A batch printed its aggregates and kept nothing they could be recomputed
from, so a published figure could not be tied to the runs behind it after
the temp directory was gone. -batch DIR now writes results.jsonl (one row
per attempt) beside batch.json (manifest v1), and -mode audit re-derives
the report's headline figures from those rows with the harness's own
aggregation, printing the claim beside the recomputation and naming every
figure that moved.
The manifest records what the figures are over (command axes, a task-id
digest, filters, and a denominator defined in tasks), which run produced
them (commit or an explicit unknown, model/provider names, seed when
known, timestamps, harness and runtime), and why they can still be
checked (results_sha256, the claimed figures, and a ledger of planned /
attempted / completed / scored / failed / unmeasured / excluded that the
audit checks for consistency). Absent figures are omitted and read as
unknown, never as zero.
This is an aggregation audit, not a replay: it proves the published
figures follow from the retained rows. It does not re-execute the model,
tools or environment, and a real replay (commit, corpus fingerprint,
model/provider configuration, seed, flags, raw results, runtime version)
is not implemented.
Problem: Accepted total-only usage remained visible to reporting but did not
cross token budgets, allowing additional tool rounds after the known spend.
Root cause: Both local and shared-root enforcement added only input/output
components, so missing or partial splits hid a known reported total.
Fix: Use one accounting helper for the larger of the nonnegative total and
component sum, and saturate both counters. Keep usage, pricing, shadow component
telemetry, context ownership and existing soft-gate/finalization behavior.
Verification: Real OpenAI HTTP/SSE adapter and Boot execute real file tools.
Final19named regressions pass; exact preceding owners fail10 events with9controls.
Root240packages once:24330named passes/215existing skips. Complete Agent race
2162pass/10existing skips; focused Boot race7pass. Vet, pinned lint, check, format,
workflow and release gates pass; four Win/Linuxamd64 checks compile only.
Problem: Linux coverage could not establish the carried-proof fixture;
CodeQL also flagged the shared inventory parser for exponential backtracking.
Root cause: the fixture inherited an unavailable host bash sandbox, and
repeated name-suffix groups could repartition the same slash sequence.
Fix: select sandbox off only in the isolated test user configuration;
parse the entire optional suffix once and add a bounded long-input test.
Verification: required builtin/Boot exit0, focused coverage four named
passes, 14-name shared contract, two Node tests, vet, pinned lint and check.
Original regex fails the bounded regression with one passing control.
No production Go, assertion, workflow, timeout, skip or baseline change.
Hosted Linux coverage, CodeQL and own-head Windows evidence remain due.
Problem: CodeQL reports exponential backtracking in the shared Windows
inventory test-name parser.
Root cause: repeated suffix groups can partition the same slash sequence.
Fix: accept the entire optional non-whitespace suffix once, preserving the
previous accepted language. Cover nested names and a long malformed suffix.
Verification: two Node tests pass; the original regex fails the bounded
regression while the existing contract test passes. Go and inventory bytes
are unchanged; prior exact-head Go qualification remains separate.
SessionLeaseHeldByOtherRuntime folded every lease-lock probe failure into
"held", so a lock it could not open became the same answer as a live holder.
Redaction then skipped a session nobody held and reported an active lease,
leaving the secret scrub silently incomplete; session and task listings
claimed active/running on the same non-evidence.
Add ErrSessionLeaseUnprovable beside ErrSessionLeaseHeld and expose
SessionLeaseOwnership as the three-state probe. Only ErrSessionLeaseHeld
proves a holder; any other lock failure is unprovable with its cause kept.
The bool wrappers keep their old "may be busy" answers, so callers that only
skip work are unchanged.
Redaction refuses an unprovable session, reports it as a skipped file with
the reason (and exits non-zero), and never calls it an active lease. Machine
session state and running task/subagent status become unknown when ownership
cannot be proven; held and free keep their current answers.
The registry declares 324 pairs now: the 202 it already had and 122 page
copies that agree field for field with the struct their handler writes. The
table moved to a second file, wireparity_table.go, rather than taking
wireparity.go past the 800-line ceiling; the checker still reads one list.
Left out on purpose, because the writer is not a named struct and a guard
over the lookalike struct would not see the response drift:
account.DeviceGrant (/account/login assembles a map), market.Page
(/market/packages assembles a map) and control's McpDraft (/mcp/parse
assembles a map). host.ts's Refusal and ViewRect are the Electron preload's
bridge, not kernel wire; workspace.ts's WorkspaceEntry has no writer.
Tests: the added half resolves, is exact today and is unique per reader; a
dropped field on createBackupRequest -> BackupCreateRequest names the field.
The terminal's hand-written HistoryMessage dropped images, via and modelRef.
A turn that was only an image has empty text, so Restore saw nothing to draw
and dropped it from the transcript, and every reload lost the device and the
model of turns it did keep. Carry the fields that decide that a turn exists
and what it was, restore the image-only row with an honest line, and register
the pair in repolint so a dropped field fails make check.
The navigation already shows the section and page, so the first card now sits
at the top of the page. The setting count and the section's value stay in the
navigation row.
Problem: Competing continuations could create two copies of one ancestor in
the same parent session, leaving later continuation ambiguous.
Root cause: The source ref lock ended before a newly allocated copy published
metadata, and each preparation acquired a different destination ref lock.
Fix: Reserve the canonical nearest-source and target-parent pair before
searching for copies. Transfer the reservation to the returned run's existing
release lifetime, and release it on failed preparation. Preserve independent
parents, sources, current-reference locks and persisted formats.
Verification: Real TaskTool publication interleavings and Store regressions;
exact old-owner overlay fails 9 named events with 5 controls. Root 240-package
coverage, delegation/sessionstore race, vet, pinned lint, make check,
actionlint and Windows/Linux compile-only checks pass. Windows named-step
inventory validation rejects missing package completion and skipped cases.
Refs #14025
Invalidate persisted Delivery proof when file mode changes without a byte edit.
Use the existing content-version reader and verify surrounding identity, mode,
size and mtime observations. Refuse links, directories and special files as
regular-file proof inputs.
Version fingerprint values so legacy byte-only proof conservatively requires
fresh evidence in both reader directions, while goals and transcripts remain
readable. Cover real Boot persistence/rebuild, compatibility and Unix kinds,
and add a required native Windows inventory with bilingual documentation.
Refs #14025
Refs #14133
Preserve every workflow step, condition, timeout and assertion. Move the
status-owned pair away from concurrently added budget inventories so both
sets compose, and collect the named evidence before broad smoke can fail.
Refs #14656
Carry the accepted UsageQuoter event owner alongside Goal usage scope capture.
Retain both optional capabilities in wrappers and their shared inventory while
keeping scope-specific fixture additions independent of the quote additions.
Verify that nested bound usage reaches the occurrence-time quote owner once,
keeps the earlier Goal attribution, and preserves the emitted JSON. Document
the two owners and retain existing Windows required-test coverage.
Refs #14661
Refs #14671
Problem: local task cost gates could count a base rate while the accounting
event charged the vendor's occurrence-time peak rate, allowing extra tools.
Root cause: Agent budgeting recomputed spend with the legacy float adapter
instead of consuming the quote owned by the existing pricing sink.
Fix: forward an optional usage-quoter capability, capture one authoritative
quote for both delivery and budgeting, and accumulate fixed-point original
currency buckets. Preserve known subtotal and incomplete-usage semantics.
Keep display currency, provider requests and cancellation separate.
Verification: actual Agent and Boot peak/off-peak/relay regressions, preceding
owner negatives with controls, full root package inventory, relevant owner
and Boot race suites, static gates, cross-compilation and a pinned Windows
inventory with verifier mutants. Hosted native and independent review remain
separate. Shared root money and reservations are not implemented here.
Refs #14025.
Every model request carries the whole conversation, so upload grows with
context size times request count and nothing in the app showed it. Each send
attempt now reports its body size, a per-process meter keeps it per provider
and day beside the usage records, and Settings > Usage shows today, the last 30
days, the average request and the providers.
Own one disabled-by-default shortcut in the local Electron profile, with immediate authorized IPC and visible registration failures. Preserve old registration and saved state when rebind or persistence fails; restore at ready and unregister during quit.
Keep browser/network viewers unsupported and in-app bindings unchanged. Cover lifecycle, authorization, failure rollback, bounded native input and settings neighbours; add a real Linux/X11 OS-key smoke and transparent platform limits.
Fixes#8290. Refs #7623; stopped-app launching is not supplied.
A session whose message payload passes the replay bound could not be opened
and the window offered no way forward. The refusal now carries a code, and
the window offers to continue from the most recent part as a new session; the
original files are only read.
The two notices a fold leaves when user turns survive only through the summary
carried no code, so a Chinese window showed the English fallback. They now
carry a code and a typed payload like the other coded notices.
Problem: background children from an older chat or Goal could charge late
usage to a newly active Goal while still retaining their original root budget.
Root cause: the usage tee selected its current recorder at event delivery,
so recorder epoch checks were applied to the wrong owner.
Fix: capture an opaque process-local accounting scope at the shared tool-call
boundary. Forward the optional capability through sink wrappers, preserve it
for nested delegation, and select the captured Goal recorder at consumption.
Keep session tally, quoting, title exclusion, provider requests and wire data
unchanged. Document between-turn accounting and persistence limitations.
Verification: real Boot delayed-background and nested fixtures, same-Goal
controls, Goal lifecycle and JSON tests, exact preceding-owner negatives,
root package inventory, owner/Boot race tests, static gates, cross-compilation
and a pinned native Windows inventory with fail-closed verifier mutants.
Hosted native execution and independent review remain separate gates.
Refs #14025.
Every inbox handler answered a store refusal with -32600 and prose, so a client
could not tell "gone" from "wrong state" from "may already have run" — and one
that retried on any error re-sent work the kernel refuses precisely because the
turn may already have run. The HTTP frontend already names 14 conditions; ACP
now answers the same names, with -32020 for "the queue refused this call" and
-32021 for "the work may already have run", and leaves -32600 to a malformed
frame alone.
docs/ACP.md states the table and each refusal's class (terminal,
retry-after-repair, do not retry). repolint's family-vs-writer check now reads
both writers, so a sentinel added to the store cannot reach ACP unnamed.
A row already records why its money may be incomplete (`cost_estimated`,
`incomplete_reason`), but RangeStats kept only a bool, so a total could say
"this folded in an estimate" and never "of these calls, N were never priced,
because …". Both read paths now count those calls and name the distinct
reasons; the projection carries the reason as schema 8, which reindexes from
the daily files; an old or reason-less row still counts, it just adds no
reason to the vocabulary.
Problem: Saved child records with missing or unknown status could reach
continuation and transcript-copy preparation.
Root cause: The shared metadata validator rejected three known bad states
but implicitly admitted every other string, including future enum values.
Fix: Permit only the existing completed/cancelled states before unchanged
identity checks. Keep metadata decoding and writers unchanged so refusal
preserves unknown fields and transcript bytes for diagnosis.
Verification: Real on-disk store/reopen matrix covers nine state shapes and
five preparation routes, artifact preservation and reference-lock release.
Exact preceding owner fails 25 named events with 31 passing controls.
Qualification results and compatibility limits are recorded in the PR body.
Problem: Explicit slash, composer and CLI profile children could exceed a
configured root token/wall budget because they bypassed the normal Agent turn.
Root cause: Those Controller entries invoked skill runners without preparing
the executor-owned scope; Goal skills installed observation without the spend
limit. Synchronous profile calls also lacked turn admission serialization.
Fix: Bind the existing Goal/root scope after blocking hooks for interactive
skills. Reuse synchronous Controller admission and root preparation for CLI
profile run/preview, retaining child pause and cancellation semantics.
Verification: Actual Boot HTTP/file-read and Controller lifecycle regressions;
exact old-owner overlay fails 13 named events with four passing controls.
Complete 239-package coverage includes the retained original Boot 15m alarm
and two disjoint remainder groups; four owner race suites and focused Boot
race pass. Root vet, full lint, check and actionlint pass. Windows/Linux
compile-only checks and the literal native gate contract pass locally.
Problem: The setting catalogue described health consent as applying on session open.
Root cause: The declaration did not match the controller's per-request global read.
Fix: Mark the guarded setting Immediate; disable and authorization changes are
checked before cached evidence is returned.
Verification: Settings package tests and the registered HTTP opt-in/disable
regression pass. No payload or request logic changed.
Problem: automatic picker probes need explicit user consent and must not
borrow credentials for project-defined destinations.
Root cause: the combined implementation used the merged configuration and
had no backend-enforced default-off setting.
Fix: use a strict user-global configuration snapshot, require the opt-in and
provider access before every cache read, refuse all redirects, and expose a
guarded settings API. Cache only fixed-size digests and status/timestamps;
never retain upstream error bodies or log credentials.
Verification: authorization/default-off/disable and malformed-config HTTP
regressions, previous-reader save overlay, desktop setting round trip,
provider-health race tests, and frontend route contracts passed. Full lint
and inventory-partitioned boot validation are tracked separately in the PR.
Refs #14617
Problem: Directory labels were inconsistent across the header and composer,
and the default green badge looked like a success status.
Root cause: Separate presentation strings described the same workspace
state, including an unnecessary badge for the ordinary directory case.
Fix: Show neutral labels only for isolated worktrees, retain explicit
current-folder/worktree choices with checkmarks, and share explanatory
tooltips while preserving full paths and branches. Update translations,
usage documentation, and the release note.
Verification: Frontend interaction, translation and style checks;
TypeScript; make check; gofmt; go vet; make lint; browser comparisons in
light/dark themes at wide and phone widths. Required Go package and
production build results are recorded in the pull request.
The boot package ran 726s on that job against 264s here, and this change does
not reach it: projectiondb.OpenReadOnly/Inspect have one production caller and
it is the doctor command. No code change.
Problem: `reasonix doctor catalogs` printed the projection's integrity_check
string as the catalog's state, so a history catalog whose history_sources row
recorded health='corrupt' printed `ok`, and a future-schema database -- which
the runtime opens as StateDegraded -- printed `ok` too. The command a user is
told to run to check their data agreed with neither the catalog nor
`catalogs reindex --json`, which reports the same state as "failed": 1.
Root cause: the state column came from Inspection.Integrity and nothing else.
The command never asked the catalog what it recorded, and never compared the
recorded schema against the schema this build migrates to.
Fix: read the facts reindex already reads. refreshStatus and refresh now share
one readStatusCounts with the new read-only Inspect readers, so the
diagnostic's failed is the number reindex reports, and the reason is the one
the catalog recorded. A newer schema reports degraded against
projectiondb.SupportedVersion, which ApplyMigrations also uses. --json gains
state, failed, pending, corruptLines/lagBytes and failures only when there is
something to say, so a healthy catalog's output is unchanged. Text stays one
line per catalog. Exit stays 0: like `catalogs reindex --json`, this command
reports rather than repairs.
Verification: red-first. A reindexed undecodable transcript printed
`history ok`; it now prints `history failed ... failed=1 source=<path>:
<recorded reason>`. A schema-2 database printed `ok`; it now prints
`degraded ... supported=1`. Controls: healthy line and exit unchanged, missing
line unchanged, healthy JSON keys unchanged. Mutation: dropping the failure
read turns the corrupt-source test red, dropping the degraded read turns the
future-schema test red. projectiondb.OpenReadOnly reads a live WAL, which
immutable=1 skipped. No projection schema, history index or reindex path
changed. Refs docs/USAGE_CATALOG.md:25, docs/HISTORY_SEARCH_CATALOG.md:42-43.
Problem: A dual-model Goal resumed after a spend-budget pause can stop its
new planner research after one tool round despite the user renewing the slice.
Root cause: ResumeGoal resets only the executor's local budget. The tool
planner keeps its own Agent accumulator and stricter local gate, so renewing
the shared executor root does not clear that old planner spend.
Fix: On a successful budget-spend resume, also reset the existing companion
planner Agent through the runner's established accessor. Preserve manual
pauses, refused resumes, cumulative Goal statistics and delivery evidence.
No new API, default, cancellation authority, prompt or persisted format.
Verification: Actual Controller/Coordinator loop reproduces two resumed
planner requests instead of four on the preceding owner. Actual Boot plus
HTTP/SSE reads three fixture files per renewed slice; the exact old owner
overlay compiles and fails both regressions with four passing controls.
Full 239-package coverage reconciles 24061 unique test/subtest passes and
215 existing skips. The original Boot command hit its 15m package alarm;
all 846 top names close via 842 original terminals and disjoint 2/2-name
remainder groups, without raising timeouts. Complete control/coordinator
race1758 passes/4 existing skips, actual Boot race1 pass, static gates and
Windows/Linux compile-only checks pass. New-head hosted/native gates remain.
Problem: The Windows fallback control missed its recorder marker and could
not remove the still-running taskkill test executable.
Root cause: exec.Command("taskkill", ...) preserves the bare command in
argv[0]. The recorder compared that argument with taskkill.exe, missed its
special entry point, and entered the complete package suite instead.
Fix: Identify the recorder with os.Executable, require LookPath to select
the fixture's same file, and fail closed on an unexpected recorder entry.
Add a bare-command alias probe whose fallback selector runs only the inert
helper. Run/Wait completes before temporary-directory cleanup. Keep the
production ownership repair unchanged and extend the native inventory to
seven pass events versus four expected failures and three passing controls.
Verification: Windows cross-compile, Windows vet and pinned Windows proc
lint pass; full root vet/lint and make check pass at unchanged baseline.
All 6268 selected Darwin inputs match the previous qualified commit; this
append changes only the Windows fixture, verifier inventory and docs.
The complete original Windows failure log is preserved. Fresh native
Windows execution on this new commit remains required, not inferred.
Problem: Repeated cancellation or shell reaping can discard an established
Windows Job Object owner and perform a new PID-based stop attempt.
Root cause: TrackedCommand removes its one-shot TrackedJob on the first kill
and also starts a redundant PID tree observer after successful job assignment.
Fix: Keep the published owner across repeated and late cancellation, and let
an established Job Object own its tree. Preserve existing non-job fallbacks.
Add native Windows owned-child probes with a harmless taskkill recorder and
an exact preceding-owner counterfactual inventory gate.
Verification: Root inventory reconciles 239 packages, 24060 test/subtest
passes and 215 existing skips. Original Boot command hit its 15m package
alarm; all 845 top-level tests close through the recorded disjoint remainder
groups without a timeout increase. Complete proc and shell race suites pass.
Root vet, full lint, make check and actionlint pass. Windows new/old owners
compile and Windows vet passes; actual native regressions remain a hosted gate.
No full PID-identity, descendant-quiescence or physical-scope claim.
Problem: semantic routing runs before Agent/Coordinator root preparation, so its reported usage can permit an extra executor tool round.
Root cause: Controller binds no root before routing, and SemanticRouter does not report finalized usage to the accepted context observer.
Fix: share executor root preparation after admission/hooks and before routing in headless and interactive turns. Observe semantic terminal usage once, independent of audit/sink and answer success. Preserve cache hits, new-task/scope/reset lifetimes, normal accounting and cancellation authorities.
Verification: disjoint required/rest commands pass all 239 root packages; four complete owner race suites pass 3939 events with 14 existing skips; actual Boot race passes 12 events. Exact preceding owners compile and fail 15 regression events with six passing controls. Format, vet, full lint and make check pass at byte-identical baseline. Six Windows/Linux binaries compile; local literal Windows gate verifies 21 required events/two package terminals and rejects skip/missing-terminal controls, not native execution.
Problem: ambient Git index/object variables can redirect host worktree preparation and object access.
Root cause: pinned Git/common/worktree directories do not override GIT_INDEX_FILE, GIT_OBJECT_DIRECTORY or GIT_ALTERNATE_OBJECT_DIRECTORIES. The shared constructors retained them.
Fix: remove those three ambient overrides in Env and Detached. Keep deliberate host private-index/object overrides and repository-declared alternates. Cover actual delivery/candidate preparation and a Boot isolated-write/apply flow.
Verification: all 239 root packages pass across disjoint required/rest commands; complete gitcmd/worktree/isolation race 306 passes; Boot race four passes. Exact preceding owner compiles and fails 18 regression events while seven controls pass. Full lint, vet and make check pass with byte-identical baseline. Windows/Linux builds compile; local literal Windows inventory verifies 25 required events and three package terminals, not native execution.
Problem: configured tasks used independent planner and executor root
counters, allowing an extra executor tool round after planning spend.
Root cause: Agent root binding began only inside each role's Run. A
plan-only turn never opened executor Run and could lose its accounting.
Fix: prepare the executor-owned root before routing, bind companion Agent
runs to it, and account legacy raw-planner usage at its logical terminal.
Track preparation by session object and host task scope. Adopt only the
preparing executor's binding after its own evidence reset; preserve clock,
immutable limit snapshots and captured old-job counters across continuation.
Keep local per-Agent gates/telemetry and all approval/cancellation boundaries.
Verification: required2 plus disjoint rest237 cover239 packages exactly
once,23980 test/subtest passes215 existing skips. Four full owner race
suites2786pass10skip; actualBoot shared and delegated token/wall16racepass.
Precise preceding Coordinator overlay compiles9fail4controls/2pkgfail.
LiteralWindows20named events/3pkg passes locally;6Win/Linux builds compile
only. Rootvet/full lint/makecheck/gofmt/diff pass,2663 baseline unchanged.
Refs #14025
Problem: Anthropic error, EOF and cancellation exits could discard usage
already reported by message_start or message_delta. Bounded consumers lost
the counts, while normal recovery reconstructed counts it already knew.
Root cause: the cumulative usage emitter was reachable only after
message_stop; consumers may stop immediately upon receiving ChunkError.
Fix: give each response one cumulative accumulator and delivery attempt.
Flush partial usage before explicit errors with existing Estimated identity,
retain cancellation-aware best-effort delivery on other exits, and compute
the original error classification before accounting changes sink state.
Keep exact committed usage, request construction and retry policy unchanged.
Verification: required builtin/boot plus disjoint rest cover all239 root
packages once,23966 test/subtest passes215 existing skips. Six full owner
race suites2615pass10skip; actualBoot4race events0skip. Exact preceding
decoder compiles and fails14events with4controls passing. Native-step
contract18events/2pkg passes locally;4Win/Linux binaries compile only.
Rootvet/full lint/makecheck/gofmt/diff pass; baseline2663 byte-identical.
Refs #14025
The prefix is read off the entry's own model list up to its last slash and left
out of how models are drawn, in every listing; ids and refs never change, an
alias wins, and models that would read the same keep their prefix. Fixes#14573.
Problem: an explicitly bounded Agent root omitted known native search usage,
allowing an extra normal tool round after a reported search answer.
Root cause: the search terminal finalized its independent request count and
reported usage without calling the accepted optional root observer.
Fix: observe the finalized record once before optional reporting. Preserve
query isolation, request counts, callback copies, unknown usage, native
completion, paid-backend consent and zero-limit behavior.
Verification: all 239 root packages pass through disjoint commands; four
complete owner race suites and real boot OpenAI/Anthropic HTTP regressions
pass. The compiled preceding owner fails seven events while nine controls
pass. Static gates, six cross-compiles and the literal 16-event Windows
contract pass locally; own native execution remains a hosted gate.
The sheet now covers the title bar too, its header standing in for the bar
(drag region, macOS light clearance, Windows buttons); the enlarge button's
arrows turn inward while enlarged. Refs #14418.
When a turn that ran a task from the session that dispatched this one ends,
its outcome and final answer go back as a message to that session. Only the
opener is answered, so a reply never starts a reply. Refs #13341.
A newly held message stamps the receiver through the existing unread fact and
announces itself with a quiet operator notice, which makes the window read the
session list again. Refs #13341.
Problem: a Windows run rejected three correctly refused capability calls
because the child conversation took 2.138s against a 2s wall-time check.
Root cause: the metric spans provider rounds and persistence, so it does
not establish whether target resolution refused before execution.
Fix: assert error/refusal guidance, Executed=false and no shell receipt for
all three results. Remove unused timestamps and preserve the existing 60s
completion guard. Production behavior and permission policy are unchanged.
Verification: the single root command passes all 239 packages, selected
boundary race checks pass 17 events, and a compiled executed-receipt test
mutant is rejected. Formatting, vet, lint and repository checks pass;
Windows/Linux boot binaries compile, without a native-execution claim.
Problem: an explicitly bounded root omitted billed command classifications
and external-content screening, allowing extra normal tool rounds.
Root cause: the shared triage terminal emitted usage without charging the
existing context-bound root counter.
Fix: observe terminal usage through the accepted private root helper before
the optional display guard. Preserve verdicts, screening notices, event
conditions, missing usage, normal telemetry and cancellation authority.
Verification: all 239 root packages pass in disjoint commands; full Agent
race and real boot HTTP/SSE matrices pass. The compiled preceding owner
fails ten events while five controls pass. Static gates, four platform
cross-compiles and the literal 15-event Windows contract pass locally.
Problem: an explicitly bounded Agent root omitted reviewer/advisor usage
already emitted by the shared bounded-call terminal, allowing extra rounds.
Root cause: normal-round accounting does not own auxiliary request usage.
Fix: connect the bounded terminal to one copied, replaceable context
observer bound to the original root counter. Preserve job cancellation,
normal telemetry, missing usage and zero-limit behavior.
Verification: all 239 root packages pass in disjoint commands; nine complete
owner/caller race suites and real boot HTTP/SSE regressions pass. Exact old
owners reproduce the gap; selected accepted auxiliary composition passes.
Formatting, vet, lint, repository checks and eight cross-compiles pass.
The 20-event Windows contract is locally verified, not native execution.
A local pane rides the hub's multiplexed stream, which attaches it only
once that stream is open. Until then the subscription has no position,
and the attach's watermark became its baseline as if nothing before it
mattered. A pane that read its transcript mid-turn and was attached after
the turn ended never received the reply or turn_done, and went on showing
the person's message alone.
The subscription now reads where the stream stands when it subscribes,
with the attach read of /events/replay. An attach that finds the stream
has moved on since reports what went by as a gap, and the pane rebuilds
from /history as it does for any other hole.
open_projects raises a fresh card naming every folder; approving adds the
folders to the project list and trusts them, vetted per folder like an extra
workspace root. Off unless the user's dispatch switch is on. Refs #13341.
Fold summary and failure-diagnosis terminal usage into the existing
context-bound root scope, including billed failures. Preserve local
normal-round telemetry and manual maintenance outside a root context.
Presets declare whether they relay other makers' models; the model menu
heading and the settings service list draw the service mark, while a relay's
model rows fall back to the letter chip. Refs #14538.
Pass the existing TaskBudget conversion to planner Agent Options. Cover
real HTTP planning reads, context overrides, disabled and unknown-cost
cases, executor fallback and request-prefix stability.
both ignored everything after the subcommand (or the name), so a typo like --jsno printed the human output and exited 0. now they exit 2 like mcp browse does.
The kernel answers POST /agent-presets/default with 204 and no body, but the
port read the answer as JSON, so a change that had landed reported
'Unexpected end of JSON input'. The call is status-only and uses post().
The plan bar sat inside the composer box and the open list had no obvious way
back out. It is now a card floating just above the composer; the open list has
a header (running and pending counts) and a collapse button at its top right,
and Esc or a press outside closes it through the shared dismiss path. The jump
buttons move up so the card does not cover them, and the conversation gets a
little bottom room while a plan exists.
