string) => until(() => { const run = runFile(h, runId); return run.endedAt ? run : undefined; });
+const notes = (h: Home) => readRows(h.queueFile)["qitem-task"]!.notes;
+const runIds = (h: Home) => readdirSync(join(h.home, "acp", "runs")).filter((f) => /^exec-[^.]+\.json$/.test(f)).map((f) => f.slice(0, -5));
+const startedId = (out: string) => /\[rig exec\] (exec-\S+) for qitem-task/.exec(out)?.[1];
+const withEnv = (h: Home, env: Record<string, string>): Home => ({ ...h, env: { ...h.env, ...env } });
+function withHome<T>(h: Home, fn: () => T): T {
+  const previous = process.env.OPENRIG_HOME;
+  process.env.OPENRIG_HOME = h.home;
+  try { return fn(); } finally { if (previous === undefined) delete process.env.OPENRIG_HOME; else process.env.OPENRIG_HOME = previous; }
+}
+/** A run record as rig exec writes one, for the states a test needs. */
+function record(h: Home, runId: string, fields: Partial<ExecRun>): string {
+  const runs = join(h.home, "acp", "runs");
+  mkdirSync(runs, { recursive: true });
+  const file = join(runs, `${runId}.json`);
+  writeFileSync(file, JSON.stringify({
+    runId, qitemId: "qitem-task", harness: "claude", agent: ["claude-agent-acp"], cwd: h.home, self: SEAT, pollMs: 1000, native: "x",
+    state: "started", pid: 1, requestedAt: "2026-11-06T00:00:00Z", transcript: join(runs, `${runId}.jsonl`), connectorLog: join(runs, `${runId}.connector.log`), ...fields,
+  }));
+  return readFileSync(file, "utf8");
+}
+
+describe("rig exec", () => {
+  it("finds each harness's adapter and native versions, and names the install when it's missing", () => {
+    const h = setup();
+    const claude = resolveHarness(HARNESSES.claude!, { PATH: h.env.PATH });
+    expect(claude).toMatchObject({ adapterVersion: "0.101.0", native: "Claude Code 2.1.318, from @anthropic-ai/claude-agent-sdk 0.3.318" });
+    expect(claude.command).toBe(join(h.home, "node_modules", ".bin", "claude-agent-acp"));
+    expect(resolveHarness(HARNESSES.claude!, { PATH: h.env.PATH, CLAUDE_CODE_EXECUTABLE: "/opt/claude" }).native)
+      .toBe("the Claude Code at /opt/claude (CLAUDE_CODE_EXECUTABLE; its version isn't read here), not Claude Code 2.1.318, from @anthropic-ai/claude-agent-sdk 0.3.318");
+    expect(resolveHarness(HARNESSES.codex!, { PATH: h.env.PATH })).toMatchObject({ adapterVersion: "2.5.0", native: "Codex 0.165.0 (@openai/codex)" });
+    expect(resolveHarness(HARNESSES.codex!, { PATH: "/nonexistent" }).problem)
+      .toBe("codex-acp isn't on PATH; install it with npm install --prefix <dir> @agentclientprotocol/[email protected] and put <dir>/node_modules/.bin on PATH");
+  });
+
+  it("reads a missing session only from each adapter's own evidence, never from a code alone", () => {
+    const { claude, codex } = HARNESSES;
+    const rollout = { code: -32603, message: "Internal error", data: { details: "no rollout found for thread id s1" } };
+    expect(claude!.missingSession({ code: -32002, message: "Resource not found", data: { sessionId: "s1" } }, "s1")).toBe(true);
+    expect(claude!.missingSession({ code: -32002, message: "Resource not found" }, "s1")).toBe(false);
+    expect(claude!.missingSession({ code: -32002, message: "Resource not found", data: { sessionId: "s2" } }, "s1")).toBe(false);
+    expect(claude!.missingSession(rollout, "s1")).toBe(false);
+    expect(codex!.missingSession(rollout, "s1")).toBe(true);
+    expect(codex!.missingSession(rollout, "s9")).toBe(false);
+    expect(codex!.missingSession({ code: -32603, message: "Internal error", data: { details: "context length exceeded" } }, "s1")).toBe(false);