Problem: restart-style MCP tests keep prior controllers alive while changing
server definitions or constructing a later session. Their background handshakes
and schema writes are therefore outside the test's advertised phase boundary.
The macOS job on #14531 observed a missing followed-tools notice, although its
precise interleaving was not captured and local old/new stress both pass.
Fix: close the owning controller before each transition in follow/unfollow and
accepted-tool restart fixtures. Existing idempotent Close drains the fixture's
MCP host and queued schema writes. Retain every notice, held-tool, provider
surface, actual-call and refusal assertion without retry, sleep or weaker gate.
Verification: six MCP effect test/subtest events and four existing deterministic
Close barrier tests pass under race. Forty fresh compiled processes pass each
before and after; they do not reproduce the hosted nil notice. Full lint, vet,
make check and four Windows/Linux amd64 builds pass. Full builtin/boot plus
disjoint root inventory continue separately; no production source changes.
Refs #14025
Refs #14133
Refs #14531
Problem: broad validation found three adjacent fixtures requiring selected
cost aliases or complete quotes even when their usage was reconstructed.
The shared quote repair correctly takes the existing incomplete-display path.
Fix: assert preserved original amounts, incomplete/partial coverage and absent
selected aliases for estimated wire, ACP status and persisted stats rows. Keep
reported-usage alias and legacy-row controls. Expand the existing Windows
inventory to all eighteen named events across seven packages.
Verification: affected consumer tests pass, with complete eventwire/ACP/stats
race suites at 351 passing events and no skips. The compiled preceding pricing
owner now fails fifteen regression events with three controls passing. Final
local Windows-step contract passes eighteen events; required-skip and missing
package-terminal controls fail. Full lint, vet and make check remain clean;
final original root validation continues against unchanged frozen Go sources.
Refs #14025
Refs #14133
Interrupted provider streams can reconstruct positive usage without terminal
reported counts. The shared quote owner forwarded that provenance but ignored
it when certifying complete cost facts, allowing recovered usage to make a
fully covered ledger.
Use the existing incomplete-usage result while retaining fixed-point amounts
and occurrence-time valuations. Reported usage, including rate-card estimates,
keeps its completeness. Existing quote fields, reasons and ledger versions stay
unchanged; no provider-visible request, permission or budget default changes.
Cover shared event/config adapters, persisted ledger coverage and real boot SSE
recovery with an actual file read. Add a Windows fourteen-event inventory in the
existing job. This is a monetary-fact prerequisite of the full concurrency plan,
not shared root-money admission or hard reservation.
Verification: fourteen focused events and the real boot race pass; compiled
preceding pricing fails twelve regression events with two controls passing.
The prior reader accepts the incomplete quote fixture. Root vet, full lint and
make check pass with the baseline unchanged. Eight Windows/Linux amd64 owner
builds compile; full root inventory and complete owner race continue separately.
Refs #14025
Refs #14133
Problem: a host wall budget could land the root Agent while delegated Agents
and Manager-owned background jobs kept independent start times or no wall bound.
Root cause: the opaque inherited root scope carried only token accounting; the
existing task time gate read each Agent's own clock after its tool rounds.
Fix: attach an immutable root task start/limit pair to that existing scope.
Preserve the start across continuations and snapshot new root overrides without
rewriting old job bounds. New tasks/reset use a new clock. The existing task and
fleet projection helper also carries this sidecar while keeping Manager-owned
cancellation/deadlines. Use existing time soft landing and finalization; keep
defaults unlimited, with no hard deadline, new timer, schema or permission.
Verification: frozen6730Go/fourmodules; required2+rest237 cover239packages,
23754test/subtestpasses/215existingskips. Complete Agent/delegation/jobs race
2652passes/10existing skips; actual boot background task/capability/fleet4race
passes; accepted14523sixGo selectedoverlay54racepasses; prefixguards4passes.
Exact preceding owner compiles and fails9events/2controls across2packages.
Windows17-event step passes locally and rejects skip/missingterminal controls;
six Windows/Linuxamd64 test binaries compile, not native execution.
Rootvet/full lint/makecheck/format/diff pass; baseline byte-identical2663.
Refs #14025
Refs #14133
Problem: a priced Agent could report complete task costs after a round omitted
usage, and estimated usage could certify a cost threshold.
Root cause: the shared HTTP attempt helper synthesized request-only usage
without its existing Estimated identity; the budget observer also treated
missing or estimated token accounting as a fully priced reading.
Fix: preserve exact HTTP counts while marking synthesized usage incomplete.
Keep only reported priced costs in the Agent subtotal, retain uncertainty for
the task lifetime, and land when the known subtotal alone reaches the limit.
Keep defaults, schemas, cancellation and finalization unchanged. Add real
OpenAI HTTP boot coverage and an explicit Windows test inventory.
Verification: frozen 6729 Go files/four module files; required2 +rest237 close
239 packages, 23754 test/subtest passes and215 existing skips. Complete five
owner race2407passes10existing skips; real HTTP boot race4passes; cache guards4.
Exact preceding owners compile with13failing events/5controls; producer-only
rollback4fails/1control; deliberate unknown-erases-known mutation1fail.
Root vet, full lint, make check and formatting pass with baseline unchanged;
six Windows/Linux amd64 owner/boot binaries compile (not native execution).
Refs #14025
Refs #14133
Problem: configured task token limits counted only each Agent's own provider
usage. Delegated Agents could continue spending outside the parent's reading.
Root cause: local runBudget instances never shared a root token total, while
background task/fleet closures start from Manager-owned contexts that omit
caller values.
Fix: bind a private atomic root counter to the existing task ledger and pass
its immutable per-run threshold through derived contexts. Project only this
scope into both background producers, preserving job cancellation authority.
Keep local telemetry, cost/wall gates, defaults, prefixes and soft landing.
Verification: compiling preceding owners9fail/2controls; preceding job-context
only4fail/2foregroundcontrols. Complete3owner race2639pass/10existing skip;
realboot5scenes+parent6racepass, explicit cacheguards4pass. Final239 package
inventory23710pass/215existing skip, required builtin/boot, rootvet, full lint,
makecheck, formatting and6Windows/Linux builds pass. New native inventory
requires20namedpasses/no skips; local jq and required-skip negative verified.
No hard reservation, root money aggregation, process kill, remote/durable
accounting or Coordinator cross-root scope is claimed. Refs #14025, #14133.
Cause: internal/runtime/isolation and the isolated task tool refuse with
isolation.unknown_id, unavailable, not_git, apply_conflict and
unsupported_arg, but the kernel catalogue words none of them. A tool card
renders say(code), so these refusals showed no reason in either language.
Fix: word the five codes in kernel.ts and en_kernel.ts, in the terms of the
tool path that emits them.
Refs #14025
Problem: whole-file rewrites can discard changes hidden by a later overlay
version, a paged read, or an oversized read that clears the previous guard.
Root cause: read_file renders one read and records another. Missing and
incomplete observations also collapse into the permissive never-seen path.
Fix: observe the renderer's decoded stream with bounded capture, refresh only
complete returned reads, retain older guards, and track first partial reads
explicitly. Keep never-seen compatibility and disabled protection behavior.
Verification: compiling old-owner controls fail five behavior regressions;
fixed focused race tests and real boot.Build checks pass. Full builtin,
root-package tests, go vet, make check and full lint are recorded in the PR.
Written bilingual security bounds and release notes accompany this slice.
A first partial read refuses replacement with the file's line count and the
read_file limit that certifies it. Without that, a default-limited read of a
file longer than 2000 lines can never be certified by rereading as told.
Refs #14025
Refs #14133
Problem: the no-index history search test could inherit a process catalog from
a temporary runtime fixture, so its asserted precondition was not established.
Root cause: shared catalogs deliberately outlive controllers. Several actual
boot fixtures and the shared unread harness omitted their existing fixture
cleanup, allowing the process pointer to outlive the temporary home.
Fix: use the existing shared cleanup in those owners, establish no-index state
explicitly, and preserve coverage for real active-catalog failures. Production
code, workflow gates, timeouts, defaults and baseline remain unchanged.
Verification: preceding fixtures compile and fail the lifecycle regression;
the provider-editor sibling control passes. Complete serve1798pass/5existing
skip and complete disjoint race inventory1798pass/5skip; root239 package
inventory23696pass/215existing skip, required builtin/boot, rootvet, full lint,
makecheck, formatting and Windows/Linux amd64 compile pass. Original Linux500
and the local fullrace5m alarm remain separate evidence, not claimed fixed by
an external rerun. Refs #14025, #14133; followup to #14505.
A mark that kept focus after a press set the rail's focus again whenever the
window handed focus back, so its neighbours stayed magnified and its peek stayed
open with the pointer elsewhere. Focus now moves the rail only when it is
focus-visible.
Foreground panic unwinding could publish a completed worker graph before its error was known, while legacy nil panics were swallowed by both foreground and background exits.
Classify return versus panic at one private delegation boundary, persist the failed transcript while owned, and release before terminal callbacks without letting old cleanup retire a reentrant replacement. Keep foreground panic propagation and accepted Manager finalizers. Cover real boot direct/capability dispatch, compiled nil-panic modes, error identity and ordinary outcome controls.
Settings and feedback refused with 'add a folder first' when no pane was open and
no folder was known, and the empty main area offered only Add folder. The kernel
already opens a session with no root in the app's own workspace, so the sheets
and the empty state now ask for exactly that.
Problem: legacy nil-panic recovery can publish a successful job after its
invocation or cleanup panics, and error-valued panics lose their cause identity.
Root cause: two boundaries treat a non-nil recover value as the only panic
signal and format panic errors without wrapping the original error.
Fix: share a private normal-return guard across invocation and finalization,
retain typed panic causes, preserve finalizer error classification and sibling
cleanup, and keep the existing invocation stack message without adding stacks
to finalizers. Move invocation recovery into its cohesive owning file.
Verification: compiling preceding owners fail six test/subtest events with two
controls passing. Actual compiled Manager child tests cover explicit panicnil
modes zero/one, invocation, cleanup and no-launch. Frozen required2/rest237
partitions close all239 packages with23656 passes/215existing skips; complete
jobs/delegation race549 passes and accepted14464 three-owner composition671
passes have zero skips. Independent required command has2349passes/66existing skips.
Vet, complete lint, make check and formatting pass at unchanged2664baseline.
Windows/Linuxamd64 owner builds are compilation only; exact-head hosted/native
and independent review remain separate gates. No Goexit, other-goroutine panic,
process-scope completion or crash recovery is claimed.
Problem: accepted job finalization in #14469 conflicts with the pending
cancellable writer registration paths in task and fleet.
Root cause: both sides own background cleanup. Closure-only cleanup misses
proven no-launch, while ID-only cleanup cannot protect replacement writers.
Fix: retain Manager-owned terminal callbacks and pass each returned writer
registration token through the shared jobCleanup option. Remove the obsolete
ID-only helper. Preserve the foreground token lifetime, accepted transcript
and progress finalizers, and every other upstream change in this ordinary merge.
Verification: the original full 239-package root command passes with 23668
test/subtest passes and 215 existing skips. Complete checkpoint/delegation/jobs
race passes 663 events, zero skips (one unchanged package cached); real boot
handoff and no-launch race passes ten events. Independent required builtin/boot
passes 2354 events with 66 existing skips. Compiling preceding callers fail
four no-launch scenarios plus parent; the Manager/token generation regression
passes. Vet, full lint, make check, formatting and diff gates pass, baseline
2664 unchanged. Six Windows/Linux amd64 owner compilations are compilation
only. Exact new-head hosted/native validation and independent review remain.
The reader menu opens from a promise callback, so its commit runs at
default priority and the effect that focuses the first item is a
separate Scheduler task (setImmediate). findByRole resolves from the
MutationObserver that sees the menu, then RTL drains with setTimeout(0);
when a millisecond passes before the next loop turn, the timer wins and
the assertion reads the reader, which still has focus.
Wait for the focus the way the other focus assertions in the suite do.
Problem: candidate groups need a common baseline while preserving the result
Store's source, managed-root and identity ownership.
Root cause: ordinary Begin captures afresh, while arbitrary snapshot/path
adoption would bypass the owning Store and its existing settlement contract.
Fix: add an opaque process-local Preparation, recheck source/workspace/Git
directory identities, and create independent results from the captured
snapshot. Extract the existing Begin bodies into private shared helpers.
No best-of, runtime, default or persisted-format activation is included.
Verification: complete 239-package root inventory, full isolation race,
independent builtin/boot command, vet, full lint and make check pass with the
2664 baseline unchanged. Compiling unsafe controls distinguish baseline
refresh and replaced-directory admission; the preceding Store implementation
passes the new preparation tests. Linked-source overlay passes; Windows and
Linux amd64 compile only. Native/hosted validation and independent review
remain separate gates.
Problem: local settings cannot load the file-backup catalog because both
machine settings URL forms return an untyped 404.
Root cause: file-backup handlers are registered only with account backup
routes, while the machine settings router registers a separate route set.
Fix: register the existing file-backup handlers with machine settings.
Normal runtime routes still register that same set once. Preserve host-only
checks, paired-device rejection, local-owner selection and mutation gates;
do not change payloads, encryption or persisted formats.
Verification: new regressions fail before the fix with plain 404 responses
and pass afterward for both machine prefixes. Focused machine settings,
file backup, account backup and mutation-gate tests pass. gofmt, go vet,
make lint, make check and the required builtin/boot tests pass.
Problem: switching settings categories retains an unrelated scroll offset, while a searched setting that loads later never receives navigation feedback.
Root cause: categories share one scrolling element and landing resolves the anchor once before asynchronous content arrives.
Fix: key the panel by category, clear pending landing only when a guarded category change succeeds, and observe target arrival until its feedback lifetime begins. Keep existing categories, layouts, disclosures and unsaved-change guards.
Verification: five regressions fail on the unmodified baseline; all 28 settings test files and 215 tests pass with the fix. TypeScript/Vite and repository checks pass. Browser verification confirms a tools panel at scrollTop 5747 switches to network at 0 and search opens the existing advanced contrast disclosure.
Problem: Network settings could be saved without edits, and pending reads,
saves or diagnostics could overwrite the state of a newly selected connection.
Connection tests also appeared to exercise unsaved form values.
Root cause: The form used one settings object for both saved state and edits,
with async completions that were not scoped to the owning connection.
Fix: Keep saved settings and the editable draft separately, disable no-op
saves, freeze submitted drafts, and identify diagnostics as testing saved
settings. Remount the form on connection changes and ignore completions from
unmounted owners. Preserve the existing expanded connection details layout.
Verification: The seven extracted regression cases failed on the baseline.
Network and i18n tests pass (10 files, 181 tests), including edit reversion
and loading/error coverage; tsc -b, make check and git diff --check pass.
Problem: the ordinary test synchronization append conflicts with upstream's
independently accepted fix for the same passive draft-persistence boundary.
Root cause: both sides replace the immediate heldDraft assertion with waitFor;
upstream additionally explains why the effect may follow the thumbnail commit.
Fix: resolve the sole conflict to the exact accepted upstream test including
its comment. Preserve all incoming dependency/UI changes and every original
runner-panic implementation Go file. This merge is required by a verified
content conflict, not unrelated base drift; no rebase or force-push.
Verification: final inventory of 239 root packages is reconciled from disjoint
required2 and rest237 commands, both exit0, with frozen Go/module inputs.
Isolation race covers all 51 top-level tests through the original completed
39 trees plus the remaining12 under the same5m budget; original full owner
command hit its cumulative alarm and is retained as exit1, not called pass.
Five affected frontend files have99 tests pass; production typecheck/build,
formatting, root vet, full lint and make check pass without baseline widening.
Exact merge-head hosted/native checks and independent review remain pending.
NewStreamDialer is how SSH reaches a remote host under the user's proxy
(remote/attach), and none of dialer.go was exercised: netclient sat at
47% against a 47.2% floor on a declared-sensitive path.
Four tests drive it against fake proxies on loopback:
- an http proxy gets CONNECT to the target with Proxy-Authorization, and
the credentials never also travel as Authorization;
- a proxy that answers 407 yields an error, not an open stream;
- an https proxy sees a TLS handshake first, so CONNECT and credentials
never cross in cleartext;
- a SOCKS5 proxy gets username/password auth and the target by name.
Each fails when the matching guard in dialer.go is removed. The
netclient floor rises from 47.2 to 65.8.
Problem: Studio job 121205496051 fails after the pasted image preview appears,
because the retained draft still contains no image.
Root cause: findByAltText observes the DOM commit, while holdDraft runs in a
passive effect. The test assumes those distinct boundaries complete together.
The unchanged merge passed a later external rerun, which does not repair the
scheduling assumption.
Fix: keep the exact single-image assertion for both heading and tab targets,
but await the draft condition with waitFor. Production UI and Go source stay
unchanged.
Verification: full Feedback interaction file 65 pass, restored focused cases
2 pass, typecheck pass. Two independent no-op-save controls each fail both
targets, proving persistence is still checked. Root vet, full lint, make check
and required builtin/boot pass at unchanged baseline 2664. The preceding
native/hosted evidence remains attached to af958f; the new head needs fresh CI.
Problem: ambiguous private registry records can be admitted or scavenged as
known dead-owner state.
Root cause: struct JSON decoding merges repeated members and folds case;
unknown-field rejection alone does not preserve the original declarations.
Fix: validate unique decoded names and exact schema spelling for records and
capacity objects before typed decoding and liveness checks. Keep writer
encoding, legal spacing/order, escaped unique names and existing resources.
Join a killed process before closing cooperative input in death fixtures.
Verification: complete 239-package root inventory and required builtin/boot
commands exit 0; mutation race 88 pass, pending wait/opaque composition 116
pass. The compiling preceding decoder fails 22 behavior test/subtest events,
with two controls passing. Formatting, vet, full lint and make check pass
without changing the 2664 baseline. Windows/Linux amd64 compile only; decode
microbenchmark overhead is measured and disclosed separately from product
performance. No runtime or default activation.
`.railhost` sets pointer-events: none, and only `.srail` opted back in. The
unloaded-history marker and the opened run inherited the none, so presses fell
through to the transcript. A style test now holds the rule.
Problem: an isolated task could write files, then lose its discoverable result
identity when a host runner or provider callback panicked.
Root cause: Store.Execute called Finish only after an ordinary runner return;
panic unwind bypassed both completed-result and inspection publication.
Fix: recover at the runner invocation boundary, retain an error-valued panic
as a cause, and publish the existing inspection identity under the result
operation owner. Do not infer physical completion, inspect for emptiness, or
delete work on a panic. Preserve ordinary outcomes and truthful live-only
recovery when bookkeeping cannot be admitted or persisted.
Verification: final 239-package inventory closes using 238 unchanged broad
packages and the final isolation race package; 111 owner race and 3 real boot
race cases pass without skips. The compiling preceding owner fails 11 cases
with 3 ordinary-error controls passing. Required builtin/boot, root vet, full
lint and make check pass at unchanged baseline 2664. Linux/Windows amd64 are
compile-only; hosted/native and independent review remain separate gates.
Refs #14025
Refs #14133
Problem: Windows PR smoke did not run jobs/delegation unit tests, and pure
delegation changes could skip native platform legs. Linux PR race coverage
also omitted the delegation owner tests.
Root cause: package routing predates the shared background resource lifecycle.
Compiling these owners as boot dependencies does not execute their tests.
Fix: run both complete packages on Windows with a retained JSON inventory,
require all15 portable lifecycle cases to pass, route pure delegation changes
to native legs and include delegation in the existing Linux race sweep.
Verification: platform-scope tests and pinned actionlint pass; the inventory
gate accepts actual local owner output and rejects missing-case, skipped-case
and missing-owner controls. Make check passes at unchanged baseline2664.
All6699 Go files match the qualified implementation commit; no product Go
source or deadline changed. Exact-head native/hosted qualification is pending.
Problem: background task/fleet can strand checkpoint writers, transcript locks
and progress resources when the Manager never invokes the supplied closure.
Root cause: resource ownership moves to closure-only defers before the Manager
checks invalid identifiers or its closed root context. Those branches bypass
the closure and its cleanup, while the caller suppresses its own ending.
Fix: register optional host finalizers at the Manager's managed-execution
boundary. Run all callbacks once before completion after actual closure return
or proven no-launch, preserve cleanup failures/panic and transfer the actual
task/fleet resources and uninvoked terminal projections to that owner.
Verification: final239-package inventory uses238 broad packages plus the final
jobs package after a deterministic liveness-test-only change:23596 test/subtest
passes/215 existing skips. Full owner race648 and real boot race5 pass without
skips; required builtin/boot2346 passes/66 skips. Original public regressions
and compiling predecessor controls fail. Selected token-adapter composition
passes40 race cases. Vet/lint/check pass at unchanged baseline2664.
The liveness pair uses a controlled clock; production thresholds are unchanged.
Managed-closure completion is not descendant scope or durable crash recovery.
Refs #14025
Problem: cancelled background task/fleet calls can remain behind checkpoint
admission while the shared writer registry also blocks queries and cleanup.
Root cause: registration holds the registry mutex across the barrier, durable
undo retirement, notification and Store persistence. Callers discard their
admission context and defer cleanup by reusable string ID alone.
Fix: reserve identified pending registrations, prepare outside the registry
mutex and serialize latest snapshot publication separately. Pass context at
all task/fleet sites and transfer token-bound cleanup to admitted jobs. Preserve
legacy APIs, completed undo retirement and active work across parent turns.
Verification: original root239 passes with23590 test/subtest passes/215 existing
conditional skips; required builtin/boot2346 passes/66 skips. Complete owning
race487 passes/zero skips; real boot handoff race5 passes/zero skips. Compiling
preceding-owner/caller controls fail the intended assertions. Vet/lint/check
pass without raising baseline2664. Linux/Windows compilation is not native QA.
The jobs-owner no-launch finalization gap remains explicitly separate.
Refs #14025
Problem: A foreground command can return bounded cancellation while cmd.Wait
is still blocked, so call return cannot establish resource-lifetime completion.
Root cause: The tracked wait owner exposes its logical result and cleanup
diagnostics but no separate observation of the still-running root Wait.
Fix: Register a stable root-wait channel before the waiter starts and close it
after Wait returns, before result publication. Expose it with the explicit
RootWaitCompleted name. Preserve existing execution, cancellation, retry and
error behavior. No production lease consumer or whole-scope claim is added.
Verification: Deterministic cancellation/retry/publication tests, portable Go
child execution and a real blocked-output process/lease regression pass. Root
Wait remains distinct from descendant or external-scope completion; nil means
no registered observation. Full local/static and platform evidence is recorded
in the PR without inheriting previous heads' native results.
Problem: Directory-scoped coordination cannot exclude writers that address a
source file through a hardlink outside that directory.
Root cause: Path containment and named inode identity do not represent an
unresolved write effect; reinterpreting old directory records would silently
change their resource promise and nested admission could reuse a weaker hold.
Fix: Publish explicit version-3 opaque effects in the existing registry,
coordinate them symmetrically against all path-bearing transactions, and
require effect coverage when reusing a parent lease. Preserve ordinary named
parallelism, model capacity, previous directory semantics and final OS-locked
grant authority. No runtime or default requests the new effect here.
Verification: Full 239-package root inventory and independent builtin/boot
commands exit 0; mutation race 82 passes/zero skips; preceding-reader binary
preserves live/dead new records; weakened-owner control fails eight cases;
selected wait-snapshot composition has 36 race passes/zero skips. Root vet,
full lint, changed-file checks and formatting pass at unchanged baseline 2664.
Windows/Linux amd64 compile only; local execution is macOS.
Problem: unfamiliar coordination records can be ignored during admission and
deleted after their OS owner exits.
Root cause: registry decoding discarded unknown fields, and version-1 resource
validation accepted every effect except capacity.
Fix: reject unfamiliar object fields and unsupported path effects before
liveness/scavenging, retain original bytes, and include record identity in the
typed protocol error. Preserve known version-1/2 records and omitted-effect
legacy semantics without changing writers or activating the runtime protocol.
Verification: complete 239-package root inventory, full mutation race,
compiling preceding-owner counterexamples, actual owner-process death,
known-record controls, selected waiting-observation composition, required
builtin/boot and static gates. Windows/Linux builds are compile-only local
evidence; exact-head hosted/native review remains separate.
Problem: A durable future settlement record must stop ordinary result
writers, including stale queued operations and destructive siblings.
Root cause: The per-ID mutex serializes operations but carries no durable
knowledge that a recovery owner has published a settlement fence.
Fix: Check host-rooted external fence metadata after operation ownership,
refuse any existing entry without interpreting its bytes, recheck context
and release only this attempt on refusal. Cover tools, Finish/Abort, empty
inspection, sweep, a real child process and actual boot/controller rebuild.
No journal writer, schema or complete recovery is introduced. Preceding
writers ignore this fence; future activation needs a cooperating protocol.
Verification: Full isolation race100 passes, zero skips; real boot race3
passes. Preceding owner compiles and fails15 test/subtests, including a
child process; predecessor boot owner actually writes source in both cases.
Final root inventory/required and static qualification recorded in PR.
Linux/Windows amd64 compilation is separate from native evidence.
Refs #14025
Refs #14133
User turns, retained failures and [[keep]] turns were budgeted as shares of
the window. On a 1M window under the 100K cost trigger those shares plus the
tail passed the trigger, so every fold was refused as still at or above it.
Each budget is now also held to a share of the trigger.
The folder and host labels on a pane tab were flex: none, so a long one
could not shrink: the title, the only part that could, went to zero width
and the close button was pushed past the tab's edge onto the next tab,
where a press activates that tab instead.
Both labels now shrink to an ellipsis and take at most 40% of the tab, and
the title keeps at least 4em. perf/tab-close-inside.mjs checks, in the
real stylesheet, that the close button stays inside its tab and is what a
press at its centre reaches, with a long title, a long folder, and both a
long folder and a long host.
Only a serve with --port-file or --open listened first; a plain serve printed the
banner, share link, token path and the non-loopback warning and bound afterwards, so
an address that can never work still looked like a running server. Every command now
binds first, the report uses the bound address, and the now-unreachable
bind-inside-RunGraceful branch is gone.
.prefs-col had a max-width and no inline margin, so once the settings window was
enlarged the column hugged the navigation and every extra pixel piled up on its
right. In the enlarged mode the column is now centred in the room beside the nav.
Problem: a candidate filename could act as a Git wildcard or magic pathspec,
writing an undeclared source file or dropping the actual literal result.
Newline-delimited hashing could also reinterpret a single filename.
Root cause: option separation was mistaken for literal path selection, and
ambient Git pathspec flags could override the host's intended selection.
Fix: centralize literal pathspecs for tree/index inspection and the shared
restore owner; C-quote hashing input by bytes; remove inherited selection
flags in the common host Git environment while preserving explicit flags.
Whole-workspace application reuses the shared writer. Keep the portable
stdout/stderr Git fixture repair aligned with the native CI follow-up.
Verification: complete 239-package root inventory, three-owner race, real boot
race, uncached required builtin/boot, compiling predecessor controls, six
Windows/Linux owner cross-compiles, formatting, root vet, lint and make check.
Native hosted checks and independent review remain separate gates.
Refs #14025
Refs #14133
withinExact used filepath.Rel, which ignores case on Windows, so a different
spelling of an absent file was reported as held there while heldWithin's comment
and TestHeldPathsCoverOnlyTheSameFileAcrossCaseSpellings require the filesystem to
prove two spellings are one file. Require the prefix bytes to agree as well.