+    expect(codex!.missingSession({ code: -32603, message: "Internal error" }, "s1")).toBe(false);
+    expect(codex!.missingSession({ code: -32002, message: "Resource not found", data: { sessionId: "s1" } }, "s1")).toBe(true);
+  });
+
+  it("--plan shows the versions and what would run, and starts nothing", async () => {
+    const h = setup();
+    const plan = await rig(h, ["run", "qitem-task", "--harness", "claude", "--plan"]);
+    expect(plan.code).toBe(0);
+    expect(plan.out).toContain(`claude: @agentclientprotocol/claude-agent-acp 0.101.0 at ${join(h.home, "node_modules", ".bin", "claude-agent-acp")}`);
+    expect(plan.out).toContain("runs: Claude Code 2.1.318, from @anthropic-ai/claude-agent-sdk 0.3.318");
+    expect(plan.out).toContain("answers: a reject refuses that one action, and the turn continues");
+    expect(plan.out).toContain("would run:");
+    expect(existsSync(join(h.home, "acp", "runs"))).toBe(false);
+    expect(notes(h)).toEqual([]);
+  }, 30_000);
+
+  it("keeps running after the calling CLI exits, and records the outcome as the agent's report", async () => {
+    const h = setup("scenario:slow");
+    const started = await rig(h, ["run", "qitem-task", "--harness", "claude"]);
+    expect(started.code).toBe(0);
+    const runId = startedId(started.out)!;
+    expect(started.out).toContain(`${runId} for qitem-task on claude: started`);
+    const live = runFile(h, runId);
+    expect(live.state).toBe("started");
+    expect(() => process.kill(live.pid, 0)).not.toThrow(); // the connector outlives the CLI that started it
+    const run = await ended(h, runId);
+    expect(run).toMatchObject({ state: "finished", outcome: "completed", summary: "the agent ended its turn (end_turn)" });
+    expect(run.settle).toContain("settled: no process carrying this run's marker remains");
+    expect(readRows(h.queueFile)["qitem-task"]!.state).toBe("in-progress"); // accepting the work stays with the custodian
+    expect(notes(h).join("\n")).toMatch(/executor fake-acp-agent 0\.0\.1 \(protocol 1, loadSession not advertised\), session fake-session-1 new, custodian builder@studio/);
+    const status = await rig(h, ["status", "qitem-task"]);
+    expect(status.out).toContain("finished, outcome completed: the agent ended its turn (end_turn)");
+    expect(status.out).toContain("The outcome is the agent's report, not acceptance: the row stays open for its custodian to judge and close.");
+    expect(status.out).toContain("Usage as the agent reported it: tokens unknown");
+  }, 40_000);
+
+  it("leaving --wait, even by a Ctrl-C to its whole process group, stops only the wait", async () => {
+    const h = setup("scenario:slow");
+    const waiter = start(h, ["run", "qitem-task", "--harness", "codex", "--wait"], { group: true });
+    await until(() => (/for qitem-task on codex: started/.test(waiter.out()) ? true : undefined));
+    process.kill(-waiter.child.pid!, "SIGINT"); // what a terminal's Ctrl-C sends: the connector is in its own session
+    expect(await waiter.exited).toBe(3);
+    expect(waiter.out()).toMatch(/stopped waiting; exec-\S+ keeps running/);
+    const run = await ended(h, startedId(waiter.out())!);
+    expect(run.outcome).toBe("completed");
+    expect(notes(h).join("\n")).not.toContain("cancel requested");
+  }, 40_000);
+
+  it("a Ctrl-C that lands just after the run is announced, while the CLI is stalled there, still stops only the wait", async () => {
+    const h = withEnv(setup("scenario:slow"), { EXEC_TEST_STALL_MS: "1500" });
+    const waiter = start(h, ["run", "qitem-task", "--harness", "codex", "--wait"], { group: true });
+    await until(() => (/for qitem-task on codex: started/.test(waiter.out()) ? true : undefined));
+    process.kill(-waiter.child.pid!, "SIGINT"); // arrives during the stall, before the CLI's next step
+    expect(await waiter.exited).toBe(3);
+    expect(waiter.out()).toMatch(/stopped waiting; exec-\S+ keeps running/);
+    expect((await ended(h, startedId(waiter.out())!)).outcome).toBe("completed");
+  }, 40_000);
+
+  it("starts a row's task once: a second start or a retry returns the same run, and --again starts another after it ends", async () => {
+    const h = setup();
+    const args = ["run", "qitem-task", "--harness", "claude"];
+    const [a, b] = await Promise.all([rig(h, args), rig(h, args)]);
+    expect(runIds(h)).toHaveLength(1);
+    const [runId] = runIds(h);
+    expect(`${a.out}\n${b.out}`).toMatch(/started by another caller|already has a run, so the task isn't sent twice/);
+    await ended(h, runId!);
+    const retry = await rig(h, args);
+    expect(retry.out).toContain(`qitem-task already has a run, so the task isn't sent twice:\n${runId} for qitem-task on claude`);
+    expect(runIds(h)).toEqual([runId]);
+    expect(notes(h).filter((note) => /^ACP run \S+: executor/.test(note))).toHaveLength(1);
+
+    const again = await rig(h, [...args, "--again"]);
+    const second = startedId(again.out)!;
+    expect(second).not.toBe(runId);
+    await ended(h, second);
+    expect(notes(h).filter((note) => /^ACP run \S+: executor/.test(note))).toHaveLength(2);
+  }, 60_000);
+
+  it("claims a row for one run only: a second claim loses, and a replacement must name the run it replaces", () => {
+    const h = setup();
+    const previous = process.env.OPENRIG_HOME;
+    process.env.OPENRIG_HOME = h.home;
+    try {
+      expect(claimRow("qitem-task", "exec-a")).toBe(true);
+      expect(claimRow("qitem-task", "exec-b")).toBe(false);
+      expect(rowRun("qitem-task")).toBe("exec-a");
+      expect(claimRow("qitem-task", "exec-c", "exec-b")).toBe(false);
+      expect(claimRow("qitem-task", "exec-c", "exec-a")).toBe(true);
+      expect(claimRow("qitem-task", "exec-d", "exec-a")).toBe(false);
+      expect(rowRun("qitem-task")).toBe("exec-c");
+    } finally {
+      if (previous === undefined) delete process.env.OPENRIG_HOME; else process.env.OPENRIG_HOME = previous;
+    }
+  });
+
+  it("lets only one of two callers that read the same current run claim the row, and a fresh start never claims during a replacement", () => {
+    withHome(setup(), () => {
+      expect(claimRow("qitem-task", "exec-p")).toBe(true);
+      let inWindow: boolean | undefined;
+      // A and B both read exec-p as the row's run; B publishes inside A's window, so A's publish fails.
+      expect(claimRow("qitem-task", "exec-a", "exec-p", () => { inWindow = claimRow("qitem-task", "exec-b", "exec-p"); })).toBe(false);
+      expect(inWindow).toBe(true);
+      expect(rowRun("qitem-task")).toBe("exec-b");
+      // A fresh start inside a replacement's window still sees the row's run, so it never claims.
+      expect(claimRow("qitem-task", "exec-c", "exec-b", () => { inWindow = claimRow("qitem-task", "exec-x"); })).toBe(true);
+      expect(inWindow).toBe(false);
+      expect(rowRun("qitem-task")).toBe("exec-c");
+      // Two fresh starts on a row with no run: the one that publishes first wins.