The terminal's copy of /history's tool call had no argumentNormalization
field, so a restored transcript rebuilt the card from Arguments alone —
the proxy's effective text — and drew it as what the model wrote. The
original was not merely hidden; it never left the decoder.
Carry the host's normalization through HistoryToolCall into the
eventwire.Tool the restored transcript draws and inspects, following the
serve payload and the desktop copy that already do. The card still draws
only the effective arguments: showing the original is a UI decision for a
later change, and changing the card would also change a live call.
Register the pair as a declared subset of the serve historyToolCall in
wireparity, naming the six fields the terminal decides from. The exact
table is the desktop's and is read as a TypeScript interface, so a Go
reader belongs in the partial table, which compares both ways over the
declared names.
The helper used CombinedOutput, so on Windows, where core.autocrlf is true, git's
'LF will be replaced by CRLF' warning on stderr was returned with the blob id and
TestApplyPathsWithUntrustedModesPreservesTheIndex/candidate_type passed the pair to
update-index --cacheinfo. Standard output is what a caller reads; standard error
now only joins a failure message.
Problem: the newly executed native Windows owner inventory failed while
constructing its candidate-type fixture, before product application ran.
Root cause: gitIn returned CombinedOutput, so a successful hash-object CRLF
warning on stderr was concatenated with its object ID and fed to update-index.
Fix: keep stdout and stderr separate at the shared fixture owner, retaining
both in failure diagnostics. Pin autocrlf=true in the compatibility regression
so this contract is exercised on every host. No production or gate changes.
Verification: the preceding helper compiles and reproduces the same diagnostic
pollution locally; the complete worktree race, Windows compilation, required
builtin/boot, vet, lint and make check pass. Hosted native revalidation remains
separate from local and earlier full-root evidence.
A live holder keeps the OS lock for its whole lifetime, and its `.lease.json`
sidecar can be deleted by a person or a file scanner while the process still
owns the session. The probe answered free on a missing sidecar without looking
at the lock, so session list said idle, task list flipped a running item to
interrupted, and redact-sessions would rewrite a live transcript. It now asks
the lock whenever a lock file exists, and still creates nothing on the common
uncontended path.
A live holder keeps the OS lock for its whole lifetime, and its
`.lease.json` sidecar can be deleted by a person or a file scanner while the
process still owns the session. The probe answered free on a missing sidecar
without looking at the lock, so session list said idle, task list flipped a
running item to interrupted, and redact-sessions would rewrite a live
transcript. It now asks the lock whenever a lock file exists, and still
creates nothing on the common uncontended path.
Problem: green PR checks could omit complete isolation/worktree owner tests on
Windows and under Linux race; pure isolation changes could skip native legs.
Root cause: the smoke and race package allowlists lacked both owners, while
the platform classifier treated untagged isolation Go as neutral.
Fix: classify isolation paths as platform-relevant; execute complete Windows
owners with an uploaded JSON inventory, require both package terminals and
portable boundary tests, and include both packages in PR race. Replace the
redundant rollback subset. Preserve existing permissions, deadlines and cache
rules; do not alter product code.
Verification: compiling old-workflow/classifier regressions, exact inventory
positive/negative controls, final classifier/testenv suites, complete 238-package
root inventory, three-owner race, required builtin/boot, vet, lint and check.
Local execution is macOS; native hosted outcomes remain a separate gate.
Refs #14025
Refs #14133
The width a hand set is kept in --flow-w and outranks the stylesheet, so once the
reading column had been dragged, choosing the full-width preset changed --doc but
not the column and did nothing. Choosing a preset now drops the dragged width.
Problem: executable-only changes and equal-byte regular/link changes could be
discarded as already applied, and source metadata edits could be overwritten.
Root cause: baseline, candidate and live file versions retained only blob IDs,
collapsing distinct Git modes and file kinds into one identity.
Fix: compare Git mode and blob together, follow the checkout's trusted/index
mode rules, hash real links and emulated links with their actual Git semantics,
and reject unsupported transitions before any source write. Keep the user's
index and configuration unchanged.
Verification: complete 238-package root inventory, worktree/isolation race,
real boot race, required builtin/boot, compiling predecessor controls and
operation-lease composition. Formatting, root vet, lint and make check pass.
Cross-compilation is recorded separately from pending hosted/native review.
Refs #14025
Refs #14133
Problem: the rooted-lock rename regression failed on Windows before reaching
the lock operation.
Root cause: Go's direct os.OpenRoot uses a Windows handle without delete
sharing, while production opens the lock directory relative to its managed
Root using an Openat handle with delete sharing. The fixture conflated them.
Fix: open the test directory through its parent Root, matching the production
constructor. Retain the real rename/replacement and physical lock-location
assertions on every platform. Document the direct-root Windows limitation.
No production code, timeout or platform skip changes.
Verification: complete filelock race (25 passes), Windows amd64 compilation,
required builtin/boot, root vet, full lint and make check. The prior full root
qualification is retained separately; new hosted Windows evidence is pending.
Problem: concurrent Stores could inspect one retained result and then write
or remove it while another operation was still using its candidate.
Root cause: reinspection checked authority without retaining a common owner.
Paths, first Finish, Abort and expiry cleanup could bypass Apply/Discard too.
Fix: add directory-handle OS file locks and one persistent per-result mutex
namespace. Hold admission through reinspection, mutation and synchronous
cleanup; preserve lock identity, active owners and live-only Finish retries.
Forward path-query cancellation and keep legacy writer limitations explicit.
Verification: complete 238-package root inventory, owner and real boot race,
compiled predecessor regressions, actual prior reader/sweeper compatibility,
Windows/Linux owner cross-compilation, formatting, vet, lint and make check.
No new native-platform or independent-review result is claimed here.
Refs #14025
Refs #14133
actions.ts sat exactly at the 800-line ceiling, so any pull request that added a
single action id failed repolint. The 44 feedback.* ids are one concern and move
to actions_feedback.ts, the way the market, schedule and document groups already
live in their own files; the registry content is unchanged.
Problem: cancelled calls remain parked behind checkpoint exclusivity, delay
their provisional resource cleanup, and may later enter hooks or execution.
Root cause: condition waits do not consume the invocation context. Observer
cleanup also exposes a shared writer-count decrement that can run twice.
Fix: centralize context-aware barrier admission, recheck cancellation after
durable preparation, return idempotent permits and bind the Agent caller.
Preserve completed undo retirement and every other active owner's admission.
Verification: deterministic cancellation and owner-preservation schedules,
compiling predecessor behavior failures, real boot direct/deferred coverage,
durable undo reloads, full root inventory, owner/boot race and required suites;
gofmt, vet, full lint and make check with unchanged gate baseline.
The literal branches of applyOldStringEdit wrote the replacement verbatim, so a
multi-line new_string inserted bare LF into CRLF files while only the fuzzy paths
converted it. Each literal replacement now takes the ending of its own line, which
also keeps an LF region of a mixed file free of stray CR.
Problem: a Store could apply an old cached Git tree after another Store had already applied or discarded that result, resurrecting its bytes in the source workspace.
Root cause: cache hits skipped durable publication and candidate admission; Git objects survive removal of the worktree. Refreshing without comparing the previous observation would also silently widen an earlier path declaration.
Fix: always reuse durable recovery/inspection for lookups, keep finished caches out of Pending, and refuse changed baseline/candidate/target trees until fresh path admission. Preserve first-Finish live-only and unsupported-version checks. No persisted, permission, tool-definition or default changes. Cross-process settlement locking and journals remain separate.
Verification: final complete 238-package root inventory passes (217 pass/21 no tests;23310 test/subtest passes,215 existing skips,194 valid cached packages). Full isolation race52 passes and real boot focus31 passes with no skips. The accepted predecessor Store compiles and fails9 behavior events. Gofmt, vet, full lint, make check and diff checks pass. Additional required builtin/boot command is running; new hosted/native evidence remains pending.
Problem: worktree task dispatch silently discarded effort, step limits and configured subagent models; parent model aliases also lost live session effort overrides.
Root cause: isolation branched before ordinary task runtime resolution, and alias resolution reloaded config instead of retaining the canonical parent runtime entry.
Fix: reuse shared task/model/effort/budget resolution, pass ephemeral options to the unattended Build, and canonicalize parent aliases to their runtime entry. Existing unsupported isolation restrictions remain refused; best-of keeps its prior policy. No persisted, wire, tool-definition or default changes.
Verification: complete 238-package root inventory passes (217 pass/21 no tests; 23289 test/subtest passes, 215 existing conditional skips, 198 cached packages). Owning delegation/isolation race has 400 passes; real boot focus has 39 passes, no skips. Compiling old owners fail 18 runtime scenarios plus parent, with two automatic-budget controls passing. First-Finish composition race has 32 passes. Gofmt, vet, full lint, make check and diff checks pass. Supplemental required builtin/boot is still running; hosted/native results remain pending.
Problem: Windows CI rejected successfully recovered isolated output because the new boot assertion required LF bytes.
Root cause: Git checkout may apply core.autocrlf conversion; recovery preserves the expected text with CRLF.
Fix: explicitly enable autocrlf in the disposable repository and normalize CRLF for the content assertion, retaining same-ID recovery and source-before-apply checks.
Verification: the old assertion compiles and deterministically fails with CRLF on macOS; the repaired four boot regressions pass under race (6 test/subtest events). Required builtin/boot pass (2291 passes, 66 existing conditional skips); vet, lint, make check and formatting pass. Native Windows replacement CI remains pending.
Problem: isolated work could lose its recovery ID or be removed when the
first candidate inspection or completed-result publication failed.
Root cause: Finish treated inspector/persistence errors like successful
empty results, and Execute hid Finish errors behind a runner failure.
Fix: retain the candidate and original ID, publish a separate versioned
inspection marker without overwriting existing records, reuse rooted
admission and same-ID inspection, and report live-only retry explicitly
when recovery publication is unavailable. Join both errors. Keep normal
completed records unchanged and refuse unsupported inspection versions.
Verification: complete 238-package root inventory, isolation and real boot
race tests, compiling preceding-owner behavior failures, actual previous
reader/sweeper compatibility controls, formatting, vet, lint and check.
Local evidence is macOS only; hosted checks and independent review remain.
Problem: A reader's hooks can change files without participating in rewind
exclusion, retiring undo, or recording the checkpoint's coverage gap.
Root cause: The invocation caller selects checkpoint admission using only the
concrete tool's read-only flag, even though it already knows that its hook set
may write. The checkpoint owner is correct but this caller bypasses it.
Fix: Include possible tool-hook writes in the two existing admission/coverage
conditions. Reuse the existing writer permit and deferred cleanup. Keep empty
hook readers unchanged and keep tool execution, coverage, and effects separate.
Verification: Compiling original owners fail the actual callback/undo and four
boot regressions; fixed owner/checkpoint and boot race tests pass. Full root
inventory, required builtin/boot, formatting, vet, lint and make check are
qualified separately in the PR. Asynchronous hook process lifetime and the
barrier's existing cancellation limits remain explicit follow-up work.
Problem: A hook can change files before refusing a tool or outside a reader's
declared paths without notifying an opted-in workspace sink. Nested writers
also omit publication until an outer writer happens to complete.
Root cause: Invalidation is gated by concrete tool execution and classification,
and publication belongs to the batch caller rather than the invocation owner.
Fix: Track possible tool-hook effects separately from Executed, publish through
the shared invocation exit after cleanup, and share result metadata projection
between batch and nested calls. Unknown hook scope invalidates content, tree and
Git metadata without claiming a confirmed write or widening authorization.
Verification: Full root 238-package inventory passed; complete agent/event race,
eight boot direct/capability hook cases under race, original-owner compiling
negative controls, and provisional-lease composition passed. Formatting, vet,
full lint and make check passed without changing the baseline. Local evidence
is macOS only; optional sink wiring and evidence accounting remain separate.
Problem:
Aborting provisional admission can expire an unrelated live user wait or
retake paths deliberately yielded to a user. A pending wait is also lost
when the last writer or an earlier restoration finishes afterward.
Root cause:
Physical release reset run state, and one retained extent represented both
active and parked ownership. Restoration dropped its temporary protection
before reholding the original caller. Expired queued grants and retention
callbacks could contribute resources to the next run.
Fix:
Separate active and parked retention, advance generations at the final
participating run end, and reconcile outstanding yields whenever writing
holds drain. Keep restoration protection through rehold, fence late
admission and retention, and clean expired grants without dropping current
writers. Actual in-flight writes stay protected after their run ends.
Verification:
The submitted owner compiles and fails all 14 selected test/subtest failure
events across the new public-API schedules. Corrected full workspacelease
and agent race pass, as do actual boot race, all 238 packages in disjoint
required/remaining partitions, required builtin/boot, vet, lint and check.
Local evidence is macOS only; hosted checks and review apply to the new head.
Refs #14025
Refs #14133
Problem:
A refused write can pin a never-executed path through later model requests.
Repeated failed restoration can also report success to an expired caller.
Root cause:
Admission merges every request into the run-retained extent before local
preparation succeeds. Call cleanup only decrements in-flight accounting.
Resume keeps its once-only result in per-invocation locals, and refused
upgrades omit the terminal wait notice after registrar contention.
Fix:
Track identified provisional calls separately from retained extents and
publish their remaining union through the existing claim registrar. Retain
before possibly writing hooks or execution. Preserve earlier writes and
parallel claims, report failed cleanup conservatively, validate lifetimes
before and after reacquisition, and cache terminal callback outcomes.
Close registrar waits on refusal and clear completed cleanup's wait state.
Verification:
Complete 238-package inventory and required builtin/boot pass on the final
source; full workspacelease/agent race and focused real boot race pass.
Compiling original execution, resume and upgrade-wait owners fail the
selected deterministic regressions. Formatting, vet, lint and check pass.
Local platform evidence is macOS only; hosted checks and review are separate.
Refs #14025
Refs #14133
Gathers the 38 fragments since 2.50.0 into one note: video wallpaper, find
case/regex switches, shortcut sheet, paged transcript loading, a narrower
reading column, dispatch trust and the delegation write-safety fixes.
Problem: standalone UI route checks require the paging API to have landed.
Maintainer guidance confirms that this check should remain a landing
dependency until #14365 is merged and the UI branch updates its studio base.
Restore the original route scanner and keep the HTTP fallback integration
tests, which prove runtime compatibility independently of landing order.
Verification: the original route test passes in the combined API/UI tree.
The standalone failure remains expected until the API base update. Vet,
complete lint, check and builtin/Boot validation pass without changing
production code, timeouts or coverage floors.
Problem: optimistic children can write paths held by live sibling tasks,
including through dynamically resolved local capabilities.
Root cause: whole-workspace optimistic grants bypass explicit-fence wrapping,
child options omit the shared scheduler, and parent admission skips children.
Fix: inherit the existing scheduler and admit actual named targets at execution,
using the actor grant's root and one complete resource set. Retain parent and
explicit-fence ownership, opaque tool availability and existing permissions.
Leave observation roots, provider prefixes and tool definitions unchanged.
Verification: compiling old owners write held files in four real boot task/fleet
direct/capability cases; fixed cases refuse overlap and preserve disjoint writes.
Full238-package root, required builtin/boot, complete three-owner race and focused
real assembly race pass. Vet/lint/check pass without raising the baseline.
Native hosted checks and independent review remain separate delivery gates.
Problem: the API-only PR failed the English refusal catalogue check.
Root cause: splitting the implementation retained the refusal codes but left
their English wording only in the UI branch.
Fix: include the same two English translations in the API branch, without
adding UI implementation.
Verification: reproduced the catalogue failure; all 170 i18n tests, typecheck,
vet, full lint, check and builtin/Boot suites pass. Existing catalogue
assertions and budgets are unchanged.
Place the shared refusal translations in EN_KERNEL on both branches so the
independent API and UI changes compose without a translation conflict. All
170 locale tests and typechecks pass on each branch.
Problem: standalone paging UI failed the CI serve route contract.
Root cause: the scanner assumed every frontend endpoint was mandatory, while
paging explicitly supports older servers through full-history fallback.
Fix: declare only the exact paging call optional, require its fallback route,
and exercise the HTTP adapter and hook across unsupported and real failures.
Verification: reproduced the original failure; route tests, 19 frontend tests,
typecheck, vet, lint, check and builtin/Boot suites pass. Covergate tests pass;
unchanged computer coverage is 88.2% on macOS versus the 89.8% floor, also
reproduced on the sibling branch. Ubuntu coverage remains for CI.
Place the shared refusal translations in EN_KERNEL on both branches so the
independent API and UI changes compose without a translation conflict. All
170 locale tests and typechecks pass on each branch.
Problem: CodeQL found two new uncontrolled-path stat alerts in record admission.
The first validator also read a claimed Git administrative backlink before
confining that path to the host-opened common directory.
Root cause: comparing embedded paths with global os.Stat did not itself provide
a filesystem namespace boundary, and marker discovery still used path globs.
Fix: enumerate and read slots through the opened managed root, perform identity
stats through root-relative handles, and constrain candidate/common metadata
reads before opening a backlink. Derive effective identities from checked host
roots, retain unknown fields unchanged, and treat brackets in root names literally.
Existing alias resolution remains metadata-only and grants no content-read scope.
Verification: final 238-package root inventory and required builtin/boot pass;
complete isolation race has 32 passes and real boot isolation race has five,
without skips. Six old-owner behavior groups compile and fail, namespace and
literal-directory regressions pass, and vet/lint/check retain the same baseline.
The replacement hosted CodeQL verdict and independent review remain pending.
Problem: the transcript paging UI conflicts with the upstream first-save bookmark fix. Root cause: both changes update the Pane-to-Transcript boundary, and retaining the runtime-id fallback loses unsaved stars on first save. Fix: merge studio f96ec2f869, retain session-path-only bookmark ownership alongside paging, and exercise both full and paged restoration. Verification: the upstream regression fails with the old scope and passes after resolution; 37 focused tests and 791 frontend files / 6343 tests pass, as do the production build, go vet, complete make lint and builtin/Boot suites. No lint baseline or deadline was changed.
Problem: a rebuilt store can trust a copied or mismatched durable record and
use its embedded candidate paths outside that record's actual managed slot.
Root cause: discovery checks the claimed workspace but does not bind snapshot
and candidate identity to the host-opened repository and observed marker slot.
Comparing Git markers only by inode also accepts cross-candidate hard links.
Fix: centralize record admission before every durable consumer; verify managed
directory membership, source identity, prefix roots and the linked candidate's
actual marker/backlink. Retain rejected records without rewriting metadata,
and resolve legacy identity through filesystem checks without launching Git.
Verification: five compiling old-owner negative-control groups fail as expected;
final 238-package root inventory and required builtin/boot pass, owning isolation
race has 30 passes and real boot isolation race has five, all without new skips.
Formatting, root vet, full lint and make check pass with the baseline unchanged.
Native hosted checks and independent review remain separate delivery gates.
An unsaved conversation scoped its rail stars by the pane's runtime id, and
the first save switched the scope to the session path, which reloaded from
an empty key: the star disappeared. The id is also a per-launch counter
(r1, r2, ...), so the key it left behind was read again by whichever unsaved
conversation pane r1 held after the next restart.
An unsaved conversation now keeps its stars in memory, and the first save
writes them under the session's path, merged with anything stored there.
The pane passes the session path alone as the scope.
TestFinalObservationIsBoundedAfterACompletedAnswer gave Run one second
after cancel (three after the deadline) to come back. Run still seals the
contract and the bundle after its context ends, because the first lock
attempt does not wait; locally that takes about 18ms, but on a loaded
windows-latest runner it overran the window and the test reported a held
observation that was not there (studio runs 40110025629, 40064086851).
The observation the test holds never returns until it is released, so any
finite wait catches the regression; use testenv.Budget for both waits.
Recovery conflated a failed Git inspection with an empty result and erased
the durable marker before attempting worktree removal. Return inspection
errors without cleanup and report the existing unavailable refusal instead
of forgetting a valid result ID.
Exercise apply path discovery, Apply and Discard against real retained
worktrees after restarting the store and temporarily hiding Git. Preserve
marker bytes and work, then retry the same ID after Git returns. Keep age
expiration and pre-publication Finish recovery as separate lifecycle units.
Separate review-role FileViews from the executor and bind direct/deferred
local tools to one fresh record for each role. Install the planner frontend
before binding so its ledger/audit and view ownership survive boot wiring.
Keep the generic read-only filter a same-actor operation, and preserve the
existing nil-runtime planner adapter without changing its tool surface.
Add four real-boot stale-overwrite regressions, runtime/clone accounting
checks, a generic-filter compatibility case, and written security bounds.
The compiling old-owner overlay reproduces all four overwrites.
Guardian's visible registry was filtered, but its read-only capability proxy
could resolve an executor writer or a declared host mutation. Set the
permanent execution flag in the session constructor so actual targets are
checked before they run, without changing the review's executor approval.
Add local resolver and real boot regressions, with compiling old-owner
failures for both mutation classes and an actual file overwrite. Document
the authorization boundary and the separate file-view ownership concern.
The drag floor was the readable measure itself (760), so the two presets were the
only widths a hand could reach on a typical window. A hand now reaches down to
640; the readable measure stays the untouched default and what Home returns to.
Problem: a child's read can refresh the parent's stale overwrite guard,
allowing the parent to discard an edit it has never read.
Root cause: copied file tools share FileViews, while deferred local tools
resolve through the root registry even inside nested children.
Fix: bind one fresh view per child to its direct and deferred file tools.
Keep disabled protection, tool identity, grants, read-only execution and
stable schemas intact. Install bindings on fresh capability frontends only.
Verification: original-owner controls reproduce the real boot overwrite;
a parent-pointer mapping control fails four nested cases, while the actor
binding passes all six shared-runtime/clone combinations. Root package,
owning race, composition, vet, check and lint evidence is recorded in the PR.
Refs #14025
Refs #14133
A dispatch target is resolved against the user's sidebar and the dispatch
switch is already the user's consent, so stopping on the folder's trust card
asked the same person twice. The folder is now recorded as trusted and the task
goes through; a folder that cannot be trusted whole still keeps its pane open on
the card with the task unsent.
Problem: Both extracted PRs became modify/delete conflicted while their local
qualification was running.
Root cause: Upstream removed FIRST_RUN_WORKSPACE.md in #14367 after the slices
had repaired its inherited lint metadata and paragraph violations.
Fix: Accept the upstream deletion and remove the now-obsolete prerequisite
repair. Retain every feature, regression and security-document change.
Verification: make check passes. The staged merge changes no Go file relative
to the independently qualified slice heads; their full package/race evidence
remains explicitly scoped to those source trees. Hosted checks must run on
the new merge heads. No unrelated production change or baseline relaxation.
Refs #14025
Refs #14133
Problem: Both extracted PRs became modify/delete conflicted while their local
qualification was running.
Root cause: Upstream removed FIRST_RUN_WORKSPACE.md in #14367 after the slices
had repaired its inherited lint metadata and paragraph violations.
Fix: Accept the upstream deletion and remove the now-obsolete prerequisite
repair. Retain every feature, regression and security-document change.
Verification: make check passes. The staged merge changes no Go file relative
to the independently qualified slice heads; their full package/race evidence
remains explicitly scoped to those source trees. Hosted checks must run on
the new merge heads. No unrelated production change or baseline relaxation.
Refs #14025
Refs #14133
Problem: A snapshot-only regression does not establish that the shared tree
reader also protects the source workspace before Apply publishes a candidate.
Root cause: The old cached index can hide a user's intervening edit from the
whole-tree drift check, allowing Apply to return success and overwrite it.
Fix: Exercise actual snapshot/candidate/Apply assembly on disposable real Git
repositories. Require ErrWorkspaceMoved and unchanged user bytes. Document the
existing Git driver confinement and the additional rehashing cost explicitly.
Verification: A compiling overlay of the original production file fails for
both wrong success and actual overwrite; the repaired owner passes. Full final
worktree race passes (48.247s). Broader root and static qualification records
are retained separately; production code is unchanged by this test follow-up.
Refs #14025
Refs #14133
Problem: A candidate snapshot can contain old bytes after the working file
changed while preserving its size and mtime.
Root cause: Copying the source index carries Git's trusted stat cache and
assume-unchanged state into StageAll, so staging can skip the actual bytes.
Fix: Preserve the copied index's staged tree, rebuild the private index from
that tree, then stage current content. Handle an unmerged private index first.
Keep staged ignored inputs, the real conflicted index, HEAD and refs intact.
Document concrete Git/path/authority boundaries and the unchanged need for
source coordination. Include the inherited documentation-lint prerequisite.
Verification: All three cache-policy variants fail against the original
function through a compiler overlay and pass with the repair. Full worktree
race passes, plus final expanded cache and conflict regressions. Check, lint
and root vet pass. Full builtin/boot exits zero (22.243s / 444.076s tests),
using a measured 15-minute package alarm with no product-deadline change.
The exact-commit full root sweep and hosted checks remain separate evidence.
Refs #14025
Refs #14133
Problem: Task-length write claims cannot independently represent permission,
assignment and the physical mutation lifetime across cooperating runtimes.
Root cause: Shared admission, input preconditions, cancellation and release
have no common short-transaction owner that callers can adopt independently.
Fix: Extract the mutation resource/queue/capacity owner and content versions
from the preserved concurrency implementation. Add minimal non-blocking legacy
claim compatibility and exception-safe publication cleanup. Keep all runtime,
isolation, hook, controller and default activation outside this review slice.
Document authority, actual sinks, compatibility and unenforced boundaries,
and map the remaining source into dependent review units. Normalize an
inherited design-document lint prerequisite without relaxing its baseline.
Verification: make check, full lint and root vet pass. Owning-package race
passes, including real subprocess death and a strengthened cleanup regression
that fails without the cleanup via compiler overlay and passes with it.
Builtin passes. The full boot invocation reaches its 600-second package alarm;
all 811 top-level tests close in four inventory-verified groups (808 pass,
3 conditional skips, no missing/unexpected/failing selected tests). Linux and
Windows amd64 test binaries cross-compile; native platform and hosted CI
evidence are separate gates. The remaining root-package sweep is in progress.
Refs #14025
Refs #14133
Problem: searching for a message outside the initial page could remain incomplete after one fast response. Root cause: the search effect depended only on busy/hasMore flags that React may batch back to their previous values. Fix: advance search on committed page revisions and release the completed single-flight request before publishing that revision. Verification: failing-before revision regression, 28 focused tests including scroll and anchor neighbors, and a browser replay fetching four pages and finding the first message.
Problem: full transcript projection and transfer scale with conversation length. The session store now owns stable page snapshots and cross-page context, exposed through optional controller reads and an additive HTTP route. Existing full-history clients are unchanged. Validate through real Boot/resume, full-projection equivalence, cursor invalidation and append stability, request authority and foreground preemption. Focused history tests, race checks, Go lint and release-note tests pass; broad suites are recorded in the PR.
Rebase the UI portion of #14087 onto current studio, preserving regex/case search and session-scoped navigation. Use optional paging with explicit unsupported-route fallback and retain mounted controls when prepending. Browser verification exposed zero-box display-contents wrappers; share rendered-box geometry with the landing writer and add a failing-before regression. Verification: 37 focused frontend tests, TypeScript checking, production build and wide/narrow light/dark browser renders; full-suite results and backend dependency are documented in the PR.