+      expect(claimRow("qitem-new", "exec-1", undefined, () => { inWindow = claimRow("qitem-new", "exec-2"); })).toBe(false);
+      expect(inWindow).toBe(true);
+      expect(rowRun("qitem-new")).toBe("exec-2");
+    });
+  });
+
+  it("exits by the run's outcome on every path: an ended run returned again, a run that ended before the first look, and a losing caller", async () => {
+    const failed = setup("scenario:error");
+    const first = startedId((await rig(failed, ["run", "qitem-task", "--harness", "claude"])).out)!;
+    expect((await ended(failed, first)).outcome).toBe("failed");
+    const again = await rig(failed, ["run", "qitem-task", "--harness", "claude", "--wait"]);
+    expect(again.out).toContain("qitem-task already has a run, so the task isn't sent twice");
+    expect(again.code).toBe(2);
+    expect((await rig(failed, ["run", "qitem-task", "--harness", "claude"])).code).toBe(2);
+
+    const quick = await rig(withEnv(setup(), { EXEC_TEST_CONNECTOR: "instant" }), ["run", "qitem-task", "--harness", "claude"]);
+    expect(quick.out).toMatch(/for qitem-task on claude: ended, outcome completed/);
+    expect(quick.code).toBe(0);
+
+    for (const [outcome, code] of [["failed", 2], ["completed", 0]] as const) {
+      const lost = withEnv(setup(), { EXEC_TEST_WINNER: outcome });
+      const loser = await rig(lost, ["run", "qitem-task", "--harness", "claude", "--wait"]);
+      expect(loser.out).toContain("qitem-task was started by another caller as exec-winner, so the task isn't sent twice");
+      expect(loser.out).toContain(`finished, outcome ${outcome}`);
+      expect(loser.code).toBe(code);
+      expect(runIds(lost)).toEqual(["exec-winner"]); // the losing caller's own record is gone, and nothing else started
+    }
+  }, 60_000);
+
+  it("ends a run interrupted when its connector dies before recording itself, even while a waiter or requester lives", async () => {
+    const h = setup();
+    const died = await rig(withEnv(h, { EXEC_TEST_CONNECTOR: "die" }), ["run", "qitem-task", "--harness", "claude", "--wait"]);
+    expect(died.code).toBe(2);
+    expect(died.out).toContain("interrupted: its connector exited (code 3, signal none) and never recorded itself, so no turn was sent");
+    expect(runFile(h, startedId(died.out)!)).toMatchObject({ state: "interrupted", outcome: "unknown" });
+
+    // A requested record names its requester; a live requester with no connector is not a live run.
+    const other = setup();
+    record(other, "exec-orphan", { state: "requested", pid: process.pid, processStart: processStart(process.pid) });
+    const status = await rig(other, ["status", "exec-orphan"]);
+    expect(status.out).toContain("interrupted: its connector isn't running and never recorded itself, so no turn was sent");
+  }, 40_000);
+
+  it("settles a run's start and its end exclusively against a concurrent reconciler, in either order", async () => {
+    const connectorLog = (h: Home, runId: string) => readFileSync(runFile(h, runId).connectorLog, "utf8");
+    const binding = (h: Home) => notes(h).filter((note) => /^ACP run \S+: executor/.test(note));
+
+    // The reconciler first: it settles the run after the connector's guard and before its start, so no turn is sent.
+    const first = withEnv(setup(), { EXEC_TEST_HOOK: "after-guard" });
+    const a = await rig(first, ["run", "qitem-task", "--harness", "claude", "--wait"]);
+    const aId = startedId(a.out)!;
+    expect(a.code).toBe(2);
+    expect(runFile(first, aId)).toMatchObject({ state: "interrupted", outcome: "unknown", summary: "interrupted by a test reconciler after-guard" });
+    await until(() => (/no turn is sent/.test(connectorLog(first, aId)) ? true : undefined));
+    expect(connectorLog(first, aId)).toContain("reconciler after-guard: won");
+    expect(binding(first)).toEqual([]);
+    expect(notes(first).filter((note) => note.startsWith("test reconciler"))).toHaveLength(1);
+
+    // The connector first: the reconciler still holding the requested copy loses, and the turn completes.