Define the optional history-page response and typed refusal translations shared by the API and UI splits. No caller switches to paging in this commit. Verification is performed on each independent split before publication.
Documents the last coupled first-run decision: with nothing remembered,
the first pane's workspace root falls back to the user's home directory
(pickLaunchWorkspace), so a session that was never given a chosen project
reads as temporary. Proposes a kernel-provisioned auto workspace instead,
modelled on MiMo Desktop's session-first behaviour (auto-created session
directories, global provider config, optional projects), and maps it onto
the existing launch chain, session store and workspace switch.
Keep the install slot claimed until a cancelled download finishes cleanup, preventing it from cancelling a new install or deleting its files. Add regression tests and a release note.
Expose download cancellation through the update API and Versions UI. Cancellation waits for the transfer to stop, removes its partial and cached artifacts, and reports a distinct non-failure state. Automatic updates do not retry a cancelled release but can proceed when a newer version is offered.
Off means nothing sleeps, cap included: a person who never turned the
setting on kept losing the least recently used page's renderer when a
seventh page opened. The cap now runs only on a finite threshold, the
same one the sweep reads.
The rx-page-sleep desktop preference decides it, default off: prefs.js
owns the threshold, the shell reads it at boot and on every preference
save, and Settings -> Appearance carries the toggle. A woken page also
comes back fitted into its panel at the scale it had before the sleep.
Five real Chromium pages through BrowserViews at a tiny threshold: the
working set they cost, the working set the sleep returns, a wake that
must answer against the page as it loaded, and a command the sleep
outlives that must fail rather than hang. Run with pnpm sleep-live.
An agent page's renderer is hundreds of megabytes the Studio pays for
whether anyone is watching. A page untouched for fifteen minutes and not
on screen is closed whole — address and a picture kept — under the same
targetId, so the kernel's handle survives; the next command, show, or
typed address rebuilds the view, replays the domains the kernel last
enabled (a domain switched off mid-session stays off), and answers only
once the page has loaded — a mid-load answer would describe the page it
was, and the kernel could not tell. A command the sleep outlives is
failed fast, for the session that dies never settles it and the kernel,
told nothing of a sleep, would wait on a page that will not speak. Six
awake pages at most — a wake yields no room: the page asked for
outranks the pages left alone — and eighteen asleep ones; past that the
oldest is buried and the kernel is told, the way a browser discards
abandoned tabs. REASONIX_NO_PAGE_SLEEP puts the feature away.
It sized itself 42px inside the pane on purpose; pulling that to 34px to match
the text moved the mirror's wrapping off the caret's in composer-reflow.
Opening the todo list grew the composer, which shortened the scroller, which
re-measured the rail: every mark moved up and the two jump buttons went with
them. The panel now floats above the composer instead of taking its room.
`--flow-w` went to `:root`, where a committed value on `.app` outranks it;
the sidebars only gave way on release, and the composer never followed at all.
Each key cell was as wide as its own text, so the descriptions started at six different positions. A shared minimum on the key cell gives them one column.
`groups` is a state slice; after a rewind shortens `marks`, a stale fold still pointed past its end and `marks[v.from]` was undefined. The fold is filtered to what exists.
`useSidebarOpen` dropped the promise with `void`, so `Workspaces` never saw the rejection and the row stayed on its old selection. It now returns the open's own promise.
`en_export.ts` repeated 我 and 会话, and `i18n.test.ts` rejects a key that lives in two catalogue files. The headings now use the entries already there. The menu test follows the second export row.
`^` and `a*` match the empty string, so the counter reported hits the painter had no range for. Both sides now skip them. Transcript.tsx keeps its 800-line ceiling by typing the mode inline instead of importing it.
The painter still searched the DOM for the literal text while the counter honoured the new modes, so a regex hit had no range to scroll to and a case-sensitive count did not match the painted one. Both now share one finder, a regex match advances past its own length, and a pathological pattern is cut short instead of freezing the window.
prompt_tokens_details with cached_tokens: 0 is the endpoint reporting the
split, so miss is derived from prompt - hit instead of staying zero and
vanishing into the unknown bucket; a hit above prompt_tokens is a broken
report and clamps to it. The overview and the meter floor the rate, so 100%
means every split token hit.
A shortcut sheet in the settings.
Every chord the window answers to, on one page: find, search, the three views, stepping through
your questions, the workbench toggles, send, close, and the width keys. Chords are written the way
the platform writes them, so macOS shows ⌘ where Windows and Linux show Ctrl. It is a page of the
settings table of contents, so searching for "shortcut", "Esc" or "运行分析" lands on it.
The strings live in `en.ts` with every other catalogue entry: the i18n guard requires one
catalogue file, and an earlier revision of this branch split them out and failed it.
Make run analysis easy to enter and to leave.
- The entry in the window bar now carries its name. It needs an inline layout for that: the
button is a 30px grid cell, so the icon and the label stacked and the text broke one character
per line. `.chrome .thbtn.top-action` takes over the layout for this variant and keeps the
label on one line; below 1180px the label steps aside so the other controls stay put.
- The analysis page has a back arrow beside its title.
- Picking a conversation that is already on screen returns its pane to the reading view instead
of re-opening it. That rule now lives in `sidebaropen.ts` — partly because it is a rule about
panes rather than window wiring, partly because `App.tsx` was 11 lines over its ceiling with
it inline. It has its own test: open pane returns to reading, unopened opens a pane, a folder
request opens a pane, and asking twice lands twice.
`Pane.tsx` differs from `studio` in one line (`onBack`), so the `useAgentRequest` wiring from
#14291 stays intact.
Match case and regular expressions in the find bar.
`usefind` now hands back `regex`, `caseSensitive` and `setMode` with the rest of its state —
the previous push declared them on `Finding` and never returned them, which is what
`frontend-next` caught. Changing either reading restarts from the first hit.
Rewritten against the shape the kernel actually writes. A transcript line is a provider message
(`role`, `content`, `tool_calls`, `host_authored`), so the export keys on `role`: user and assistant
only, skipping host-authored user lines, tool results and system lines, and skipping assistant turns
that carry only tool calls. Headings go through `t()`. Native archives keep their `records` instead of
exporting an empty list. The test now uses real JSONL lines and covers the native-archive case.
The controller runs only `/mcp connect` (#14167), so `/mcp add x` now
answers "unknown /mcp subcommand". Both guides and `reasonix mcp help`
still sent a running chat there to save and connect a server. They now name
`n` in the terminal /mcp panel, which installs through the same
POST /mcp/install as Desktop's Add and connect. The guides also name `p`,
the panel's project install, which InstallMCPServer writes to the
project's `reasonix.toml` (MCPConfigPathForEntry -> ProjectConfigPath).
TestTheDocsNameOnlyMCPAndSkillsSubcommandsThatRun reads docs/*.md and the
READMEs and holds every `/mcp <sub>` and `/skills <sub>` they write to the
subcommands completion offers, which are the ones the controller runs.
TestTwoAsksRetakeTheClaimOnlyAfterTheLastIsAnswered answered the first ask
as soon as two asks were registered. Ask registers before it yields the
workspace claim, and Yield reads the claim log from disk before it counts
the waiter, so on a slow Windows runner the second ask could still be short
of the yield when the first one resumed. The first resume then saw no other
waiter and took the claim back, and the test reported the claim retaken while
the second ask waited (studio 39941591112, windows-latest).
The test now waits for the second ask's prompt-queued notice. Ask reaches
lockPromptFor only after its yield, so the notice is the point at which the
second ask is waiting on the user.
CI on Windows failed both new assertions: the detail named
`c:\outside-watch\reasonix cli.desktop` and the temp deliverable came
back lowercased, while the tests compared against the path spelled as
constructed. `evidence.normalizePath` lowercases on Windows, because the
ledger needs one identity per file and Windows hands back either case.
Compare against that identity instead of the constructed spelling: the
rule under test is about which paths are reported, not how the host
happens to capitalise them.
The gap names writes the host watches nothing of. A folder the person
granted as workspace is walked with the root and observed the same way,
so a write there is work product and must not carry the caveat.
This pins the rule at the boundary extraRoots decides, so a later change
cannot start reporting exactly the deliveries the grant exists to allow.
`Caches` read the process's home directory itself, which put the package two
reads over its `process-input` budget and the repo one over its ceiling. Both
lists answer from the same fact, so it is read in one place and passed to each.
The cache list also gained the two entries `Dirs` already had on the platforms
that have them: `$HOME/.cache` everywhere, and `Library/Caches` on macOS. A
cache filter that named only `go`, `.cargo` and `.npm` would report a Unix build
cache as out of sight on the one platform whose caches live under a different
directory name.
Two defects, both found by CI on Windows and both in this PR's own new code.
`underScratchOrCache` asked `hostwrite.Dirs()` for the toolchain caches. That
function answers what a jail grants a command, and Windows jails nothing, so it
returns nil there: every build cache write was reported as out of sight on the
one platform where the list is empty. `hostwrite.Caches()` answers the question
actually being asked — where does a build write through — and is defined on every
platform. The two questions stay apart because conflating them is what made a
Windows build produce a caveat about its own module cache.
The one-off directory `sessiontemp` hands a turn is recognised by its position,
not by its name. It is a direct child of the host's temporary root, so that root
is read from the host and the first component below it is compared against the
prefix. The previous version walked every component, which exempted any folder
deeper in the tree that happened to spell the prefix, and it assumed the root was
named `/tmp`, which is not where Windows puts it.
The tests that pinned these edges set `HOME` and `TMPDIR` and then read the
result, which is why they agreed with the code on Linux and disagreed with it
everywhere else. `HOME` and `TMPDIR` are POSIX-only names: on Windows the
temporary root comes from `TEMP`, so `t.Setenv("TMPDIR", …)` changed nothing and
the assertions ran against the real root. They now call `disposable` with the
roots as arguments, and build paths from the host's own temp root, so each
assertion is about the rule rather than about an environment.
Tested: the thirteen out-of-scope tests, the packages this PR touches, and
`GOOS=windows go vet` on `internal/runtime/agent` and `internal/base/hostwrite`.
A turn that wrote a file outside the workspace and then answered ended with a
`done` receipt and no gaps. The host saw the tool call succeed, but its
observation only covers the workspace and its granted extra roots, so the write
was dropped before the report was built: no mutation, nothing to verify, nothing
to say. `final_readiness` agreed — it judges a turn by whether it has a writer,
and a write it cannot see is not one — so the turn ended as a quiet success
while the work it claimed had landed nowhere the host could check.
Record those writes as a gap instead. `outsideObservedWrites` reads the turn's
own successful write receipts and keeps the paths that are outside the workspace
and outside what the host owns: the session scratch directory, toolchain caches.
Two exemptions carry the weight. A public temporary root is not exempt — a
person who asked for a file there asked for something to keep — and neither is
`bin/` under a cache, because `go install` puts the user's own executable in
`~/go/bin`. A scratch directory earns its exemption only as a direct child of a
temporary root, so a folder is not exempt for spelling the prefix.
Reported, never enforced. The verdict drops from done to partial and the receipt
names the gap; the turn ends exactly as it did before. A person who asked for a
file outside the workspace got what they asked for, and what they are owed is
the line saying that part is out of sight, not a turn that will not end. For the
same reason the gap is a notice, not a warning.
CompletionReceipt also stops discarding the report's own gaps when the verdict is
unknown. It returned only the claim-yield and mod gaps there, so a turn the host
could not judge — a plan with no criteria, writing only outside the workspace —
went back to saying nothing at all.
Blast radius: the completion report on turns that write outside the workspace,
and the receipt the host renders for them. The readiness gate, Delivery, Goal,
cancel and budget paths do not read the report, so none of them is affected.
Neighbouring behaviours tested: a turn that writes outside the workspace is not
reported as finished; one that stays inside keeps its quiet ending; a plain
answer and a read-only investigation end with no caveat and no continuation; a
scratch probe carries none; a deliverable in the public temp directory does; a
lookalike directory and an installed executable are not exempt; an unfinished
todo list still holds the turn; an unjudged turn keeps its gap. Desktop
`ReceiptCard` gained one case, with the three catalogue entries the i18n gate
requires; typecheck and the full vitest run pass.
Scope: this reports a write the host cannot see. It does not decide that a turn
which never wrote anything was unfinished — a read-only multi-step turn that
stops early still ends as it did.
A background sub-agent writes its work into its own transcript and its tool
calls, never into the job that started it. The job's activity stamp therefore
never moved, and the watchdog judged the run by bytes on streams that stay
empty: a busy sub-agent read as a job that had been silent since it started, and
a stalled notice went out about work that was progressing.
Bind the child's progress to its job at beginExecution — the one boundary every
delegated run crosses, foreground included, where a context with no job binds
nothing. A finished tool call is one reported step; token deltas refresh the
activity stamp at most once every two seconds, because the watchdog reads
minutes and the child emits far more often than that.
Reported steps are kept apart from output text rather than folded into it. They
answer different questions — whether the child advanced, and whether it wrote
here — and a caller decides from the pair. `wait` says both, and a stalled
notice says what the job had reported before going quiet, since a job that
worked and stopped needs a different next move than one that never got going.
Tested: internal/tools/jobs (a job that writes nothing stays alive while it
reports progress; wait reads steps separately from output), internal/tools/builtin
(progress wording per reading), internal/runtime/delegation (binding at
beginExecution, a finished tool call reports one step, deltas are paced, a
foreground run stays unbound), internal/contract/event, and desktop
frontend-next (typecheck; full vitest run, 6120 tests). The desktop stalled-notice
sentence had no English for its new wording and the i18n gate caught it; the
three entries are added here.
`wait` accepted any `timeout_seconds`, and omitting it meant blocking until
every job finished. A sub-agent that never exits then holds the turn open with
no way for the caller to tell a long wait from a hung session: the tool result
is the only thing that returns, and it returns only when the job does.
Cap one wait at 300 seconds and apply that cap when the timeout is omitted too,
so omitting it means "wait for the jobs" rather than "wait forever". The result
says when the ceiling was reached, and only then: a job that finishes first
returns before it, and claiming a 300s wait that never happened would be a
second false report in a message whose whole job is to be accurate.
The schema states the same bound the runtime applies. It previously promised an
unbounded wait, which the tool no longer performs.
Tested: internal/tools/builtin (wait ceiling applied and reported, a 1s wait
returns progress instead of holding), internal/tools/jobs, and the boot golden
recapture for the changed tool schema. internal/assembly/boot has one
pre-existing failure on this host (TestEffectSessionScratchIsReachableByItsSpelling)
that reproduces on the unmodified base: bubblewrap cannot create a namespace
here, so the sandboxed bash call the test relies on refuses to run.
absorbInterruptedCopy called os.Lstat itself, which CodeQL reports as a new
path sink in changed code. sessionArtifactExists is the store's pre-existing,
baselined reader for the same question: the copy exists when its transcript or
an owned sidecar does. It answers bool, so missing and unreadable look alike;
the import's callers already accept that, and a copy that is there but cannot
be read still reports its error where it is loaded.
CodeQL flagged the two sidecar paths as new sinks once copyBranchMetaSidecar moved into migrate_copy.go: the os.ReadFile(BranchMetaPath(srcPath)) body was baselined in migrate.go, and a moved copy is not. The function moves back to its original position with only the error return changed, and completeBranchMetaSidecar now asks the pre-existing LoadBranchMeta reader whether the destination sidecar is there instead of calling os.Stat itself: ok means present and never rewritten, an error means unreadable and never clobbered, and only a clean absence proceeds. sameDirPath (a pure path comparison, no filesystem access) moves to migrate_passes.go so migrate.go stays inside its ratcheted file-size budget.
An import publishes the transcript first and the branch sidecar that carries
what it means second. A run stopped in that window left a transcript with no
sidecar, and the retry treated the transcript as completion: the flat import
healed only subagent artifacts and stamped the marker, and the v4 import only
healed the title. Both now finish the sidecar, refuse a failed sidecar write
as incomplete so the next run retries, and leave a finished copy untouched.
A v4 copy that is still a prefix of 1.x is brought up to the source; a copy
continued in 2.x keeps its own turns.
A power loss on NTFS can leave a .meta sidecar extended but never flushed, so it holds only NUL bytes. LoadBranchMeta reported that as a decode error, which made the recovery-lineage walk refuse to archive anything in the folder and kept UpdateBranchMeta from ever rewriting the file. The file records nothing, so it now loads as a missing sidecar and the next save replaces it.
Signed-off-by: ShryukGrandhi <[email protected]>
loadMemory returned (false, nil) for a read error, so every caller treated a
present-but-unreadable fact file as missing: it vanished from /memory, the
desktop panel, the memory tool and recall, and a save by name then minted a new
id at revision 1 over it with no snapshot, losing the old identity and history.
loadMemory now returns a typed FactReadError. ListActive reports unreadable
facts beside the readable ones, and the listing, read-by-name, panel, tool and
recall audit all name them; recall still serves the readable facts and keeps
its old sentence for a whole-directory failure. A save by that name refuses
before snapshotting or writing anything.
The terminal counted and drew a host-authored inbox entry as the person's
own pending line, so `d d` on it reached DELETE /inbox/items/<id> and the
kernel dropped the runtime's continuation. The desktop has always filtered
these out; the terminal never decoded the field it filters on.
The mirror now carries origin, readonly and capacity, and a declared-subset
wire-parity check holds it (and the page's copy) to the fields a decision
reads, so dropping one is a `make check` failure instead of a silent zero.
A host continuation is out of the pending rows and out of the count, except
where it is blocked or uncertain and needs a decision only the person can
make; there it is drawn marked as the runtime's. A read-only snapshot offers
no mutation key and says why.
The force-push to the restacked head did not start the test/build workflows
(closing and reopening only re-ran the labeler), so this empty commit emits a
synchronize event. No code change.
A line that carried a currency and a quote of zero was priced by the exact
JSONL read but discarded by the projection, which treated amount 0 as "no
price". A single-day range reads JSONL while a multi-day range reads the
projection, so one session showed [{CNY 3},{USD 0}] on 今日 and [{CNY 3}] on
30 天, lost costEstimated on a day whose only quote was zero, and could emit
the false "usage but no cost record" note.
A currency is the whole test for "priced"; zero is a genuine free quote. Both
build paths now keep it: projectRecord no longer drops amount 0, the rederive
view no longer filters cost<>0, and the rollup read no longer skips amount 0,
while a row with no currency still carries no money. No schema column was
needed, so SchemaVersion is unchanged; every launch reconciles the day files
through rederive, which repairs a projection built by the old filters.
Tests cover the fixed two-day history through the exact read, the incremental
append projection and the rebuilt projection, the zero-amount-only day's
costEstimated in both, and controls for a currency-less row and a non-zero
amount.
<active-goal contract="full"> carries the whole goal-mode contract into the
model-visible history, and nothing retired it when the goal stopped: the next
request carried no goal block while the contract still said to keep pursuing
it, and active-goal being a superseded tag made a fold keep that stale copy.
A stop the model did not report now renders a tombstone under the same tag:
reason paused, cleared, budget, or host (the completion check could not run).
A model-reported block and a completed goal compose nothing, because the model
already knows. Delivery is asked of the visible history through the delivery
PR's projectionDebt: the tombstone is owed while the newest active-goal block
in view is a contract or reminder, and not once it is the tombstone itself, so
a rewind or fold that drops it owes it again with no remembered state.
SupersededUserBlockTags is unchanged, so the tombstone is the newest copy a
fold keeps.
The boot harness's blockOf now matches a tag prefix, since a variant block
opens with attributes after its tag.
Retain legacy DAG payload locations instead of discarded message bodies, preflight JSON without copying large strings, and refuse oversized records or retained histories without modifying source data or falling back to stale checkpoints.
Refs #14116
A call that spent tokens but recorded no money was counted in the token
and request totals and contributed 0 to Cost, with nothing in the payload
saying its money was unknown. Two paths produced that shape.
Title generation wrote such a row live: the serve title model builds its
own stats.Recorder instead of emitting through the assembled sink, so no
CostQuoteSink sits in front of it. The title event carried a rate card but
no quote, and the recorder writes cost fields only for a non-nil quote.
It now prices the call at emit time with event.EnsureCostQuote, using the
same QuoteContext construction the assembled sink uses
(boot.QuoteContextFor), so it is quoted like any other call.
Rows written before the cost fields existed (and any row that simply
lacks them) had the same effect on reads. The incompleteness trigger is
now "no cost fact": an explicit cost_complete=false, or tokens counted
with no cost_amount. One predicate (usagecatalog.CostIncomplete) is
applied on all three read surfaces, so they agree.
No price is invented and no legacy row is backfilled.
The flat-import re-home took the save locks but never asked the session
lease, so it could move a WAL-backed transcript out of a live session root
and delete the old path while another runtime (or this one) still owned it.
Gate every move/remove in the migrate family on the same predicate
renameOverlongSession uses: sessionLeaseHeldLocally OR
SessionLeaseHeldByOtherRuntime. A deferred base is not counted as imported
and leaves the one-time marker unwritten, so a later run settles it.
Tests: audit red/green pair, unleased control, same-process lease control.
Problem: the post-answer finalization fix note would be missed by the current release fragment collection.
Root cause: it used the retired shared unreleased.md path and did not name the fixing PR.
Fix: move the bilingual entry into unreleased/14140-bound-post-answer-finalization.md, preserving #13953 and adding #14140 to both bullets.
Verification: release-note and Markdown-reference tests pass (12 tests); make check, fragment discovery, bilingual references, release rendering, and whitespace checks pass. Only release-note documentation changes.
Problem: the compaction fix note was outside the release fragment glob and cited the related finalization PR instead of its own fix.
Root cause: the note used the retired shared unreleased.md layout.
Fix: move the bilingual entry into unreleased/14144-compaction-eof-vs-cancellation.md and reference #14144 in both bullets.
Verification: release-note and Markdown-reference tests pass (12 tests); make check, fragment discovery, exact bilingual references, release rendering, and whitespace checks pass. Only release-note documentation changes.
Problem: compaction could report success and accept a partial digest after
the summary deadline or caller cancellation.
Root cause: providers close their channel when context ends. Selecting EOF
instead of the simultaneously ready cancellation accepted nonempty output;
the deferred ceiling classifier only corrected non-nil errors.
Fix: keep EOF validation at one result boundary and check the request context
before accepting a digest or classifying truncation. Preserve typed ceiling
and cancellation failures, usage accounting, and stable provider requests.
Verification: force EOF/cancellation overlap with a progress barrier and fake
clock, cover normal completion/truncation and charged usage, and exercise
the real Build-to-notice path. Baseline production source fails on macOS
and Linux; fixed Linux cases pass 20 repetitions. Focused race/cache tests,
full root tests, required tool/boot tests, vet, lint, make check, and release
script checks pass. A separate combined tree also validates this fix with
the post-answer finalization change from PR #14140.
The seer controller.New bound was the executor's method value, and that value
reads the session in place: a SetSession from /new, resume, a branch switch, an
extension system prompt or a fold is already visible to it. Those sites answer
from the current session. A model switch is different — it builds a new
controller, and ApplyRuntimeMigration adopts the old conversation into it. The
carried turns still hold the standing blocks the old model was sent, so the
derived rule reported them delivered and the new model's first request carried
no listing.
The seer now resolves through the controller and answers "nothing is visible"
for a generation that has delivered nothing yet. ApplyRuntimeMigration marks
that state when the replacement's model ref differs from the outgoing one; the
hand-off to the runner clears it, so exactly one request restates the standing
blocks and the visible history answers after that. A rebuild that keeps the
model keeps the derived answer.
Problem: A completed model answer could leave the controller running while
workspace observation or an evidence stream lock kept finalization alive.
Root cause: Post-answer observation and sealing detached their deadlines
from the turn context. Observation called filesystem operations synchronously,
so its existing timeout did not actually bound the caller's wait.
Fix: Carry the run context through every finalization entry point. Bound
observation with a buffered result that captures no mutable turn state,
record missing observations as gaps, and cancel unused initial snapshots.
Keep optional finalization cancellation from discarding completed answers.
Verification: Deterministic cancellation, deadline, observation-budget,
held-seal-lock and real boot/controller completion regressions fail before
the fix and pass on macOS and Linux arm64 after it. Related race checks,
121 frontend tests, formatting, vet, lint and changed-file checks pass.
The required builtin-tool and full boot package suites also pass.
Two rows of the picker table described the 1.x CLI: an "MCP import picker"
the 2.x terminal does not have (`reasonix mcp import` imports without one),
and an "MCP manager" whose `r` refreshed and whose number keys chose actions.
The 2.x /mcp panel (mcppanel.go, mcpmanage.go) moves with Up/Down or j/k,
opens a server's tools with Enter, toggles with Space, adds with n through a
review that installs with y, l or p, removes with d then y, signs in with a,
grants requested access with u, and reconnects a failed server with r.
The chat table also missed keys the terminal handles: Ctrl+S (steers a
running turn, sends an idle draft), Shift+PageUp/PageDown, Ctrl+Insert to
copy a selection, Shift+Insert to paste text where the terminal hands it
over, Ctrl+G to keep folded pastes as files, and Ctrl+D twice in the resume
picker. The tool approval row now says its numbers follow the rows shown and
that Esc does not answer in vi mode, and plan approval gets its own row. The
Chinese table also gains the Ctrl+Enter / /steer row it never had.
Pin the three behaviours no test held: Ctrl+S on an idle draft, an approval
number with a row missing, and Esc on a plan approval.
The English remote section gained two blocks the Chinese guide never got: a WSL
distribution as a machine (kind = "wsl", what it accepts, how Studio adds and
keeps it, where its model credentials come from, which subcommands stay
SSH-only), and the rule that an agent's own shell may run only `list`, `help`
and `wsl list|test` of `reasonix remote` (exit 77 otherwise). Translate both into
the same place in GUIDE.zh-CN.md, between the config example and the CLI block,
using the Studio labels the Chinese window shows (设置 → 远程, WSL 发行版).
The guide says `wsl test` runs for an agent, but the test pinned only
`wsl list`; add the `wsl test` row. A distribution that does not exist answers
after the same `wsl --list` the `wsl list` row already runs, so no distribution
is started.
Both were raised as non-blocking notes on #13784.
validPriceSource and its page mirror sourceProblem refused control and blank
characters but not format characters (Unicode category Cf: zero-width, soft
hyphen, bidirectional controls). A source carrying one was stored as written
while the page drew something else: with U+200B or U+00AD inside the host the
kernel stores a host that is not "api-docs.deepseek.com", and the page's host
line, read through the browser's URL parser, shows "api-docs.deepseek.com";
U+202E in the path reverses the tail in the hover title.
Refuse Cf in the same loop on both sides, the check validElicitURL already
makes for a link a person is shown. The page uses \p{Cc}\p{Cf} so it also
refuses the C1 controls Go's unicode.IsControl already refused.
Raised as a non-blocking note on #13864.
Problem: CodeQL flagged direct path opens in the cancellable checkpoint readers.
Root cause: The normal and torn-tail JSONL readers followed file symlinks beyond
the parent directory supplied by their caller, even though frontend request
authorization is owned separately at the HTTP boundary.
Fix: Open both checkpoints through os.OpenInRoot, retaining cancellation and
recovery semantics while rejecting links outside that directory. Preserve
relative in-directory links. Add a deterministic boundary regression and
bilingual release notes, and separate the sidebar Retry action from its note.