+    const second = withEnv(setup(), { EXEC_TEST_HOOK: "after-start" });
+    const bId = startedId((await rig(second, ["run", "qitem-task", "--harness", "claude"])).out)!;
+    expect(await ended(second, bId)).toMatchObject({ state: "finished", outcome: "completed" });
+    expect(connectorLog(second, bId)).toContain("reconciler after-start: lost");
+    expect(binding(second)).toHaveLength(1);
+    expect(notes(second).some((note) => note.startsWith("test reconciler"))).toBe(false);
+
+    // The end: a reconciler that settles the started run before the connector records its end wins, and the
+    // connector's own outcome is told to the row without replacing that record.
+    const third = withEnv(setup(), { EXEC_TEST_HOOK: "before-end" });
+    const cId = startedId((await rig(third, ["run", "qitem-task", "--harness", "claude"])).out)!;
+    await until(() => (notes(third).some((note) => note.includes("that record stands")) ? true : undefined));
+    expect(runFile(third, cId)).toMatchObject({ state: "interrupted", summary: "interrupted by a test reconciler before-end" });
+    expect((await rig(third, ["status", cId])).out).toContain("interrupted by a test reconciler before-end");
+    expect(notes(third).join("\n")).toContain(`rig exec run ${cId}: its connector finished with outcome completed (the agent ended its turn (end_turn)) after the run had been recorded interrupted; that record stands`);
+  }, 60_000);
+
+  it("recovers a run from its slots when a crash left the copy behind", async () => {
+    const h = setup();
+    const gone = spawn(process.execPath, ["-e", ""]);
+    await new Promise((resolve) => gone.once("exit", resolve));
+    const runs = join(h.home, "acp", "runs");
+    // The connector claimed its start and died before writing the copy, which still says requested.
+    record(h, "exec-crashed", { state: "requested", pid: process.pid, processStart: processStart(process.pid) });
+    writeFileSync(join(runs, "exec-crashed.start.json"), JSON.stringify({ by: "connector", pid: gone.pid, processStart: "Thu Jan  1 00:00:00 2026", startedAt: "2026-11-06T00:00:01Z" }));
+    const status = await rig(h, ["status", "exec-crashed"]);
+    expect(status.out).toContain("interrupted: its connector ended without recording an outcome, so the outcome is unknown");
+    // The connector recorded its end and died before writing the copy, which still says started.
+    record(h, "exec-ended", { state: "started", pid: gone.pid, processStart: "Thu Jan  1 00:00:00 2026" });
+    writeFileSync(join(runs, "exec-ended.start.json"), JSON.stringify({ by: "connector", pid: gone.pid, startedAt: "2026-11-06T00:00:01Z" }));
+    writeFileSync(join(runs, "exec-ended.end.json"), JSON.stringify({ by: "connector", ended: { state: "finished", outcome: "completed", summary: "the agent ended its turn (end_turn)", endedAt: "2026-11-06T00:00:09Z" } }));
+    expect((await rig(h, ["status", "exec-ended"])).out).toContain("finished, outcome completed");
+    expect(notes(h).filter((note) => note.startsWith("rig exec run exec-ended"))).toEqual([]);
+  }, 40_000);
+
+  it("the connector entry runs only the requested run it was started for, and leaves any other record untouched", async () => {
+    const h = setup();
+    const done = record(h, "exec-done", { state: "finished", outcome: "completed", endedAt: "2026-11-06T00:00:05Z" });
+    const waiting = record(h, "exec-waiting", { state: "requested" });
+    const connector = async (runId: string, marker?: string) => {
+      const child = spawn(process.execPath, ["--import", "tsx", CONNECTOR_ENTRY, runId], { env: { ...h.env, ...(marker ? { OPENRIG_EXEC_CONNECTOR: marker } : {}) }, stdio: "ignore" });
+      return new Promise<number | null>((resolve) => child.once("exit", (code) => resolve(code)));
+    };
+    expect(await connector("exec-done", "exec-done")).toBe(1);
+    expect(await connector("exec-waiting")).toBe(1);
+    expect(await connector("exec-waiting", "exec-other")).toBe(1);
+    expect(readFileSync(join(h.home, "acp", "runs", "exec-done.json"), "utf8")).toBe(done);
+    expect(readFileSync(join(h.home, "acp", "runs", "exec-waiting.json"), "utf8")).toBe(waiting);
+    expect(notes(h)).toEqual([]);
+  }, 40_000);
+
+  it("reads a run whose connector died, or whose pid now names another process, as interrupted, with one note", async () => {
+    const h = setup();
+    const gone = spawn(process.execPath, ["-e", ""]);
+    await new Promise((resolve) => gone.once("exit", resolve));
+    const other = spawn(process.execPath, ["-e", "setInterval(() => {}, 1000)"], { stdio: "ignore" });
+    cleanup.push(() => other.kill("SIGKILL"));
+    const runs = join(h.home, "acp", "runs");
+    mkdirSync(runs, { recursive: true });
+    const record = (runId: string, pid: number) => writeFileSync(join(runs, `${runId}.json`), JSON.stringify({
+      runId, qitemId: "qitem-task", harness: "claude", agent: ["claude-agent-acp"], cwd: h.home, self: SEAT, pollMs: 1000, native: "x",
+      state: "started", pid, processStart: "Thu Jan  1 00:00:00 2026", requestedAt: "2026-11-06T00:00:00Z", startedAt: "2026-11-06T00:00:01Z",
+      transcript: join(runs, `${runId}.jsonl`), connectorLog: join(runs, `${runId}.connector.log`),
+    }));
+    record("exec-dead", gone.pid!);
+    record("exec-reused", other.pid!);
+    for (const runId of ["exec-dead", "exec-reused"]) {
+      const status = await rig(h, ["status", runId]);
+      expect(status.out).toContain("interrupted: its connector ended without recording an outcome, so the outcome is unknown; no process carrying the run's marker remains");
+      expect(runFile(h, runId)).toMatchObject({ state: "interrupted", outcome: "unknown" });
+      await rig(h, ["status", runId]);
+      expect(notes(h).filter((note) => note.startsWith(`rig exec run ${runId} interrupted`))).toHaveLength(1);
+    }
+    expect(other.exitCode).toBeNull();
+    expect(other.signalCode).toBeNull(); // the process that reused the pid was never signalled
+  }, 40_000);
+
+  it("cancel asks the agent to stop; the acknowledgement and what still runs are kept apart", async () => {
+    const h = setup("scenario:slow");
+    const runId = startedId((await rig(h, ["run", "qitem-task", "--harness", "claude"])).out)!;
+    await until(() => (notes(h).some((note) => note.includes("executor")) ? true : undefined));
+    const cancel = await rig(h, ["cancel", runId]);
+    expect(cancel.code).toBe(0);
+    expect(cancel.out).toContain("An acknowledgement isn't settlement");
+    const run = await ended(h, runId);
+    expect(run.outcome).toBe("cancelled");
+    expect(run.summary).toMatch(/^cancel requested at \S+, acknowledged by the agent at \S+ \(stopReason cancelled\); upstream spend is unknown$/);
+    expect(run.settle).toContain("settled: no process carrying this run's marker remains");
+    expect((await rig(h, ["cancel", runId])).code).toBe(1); // an ended run isn't signalled again
+  }, 40_000);
+
+  it("records a failed turn as failed, and a completed turn that left a process running as both", async () => {
+    const failed = setup("scenario:error");
+    const failedId = startedId((await rig(failed, ["run", "qitem-task", "--harness", "codex"])).out)!;
+    expect(await ended(failed, failedId)).toMatchObject({ outcome: "failed", summary: 'session/prompt returned error -32603 Internal error {"details":"context length exceeded"}' });
+
+    const pidFile = join(tmpdir(), `rig-exec-companion-${process.pid}`);
+    const leftBehind = setup("scenario:reply PONG", { FAKE_ACP_COMPANION_PID_FILE: pidFile });
+    const runId = startedId((await rig(leftBehind, ["run", "qitem-task", "--harness", "claude"])).out)!;