Verification: The escaping-link regression fails before the change and passes
after it. Complete storage/replay/catalog/scheduler suites, focused race and
HTTP boundary tests, 13 frontend tests, TypeScript checking, release-note tests
and repository/Go lint pass. Hosted CodeQL must confirm the new head separately.
An audit found that a fold which keeps the old block used to leave a
removal unannounced (the tombstone was gated on a bookkeeping flag the
fold cleared). Deriving delivery from the visible history fixes it; this
pins that a removed instruction is still announced while the model can
read the old block.
An audit found the same hole through two more entry points: /rewind
truncates the turn that carried the block, and /switch replaces the
transcript with another branch's prefix. Deriving delivery from the
visible history fixes both, so the audit's two providers-boundary tests
now guard that they stay fixed.
A standing projection was settled as delivered when a turn was handed to the
runner, so anything that refused the turn before the provider (an extension
block, a cancel, a failed send) or a /new left the debt settled although no
request had carried the block.
Delivery is now a fact about the model-visible history: owed compares the
newest block of its tag in the visible user records, so a LocalOnly record or
compaction summary never answers for it. A refusal, cancel, fold, rewind or
session rotation changes the answer by itself; nothing is recorded at the
hand-off.
Removed: projectionDebt.{pending,delivered,settle,forget,sent,inherit},
promptFailureDebt.{composed,settle}, settleTurnProjections and runSettled, the
CompactionDone→forget handler, InheritScratchFrom and the per-owner forget
hooks. One-time failure/refusal notes are filtered by visibility instead of
settled, so a turn refused before the provider still owes them.
internal/state/memory guarded its read-modify-write sequences with a
process-local mutex only, so two processes sharing a memory directory could
duplicate a subject key, lose a rename, roll a withdrawal over a newer
revision, or drop an index line.
Take one filelock per directory the store mutates (Dir and GlobalDir, sorted
by canonical absolute path) under the user lock root, bounded by a short wait
that returns a retryable StoreLockedError. The in-process mutex is still taken
first and the file lock second, one order everywhere, so the two cannot
deadlock.
CodeQL flagged the new snapshot parser for stat'ing a caller-supplied path
(store_snapshot_identity.go:44, 'This path depends on a user-provided value').
The parser now takes the name only; loadMemory keeps the stat where it always
lived and fills the missing timestamps itself.
snapshotMemoryRevisionInDir re-read the fact file and filed those bytes
under memory.Revision without checking that the file still carried that
id and revision. Another process publishing revision N+1 in that window
made slot N hold revision N+1's bytes: Restore(id, N) then returned a
body revision N never had, and revision N became unrecoverable. The same
file name re-created with a different id was filed the same way.
The bytes the function already reads are now parsed by loadMemory's own
parser (parseMemory, factored out rather than duplicated) and the snapshot
is refused with the existing "memory revision conflict: expected %d,
found %d" wording when the id or revision no longer match. The check runs
before any slot is created, so a refusal files nothing. SaveWithOptions
and RestoreArchived already return a snapshot error before publishing, so
no caller changes. Single-process behaviour is unchanged: the
process-local mutex makes the mismatch impossible there.
Tests cover the two cross-process probes (refusal plus nothing filed),
archive-and-recreate with a new id, an ordinary save/restore control over
several revisions, and a mutation that removes only the check.
Problem: PR #14086 conflicted with studio after the upstream listing-test repair.
Root cause: Both branches changed the same cache test. Upstream #14097 replaced
timing-dependent large fixtures with direct observation of cached replay, while
this branch added an assertion that derived summaries were persisted.
Fix: Merge studio at 58c80e3f9f, retain the upstream three-conversation corruption
fixture and preview checks, and assert the matching number of persisted entries.
Keep upstream commit history and the independent progressive-loading PR scope.
Verification: Complete sessionstore/sessionv4/historycatalog/backgroundwork
suites, focused cache/cancellation race checks repeated three times, real
resume/tree HTTP effects, 27 frontend interaction tests, and repository/Go lint.
TestV4ConversationsWithoutAListingCacheAreReplayedOnce called a second
listing a replay unless it ran five times faster than the first. A busy
macOS runner measured 411ms against 1.35s and failed a PR that does not
touch sessionstore.
Garble each conversation's log in place after the first listing, keeping
the size and modification time the remembered listing is keyed on: a
listing that remembered returns the same rows, one that replays loses
them. Three small conversations now do what sixty large ones did.
userTexts, the strings a copied command is exempt for because the user
wrote them, returned each user turn's provider-visible Content. That is
the composed turn, with the host's recalled memory, instructions and
policy around the person's words, so a page's command that reached the
turn only through a host block was taken for one the user typed and was
neither counted nor put to the person.
Read each turn through sessionstore.UserMessageText, the definition the
history readers share. This only narrows the exemption.
Problem: The workbench drag change failed the action census and sidebar
text-drop regression in the frontend CI suite.
Root cause: The new dragend listener had no action identity, and the drop
router read the workspace-path payload without checking its advertised MIME
type. Transfers returning unadvertised data could be routed as file paths.
Fix: Register file-drop.finish through listenAction, require the workspace
MIME type before reading its path, and cover plain text and URI-list drops.
Verification: Reproduced the action census and both MIME regression failures
before the fix. All 752 frontend test files and 6027 tests pass, including
77 focused action/drop/intake tests. Typecheck, census, census:verify,
production build, bundle fixture exclusion, gofmt, go vet, make lint,
builtin/boot tests, and make check pass.
Refs #12771
mechanicalFoldDigest, the stand-in when the summarizer cannot digest a
fold, listed each user turn's provider-visible Content. The oldest turns
a fold takes carry the host's date and standing-instruction blocks, so
the 240-character line went on <current-date> and
<project-instructions> and the person's words were cut off, and a turn
of host blocks alone was listed as something the user said.
Read each turn through sessionstore.UserMessageText, the definition the
history index and its snippets already share, and skip a turn with
nothing left.
Add the missing bilingual release-note fragment for PR #14082 so users
can discover that workbench files and folders can be dragged into chat.
Reference the related file-to-chat issue in both languages.
Verification: release-note rendering and Markdown reference tests passed
(12 tests), alongside the required pre-commit checks.
Refs #12771
Problem: the split listing and scheduling change needs its own user-facing release fragment.
Fix: add bilingual notes for progressive summaries, incomplete coverage and foreground priority.
Verification: release-note rendering and Markdown reference tests, make check and git diff --check pass; runtime code is unchanged.
Problem: the user-visible startup behavior had no release fragment.
Fix: document in Chinese and English that ordinary resource reads wait for the authorized connection, as requested in review of #14084.
Verification: release-note rendering and Markdown reference tests pass; make check and git diff --check pass. No runtime code changed.
Problem: Files and folders in the workbench could not be dragged into chat.
Root cause: Explorer and Git rows emitted no path payload, and composer
intake only recognized external files and ordinary text.
Fix: Emit a typed source-workspace path and route it through existing
reference validation. Preserve nested repository prefixes, clear cancelled
drag hints, and keep internal drags outside external-file deduplication.
Verification: Typecheck and 689 focused frontend tests passed, including
native-drop compatibility. Browser drag checks, make check, go vet, make
lint, builtin/boot tests, and both drop-handler Go suites passed.
Refs #12771
Problem: Sidebar discovery waited for full foreign transcript replay while
history indexing competed with interactive session restoration.
Root cause: Complete-list reads owned metadata replay, identity replacement
scanned growing message slices, and derived-data producers lacked shared admission.
Fix: Return explicit progressive coverage, share single-flight summaries and
validated caches, fold replay identities linearly, and retain cancellable scan
generations. Admit background work through one foreground-aware scheduler.
Apply foreground admission at the history route group so independent readers
inherit it without coupling their implementation to background work.
Verification: Storage/scheduler/catalog suites, real Boot/resume and tree/HTTP
effects, sidebar interactions, production frontend build, race checks and lint.
Keep benchmark cleanup registered even when its bounded deadline fails.
Split from PR #14077; no startup fixes or transcript pagination are included.
sessionstore.UserSaidText stripped only the host blocks prepended to a
persisted user turn, while UserMessageText also stripped the blocks the host
appends after the person's words (<execution-policy>, <memory-recall>). The
history index and its snippets therefore indexed recalled facts and the frozen
policy as if the user had said them, so a query for a fact matched a row whose
snippet could not contain it: the same divergence #14022 fixed for the leading
blocks, one definition still giving two answers.
Move UserPreviewText's pipeline down into userAuthoredText and let UserSaidText
add the plan-mode marker strip; UserMessageText delegates to UserSaidText for
both RawContent and Content. The appended cut keeps the existing byte-exact
suffix rule, so a complete block the person quotes at the end of their own turn
is still indistinguishable from a host-appended one and is stripped; the test
pins that limit instead of hiding it.
Bump historycatalog.TokenizerVersion to 4 because the indexed vocabulary
changes again. ensureTokenizerVersion wipes history_fts/documents/sources/roots
when the stored stamp differs and the manager re-registers its roots, so the
next open re-indexes them; Rebuild stamps what it built with. Covered by
TestTokenizerVersionMismatchClearsMixedProjection and
TestRebuildSurvivesTheVersionCheck.
Tests: TestFTSIndexesWhatTheUserSaidNotTheAppendedBlocks and the extended
TestIndexAndSnippetAgreeOnTheUserText / TestSnippetBuildersReadTheIndexedUserText
pins were red before the change. TestAutoRecallNormalTurnIsUnchanged and
TestUserSaidTextQuotedExecutionPolicy are controls, green before and after.
Problem: File delivery requires explicit opt-in even when cards are useful.
Root cause: The user-only configuration key defaults to false.
Fix: Default it on while preserving explicit off choices, read boundaries
and tool exclusions. Keep this policy separate from the settings UI PR.
Verification: Config round trips, complete config/settings/builtin/boot
suites, exact request goldens, schema stability, vet and lint.
The default request fixture gains 143 schema tokens. A newly opened session
may have one cold prefix; this does not mutate the live session's schemas.
Problem: The settings PR also changed the default tool surface.
Root cause: UI discoverability and default-on policy were bundled together.
Fix: Restore the existing default, registration fixtures and configuration
rendering, retaining the searchable user-only switch below Tool approval.
Keep UI wording and route tests independent of a future default decision.
Verification: Settings and transport tests, browser captures, Go suites,
frontend build, vet, lint and repository policy checks.
Problem: Preset modules could only be added one at a time, with two different actions both labeled Add module.
Root cause: A native single-choice picker shared its label with manual blank-ID insertion, and catalog choices did not carry source provenance.
Fix: Add a shared searchable grouped checkbox picker, authoritative catalog provenance, batch selection and parameter-safe draft updates. Keep manual IDs under an explicit advanced entry and document selection semantics.
Verification: 219 frontend tests, typecheck, production build, census, owning Go packages including boot and builtin, focused Desktop preset/composition tests, vet, lint, make check and documentation checks passed. Browser checks covered themes, responsive layouts and save/reopen with MockHub.
Problem: AI file delivery cards required a manual, default-off config edit.
Root cause: present_files had no settings surface or read/write endpoint, and
the config delta renderer treated false as the default.
Fix: Add a searchable user-only switch immediately below Tool approval.
Default the tool on, preserve explicit off choices, and keep registration
changes at session opening. Update wire contracts, documentation and the
provider-visible tool fixtures for the additional 143 schema tokens.
Verification: Config round trips, route grants, UI persistence/search/order,
real boot tool-surface effects, frontend tests/build, Go tests/vet/lint and
browser checks in light/dark wide and phone layouts.
Problem: A redundant arguments envelope makes usable tool calls fail required
field validation and can cause repeated model retries. The current CI security
scan also reports GO-2026-6809 in the pinned Go 1.26.9 standard library.
Root cause: Scheduling, policy checks, execution, and history had no shared
representation of effective arguments inside a schema-invalid wrapper. The
root and Desktop modules still selected the affected Go patch release.
Fix: Unwrap one unambiguous envelope only after full target-schema validation.
Use effective arguments for host execution, evidence, decision barriers, and
surface actions while retaining the original provider call. Select Go 1.26.10
for both modules so CI and packaged Desktop use the patched standard library.
Style the compatibility notice and original-input disclosure like adjacent
tool metadata, with a visible expand marker and keyboard focus outline.
Verification: The envelope regression suites, provider history, permission,
proxy, and interactive surface tests passed. With Go 1.26.10, govulncheck ./...
reported zero reachable vulnerabilities and go vet ./... passed. Frontend type
and interaction checks passed on the rebased PR. The final card style passed
frontend type checking, 16 relevant UI tests, make check, and local browser
inspection in light and dark themes with its disclosure closed and open.
reconcileRoot treated seen_generation<>generation as proof that a transcript is
gone. A row indexed by another process's persist event carries generation 0, so a
conversation indexed after this scan listed the directory was swept health=missing:
it vanished from history search and the incomplete-index warning stayed silent,
because the root's counts still looked complete. Measure absence with stat and
adopt live rows into this generation; only a transcript confirmed gone is swept.
Problem: the release fragment mentioned static HTML support without explicitly
calling out the change from automatic live rendering.
Fix: state in Chinese and English that HTML opens with scripts disabled and
desktop users select Interactive run to execute the page, as requested in review.
Verification: release-note script tests, make check and whitespace validation.
No runtime behavior, dependencies or test expectations changed.
Problem: frontend-next CI failed 25 assertions across action discovery,
orphan detection, HTML preview races and copy feedback consumers.
Root cause: TooltipButton was not declared as an interaction primitive; new
preview controls had no action identities, and the superseded image component
remained orphaned. Neighbouring tests still expected native title tooltips and
automatic HTML execution rather than the explicit Interactive run action.
Fix: certify and name the actual UI inputs, restrict tooltip blur to local
dismissal, remove the unused image renderer, and align consumer tests with the
new contracts while retaining copy failure/retry and browser race coverage.
The unnamed-input ceiling falls from 218 to 217 rather than being relaxed.
Verification: full frontend suite (749 files, 6010 tests); supplemental action,
tooltip and browser guards; typecheck and production build; go vet, make lint
and make check. No workflow skip, timeout increase or alert suppression.
Problem: delivery-card Add to conversation bypassed the workbench's reference
validation, allowing names with token boundaries or MCP ambiguity into drafts.
Root cause: Pane independently formatted @ references from display paths.
The kernel grammar preserves literal backslashes and only escapes space/tab,
so generic backslash doubling would change the selected file.
Fix: share one grammar-specific formatter, use structured relative paths for
workspace cards, disable unrepresentable references while keeping preview and
path copying, and remove the unused intake refToken helper. Keyboard navigation
and initial focus skip disabled menu actions.
Verification: 170 related frontend tests, production build/typecheck, real
browser insertion, kernel backslash round-trip and scoped-resolution tests,
go vet, make lint and make check. Four new rejection cases fail before the fix.
Problem: the presenter was registered for every install, and review requested
default-off model behavior, repository screenshot removal and PDF hardening.
Root cause: present_files used the unconditional builtin path rather than the
existing user-scoped conditional registration and provider-surface mechanism.
Fix: gate registration with tools.present_files, protect user scope across
project merges, preserve the setting on save, and expose the tool only after
opt-in. Restore the unconditional schema budget and remove PR screenshots.
Document PDF.js 6's removal of eval/isEvalSupported, keep scripting disconnected
and explicitly disable XFA rather than passing an unsupported no-op option.
Verification: boot.Build direct/proxy off/on/forbidden/disabled/excluded effect
matrix; user-scope and save/reload tests; document tests and typecheck; real
browser PDF navigation; go vet, make lint, make check and release-note tests.
Record the user-visible preview formats, responsive delivery cards and refresh fixes for PR #14031. Verified with make check and git diff --check; no runtime behavior changed.
Problem: AI-created documents could not be handed off as validated, durable
file cards, and binary files could not be read through the text editor route.
Root cause: delivery identity lived only in ordinary tool output, while file
reading, view selection and auxiliary actions lacked a shared preview path.
Fix: add a bounded root-confined reader, Controller and raw-byte preview APIs,
and host-confirmed present_files results projected through live and stored
history. Add lazy PDF, worker CSV and sandboxed HTML viewers, responsive file
cards, type icons, current-file state and accessible shared tooltips. Preserve
dirty drafts while surfacing deletion/revocation and avoid metadata reads per
keystroke. Keep the current workbench request-consumption contract intact.
Verification: root and desktop Go tests, go vet, make lint/check, frontend
typecheck, 498 related tests and production build. Browser checks cover PDF
navigation, CSV literals, menus, request consumption and responsive themes.
The builtin schema inventory grows by 143 measured tokens; boot effect tests
verify stable schemas within a turn for direct and capability-proxy calls.
Cause: candidateTexts.of loaded a candidate's transcript only to build a
snippet. When that load failed it set failed[path] and returned, dropping the
candidate with no error, no repair, and no status change. The row's ledger
digest and size/mtime still matched, so Indexed == Total kept certifying the
projection complete and the history tool answered "No saved session history
matched ..." for a query the session did match.
Fix: record a deterministic read failure with the shape the indexer already
uses for the same fault. RecordReadFailure flips the row from ok to corrupt
with last_error (existing columns), publishes the revision so Status reports
Indexed < Total, and re-enqueues the source. The periodic scan (#13923: a
non-ok source is always revisited) then re-reads it, so a restored mode heals
without a save.
Transient failures are not distinguishable from permanent ones at this layer,
so every load failure is recorded; retries stay bounded because a row that is
already corrupt is not re-queued per search, and the scan retries it once per
interval (five minutes by default). No wording, persisted field, schema, or
tokenizer version changed.
Tests: the audit's red test now passes (the tool prefixes the existing
"history index is still building" notice and status reports Failed 1,
Indexed 0/1); a readable control that genuinely matches nothing still answers
no matches with no alarm; and after chmod is restored the next ReconcileRoot
scan serves the content again. Mutation check: reverting only the fix leaves
the first and third red with the quoted false claim.
The FTS projection indexed raw user message Content while every display path
stripped the host's prepended blocks (<response-language>, <current-date>,
<project-instructions>, <available-skills>, <workspace>): a query for a
block-only phrase matched a row whose snippet could not contain it.
Move the strip into sessionstore.UserSaidText, the one definition of what a
user turn says, and call it from both sides: the indexer
(historycatalog/documents.go) and the snippet builders (history/search.go,
history/indexed.go). The history package's private stripComposePrefixes is
gone, so the two cannot drift again.
Scope is today's snippet behaviour: the leading blocks stripComposePrefixes
took, plus the plan-mode marker. Trailing <execution-policy>/<memory-recall>
stripping is deliberately not added; whether "the user said" also excludes
host-appended blocks is a product decision left for the PR body.
Bump historycatalog.TokenizerVersion to 3 because the indexed vocabulary
changes. ensureTokenizerVersion wipes history_fts/documents/sources/roots when
the stored stamp differs, and the manager re-registers its roots, which
re-indexes them; Rebuild stamps the version it built with. Covered by
TestTokenizerVersionMismatchClearsMixedProjection and
TestRebuildSurvivesTheVersionCheck.
Tests: TestFTSIndexesWhatTheUserSaidNotTheHostBlocks (red before the change),
TestIndexAndSnippetAgreeOnTheUserText, TestIndexKeepsPlainUserTextUnchanged,
TestSnippetBuildersReadTheIndexedUserText.
Problem: selecting a custom preset captures unused skill templates and
extension packages and waits for optional MCP startup.
Root cause: resource snapshot creation also served as the runtime permission
model, making selection cost scale with installed resource payload.
Fix: add live capability-grant bindings while preserving legacy snapshots.
Resolve current skill/MCP/extension/hook/command authority at use, defer
optional startup, retain stable provider contracts, and fence new event
and archive formats from older readers. Use lightweight status summaries
and per-instance extension call leases for safe concurrent replacement.
Verification: root and Desktop suites, full boot/builtin suites, affected
storage/controller/frontend tests, lifecycle race checks, frontend interaction
tests and typecheck, lint/vet/check, actionlint, and Windows cross-compilation.
Twenty-sample real switch measurements include build, persist and publication;
64 MiB unused templates remain outside the binding and stay below the local
500 ms P95 target. Native Windows execution is delegated to hosted CI.
#13954 let /resume delete a saved session with a confirming second
Ctrl+D. The merge of #13755 (d87d0d959) reverted it: its branch's
"Merge studio into a5/schedule-api" (ba48dfe0b) resolved files only
studio had changed to the branch's older copies, taking out the picker
code, its tests, the i18n catalogue and the release note.
Restore #13954's eleven files to their content on the studio side of
that merge; none has changed since.
#13956 gave user ! commands the login shell's PATH through
shellrun.CommandEnv, shared with the bash tool. The merge of #13792
(ac1a737ee) reverted it: its branch's "Merge studio into
feat/13111-origin-ask" (ed1dfebb9) resolved files only studio had changed
to the branch's older copies, deleting shellrun/loginenv.go, the Env
line in RunShellWith, the regression test and the release note, and
moving the probe back into bash.go. A ! command again ran with only the
host's PATH, so tools from nvm or Homebrew were not found when Studio
was started from a launcher.
Restore the nine files to their content on the studio side of that
merge; none has changed since.
Studio now requires every config key to be declared once with who may change it
and when it takes effect; the two keys this branch adds were not in that table,
so loading a config panicked with `setting.undeclared: network.dns`.
Both are `Guarded` / `OnOpen`, beside `network.proxy_url`: which resolver a
request goes through is a network boundary, so the model must not change it by
itself, and a new value applies on the next start rather than mid-session.
Also rebased onto studio, which the same push asked for.
Problem:
The native Windows link gate fails when capturing a junction source with
an internal junction, despite the ordinary symlink cases passing.
Root cause:
filepath.EvalSymlinks can retain a Windows junction alias and fail when
walking its child. It is not the repository's physical-identity resolver.
Fix:
Use fileutil.ResolveExistingPath for source roots, contained entries, and
single-file hook parents. Preserve containment checks, rooted content reads,
strict retained-directory verification, and filesystem error identities.
Verification:
Local source-capture and composition regressions and Windows amd64 test
compilation pass. Native verification remains owned by the Windows
composition-links gate, including the previously failing junction case.
The Windows runner's Git Bash has no shasum, and the grader's
`shasum -c ... || exit 1` turned "command not found" into the same exit 1 as
a changed file, so the reference solution failed there with nothing printed.
The SHA-256 manifest is now checked with hashlib inside the grader's Python,
over raw bytes, and a mismatch names the files that differ.
gradeSeed now carries the grader's output in its error: the CI log said only
"exit status 1", which cannot tell a missing tool from a failed check.
Verified locally: on a PATH without shasum the reference solution grades
clean; a CRLF Makefile and an edited Go file each fail and are named; the
pristine seed still fails. snapshot.sh --check, the corpus tests and
make check pass.
Problem:
The Windows smoke suite did not include agentpreset's snapshot tests and
symlink fixtures could be skipped without proving the reported platform.
Root cause:
The platform package list omitted agentpreset, and no fixture exercised
directory junctions. One portable assertion assumed Unix executable bits.
Fix:
Add native junction capture/restore, hook bundle, escape, cycle, and retained
snapshot rejection cases. Run nine focused source-link tests on Windows and
fail the gate for skipped or missing cases. Compare retained permissions to
the actual source filesystem mode.
Verification:
agentpreset and testenv suites, Windows amd64 test compilation, go vet,
make lint, make check, and actionlint pass locally. Native Windows execution
is explicitly pending the hosted composition-links step.
Problem:
Saved personas can fail when selected with inherited skills, extensions, or
hooks. MCP startup failures lack structured server identity and guidance.
Root cause:
Ordinary discovery accepts linked sources but composition snapshot capture
rejects them. MCP startup errors carry the server only in prose instead of
the module metadata consumed by the localized refusal boundary.
Fix:
Flatten contained source links through rooted reads, retain strict snapshot
verification, and preserve module identities and sanitized failure causes
through HTTP and Chinese/English rendering. Keep required MCP failures
blocking and preserve the original session on a failed selection.
Verification:
Full root Go tests, go vet, make lint, make check, frontend typecheck, and
198 relevant frontend tests pass. Browser fixtures cover failure, retry,
management access, and light/dark wide/phone error presentation. Regressions
cover source deletion and restart, prefix stability, revocation, containment,
snapshot tampering, and credential redaction.
The e2e fixtures finish in 5-15 requests, while a real task in this tree took
deepseek-flash 38. A change aimed at cutting round trips cannot show up on
the short suite, so this adds four long tasks whose workdir is a git archive
of commit 42636ddcb (pricing, ablation, agentgraph and five stdlib-only
leaves) plus a trimmed REASONIX.md, a Makefile, a module README and a small
benchctl consumer:
- rt-cache-write-rate: a CacheWrite rate threaded through cost, fingerprint,
card identity and the catalog, plus an Anthropic row that uses it
- rt-stale-skills: Markdown-only repair of three project skills with no
recognised verifier, the shape behind #13293's lingering stale_verification
- rt-rate-lookup-signatures: five lookups moved to (RateCard, bool),
CatalogKey and ModelRef across about 20 call sites, most in _test.go
- rt-plan-ablation-module: a five-step plan-first task signed off per step
snapshot.sh rebuilds every workdir from the pin and --check detects drift.
One pinned test compares CheckedOn with the wall clock and would fail every
workdir from 2026-11-20; base/ anchors it at the snapshot day instead.
Each task ships a solution/; the corpus is registered with the authoring
guards, which check that every seed fails its grader and every solution
passes. A deepseek-flash baseline (2 trials) solved 8/8 at 13-31 requests
once two grader false negatives in rt-stale-skills were fixed; the README
records it, along with a multi_edit prompt arm and a complete_step
description arm that did not move the request count.
The guard in perf/branch-alignment.mjs was not listed in the Studio workflow, so nothing ran it; the fix itself is user-visible and carries a fragment.
The unavailable refusal carried the module and detail only inside a nested
diagnostics list, so the window filled a generic sentence from a code alone.
Flat module/detail params now cross the boundary and both languages use them.
StartForSession called m.wg.Add(1) without a lock while Close could already
be inside m.wg.Wait() after cancelling the root: Add racing Wait at a zero
count is the misuse sync.WaitGroup documents. Under
`go test -race -run TestStartRacingClose -count=40 -cpu 1,2,8` on studio it
reports a data race and panics with "WaitGroup is reused before previous Wait
has returned".
The launch moves into launchJob, which decides under m.mu, the lock both
Close paths now cancel under, whether the job joins the wait group or is
refused. A refused job never runs: it ends Interrupted, its metadata is
written, and it is recorded and published like any other completion, so a
task recorder that saw it start sees it end. The goroutine body moves
unchanged.
Split out of #13947 at the reviewer's request so the jobs manager is
reviewed on its own.
Dirs resolves the path on the far machine and lists it; when the listing
fails but a stat says the path is a directory, the answer is the directory with
Unlisted set instead of an error. The picker then names why the rows are
missing and 'use this folder' works on the confirmed path. A path that is not
there, or not a directory, stays an error. Listing a directory this account
cannot read is fs.ErrPermission on the script layer, as it is on SFTP.