+    const run = await ended(leftBehind, runId);
+    const companion = Number(readFileSync(pidFile, "utf8"));
+    cleanup.push(() => { try { process.kill(companion, "SIGKILL"); } catch { /* gone */ } rmSync(pidFile, { force: true }); });
+    expect(run.outcome).toBe("completed");
+    expect(run.settle).toContain(`incomplete: 1 process(es) carrying this run's marker still run (${companion})`);
+    const status = await rig(leftBehind, ["status", runId]);
+    expect(status.out).toContain("finished, outcome completed");
+    expect(status.out).toContain(`After the turn: agent exited`);
+    expect(status.out).toContain(`still run (${companion})`);
+  }, 60_000);
+
+  it("a session the adapter says doesn't exist reads missing, and no fresh session is started", async () => {
+    const h = setup("scenario:reply PONG", { FAKE_ACP_LOAD: "1", FAKE_ACP_MISSING: "codex" });
+    const runId = startedId((await rig(h, ["run", "qitem-task", "--harness", "codex", "--load-session", "sess-gone"])).out)!;
+    const run = await ended(h, runId);
+    expect(run.outcome).toBe("missing");
+    expect(run.summary).toContain("session sess-gone doesn't exist for this agent (session/load returned -32603 Internal error");
+    expect(readFileSync(run.transcript, "utf8")).not.toContain("session/new");
+  }, 40_000);
+});
@@ -0,0 +1,34 @@
+// One `rig exec …` invocation in its own process, which then exits as the CLI does, over the test's file queue.
+// EXEC_TEST_CONNECTOR=die starts a connector that exits before recording itself; =instant one that records a completed
+// outcome before the CLI's first look. EXEC_TEST_WINNER=<outcome> has another caller claim the row, with a run that
+// ended with that outcome, between this caller's claim read and its publish. EXEC_TEST_STALL_MS stalls the CLI right
+// after it announces the run, as a process descheduled there would be.
+import { writeFileSync } from "node:fs";
+import { join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { claimRow, execCommand } from "../../src/commands/exec.js";
+import { fileQueue } from "./exec-file-queue.js";
+
+const entry = fileURLToPath(new URL("./exec-connector-entry.ts", import.meta.url));
+const instant = fileURLToPath(new URL("./exec-instant-connector.ts", import.meta.url));
+const queue = fileQueue(process.env.EXEC_QUEUE_FILE!);
+const connector = process.env.EXEC_TEST_CONNECTOR;
+const winner = process.env.EXEC_TEST_WINNER;
+const stall = Number(process.env.EXEC_TEST_STALL_MS ?? 0);
+await execCommand({
+  queue: async () => queue,
+  connectorCommand: (runId) => connector === "die" ? [process.execPath, "-e", "process.exit(3)"]
+    : [process.execPath, "--import", "tsx", connector === "instant" ? instant : entry, runId],
+  pollMs: 100,
+  startWaitMs: 20_000,
+  afterAnnounce: stall ? () => { for (const until = Date.now() + stall; Date.now() < until;) { /* busy: no timer or signal callback runs */ } } : undefined,
+  beforeClaim: winner ? () => {
+    const runs = join(process.env.OPENRIG_HOME!, "acp", "runs");
+    writeFileSync(join(runs, "exec-winner.json"), JSON.stringify({
+      runId: "exec-winner", qitemId: "qitem-task", harness: "claude", agent: ["claude-agent-acp"], cwd: runs, self: "builder@studio",
+      pollMs: 1000, native: "x", state: "finished", pid: 1, requestedAt: "2026-11-06T00:00:00Z", endedAt: "2026-11-06T00:00:05Z",
+      outcome: winner, summary: `the winner's run ended ${winner}`, transcript: join(runs, "exec-winner.jsonl"), connectorLog: join(runs, "exec-winner.connector.log"),
+    }));
+    claimRow("qitem-task", "exec-winner");
+  } : undefined,
+}).parseAsync(["node", "rig", ...process.argv.slice(2)]);
@@ -0,0 +1,22 @@
+// The detached connector `rig exec run` starts, over the test's file queue instead of the daemon.