Probe took SFTP before any other step, so a server that refuses the subsystem
ended the test right after 'connection OK'. With the script layer behind
Client.Files it now completes, and the report carries sftp or exec.
pkg/sftp reports the refusal as an unnamed error string, so the subsystem is
now requested in sftpfs.New and a server reply of 'failure' becomes
ErrSubsystemRefused. Client.Volume prefers SFTP and, only on that sentinel,
serves the connection from execfs over the exec channel and remembers the
answer until the connection is replaced. Any other failure (dead link, refused
channel) is returned unchanged. The exec runner quotes the script and every
argument whole into one /bin/sh command line.
The WSL file layer was a fixed set of /bin/sh scripts taking every path as a
positional argument. It is now a package of its own that takes a runner, so
another transport can reuse it. Scripts name a missing program by exit status,
refuse a file over the transfer limit by type, and check the byte count before
publishing a write. List, Download and UploadAtomic are added, and the shared
scenario now covers names with quotes, spaces, a leading dash, newlines, shell
metacharacters and a path over 3000 bytes.
`flexString` existed byte for byte in the openai, gemini and responses packages.
One type written three times diverges at the next edge case, so it moves to
`internal/base/flexstr`, which is a leaf (it imports nothing under reasonix) and
is declared in `tools/repolint/layers.go` so the layering gate allows every model
package to use it.
No wire behaviour changes: a string code, a number, an absent field and null read
exactly as before.
Problem: The CUA readiness change expands #13720 beyond its reviewed
launcher-ownership scope.
Root cause: The native CUA startup race was repaired alongside the
launcher conflict resolution, despite having a separate runtime owner.
Fix: Revert cad3c58f8a on this branch.
The CUA fix and regressions are preserved on fix/cua-socket-readiness,
based directly on studio. Keep the original launcher fix and conflict
resolution unchanged.
Verification: The resulting tree exactly matches e67fe2ce96. Electron
tests pass (189 passed, one existing platform skip), as do go vet,
make lint, make check, and the builtin/boot package tests.
Document the user-visible startup and restart fix from #13985 in the unreleased Studio notes, in Chinese and English. Verified with make check and the Studio release-notes renderer tests.
Problem: Embedded CUA startup can fail while a valid driver is still
initializing its socket permissions.
Root cause: The driver binds its socket before chmod sets 0600, while the
host treats the intermediate permissions as a terminal startup error.
Fix: Wait within the existing readiness deadline, preserving socket-type
and identity checks and forbidding metadata requests before private
permissions are observed. Cover initial startup and abort replacement
through the shared launch loop.
Verification: Deterministic permission-ordering and deadline regressions,
file/symlink rejection, 188 passing Electron tests with one platform skip,
native macOS CUA acceptance, go vet, make lint, make check, and builtin/boot
package tests. Split from #13720 at the maintainer's request.
- The native bridge answers only the kernel's page and the one computer the
person scanned a code for; `openInApp` loads an origin once, and only that one.
- Non-web URLs are routed only for mailto: and tel: — an intent:// URL can start
any exported activity.
- The `reasonix://` filter that nothing handled is gone.
- A release with no key is unsigned instead of signed with the debug key, and the
debug build no longer wears the release key.
- The Gradle wrapper is pinned by checksum and the mirrors are out of the
repository.
- The scripts find the kernel and the page at the repository root, speak English,
and copy LICENSE from it instead of committing a second copy.
- The README is what a builder needs, and states what the permissions cost.
Problem: Native macOS CUA acceptance can reject a valid driver startup with
"Invalid private CUA socket".
Root cause: The pinned driver binds its socket before chmod sets 0600.
The host treats the intermediate mode as a terminal error. The previous
fixture creates its socket with private permissions immediately, hiding
the production ordering.
Fix: Treat non-private socket permissions as pending readiness within the
existing deadline. Keep non-socket rejection and identity validation, and
never request metadata until private permissions have been observed.
The shared launch loop also covers abort-triggered replacement.
Verification: Deterministic regressions force observation before chmod on
initial launch and restart, and cover deadline expiry, child cleanup, files,
and symlinks. Both permission regressions fail before the fix. Electron
tests pass (193 passed, 1 existing skip), as do native macOS CUA acceptance,
go vet, make lint, make check, and the builtin/boot package tests.
- The native bridge answers only the kernel's page and the one computer the
person scanned a code for; `openInApp` loads an origin once, and only that one.
- Non-web URLs are routed only for mailto: and tel: — an intent:// URL can start
any exported activity.
- The `reasonix://` filter that nothing handled is gone.
- A release with no key is unsigned instead of signed with the debug key, and the
debug build no longer wears the release key.
- The Gradle wrapper is pinned by checksum and the mirrors are out of the
repository.
- The scripts find the kernel and the page at the repository root, speak English,
and copy LICENSE from it instead of committing a second copy.
- The README is what a builder needs, and states what the permissions cost.
- The native bridge answers only the kernel's page and the one computer the
person scanned a code for; `openInApp` loads an origin once, and only that one.
- Non-web URLs are routed only for mailto: and tel: — an intent:// URL can start
any exported activity.
- The `reasonix://` filter that nothing handled is gone.
- A release with no key is unsigned instead of signed with the debug key, and the
debug build no longer wears the release key.
- The Gradle wrapper is pinned by checksum and the mirrors are out of the
repository.
- The scripts find the kernel and the page at the repository root, speak English,
and copy LICENSE from it instead of committing a second copy.
- The README is what a builder needs, and states what the permissions cost.
The kernel-side pieces this branch came with are their own PRs now, so this keeps
only `android/`. Rebasing onto studio also drops the mailbox delivery code and the
mailbox credentials that the review asked to remove.
Problem: PR #13720 conflicts with the current studio branch in the Electron
unit test suite.
Root cause: Both branches inserted independent regression tests at the same
location after the host startup failure tests.
Fix: Preserve the launcher ownership, relaunch handover, and closed-console
tests together with studio's credential-cookie restoration test. Retain the
automatically merged shell changes and updated Electron dependencies.
Verification: Electron unit and Linux packaging tests passed (189 passed,
1 existing skip). Formatting, go vet ./..., make lint, make check, and the
builtin and boot package tests passed.
TestEffectTruncatedSummaryOnPressureKeepsCodeAndRetries and
TestEffectModelSwitchKeepsTheCompactionProjection pin what happens turn
by turn around the compaction trigger, but the turn the trigger falls on
depended on how large the first request happened to be. Measured on
studio: about 550 more characters of prompt or tool schema flip the first
to "7 failed summaries" and the second to "switch spent 1 summarization
call", with no change to compaction. So any prompt or tool-description
edit, or a host whose tool surface is a few hundred characters larger,
failed them, and a change could pass on one runner and fail on another.
Each fixture now builds once, measures its first request the way the
prompt estimator does, and pads its system prompt up to a fixed size, so
the trigger and pressure turns are the fixture's. The sizes sit in the
middle of ranges where the outcome does not move: 30000-40000 characters
for the truncated-summary fixture (6 attempts throughout), 40500-48000 for
the model-switch fixture (3 folds before the switch, none after). With
0-5000 characters of extra first-request weight injected ahead of the
measurement, both now give the same result; before, both flipped near
550. The bounds are unchanged.
waitForDone returns on the turn-done event, which the controller emits before it drops its running flag, so ClearSession could answer "cannot clear while a turn is running". Under 8 parallel race-enabled processes at one CPU the unchanged test failed 3 times in 1200 runs; with waitIdle it failed 0 times in 1200.
Windows has a backend now, so Available() is true, but it enforces only after the
user turns on windows_integrity. Tests and the e2e bench that ask whether a default
configuration is confined use ConfinesByDefault (false on Windows), and a caller's
override to enforce (an observed run asks for one) is honoured on Windows only
when the user opted in, since enforcing edits permissions on their folders.
A relocation moves only the entries StateRootEntries names. The grant record sat in
sandbox/ outside it, so a move would have left the record behind while the
permission entries it lists stayed on the folders, with nothing to take them back.
The restricted token now drops every privilege but the traverse check, so backup
semantics cannot write outside the roots even from an elevated session. A SID is
synthetic only if no account resolves it. The launcher refuses to run under a
restricted token (a command's own launcher inherits it), grants only session temp
folders, hands the child just the standard handles, runs it in a job with UI
limits, and the default DACL and the grant mask are narrower. Links are not
followed at grant time and a link root is refused. REASONIX.md, the guide, the
Settings text and the release note list what is and is not bounded.
The host already confines bash's $TMPDIR to a per-session directory, the
file tools resolve a leading $TMPDIR to it, and the host deletes it when
the session ends. The model was never told the last two, so it verified
with `scratch=$(mktemp -d); ...; rm -rf "$scratch"` or wrote tests to /tmp,
and the host refused them (shell.delete_nonliteral, shell.delete_sequence,
write scope): 26 refused calls over 36 deepseek-flash benchmark runs, one
model round each.
A <session-scratch> turn block now says: put throwaway files at
`$TMPDIR/<name>`, spelled that way in bash and in the file tools, and do
not delete them. The host path is deliberately not projected: under bwrap
bash sees the directory at /tmp, and on Windows a backslashed literal does
not survive a bash command line, so $TMPDIR is the one spelling every tool
resolves. The block follows the dateAnchor/projectionDebt pattern: it rides
the turn and never the prefix, settles only when a turn reaches the runner,
is re-owed after a fold, retracts itself once if the directory can no
longer be created, is projected only for a manager the assembly supplied,
and is inherited across a rebuild. Its bytes carry no per-session value,
so two sessions over the same state project identically.
No shell-safety rule changes; the effect test asserts that a recursive
delete through $TMPDIR or a mktemp variable is still refused.
supersedeStandingState judged only the fold region's kept turns, so a
standing block the region held once survived beside newer copies in the
recent tail. A fold whose region kept a single user turn dropped nothing:
TestEffectFoldDropsSupersededStandingState went from 3 <workspace> copies
to 4 across such a fold once a turn block shifted the fold one turn
earlier. partitionFoldWith now takes the tail, and the tail's user-turn
blocks are marked seen before the kept slice is walked newest-first.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Every CompactionDone forgot what the model had been told -- the date,
the skills listing, the project instructions and rules -- so the next
turn sent them again. A truncated or refused digest also emits
CompactionDone, with Messages == 0: nothing was summarised and the turn
that carried those blocks is still verbatim, so re-sending them only
grew the prompt the next attempt had to fold, by one skills listing per
failed fold.
The producer already says which case it is: a successful fold reports
Messages = len(fold) and emitCompactionAborted reports 0. The sink now
reads that instead of treating every completion as a fold.
TestEffectTruncatedSummaryOnPressureKeepsCodeAndRetries goes from 5 to 6
failed summaries over its 10 turns: the prompt no longer grows by a
re-sent listing after each failure, so the pressure fold lands a turn
later. That is inside the test's bound of 6 and is the growth the retry
gate is meant to read.
Inside run_script, call("read_file", ...) returned the model-facing form:
every line prefixed " N→", CRLF folded to LF, the default 2000-line page,
and — past 32 KiB — a window with a "read on from here" marker. A script
that read a file, edited the text and wrote it back put all of that into
the file. With deepseek-flash on migrate-signature-3packages and
agent.code_mode on, one bulk rewrite corrupted 36 files and cost ~15 rounds
to repair.
Two layers each rendered for the model a result whose reader is a program:
- read_file: a call another tool made (tool.IsNested) gets the decoded text
verbatim — line endings and final newline kept, no default limit, offset
and limit selecting whole lines. The model's own read is unchanged.
- nestedInvoker: the calling tool reads what the tool returned, not the
bound fitted to the model's context. The nested card still shows the
bounded form; toolOutcome.keepFull records both originals in one place.
write_file already preserves the existing file's encoding, so a BOM file
round-trips once the read stops adding to it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
model_rounds and round_outcomes treated every call run_script made as a
top-level dispatch, so each inner result reopened a model gap and each inner
dispatch closed one. One recorded code_mode run with 24 executor requests
reported a median of 87.5 rounds. Nested calls carry no ParentID — only an id
derived from the caller's, inside its dispatch→result window — so the digest
had nothing structural to tell them apart.
The nested id is now declared in internal/contract/tool (NestedCallID /
IsNestedCallOf) and the agent's nestedInvoker builds ids through it, so the
producer and the reader share one definition. e2ebench's callNesting tracks
open top-level calls and treats subagent calls (ParentID) and invoker calls
(derived id) alike: neither cuts a round, joins a batch or adds tool time.
roundEnds and splitAtCorrect use the same tracker. NestedCalls reports how
many there were, and ModelRounds' doc now says what it counts.
With -trials N every trial wrote <task>.trajectory.jsonl, so only the last
survived and every trial's digest read it. Trials now write
<task>.trial-N.trajectory.jsonl (legs: <task>.trial-N.segM...); a single
trial keeps the old name. sft mode resolves the task id through
trajectoryTaskID, which strips trial and leg marks; traj, serve and barrier
modes glob *.trajectory.jsonl and label each file by its stem, so trials
read as distinct entries.
Cause: the digest had no identity for tool-made calls; the trajectory path
had no trial in it.
Blast radius: cmd/e2ebench digest and file naming; nestedInvoker's id string
is unchanged (same "<parent>.<n>" format, now built by the contract helper).
Neighbouring behaviours tested: subagent ParentID exclusion, segment paths,
answer-isolation audit across legs, existing round decomposition tests.
Why this layer: the id convention is owned where nested calls are created;
the digest reads it rather than restating the format.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Problem: Windows CI rejects the retained-hook source-deletion fixture as a
credential-bearing launch before reaching its restoration assertions.
Root cause: credential projection scans the complete quoted shell string as a
generic value, splitting a Windows drive colon inside a quoted file argument.
That creates an unmatched quote and falsely marks the command as sensitive.
Fix: reduce shell launches to static arguments before credential inspection,
preserve exec and legacy argv launches, and continue refusing literal token and
header arguments or shell forms that cannot be reduced safely. Keep source
commands and retention/restoration semantics unchanged.
Verification: regression reproduces both quoted Windows path failures before
the fix, while exec/legacy argv and token/header refusal cases stay covered.
Focused retention/security regressions and their race run pass; full builtin
and boot tests, gofmt, vet, lint and changed-file checks pass locally.
Problem: CodeQL reported new path and allocation alerts, and Windows repeatedly
failed the retained-hook restoration test.
Root cause: Checked selectors were still passed onward as raw request values;
some generated paths and capacity hints did not expose their bounds at the IO
owner. The HTTP DTO also carried an internal inherited binding. The hook fixture
fed Windows backslashes through Bash resource parsing.
Fix: Resolve import/export selections to registered values, forward checked
relative names for resume/migration, and guard the generated session slug before
joining it to state storage. Use rooted declaration/resource reads and copies,
bound resource reads, and avoid unchecked capacity addition. Pass inherited
bindings outside the wire DTO. Normalize and quote the hook fixture path, including
spaces, while retaining source-deletion and exact-binding assertions.
Verification: Hosted alert locations reproduced locally with the path-injection
and allocation queries; before/after query results and the base scan are retained
separately. Boundary regressions fail on the original head and pass with the fix.
Owning HTTP, store, config, hook, sidecar, plugin, specification and encoding suites;
full builtin/boot tests; gofmt, vet, lint and changed-file checks. Hosted Windows
and CodeQL results remain required for exact-head acceptance.
Problem: The composer exposed the new mode picker to users who had not opted in,
contrary to the maintainer's product-visibility decision.
Root cause: Removing the persistent visibility preference also removed its
settings/search action and the RunLine mounting gate.
Fix: Restore the original on/off storage key and default-off behavior. Add the
switch and concise explanation to Persona management, restore settings search
and the action, and gate both mounting and catalog reads. Preserve card defaults,
catalog/editor design, persona copy and current conversation bindings. Failed
storage writes retain the previous choice and report an error.
Verification: Missing/disabled/unknown and toggle regressions fail before the
fix. Focused Chinese/English navigation, persistence, storage-failure and binding
tests pass. Full frontend: 729 files / 5,857 tests; build/typecheck, UI census and
ten identical fixture reports; gofmt, vet, lint, make check, builtin and boot
suites. Browser visual acceptance remains unavailable under the URL policy.
Problem: Built-in preset names and descriptions read like configuration modes
rather than collaborators, and their relationship to Harness prompts was unclear.
Root cause: Technical copy was reused across the catalog, composer and persona
editor without explaining the existing short-identity prompt adaptation.
Fix: Rename the built-ins to Everyday partner and Lightweight partner, update
their bilingual descriptions and guidance, and derive the unbound picker label
from its canonical catalog definition. Keep IDs, persona bytes and selections
unchanged, retain historical translations, refresh revision metadata and document
the pinned official Harness reference and native prompt boundary.
Verification: Frontend 729 files / 5,852 tests and production build/typecheck;
contract, store and HTTP suites; gofmt, vet, lint, changed-file check, builtin and
boot suites. Fix a discovered worktree test race by waiting for the asynchronous
checkout state instead of the creation title shared by initial and checkout
phases. Browser visual acceptance remains unavailable under the URL policy.
Problem: Built-in presets use flat rows, choosing a future task mode requires
repetition, and persona fields are hidden behind an advanced disclosure.
Root cause: Management navigation and new-task launch intent have no shared
explicit preference; the editor groups basic persona fields with identifiers.
Fix: Match the selected two-column card hierarchy with badges, IDs, descriptions
and separate details, usage and copy actions. Persist an explicit built-in card
choice in a new machine-local key and inject it through the local/remote open
owners only for new tasks. Explicit request choices override it; restoring
history keeps its binding. Existing server preferences alone remain inert.
Expose both persona textareas, runtime context, scope and ID directly, rename
settings navigation to Persona, and keep literal persona fields authoritative.
Verification: Full frontend suite plus real MockHub launch effects, blank-pane
reuse, local/remote history, explicit overrides, storage compatibility/refusal,
cross-window refresh and independent footer actions. Production build/typecheck,
UI census/fixtures, format, vet, lint, changed-file and builtin/boot checks.
Browser-rendered visual comparison remains blocked by the URL policy, as
recorded in design-qa.md.
Problem: "After the first-party guidance" exposes prompt-construction jargon
without explaining what users should enter.
Root cause: Both persona editors use a position-based label; the composition
editor has no suffix helper text.
Fix: Label both suffix inputs "Additional instructions" and explain that the
text follows the built-in system instructions and may be left blank. Add the
Chinese/English copy and associate the helper with each textarea. Keep the
existing literal suffix binding and update affected interaction expectations.
Verification: 24 focused persona, preset and localization tests; production
build/typecheck; formatting, vet, lint, changed-file checks and builtin/boot
suites.
Problem: The catalog's create icon and label stack inside a cramped button.
Root cause: Studio icons render as blocks while the button lacks a horizontal
layout; the title and description also share a flex item beside the button.
Fix: Give the button an inline-flex row with no wrapping, a 36px minimum height
and 16px horizontal padding. Align title and action in one grid row and place
the description across the next row, retaining the primary theme colors.
Verification: 27 focused frontend tests, production build/typecheck, formatting,
vet, lint, changed-file checks, and builtin/boot suites. The local preview stays
running; browser observation remains blocked by its URL security policy.
Problem: Preset browsing, creation and the Standard persona were mixed in one
long Settings page, hiding the save flow and exposing configuration too early.
Root cause: Catalog selection and editable drafts shared one inline surface,
without an owner-level guard for Settings navigation.
Fix: Follow the selected full-content editor reference. Group the catalog by
scope; open a separate editor with common fields first, expandable capabilities
and advanced options, and a persistent Cancel/Save footer. Preserve literal
persona bytes, revision checks, import/export/delete and session bindings.
Guard dirty exits across return, Settings close, category and search paths.
Keep pending saves mounted and reject stale import results. Reusable switches
declare button type so toggles cannot accidentally submit the editor form.
Verification: Full frontend suite: 727 files, 5843 tests; focused editor,
navigation, pending/refused save, import ownership and binding regressions;
production build, typecheck, UI census and fixture verification; root format,
vet, lint, changed-file checks and builtin/boot suites. Browser visual acceptance
remains blocked by the local URL security gate and is recorded in design-qa.md.
Every open tab keeps its pane mounted, and a hidden pane was only
visibility:hidden, so it still took part in layout. The rail width is animated
per frame, so each frame laid out every pane's transcript again. With three
panes the median frame at 4x CPU slowdown went from 94 ms to 18 ms.
Problem: the mode picker required a separate settings toggle before users
could discover or use it.
Root cause: both the picker and its run-status row depended on a persistent
visibility preference unrelated to the conversation's selected mode.
Fix: render the picker directly, remove the retired setting, preference state,
search entry and action, and update the bilingual documentation. Standard
remains the default; displaying the catalog never selects a composition.
Verification: failing-first tests for absent and disabled legacy preferences;
default Standard with a saved Minimal preference; picker, settings, remote-open
and settings-search interaction suites; frontend typecheck, full suite, census,
fixture verification and build; browser preview; go vet, make lint, make check,
builtin tools and boot assembly tests.
Problem: Publishing Studio did not regenerate its website changelog.
Root cause: Product publication and the website deployment had no explicit handoff, and Pages only accepts runs from the website branch.
Fix: Dispatch the website workflow after stable publication with the required version and a unique request ID. Wait for that deployment and propagate failures. Register dispatch on the default branch without enabling deployment from it, and use the workflow commit when recovering old tags.
Verification: All 13 publication tests, shellcheck, signing workflow contracts, go vet, make lint, make check, and the required builtin/boot suites pass.
Problem: the custom-preset form grouped storage scope and ID under one label,
leaving their purpose unclear.
Root cause: two independent controls shared the Preset ID field container
without visible labels or usage guidance for each control.
Fix: give each control its own label and short translated description, and
associate the descriptions with the controls for assistive technology.
Verification: AgentPresets interaction tests, frontend typecheck, production
build, census and fixture verification; browser layout and console checks;
go vet, make lint, make check, builtin tools and boot assembly tests.
Problem: PR #13613 conflicted with the current studio branch.
Root cause: both branches extended session status, ACP configuration, and
project dispatch while the shared stylesheet and session-new function reached
their existing size limits.
Fix: retain composition bindings alongside tariff status, apply the ACP
approval ceiling, and reuse pending dispatch panes without dropping inherited
composition bindings. Move the picker styles to their own sheet and compose
the ACP config projections without increasing the ratchet budgets.
Verification: frontend typecheck, 5,830 frontend tests, production build,
census invariants and fixtures; go vet, make lint, make check; ACP, HTTP serve,
composition contracts/storage, session storage/control, delegation, builtin
tools, and boot assembly tests.
Merged studio 3d9f9ff74c into PR head
ae6a9de150.
The kernel refuses every request without it, and the cookie belongs to the
host rather than the port, so a removal left the window refused until a
relaunch. The shell holds the token and puts it back, logging the cause.
.gitattributes pins *.go to eol=lf but not *.golden, so a Windows checkout turns
the byte-compared goldens into CRLF and TestCompareOutputUnchangedWithoutUnaccountedRuns
fails there. Pin *.golden to eol=lf as well.
The macOS tray loaded the full-colour tray.png, so the menu bar showed the
whole app badge. A Template-suffixed black-on-transparent pair lets macOS tint
it for light and dark bars; Windows and Linux keep their files.
aggregateArm dropped an unaccounted run's wall before adding it to
s.WallMs, so compare's Solved/hour and global wall/task divided by a
smaller wall than the suite report uses for the same -json artifact:
report.go adds every run's wall, and a killed run's wall is measured
unconditionally in main.go (only its tokens never landed). The
per-class accumulator ran before the skip, so one section printed two
populations, and the sign of "which arm is cheaper" could flip.
WallMs is now the all-run wall and AccountedWallMs the accounted
subset: Solved/hour and wall/task divide the all-run total, while
"Wall seconds / solved" keeps dividing accounted wall by accounted
solves because that denominator is solved runs. No new persisted
state, and neither how Unaccounted is decided (segmentrun.go) nor the
suite report changes.
This is the residual of #13446 applied to a field that is NOT
unmeasured. That PR decided which runs are unaccounted and kept their
spend out of the token/cost ratios; it never touched wall, and this
change does not revisit that decision. It fixes the wall population
those adjacent rate lines divide, measured for every run including the
ones #13446 excluded elsewhere. Not a re-report of #13446.
Verified: both probes are red on 53a8db913 (compare Solved/hour 120.0
vs the report's 3.9; global wall/task -30.0s vs the class row's
+870.0s) and green after; a no-unaccounted fixture renders compare
byte-identically (goldens); "Wall seconds / solved" divides accounted
wall by accounted solves under either rate line reverted to the
accounted total, while the matching probe turns red; the suite
report's Solved/hour convention is pinned unchanged. gofmt,
GOOS=windows go vet ./cmd/e2ebench/, go test ./cmd/e2ebench (only the
pre-existing environment TestStateHomeIsOutsideWorkdirParent fails,
identically with this change stashed), make check (repolint: clean),
and golangci-lint run ./cmd/e2ebench/... (0 issues) pass.
Cause: a transcript that could not be read was stamped health=corrupt while its
root was left state=ready with the failing size/mtime signature. Both automatic
repair paths trusted that row: reconcileRoot returned before indexPath ran and a
restart reopened the same ready row. Equal size/mtime says nothing about
readability, so a restored mode or a remounted transcript was never revalidated;
status stayed Indexed 0 / Total 1 until a save or a manual `catalogs reindex
history`.
Fix: chose the scan's skip predicate — reconcileRoot only short-circuits when
state=ready, the signature matches, indexed==total, and no source of that root
is non-ok (health<>'ok' or missing_since<>0). That is the decision that
prevented indexPath from ever being consulted. The same rule is applied at
source granularity: sourceProjectionUnchanged now also requires the stored
health to be 'ok', because after the root is rescanned a legacy source with no
ledger digest (digest=="") would otherwise be certified healthy from its
fingerprints without being read.
Fast path kept: a root whose last scan was complete and healthy still returns
before any transcript is opened, so the five-minute tick remains a directory
walk plus stat, not a re-read of every transcript. A source that stays corrupt
is retried once per scan (one failed read per corrupt source per tick) and
scan_generation advances exactly once per scan, so a permanently unreadable
transcript cannot wedge or loop the worker.
Tests: the audit's TestCorruptSourceIsRetriedOnceTheTranscriptIsReadable (red
before: health stays corrupt after the transcript heals) plus controls for a
healthy unchanged root still skipping without a read (scan_generation, revision
and document ids unchanged; a now-unreadable transcript keeps health=ok) and for
a permanently corrupt source retried on every scan while its healthy sibling is
not re-read. Permission fixtures skip on Windows and root, and every catalog in
these tests points at a temp path.
golangci-lint's intrange rule rejects a C-style loop whose index is never
read. Both loops in the new cases just repeat an emit, so they take the
Go 1.22 form. Behaviour is unchanged.
Store.MigrateV2WithVersion returned on the first fact whose os.ReadFile
failed, so facts later in the same directory and every later directory
were never migrated. MigrationReport carried only { Migrated int } and the
boot caller discarded it, so nothing named the skipped files: the audit
observed {Migrated:1} with zeta and omega never reached.
The contract says otherwise. docs/SESSION_MEMORY_RETRIEVAL.md:259-261: a
read condition is "reported instead of returning partial lists or
calculating a revision from incomplete history". :268-269: "These
conditions ... do not establish data deletion." The pass therefore records
each failed basename in MigrationReport.Unreadable, sorted, and continues
with the rest; boot's warning names them instead of printing only the
first path.
A per-fact read failure returns a nil error on purpose: the field is the
channel for "skipped", and folding it into err would make a skipped fact
indistinguishable from a pass that could not run. Directory and index
failures still return an error, because those leave the pass's own work
incomplete.
Tests cover: the audit fixture (alpha readable, beta chmod 000, zeta after
it, omega in the second directory) now migrating alpha/zeta/omega with
Unreadable=[beta.md], cross-directory sort order, a fully readable control
with no error and an empty list, the repair path for an already-migrated
fact behind an unreadable one, and the boot warning naming beta.md.
Review asked for comments to describe what holds now and leave how the code
got here to the commit history. The behaviour is unchanged.
The history this replaces, kept where it belongs:
* Leaving Source in the zero-probe kept ToolProgress on the passthrough
path, so the first attempt at merging progress merged nothing in
practice. Source is identity — it names the emitting party — so it
belongs in the key, and keeping it there is also what stops a planner's
text merging into an executor's.
* The e2e run over a real kernel is what caught that; the unit tests did
not, because they emitted deltas with no Source at all. Both are now
covered: TestCoalesceMergesDeltasCarryingSource and
TestCoalesceSeparatesBySource pin the key's shape.
Cause: persistCompactionStateLocked reached SaveCompactionState, a blind
atomic overwrite carrying no writer identity, while transcript saves fail
closed with ErrSessionWriteAuthorityMissing. A controller in the
stale-authority window (rebind, controller replace, lease generation change)
therefore installed and published a fold whose transcript save was refused:
it replaced the owner's installed digest and paid for the same span twice.
Enforced in the store rather than in the agent: (*Session).SaveCompactionState
presents exactly the permit save.go presents (requireWriteAuthorityForSave,
including BeginSave's in-flight accounting that lease Release waits for), so
there is one adjudication of the permit and no second check re-reading
WriteAuthority() in the agent. Every write to the owner's sidecar already
funnels through the agent's single persistCompactionStateLocked, which now
calls that method. The raw SaveCompactionState remains for paths that write a
fresh recovery lane and for low-level CAS tests that never bind authority.
The CAS on (generation, projection_version, covered_prefix_hash) is untouched:
this is an additional gate, not a replacement.
On refusal the typed authority error propagates as
errProjectionNotPersisted: commitSummaryProjection restores the previous
projection, so nothing is installed or published, and the next attempt under a
valid permit folds once.
Tests: internal/runtime/agent/compaction_sidecar_authority_test.go proves the
no-authority refusal (sidecar and digest unchanged), the authorized control
(one summary, one digest), and that a refused publish costs one normal retry,
not a second summary.
reanchorPattern stripped the backslash from \!name and \#name before
handing a root pattern to the matcher, which then read the first as a
negation and the second as a comment. Files named !name or #name that git
ignores through the root .gitignore, .git/info/exclude or the global
excludes file were therefore searched and listed. Nested .gitignore files
were not affected, since their patterns are re-anchored under /dir/.
Keep the escape on root patterns when it guards a leading ! or #.
The global ignore file was taken from the first core.excludesFile found:
$XDG_CONFIG_HOME/git/config before ~/.gitconfig, and the first value in a
file. Git reads the files in that order and keeps the last value, so a
user's ~/.gitconfig setting lost to the XDG one. The value was also used
nearly raw, so an inline comment or escaped backslashes produced a path
that does not exist, and the user's global ignore rules were silently
dropped.
Keep the last value across files and within a file, and decode the value
like git: strip unquoted # and ; comments, remove quotes, and unescape
\\ and \". Other backslashes stay as written so raw Windows paths still
work.
useRevealed paces a streamed answer on the frame clock and keeps its cursor
in a ref. The effect returned early when the cursor had caught up and only
handled a shrink inside step(), which the early return made unreachable. A
replacement shorter than what was already revealed — regenerate, rewind,
history restore — therefore left the cursor above the new length, and every
later growth below it was short-circuited: the typing animation silently
stopped for that message while the text itself still arrived in one piece.
Handling the shrink before the early return fixes it. The cursor comes back
with the text, so the next growth is paced exactly like the first one.
The existing three tests did not cover this: they only grew the text. The
new case fails against the old code with "expected 0 to be greater than 0".
A shell emits one ToolProgress per output line, so a build log crossed the
coalescer as one frame per line. Downstream that filled a subscriber's
64-frame queue and the hub then dropped answer text beside it, because text
and progress are both droppable. Measured on a real 500-line command: 285
frames unrepaired, 2 with this.
streamDelta replaces isStreamDelta and returns the key a burst merges under
plus its payload, so progress joins text, reasoning and compaction progress:
* the key carries Source. Every progress chunk an executor emits sets
Source="executor", and leaving it in the zero-probe kept the event on the
passthrough path — the first attempt at this merged nothing in practice.
Keeping it in the key also stops a planner's text merging into an
executor's.
* the key carries the call id, so two tools reporting at once never land
in one another's card.
* progress carrying more than ID+Output — a sub-agent's status channel sets
Name, ParentID and a rate — fails the probe and still crosses untouched.
* progress without an id is not buffered under an empty key.
Verified against a real kernel over SSE: a 500-line command emits 2 frames
that concatenate to the same 500 lines in order, and a 3000-line command
emits 4. Alternating text and progress is unchanged (the key switch flushes
either way, so that burst still cannot merge).
The kill test proves the leftover is removed and the live fact is untouched, but
nothing pinned the branch itself: forcing an active id down the rename-back path
now fails because ListArchived lists the live fact as archived again.
RestoreArchived moved the archive entry away only after the active file
and its index line were published. A process killed in between left one
id both active and archived, so ListArchived reported a live fact and a
retry was refused as "already active".
The entry now moves to a hidden .restoring-<id>-<stamp> name before the
active file is written. The stamp is the archive time, so the name
carries both the id and the original archive name with no journal and no
content comparison. Every restore sweeps leftovers first: an active id
means the dead restore had published, so only the redundant copy goes;
an absent id means it never published, so the entry is renamed back. A
failed sweep reports its error and publishes nothing.
Resolve .reasonix/reasonix.toml only when the root project config is absent. Share selection across loading, local edits, installs and repair; retain project-scope restrictions and reject symlink escapes.
Fixes#6007
chmod 000 does not remove read permission on Windows, so the fixtures that make
MEMORY.md unreadable would fail their own self-check there. The classification and
aside behaviour are unchanged on Unix; the directory-named case stays cross-platform.
After an unreadable index is moved aside and the derived index republished, the
store must list exactly the fact files and the index must not name the aside.
MigrateV2WithVersion is the only entry that recomputes the derived
MEMORY.md, and it returned readIndexIn's error. Every mutation path reads
that same file first, so a restored store whose index has a foreign owner
or a changed ACL froze every memory write for good: boot only warned, and
the docs' promised re-index has no other entry.
Classify the read error by identity: only fs.ErrPermission (permission
denied, a foreign owner, an ACL) and a typed not-regular-file error (a
directory named MEMORY.md) reach the rebuild. Every other error -- an I/O
failure, a resource limit, a momentary lock -- still returns, because the
file may be perfectly good and republishing would overwrite hand-written
text a retry can still read.
Before publishing, atomically rename the unreadable entry to a stamped
name in the same directory (MEMORY.md.unreadable-<unixnano>), so the bytes
survive for a later repair; a rename needs write access to the directory,
not read access to the file. If the rename fails, return the original
error and publish nothing. flushIndexIn keeps the readable path, so
hand-written text is still preserved line for line when it can be read.
No new persisted state beyond the aside copy; the MigrateV2 signature and
boot call site are unchanged, and an unreadable store directory or fact
file still aborts the pass.
Tests: the audit probe is red before and green after; controls cover the
stale readable index, hand-written survival, byte-identical healthy boot,
missing-index rebuild, and a later save succeeding after the repair. New:
a transient read error leaves the index byte-identical (red against the
previous revision), the unreadable bytes survive under the stamped name, a
directory index repairs, and a failed aside rename publishes nothing. Each
new guard was reverted alone to confirm its control goes red.
The turn-start workspace snapshot was awaited with no bound, so in a very large
workspace (a drive root: 7 s on a warm SSD for the first turn, 0.5 s after) the
first tool could not start until the walk finished, and a command's own
timeout_seconds, which only starts when the command does, never covered it.
The wait now ends at snapshotWait, the bound the end-of-turn snapshot already
has: the walk is cancelled, the snapshot records itself incomplete, and the
bundle says the turn's changes were not compared.
The row sized the name, the model chip and the status to their content and gave
the task only what was left, so with the chip present the task was squeezed to
8 px at a 360 px panel and 0 px at 280 px. Below 460 px of the list's own width
the row now takes up to three lines: name and status first, the task second,
the chip third.
profileForSkill set only the name and count, so a skill sub-agent (explore,
locate, run_skill, read_only_skill) showed no task while a task sub-agent
showed its first line. The label helper moves to the event package, next to
the Profile it fills, and both producers use it.
A task text that names a file in the host's own .reasonix/attachments directory
now gives the delegate that image (judged by the resolved path, read through the
same checks an @-reference gets), on the vision role with the note that it rides
the message, including on a follow-up turn that carried no image. The
read_only_skill runner took the turn's candidates without routing the model or
saying so, so a text-only skill sub-agent dropped them silently.
Each delta did s += piece on the row's text, so a long stream cost time and
memory quadratic in its length: 20000 progress events of one command (4.5 MB)
took about 13 s of the UI's own time. The text now grows in a buffer kept beside
the row, trusted only while the row still holds exactly what it last handed out.
The product keeps its 15 second bound on one ssh -G; the source now carries it
as a field with that default, and the test that runs the installed ssh sets it
to testenv.Budget, a bound that only reports a hang.
internal/session/control/git_push.go measured 0 from the control package's own
tests, so declaring it sensitive recorded a floor of 0 that could never fail.
The tests cover the trust refusal for both calls, a repository that is not
there, a branch with no upstream, and the commit lock released after a refused
push; the floor is now 100.
Two project folders under one git repository share a writer domain on purpose
(the lock and the visibility of each other's claims), but an undeclared writer
claimed the domain root, so sibling projects waited on each other for bash and
MCP writes, and the refusal read held X, requested X. It now claims the
selected folder; a session on the repository root still overlaps both, and a
declared path that cannot be resolved still claims the whole repository.
Problem: trusting a publisher whose plugin source has a mistyped commit
reports a misleading local-or-unsafe-source installation failure.
Root cause: plugin source validation collapsed its refusal into the generic
bad-source class before the HTTP and presentation layers could explain it.
Fix: preserve a distinct invalid-plugin-source error through publication and
installation, translate its complete-commit requirement in both languages,
and align the shared client validator with existing registry path checks.
Verification: Go vet, repository lint, owning market/serve tests, the required
builtin/boot tests, 391 frontend market/i18n tests, and TypeScript checking.
The push is the one git action Reasonix performs that leaves the machine, and
the host-only route list decides which requests a paired device may make. Both
now demand review plus security_review and are held to the coverage they have:
gitpush 84.0, serve/git_push.go 77.0, serve/host_routes.go 93.0, and
control/git_push.go 0 (its behaviour is covered by the serve and boot effect
tests, which the per-package gate does not count).
An Auto Guard card and a call that asks before it writes now give the claim
back for the wait; the lease records granted claims so the resume can name
who took the ground, the model gets workspace.claim_contended and the receipt
workspace_claim_yielded. A blocked acquisition ends at a user-set maximum wait
with workspace.claim_timeout.
closeBootTestHistoryCatalog gave the history and usage catalog closes a
fixed 5 s. Closing a SQLite catalog, and waiting out an open still in
flight, can take longer on a loaded Windows runner, and the posture
effect test failed there in cleanup with "close usage catalogs: context
deadline exceeded". The wait only exists to report a hang, so use
testenv.Budget; a close that never returns still fails the same way.
A child that shares the parent's provider starts on the depth in force now, and one
that inherits the parent's model without an effort of its own is built for it, as a
rebuild would have given. The status line reads the depth the session is on.
A provider that exposes the depths it accepts in Request.EffortOverride
(provider.RequestEffortProvider; the OpenAI-shaped adapter does) lets a session change
depth without a rebuild: the controller sets a session depth every main request
carries, updates its running selection, and accepts the change mid-turn. Serve (and so
the terminal and the desktop) and ACP both apply it in place and keep their rebuild for
a depth the provider does not take per request. The level is still saved to the user's
config so a restart opens on it.
Problem: the advancing Studio branch conflicted with the mod-turn settlement
test and changed queue pause and child-dispatch reporting contracts beside the
reviewed composition work.
Root cause: both branches added turn-completion waiting, the combined inbox
owner exceeded the file ceiling, and the effort audit still expected concrete
runtime identity in dispatch profiles after Studio restored selection profiles.
Fix: integrate exact Studio da9fbe7307 normally.
Keep both queued TurnDone and durable idle settlement. Group pause operations
in inbox_pause.go without behavior changes. Assert requested selections on
dispatch events and concrete identities on status events, retaining real
provider effort, rejection and persisted-child checks. Preserve landed queue,
billing, phone and remote protection work and the default-off composition UI.
Verification: format/vet/lint/make check; all 229 root Go packages and required
builtin/boot; affected HTTP/storage/controller and focused boot races; pause
controller regression; Desktop/SDK; 699 frontend files and 5670 tests,
typecheck/build/census; 12 browser/eight Electron renderer scenes; Electron
units (177 pass, one existing skip); 13 official ACP SDK conformance tests.
No threshold or authority expansion; no hosted check result is substituted.
The vendor's published schedule already prices every request; the status bar
now says, live, whether the model is billed at the peak or the off-peak rate
and when that changes, where the model is billed at the vendor's own rate.
Problem: repeated New session reset a blank pane's mode, copied built-in
descriptions ignored the window language, and identifier fields missed theme
styles. The picker was always visible and native archives were parsed in the
renderer. Unrelated transitions also repositioned every open popup.
Root cause: pane reuse resolved an unrequested default; editor styles depended
on a container the identifier row did not share; interface opt-in was missing;
archive transport carried parsed objects; placement had no anchor scope.
Fix: change a reused pane only for an explicit selection and preserve focus on
refusal, translate built-in copies, share field tokens, add a default-off local
picker switch that preserves bindings and the original hidden-state layout,
upload opaque files to the existing host decoder with typed localized errors,
keep the prior JSON envelope, and scope transitions to each popup's anchor or
container. Show catalog unavailability and wrap the editor components.
Verification: failing-first regressions; format/vet/lint/make check; all 228
root Go packages, required builtin/boot, HTTP/storage races and exact-binding
archive transport tests; Desktop/SDK; 691 frontend files and 5608 tests,
typecheck/build/census; 12 browser and eight Electron renderer scenes across
languages/themes/widths; Electron units (176 pass, one existing skip).
No authority, schema, safety policy or verification threshold is widened.
A balance_url on a host that is not a vendor's own now names its format
(balance_format) or is not read, and the key is not sent to it. The connection
editor edits the address and format. Existing relay entries are declared as
deepseek once, and a format the 1.x line drops from the shared file is put back
at the next start.
Problem: ungranted HTTP clients could mutate presets or switch compositions,
and empty hook selections could remove configured safety guards. A paired
remote pane could also borrow the host credential, and a selected native
extension could override the host's denied read.
Root cause: preset routes bypassed request-scoped edit grants, proxying erased
device authority, and frozen automation was treated as the whole safety policy.
Fix: gate preset mutations before store/admission access, refuse paired preset
writes before remote credential forwarding, overlay current authorized guards
outside retained automation, update their shared runner on live edits, resolve
hook inspection/save/approval/reload against one home, and mark bound runtime
packages tighten-only. Preserve ordinary unbound extension behavior, immutable
bindings, optional retained automation and the accepted Standard/Minimal UI.
Verification: failing-first HTTP, proxy and real boot/provider regressions;
format, vet, lint, make check, full root and required builtin/boot suites;
affected package and focused boot races; Desktop/SDK; frontend 689 files/5598
tests, typecheck/build and census; browser eight/Electron renderer four scenes;
Electron units. Correct an inherited mod fixture's premature teardown using
the existing durable-idle notification, with 20 independent process runs.
No test threshold, approval gate or product Close behavior is weakened.
The queued prompt starts a turn of its own, still writing under the test's temp
directory when the test returned, so the directory's cleanup failed with 'directory not
empty'. Seen as 30 failures in 40 runs with eight test processes at once, 0 in 40 after.
The test gave a cold Git Bash two seconds to write the heredoc's file; on a shared
Windows runner the start can take longer, and the failure (4.25 s) was that wait plus
the cancel. The bound is now 15 seconds, only to keep a hang from lasting forever.
The test compared the host's measured thought to the provider's 60 ms sleep minus 10,
which a loaded -race runner breaks (42 ms against 50). The size of the measurement is
the thought clock's own unit test with an injected clock; this test keeps its claim that
a measurement is taken, kept with the turn and framed identically.
A summary request the provider rejects as larger than its context (the documented
context_length_exceeded code, nothing else) is retried at 70%, 50% and 35% of the
refused input, shortened by the same deterministic steps as before; a refused request
costs nothing, so the digest is not given up until three smaller inputs were refused.
A provider that refuses every size ends as summary_input_too_large.
The inbox manifest now records why a queue is paused, as a typed code beside
the old boolean. A message approved on its card is claimed over any pause; one
let in by a standing release or dispatch switch only over a pause the host set
to guard stale input, and stays queued over one the person set. The pause and
older queued items are left as they were.
Problem: Studio's shared remote-open lifecycle conflicts with the composition
callback, and the HTTP adapter omits the selected composition from its body.
Root cause: composition selection was carried by an App-local path rather than
the shared open-ordering owner and the transport serialization boundary.
Fix: merge exact Studio 5b1c282fed, extend
useOpenRemotePane and SseHub with the selection, preserve newest-open ordering,
duplicate-click coalescing and refusal cleanup, and keep editor styles with
their existing catalog stylesheet without widening the shared CSS budget.
Align the child-effort effect assertion with Studio's concrete dispatch model
and effort reporting while retaining provider, refusal and transcript checks.
Keep Standard/Minimal cards, the mode-only left trigger and right run state,
default-off semantics and all existing source/workspace authorization gates.
Verification: failing-first remote Hook/HTTP body regressions; format, vet,
lint, make check, full root and required builtin/boot tests; affected package
and focused boot races; Desktop/SDK Go; frontend typecheck, 689 files/5598 tests
and build; census fixtures; browser eight and Electron renderer four scenes;
Electron unit suite (176 pass, one existing skip). Local review/security review
does not replace maintainer approval or hosted CI.
[tools].disabled removes a tool from every request and refuses a call to it with
the reason; tools another feature requires declare so and stay on with a note;
a project file can only add names. Settings lists every tool with a switch.
The approval carries the reason code origin_text, so Studio words it in the reader's
language, hides the full-access button (which cannot answer it), and shows the call
whole; the terminal wraps the subject and reason instead of clipping them; ACP
gets a fresh allow-once or reject. The label names up to three origins, then counts
the rest. The guide says headless YOLO refuses such a call.
PermissionLists carries askOnExternalOrigin; a save without it leaves it as it is,
the rules endpoint's rebuild applies it, and it keeps the endpoint's provider-edit
grant. The desktop type and mock mirror it.
With permissions.ask_on_external_origin on (user file only, off by default), a call
that is not a read and whose arguments repeat, byte for byte, text the task's
external results gave it is put to a person as a fresh approval when only the
writer fallback (Auto, YOLO) would have let it run. Explicit allow rules, session
allows, reads and authorized MCP servers are untouched; with nobody to ask the call
is refused naming the origin.
Problem: current Studio conflicts in the state-root inventory and adds a move
path that strips the system message containing an immutable composition binding.
That could silently reopen a bound conversation with ordinary capabilities.
Root cause: cross-project adoption intentionally omits project instructions and
permissions, but the new mover did not distinguish immutable composition metadata
from an ordinary system prompt.
Fix: normally merge exact Studio0363720bd. Retain both rules and agent-presets
state roots. Refuse bound conversations in the common MoveSession owner before
copy/delete, preserve the source, and project the typed refusal with bilingual
HTTP wording. Keep ordinary unbound moves, ACP authentication/new-conversation
contracts, directory grants, exact dispatch inheritance and the accepted mode UI.
The preceding source-authority fix remains intact.
Verification: failing-first store and HTTP move tests now refuse blank/started
bound moves and retain the original. All226 root packages pass: mandatory2 plus
the inventoried other224; boot340.765s. Affected race suites pass. The full boot
race invocation previously reached its10minute aggregate limit; final evidence
uses one compiled source snapshot and four disjoint groups covering all753
tests exactly once, without raising the timeout. Static gates, Desktop/SDK,
frontend675files5489tests/typecheck/build/census and browser8/native4 scenes pass.
Local review/security review recorded separately from required maintainer review.
The folder in a draft came from the request and went straight to Abs, EvalSymlinks
and Stat. It is now matched, as text, against the remembered projects the person
trusted (scheduler.Folders), and everything after the match uses the list's own
spelling. Anything else is refused as schedule.workspace_unknown, in the kernel,
the CLI and the board's folder list. A run re-checks trust before it starts.
Problem: selecting a repository preset could change an already authorized MCP
server's package arguments or destination and reuse its executable authority or
credentials. The same provenance issue applies to extension args and env.
Root cause: composition parameter validation checked shape and credential
references but did not distinguish repository declarations from personal ones.
Pinning an executable does not authorize the program selected by its arguments.
Fix: enforce source authority in the shared contract. Project MCP/extension
parameters allow only supported timeouts. Validate declarations, imports, old
revisions, and both definition/resolved halves of retained executable bindings.
Keep structural history validation separate so legacy unsafe bindings remain
readable without regaining execution. Align Mock save/import, document the source
table in both languages, and add the release note. Preserve Standard/Minimal,
unbound defaults, persona, additional directories and exact dispatch bindings.
Verification: failing-first boot tests prove changed args start a subprocess and
changed URL contacts another endpoint before the fix; both now refuse before
startup. User positive controls, scope/parameter table, store/revision/old-record
and event-log history tests pass. Format/vet/lint/makecheck, required builtin/boot,
full root, contract/store/sessionstore/serve/ACP race, Desktop/SDK Go and frontend typecheck/build pass;
671 frontend files / 5472 tests pass. Local review/security review recorded;
maintainer approval and new-head hosted CI remain separate requirements.
The full boot race invocation reached its existing aggregate 10-minute limit
as another test started; disjoint test-list groups are used to finish the whole
inventory without raising that timeout or skipping tests.
Problem: newer Studio changes conflict with the ACP session contract and HTTP
runtime opening after the previous assembly integration.
Root cause: Studio adds a new-conversation execution-mode flag where this branch
also owns explicit composition selection and immutable inherited bindings.
Fix: normally merge Studio 4d3c09211. Add NewConversation to the one SessionParams
definition and retain it in HTTP build options. Resolve moved transcript paths
before reading either branch settings or the original composition. Keep exact
dispatch inheritance, workspace refusal, candidate rollback, ACP directories and
the accepted Standard/Minimal composer and management design.
Verification: format, vet, lint, make check, mandatory builtin/boot, full root
Go, Desktop Go, frontend typecheck/build and census pass. Frontend 671 files /
5465 tests and eight browser/four native renderer scenes pass. Affected
boot/ACP/serve/control/remote/delegation race tests pass. Provider request bytes
equal actual Studio with and without persona. Local review/security review are
recorded separately from maintainer review. The unchanged macOS computer
coverage baseline remains 88.2% below 89.8%; no floor is changed.
Problem: Studio advancement made the Agent composition PR conflict in boot
assembly while adding trace export and a workspace-aware mod runtime source.
Root cause: wireMCP now has a composition error result, and extension preflight
filters immutable bindings where Studio introduced its package source.
Fix: normally merge Studio ed1c127d6. Keep OTLP cleanup before returning from
wireMCP and retain workspace-aware discovery before exact composition filtering.
Preserve upstream tighten-only dispatch, replay protection, ACP elicitation and
the withheld mod switch. Keep Standard/Minimal, opt-in binding, ACP directories,
exact dispatch inheritance and the accepted composer layout.
Verification: gofmt, vet, lint, make check, required builtin/boot, full root Go,
affected boot/extension/modhost/ACP/serve/control race suites, Desktop and SDK
tests pass. Frontend typecheck/build, census, 668 files / 5450 tests, Electron
unit tests and eight browser/four native renderer scenes pass. Raw provider
requests equal actual Studio with both persona and empty persona. Local review
and security review found no new issue. Coverage gate still reports only the
unchanged macOS computer baseline, 88.2% versus 89.8%; no floor is lowered.
Problem: Agent compositions still carried the persona prerequisite diff and
missed newly added session lifecycle contracts on Studio.
Root cause: the branch predated the landed persona, ACP additional directories
and project-dispatch assembly paths. Combining their independent additions
also exceeded existing file/scalar-state budgets.
Fix: normally merge Studio 896552483 after verifying #13588 actually landed.
Resolve the Persona/composition imports and keep one ACP SessionParams owner.
Clone vetted extra directories through composition switches. Pass the exact
private binding into dispatched runtime assembly; reject cross-workspace
bindings before opening, and reject fresh assembly errors without a history
fallback. Ordinary unbound dispatch remains unchanged. Extract the command
catalog and Mock memory projection, and place the MCP guard with catalog
refresh state. No budget baseline is raised.
Keep Standard/Minimal cards, the left mode-name-only picker, right run status,
default-off/unbound Standard behavior and the landed user-config persona.
Update bilingual behavior/API documentation and release notes.
Verification: format, vet, lint, make check, required builtin/boot tests, full
root count1 atomic-coverage tests, the complete CI race package group, Desktop
and SDK suites; frontend typecheck/build and 667 files / 5443 tests. Electron
unit176 pass with one existing skip. Browser8 and Electron4 controlled rendering
scenes cover locale/theme/width, selection, mode cards, locked/new flows and
left/right layout. Real HTTP requests match landed Studio byte-for-byte with
user persona29139 bytes and empty persona29110 bytes, with fixed probe PATH.
ACP/dispatch regressions fail before propagation and pass afterward. Local
review and security review cover MCP contracts and dispatch/ingress authority.
Known local gate: make coverage-gate reports the untouched macOS computer
package at88.2 below89.8. The exact command on landed Studio reproduces the
same result and uncovered lines. No floor is lowered; hosted Linux gates
remain required. Hosted results for this new head are reported separately.
One draft shape and one set of confirmation facts for every frontend; a
creation is accepted only with the token of the preview the person saw; only
the machine's owner may create or resume, any reader may pause or delete.
Problem
A development Studio shell started with `pnpm start` (`electron .`) had two
ways to outlive the launch that owned it:
- with its launcher killed outright, the shell and its kernel kept running
with nobody holding them;
- with the launcher's console closed underneath it (a terminal gone, a
background job's reader exited), the next kernel stderr line the shell
echoed raised an EPIPE that nothing handled. Electron's modal "A JavaScript
error occurred in the main process" box then held the main thread, so the
shell stopped draining the kernel and ignored SIGTERM (still alive 30 s
after it, 2/2 runs). The kernel logs a request every few seconds, so this
followed within seconds of the console closing.
Root cause
main.js echoed kernel stderr with a bare process.stderr.write and attached
no error listener, and nothing in the shell watched for the loss of the
process that launched it. The kernel already ends with the shell through its
stdin lease; the shell had no equivalent towards its own launcher.