+// EXEC_TEST_HOOK puts a concurrent reconciler at one point in the connector's life, holding the copy of the run it read
+// earlier and settling it as interrupted: after-guard (before the connector claims its start), after-start (after it
+// has, still holding the requested copy), before-end (holding the started copy, before the connector records its end).
+import { readRun, runExecConnector, settleRun, type ConnectorHooks, type ExecRun } from "../../src/commands/exec.js";
+import { fileQueue } from "./exec-file-queue.js";
+
+const runId = process.argv[2]!;
+const queue = fileQueue(process.env.EXEC_QUEUE_FILE!);
+const requested = readRun(runId);
+const interrupt = async (label: string, stale: ExecRun | undefined) => {
+  const settled = await settleRun(stale!, { state: "interrupted", outcome: "unknown", summary: `interrupted by a test reconciler ${label}`, endedAt: new Date().toISOString() },
+    `test reconciler ${label} settled ${runId}`, async () => queue);
+  console.error(`reconciler ${label}: ${settled.won ? "won" : "lost"}`);
+};
+const hook = process.env.EXEC_TEST_HOOK;
+const hooks: ConnectorHooks = hook === "after-guard" ? { afterGuard: () => interrupt(hook, requested) }
+  : hook === "after-start" ? { afterStart: () => interrupt(hook, requested) }
+    : hook === "before-end" ? { beforeEnd: () => interrupt(hook, readRun(runId)) }
+      : {};
+
+process.exitCode = await runExecConnector(runId, queue, hooks);
@@ -0,0 +1,37 @@
+// A queue kept in one JSON file, so a test, the `rig exec` CLI process and its detached connector all see the same rows.
+import { readFileSync, renameSync, writeFileSync } from "node:fs";
+import type { AcpQueue, AcpQueueRow } from "../../src/acp/connector.js";
+
+export type FileRow = AcpQueueRow & { notes: string[]; summary?: string };
+
+export function readRows(file: string): Record<string, FileRow> {
+  return JSON.parse(readFileSync(file, "utf8")) as Record<string, FileRow>;
+}
+
+export function writeRows(file: string, rows: Record<string, FileRow>): void {
+  const tmp = `${file}.${process.pid}.tmp`;
+  writeFileSync(tmp, JSON.stringify(rows));
+  renameSync(tmp, file);
+}
+
+export function fileQueue(file: string): AcpQueue {
+  const change = (fn: (rows: Record<string, FileRow>) => void) => { const rows = readRows(file); fn(rows); writeRows(file, rows); };
+  let asks = 0;
+  return {
+    show: async (id) => {
+      const row = readRows(file)[id];
+      if (!row) throw new Error(`queue show ${id} returned HTTP 404`);
+      return { ...row };
+    },
+    claim: async (id) => change((rows) => { rows[id]!.state = "in-progress"; }),
+    update: async (id, update) => change((rows) => {
+      rows[id]!.notes.push(update.note);
+      if (update.state) rows[id]!.state = update.state;
+    }),
+    create: async (row) => {
+      const id = `qitem-ask-${process.pid}-${++asks}`;
+      change((rows) => { rows[id] = { qitemId: id, state: "pending", destinationSession: row.destination, body: row.body, resolution: null, tags: row.tags, notes: [], summary: row.summary }; });
+      return id;
+    },
+  };
+}
@@ -0,0 +1,10 @@
+// A connector whose turn ends before the CLI that started it first looks: it records a completed outcome in one write.
+import { readFileSync, renameSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+
+const runId = process.argv[2]!;
+const file = join(process.env.OPENRIG_HOME!, "acp", "runs", `${runId}.json`);
+const run = JSON.parse(readFileSync(file, "utf8")) as Record<string, unknown>;
+const at = new Date().toISOString();
+writeFileSync(`${file}.instant`, JSON.stringify({ ...run, state: "finished", pid: process.pid, startedAt: at, endedAt: at, outcome: "completed", summary: "the agent ended its turn (end_turn)" }));
+renameSync(`${file}.instant`, file);
				`@@ -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>