Fix
- src/launcher.js owns both: watchLauncher ends an unpackaged shell once its
launcher is gone (POSIX reparents an orphan; Windows keeps the dead pid, so
there it is probed), and consoleEcho drops output for a console that failed
or backed up instead of throwing or queueing. shell.log and host.log keep
every line.
- Both relaunch paths (the kernel's relaunch act and the X11 relaunch) hand
the next instance over first: the launcher exits with this process, so a
relaunched shell would otherwise read its own start as the launcher dying.
- shell.log records the shell's and the kernel's pids, which is how the live
check finds them.
Verification
- node --test test/unit.mjs test/linux-packaging.test.mjs: 181 pass, 1 skip.
The new real-pipe test fails (child exits 1 on EPIPE) without consoleEcho.
- pnpm launcher-live (new, real shell + kernel, macOS arm64): launcher
SIGKILL after and before the handshake, closed console then SIGTERM, plain
SIGTERM all pass; with the fix removed the first three fail.
- pnpm smoke: 62 ok, 1 fail ("the bridge exposes verbs and nothing else"),
identical on unmodified studio.
Problem: the latest studio settings catalog must know the persona fields, and
opt-in Standard mode must retain access to the prerequisite user-persona UI.
Root cause: the new catalog landed during the composition review, while the
initial composition page had replaced persona editing with a default adapter.
Fix: integrate the corrected persona parent and current studio, reuse its
editor in Standard mode, preserve settings-search navigation, and distinguish
user persona edits from immutable composition bindings. Keep the accepted two
cards, compact left picker and right run information.
Verification: gofmt, vet, lint, make check, root Go suite including boot/builtin,
frontend typecheck/build and 656 files / 5,385 tests pass. Parent persona keys
are Guarded; the actual request-byte and old-reader probes remain unchanged.
Problem: default selections bound every new session and migrated the user's
persona, changing the ordinary path and making downgrade impossible.
Root cause: entry points manufactured a selection, persona settings were
adapted into the preset store, and blank unbound rewrites were not persisted.
Fix: keep unselected and Standard sessions unbound, preserve the user persona,
require explicit non-Standard selection, and restore durable state on failed
publication from either origin. Reuse the corrected persona parent, preserve
Standard/Minimal cards and the compact unified picker, and remove committed
composition captures. Freeze both skill discovery paths without copying locks.
Verification: root Go suite, vet, lint, make check, Desktop/SDK suites, Electron
unit tests and focused frontend suites pass. Real Build HTTP bytes match the
persona parent; its original reader can continue a new unbound transcript.
Project authorization, rollback and skill discovery regressions pass.
Problem: studio's native-video projection refactor overlaps the local-only
Agent composition metadata filter at the provider boundary.
Fix: keep the shared projection predicate and include composition metadata in
its host-only fields. Preserve stored bindings/video references and remove
answered video payloads only from model requests. Add regressions for a binding
as the sole copy trigger and for binding/video projection together.
Verification: gofmt, go vet, make lint/check, uncached root Go tests,
provider/OpenAI/boot/builtin suites, Desktop Go tests, frontend typecheck,
production build and 5,348 frontend tests. Existing mode layouts are preserved.
Problem: studio's MCP-follow changes and session compositions both own the
publication of a server tool catalog.
Root cause: the new followed-tool metadata must travel with the same catalog
whose fingerprint the immutable session binding validated.
Fix: retain follow metadata and the composition check before catalog publication.
Preserve upstream account localization and MCP-follow behavior. Add refresh,
stale-adapter and reconnect regressions with following enabled and disabled,
plus a real boot/provider test showing old bindings block drift while new
conversations can bind the changed server.
Verification: gofmt, go vet, make lint/check, root Go and boot/builtin suites,
Desktop and SDK Go suites, composition-guard race tests, frontend typecheck,
production build and 5,347 frontend tests. Browser evidence remains in design QA.
Problem: the shipped catalog and composer exposed more mode choices and status
badges than the requested Standard/Minimal workflow needs.
Fix: ship two native compositions, with a full-configuration Standard mode
and a terminal/image Minimal mode using a concise persona. Preserve retained
bindings and report retired defaults explicitly. Separate built-in cards from
custom presets, keep only the mode name in the trigger, and place run status
at the right. Keep source/lock details in the accessible label and menu.
Verification: gofmt, go vet, make lint/check, required boot/builtin tests,
complete root and Desktop Go suites, frontend typecheck/build and 5,343 tests.
Provider-boundary and custom read-only fixtures retain capability-denial
coverage. Browser checks cover light/dark, 1440px/390px and locked controls.
Problem: the session-composition branch conflicted with the latest studio
settings, composer, peer-thread and runtime-construction changes.
Root cause: both branches extended shared port declarations, provider wiring
and the composer status row; failed-build cleanup also gained a separate
cancellable construction lifetime upstream.
Fix: retain both sets of settings and providers, preserve the top-row unified
Agent menu and responsive anchoring, and combine lifetime cancellation with
composition resource cleanup. Keep upstream accessible status labels and
text-size scaling. Add provider-boundary coverage showing the new image tool
cannot exceed review/research tool selections, plus updated visual evidence.
Verification: gofmt, go vet, make lint/check, required boot/builtin tests,
root and Desktop Go suites, frontend typecheck/build and 5,342 tests, Electron
unit tests, and browser light/dark checks at 1440px and 390px.
Problem: Session composition controls competed with task actions in the
bottom toolbar, and the adjacent ellipsis did not explain how to manage
presets.
Root cause: Composition selection and management were rendered as separate
controls below the input. Floating menu placement also missed fold changes
published after window resize and the end of layout transitions.
Fix: Place one Agent presets menu beside the run status, show sources and
capability notes, and group management below a separator. Preserve locked
bindings and explicit new-conversation actions. Reuse the existing Picker
with per-choice action identities, disabled-row keyboard navigation and
current-state semantics; refresh shared floating anchors after reflow.
Update bilingual guides, release notes and before/after screenshot evidence.
Verification: All 633 frontend test files / 5,225 tests pass, including
selection/management, keyboard focus, pending/failure, lock/new-session,
source identity and responsive positioning regressions. Typecheck,
production build, gofmt, go vet, make lint, make check, and required
boot/builtin suites pass. Browser light/dark, wide/390px and resize recovery
were verified with synthetic fixtures.
Problem: an admitted turn could fail before adding its user message and unlock
its composition. Controller-owned continuations could strand ACP config changes,
and MCP disconnect could remove code needed by a subsequent reconnect.
Root cause: the started latch followed message insertion, ACP drained only its
own prompt returns, and connection cleanup owned session package resources.
Fix: durably latch composition admission before executing or acknowledging a
turn and refuse work while the latch is pending. Notify owners at real idle
transitions and drain ACP changes through its existing serialization domain.
Separate transport state from package lifetime, retain script sibling bundles,
exclude VCS metadata, reject linked transcript exports and freeze late tool names.
Add ACP exact-binding/history tests, actual old-reader evidence, bilingual notes
and the complete light/dark wide/390px before-and-after screenshot set.
Verification: deterministic async/sync admission and failed-save regressions,
ten staged background-continuation/config-switch runs, real MCP reconnect cycles,
shared lifecycle race tests, package coverage, vet, lint and repository checks.
Frontend full suite and independent Desktop/SDK/Electron suites remain green.
Refs: #13613
Problem: Studio persona settings cannot select an independent tool and module
combination for each conversation, or preserve that combination across reloads.
Root cause: discovery and runtime construction follow mutable shared configuration
without a versioned session composition or a shared admission boundary.
Fix: introduce versioned scoped definitions, retained resources and authoritative
session bindings. Assemble isolated runtime selections, enforce tool ceilings,
pin MCP contracts, and share the lifecycle gate with first-prompt admission.
Expose independent HTTP, CLI, ACP and Studio management and selection controls,
with migration, native import/export, diagnostics and read-only history recovery.
Verification: composition mechanism tests, provider-wire cache tests, lifecycle
rollback tests, sensitive-boundary regressions, frontend typecheck and all 5,219
frontend tests; gofmt, vet, lint and repository checks. Browser light/dark wide
and 390px screenshots and real Electron backend interactions are included.
Refs: #13588
2026-10-17 19:09:49 +08:00
2842 changed files with 176074 additions and 8952 deletions
go test -count=1 -timeout=3m -json -run '^TestSubagent(PersistedStatusAdmission|UnknownStatusStillReadableForDiagnosis)$' ./internal/runtime/delegation | tee "$RUNNER_TEMP/subagent-status-admission.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
go test -count=1 -timeout=3m -json -run '^Test(CaptureDirectoryMaterializesContainedLinks|CaptureDirectoryRefusesUnsafeLinks|CaptureHookFileFlattensContainedLink|MaterializeStillRejectsLinkedRetainedFilesAndRoot|WindowsCompositionJunctions|WindowsLiveSkillJunctions|CompositionRetainsLinkedSkillAcrossRestart|CompositionLinkedSkillRejectsLinkedResources|CompositionLinkedExtensionResources|CompositionLinkedHookResources)$' ./internal/state/agentpreset ./internal/ext/skill ./internal/assembly/boot | tee "$RUNNER_TEMP/composition-links.jsonl"
go test -count=1 -p 4 -timeout=18m -json ./internal/runtime/isolation/... ./internal/platform/worktree/... ./internal/state/mutation/... | tee "$RUNNER_TEMP/isolation-owners.jsonl" |
go test -count=1 -timeout=3m -json -run '^Test(UsageScope.*|DelegatedUsageKeepsGoalOwner|EffectDelegatedUsageKeepsGoalOwner)$' ./internal/contract/event ./internal/session/control ./internal/assembly/boot | tee "$RUNNER_TEMP/goal-usage-scope.jsonl"
go test -count=1 -timeout=3m -json -run '^TestEffectPlannerConfiguredTaskBudget$' ./internal/assembly/boot | tee "$RUNNER_TEMP/planner-local-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(SemanticRootUsage|EffectRoutingRootBudget)$' ./internal/runtime/capability ./internal/assembly/boot | tee "$RUNNER_TEMP/controller-routing-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/runtime/capability", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(PreparedRootBudget|CoordinatorRootBudget|EffectCoordinatorRoot)' ./internal/runtime/agent ./internal/runtime/coordinator ./internal/assembly/boot | tee "$RUNNER_TEMP/coordinator-root-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/runtime/agent", "reasonix/internal/runtime/coordinator", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(GoalResumeRenewsPlannerBudget|GoalNonBudgetResumeKeepsPlannerBudget|GoalTokenBudgetPausesAndResumes|EffectGoalResumeRenewsPlannerBudget)$' ./internal/session/control ./internal/assembly/boot | tee "$RUNNER_TEMP/goal-planner-resume.jsonl"
go test -count=1 -timeout=3m -json -run '^Test(RootWall|EffectRootWall)' ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/root-wall-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(RootToken|EffectRootToken|BudgetCountsKnownReportedTokens|ReportedToken|EffectReportedTotalTokenBudget)' ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/root-token-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(RequestOnlyUsage|RunBudgetMissingUsage|TaskBudgetKnownCost|RunBudgetExplicitZero|TaskBudgetUsageCompleteness|TaskBudgetCostUses|EffectBudgetCompleteness|TaskBudgetGateIgnoresCostWhenUnpriced)' ./internal/contract/provider ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/budget-usage.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/contract/provider", "reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(HostStorageEnvironment|HostObjectStorage|HostExplicitStorageControls|DetachedStorageEnvironment|WorktreePreparationOwnsItsIndex|EffectIsolatedTaskOwnsGitStorage)$' ./internal/platform/gitcmd ./internal/platform/worktree ./internal/assembly/boot | tee "$RUNNER_TEMP/git-storage-environment.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/platform/gitcmd", "reasonix/internal/platform/worktree", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(InterruptedUsage|EffectAnthropicPartialUsageRecovery)' ./internal/model/anthropic ./internal/assembly/boot | tee "$RUNNER_TEMP/anthropic-interrupted-usage.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/model/anthropic", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(RootWebSearch|EffectRootWebSearch)' ./internal/tools/websearch ./internal/assembly/boot | tee "$RUNNER_TEMP/root-websearch-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/tools/websearch", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^(TestProfileRootBudgetLifetime|TestProfileBudgetPreparationUsesTurnAdmission|TestEffectExplicitSkillRootBudget)$' ./internal/session/control ./internal/assembly/boot | tee "$RUNNER_TEMP/explicit-skill-root-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/session/control", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(RootTriage|EffectRootTriage)' ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/root-triage-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(RootAuxiliary|EffectRootAuxiliary)' ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/root-auxiliary-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
(["reasonix/internal/runtime/agent", "reasonix/internal/assembly/boot"] - $packages | length) == 0 and
go test -count=1 -timeout=3m -json -run '^Test(LocalBudgetUsesOccurrenceQuote|LocalCost.*|UsageQuote.*|EffectLocalCostUsesOccurrenceQuote)$' ./internal/contract/event ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/local-cost-quote.jsonl"
go test -count=1 -timeout=3m -json -run '^Test(AuxiliaryUsageObserver|BoundedUsageObserver|RootBounded|EffectRootBounded)' ./internal/contract/provider ./internal/model/boundedllm ./internal/runtime/agent ./internal/assembly/boot | tee "$RUNNER_TEMP/bounded-root-budget.jsonl"
jq -se '
[.[] | select(.Test == null and .Action == "pass") | .Package] as $packages |
[.[] | select(.Test != null and .Action == "pass") | .Test] as $tests |
all(.[]; .Action != "skip" and .Action != "fail") and
platform=$(git diff -z --name-only --no-renames "$BASE_SHA" HEAD | bash scripts/ci-platform-scope.sh)
if git diff --name-only --no-renames "$BASE_SHA" HEAD |
grep -qE '^(desktop/electron/|desktop/packaging/|desktop/cmd/update-helper/|desktop/cmd/studio-manifest/|desktop/go\.(mod|sum)$|cmd/reasonix-studio-host/|internal/platform/(update|appupdate)/|go\.(mod|sum)$|scripts/(linux-install-smoke|verify-linux-package|verify-studio-artifacts)\.sh$|scripts/(windows-signing-lib|verify-windows-package|verify-windows-authenticode|install-smoke)\.ps1$|\.github/workflows/(studio|release-studio)\.yml$)'; then
grep -qE '^(desktop/electron/|desktop/packaging/|desktop/cmd/update-helper/|desktop/cmd/studio-manifest/|desktop/go\.(mod|sum)$|cmd/reasonix-studio-host/|internal/platform/(update|appupdate)/|go\.(mod|sum)$|scripts/(linux-install-smoke|verify-linux-package|verify-linux-appimage|verify-studio-artifacts)\.sh$|scripts/(windows-signing-lib|verify-windows-package|verify-windows-authenticode|install-smoke)\.ps1$|\.github/workflows/(studio|release-studio)\.yml$)'; then
`verify-page.sh` is the one to run after a change: it starts the kernel with the
same arguments `Kernel.kt` uses (`--auth token`, `--token-file`, `--port-file`,
`-page`) and checks that it binds a port and answers 200. Without a device, that
is as deep as the chain can be verified.
## Where the code is
| File | Owns |
| --- | --- |
| `MainActivity.kt` | The WebView, the back gesture, downloads, external links, renderer restart |
| `Kernel.kt` | Unpacking the page, starting and stopping the kernel, waiting for the port, health probes, logs, the workspace directory |
| `KernelService.kt` | The foreground service that keeps a long turn alive |
| `NativeBridge.kt` | `@JavascriptInterface`: what the page may ask the phone for |
| `BrowserShellServer.kt` | Loopback HTTP server the kernel's browser tools drive the overlay WebView through |
| `Downloader.kt` | Writing a kernel-produced file through the system document picker |
The page reaches the shell through `window.reasonixHost` and `ReasonixNative`,
defined in `desktop/frontend-next/src/port/host.ts` (`AndroidHost`).
**Phone and desktop share one page build.** What differs only on a phone is
decided at runtime, not by a second set of components. `phoneShell()` in
`src/port/host.ts` and the `[data-shell="android"]` selector in `mobile.css` both
read `<html data-shell="android">`, which `index.html` sets before the bundle
loads.
One attribute feeding both is what keeps style and components from disagreeing
about what a phone is.
## What the page may ask the phone for
`NativeBridge` is reachable from whatever document the WebView holds, so every
method that touches the device asks `trusted()` first.
The activity answers by comparing the **main frame's** origin against the
kernel's and against the one paired computer the person scanned a code for. A
page that navigated elsewhere is not trusted, even though it may still hold the
object.
Five methods answer without the check — `version`, `systemTheme`, `safeArea`,
`setLayerOpen`, `setRunning` — because they report back what the page already
knows about its own window and grant nothing.
`openInApp(url)` is the pairing decision, not a navigation one: it loads a URL in
this window only for the origin the scanner just returned, and only once.
Without that, any page holding the kernel origin could hand itself the bridge.
External links leave through the browser. A non-web scheme is routed to the
system only for `mailto:` and `tel:`; everything else is dropped, because an
`intent://` URL can start an arbitrary exported activity with caller-supplied
extras.
## Permissions, and what they cost
`MANAGE_EXTERNAL_STORAGE` — opening a folder on the phone as a workspace. The
kernel is a Go binary: it opens real paths with `open(2)`, and SAF hands back a
`content://` URI it cannot open.
Requested from the folder-picker action, never at launch. Android shows no dialog
for it and opens system settings instead.
**There is no OS-level sandbox around this kernel.** The PR sets the sandbox to
off on Android, so confining the agent to the chosen folder is Reasonix's own
write-root check and nothing else. A shell tool with all-files access can reach
all shared storage.
`REQUEST_IGNORE_BATTERY_OPTIMIZATIONS` — Doze freezes the app group, which stops
a turn that is still running. Asked once when a turn actually runs, never at
launch, and the person can refuse. Refusing costs battery life during long turns
and nothing else.
**Google Play restricts this permission**: fine for a sideloaded build, a
store-policy risk if this is ever listed.
`FOREGROUND_SERVICE` / `FOREGROUND_SERVICE_DATA_SYNC` — keeping the kernel alive
while the screen is off. `dataSync` is capped per 24 hours from Android 15. It is
the closest honest description of "an agent is working".
`CAMERA` — reading the pairing code a computer shows. Requested when the person
asks to scan, never at launch. `ScanActivity` is not exported.
`READ_EXTERNAL_STORAGE` / `WRITE_EXTERNAL_STORAGE` — the same folder walk below
API 33. `maxSdkVersion`-gated; above 33 they are superseded by the permission
above.
`allowBackup=false` with explicit extraction rules, so a `keystore` or a token in
app storage never leaves the device in a cloud backup.
## Cleartext HTTP
The kernel serves its page on `127.0.0.1:<port>`, and a paired computer serves
the same page over its LAN address. Both are plain HTTP.
Neither address can be written into a `network-security-config`: the port is
chosen at start, the LAN address at scan time, and the format matches literals
and domains, not subnets.
So the config permits cleartext and the boundary is in the code instead.
`shouldOverrideUrlLoading` sends every other http(s) origin to the browser, and
the bridge answers only the kernel and the just-paired computer. A plaintext page
on the LAN cannot become the document holding the bridge.
## What goes into the APK
| Content | Size | Why it is shaped this way |
| --- | --- | --- |
| `lib/arm64-v8a/libreasonix.so` | 45 MB | Go static binary, stripped. It is a `.so` rather than a file in `assets/` because only `nativeLibraryDir` is guaranteed executable — a binary copied to app storage fails on a device whose data partition is mounted `noexec`. |
| `assets/page/` | 14 MB | The frontend build, including the Inter and Noto Sans SC fonts, `noCompress` |
| `assets/LICENSE` | 1 KB | The repository's MIT licence, verbatim. MIT requires the notice and the licence text to accompany a binary distribution. |
`assets/LICENSE` is generated by `sync-page.sh` from the repository's root
`LICENSE`, not committed a second time: a checked-in copy drifts when the root
file is updated.
It sits in `assets/` rather than `assets/page/` because that directory is removed
and rebuilt on every build.
A release APK is about 31 MB.
## Building
**The Gradle wrapper is committed**, jar and all. `build.sh` refuses to run
without `gradlew` rather than generating one from whatever `gradle` is on `PATH`:
the dependency versions are part of the build.
`gradle-wrapper.properties` pins `distributionSha256Sum`, so a mirror that serves
a different distribution zip fails verification instead of being unpacked and
run. Dependency repositories are the official ones only — a mirror belongs in a
user-level init script.
**Signing.** A release with no key configured comes out **unsigned**
(`app-release-unsigned.apk`) with a warning. It used to fall back to the debug
key, which produced an artifact that looked shippable and was signed by a
per-machine key nobody chose: it could not install over the previous build.
Point the build at a real key by writing `keystore/keystore.properties`
(`storeFile`, `storePassword`, `keyAlias`, and optionally `keyPassword`) or by
setting the matching `REASONIX_KEYSTORE*` environment variables. Key material
stays out of the repository and out of build logs.
Debug builds always use the debug key. Signing a debug build with the release key
would let anyone who can read the keystore produce a build that installs over the
shipped one under the same signature.
**Non-ASCII checkout paths.** AGP refuses to build under a path with non-ASCII
characters on Windows unless `android.overridePathCheck=true` is set. That is a
property of one machine's disk layout, not of this project, so it is not in the
repository's `gradle.properties`.
Set it in a user-level `~/.gradle/gradle.properties` or pass
// The extraction script: prefer article/main for content pages,
// fall back to body. innerText preserves visual line breaks. It
// also stamps a ref on every element a person could act on, so the
// kernel can name one in an act step the way the desktop backend
// names an accessibility node. A ref already on an element is kept,
// so it stays the same across snapshots of one document; a fresh
// document starts the numbering over, and the kernel only accepts a
// ref from the snapshot it last read.
valscript="""
(function(){
var seq = window.__rxRefSeq || 0;
var used = {};
var marked = document.querySelectorAll('[data-rx-ref]');
for (var i = 0; i < marked.length; i++) { used[marked[i].getAttribute('data-rx-ref')] = true; }
function roleOf(el) {
var tag = el.tagName.toLowerCase();
var role = (el.getAttribute('role') || '').toLowerCase();
if (role) return role;
if (tag === 'a') return 'link';
if (tag === 'button') return 'button';
if (tag === 'select') return 'combobox';
if (tag === 'textarea') return 'textbox';
if (tag === 'input') {
var t = (el.getAttribute('type') || 'text').toLowerCase();
if (t === 'checkbox' || t === 'radio' || t === 'submit' || t === 'button' || t === 'reset') return t;
return 'textbox';
}
return 'generic';
}
function nameOf(el) {
var n = el.getAttribute('aria-label') || el.getAttribute('placeholder') || el.getAttribute('name') || '';
if (!n) { n = (el.innerText || el.textContent || '').trim(); }
return n.replace(/\s+/g, ' ').slice(0, 160);
}
var sel = 'a[href],button,input,select,textarea,[role=button],[role=link],[role=checkbox],[role=radio],[role=tab],[role=menuitem],[contenteditable=true]';
var els = document.querySelectorAll(sel);
var elements = [];
for (var j = 0; j < els.length; j++) {
var el = els[j];
var ref = el.getAttribute('data-rx-ref');
if (!ref) {
do { seq++; ref = 'e' + seq; } while (used[ref]);
el.setAttribute('data-rx-ref', ref);
used[ref] = true;
}
var raw = (el.value != null && String(el.value) !== '') ? String(el.value).slice(0, 160) : '';
Five harnesses live under `benchmarks/`; `cmd/e2ebench` also exposes a
Six harnesses live under `benchmarks/`; `cmd/e2ebench` also exposes a
SWE-bench Verified mode:
-`e2e/` — the committed end-to-end task suite, driven by
@@ -19,6 +19,10 @@ SWE-bench Verified mode:
-`fanout-width/` — a shape-instructed fleet whose members share nothing, for
pricing what running them at the same time is worth. See
[Pricing the run graph](#pricing-the-run-graph) below.
-`realrepo/` — long tasks over a pinned snapshot of this repository, for
measuring changes that cut model round trips (13–31 requests per task on
`deepseek-flash`). See
[its README](realrepo/README.md).
## Directory layout
@@ -416,7 +420,7 @@ own outcome).
| Flag | Default | Purpose |
| --- | --- | --- |
| `-mode` | `suite` | `suite` \| `diff` \| `swebench` \| `compare` \| `traj` (`diff` generates tests for the PR diff; `swebench` runs the official per-instance evaluation; `compare` renders KPI/Pareto readouts from 2+ `-json` reports; `traj` re-digests recorded trajectory files without spending tokens). |
| `-mode` | `suite` | `suite` \| `diff` \| `swebench` \| `compare` \| `traj` (`diff` generates tests for the PR diff; `swebench` runs the official per-instance evaluation; `compare` renders KPI/Pareto readouts from 2+ `-json` reports; `traj` re-digests recorded trajectory files without spending tokens; `audit` checks a batch's published figures against its retained rows). |
| `-task` | *(all)* | Suite mode: run only these comma-separated task IDs (e.g. `-task fix-add-bug`); unknown IDs fail with the available list. |
| `-attempts` | `1` | Suite and diff modes: retry a task until an attempt passes, up to N; enables the `Pass@≤N` KPI, and TTCS charges a retried solve with its failed attempts' wall. |
@@ -429,6 +433,7 @@ own outcome).
| `-ablate` | *(none)* | Ablation arm: comma-separated subsystems to switch off — `evidence`, `planner`, `subagent`, `retrieval`, `compaction`, `upstream` and `recall-search`; `none` \| `all`. |
| `-json` | *(none)* | Write the JSON report here (optional). |
| `-batch` | *(none)* | Suite mode: write `results.jsonl` (one row per attempt) and `batch.json` (manifest v1) here, so the report's figures can be audited later. Audit mode: read the batch from here. See [Auditing a batch](#auditing-a-batch). |
| `-trajectories` | *(none)* | Suite mode: write one `<task-id>.trajectory.jsonl` per task into this directory (the agent's full event stream with timestamps — see `reasonix run --trajectory`). The report gains a time-attribution line (tools vs. model) and each JSON result a `trajectory` digest. |
| `-force-planner` | `false` | Suite mode: prefix each prompt with a plan-first directive so the two-model turn engages regardless of the planner gate. Use for the "with planner" arm of an A/B; results carry `plan_forced` so arms are only comparable with equal forcing. |
| `-anchor` | `blind` | Suite mode: which hypothesis the agent holds before it looks at anything — `blind` (none, the control) \| `correct` \| `wrong`. The seeded arms prefix each prompt with the task's authored seed and **skip** tasks that have none, so an unseeded control run never lands in a seeded denominator. Results carry `anchor`. See [Anchor resistance](#anchor-resistance). |
from `results.jsonl`, prints the manifest's claim beside the recomputed value,
and exits non-zero naming any figure that moved. It also refuses a batch whose
manifest is missing, records no commit, or whose rows are truncated.
This is an **aggregation audit, not a replay**: it proves the published figures
follow from the retained rows. It does not re-execute the model, tools or
environment, and cannot show the experiment would reproduce. A real replay would
need the commit, a corpus fingerprint, the model and provider configuration, the
seed, the flags, the raw results and the runtime version; it is not implemented.
`schema_version` is top level. The rest is three groups:
| Group | Fields | Answers |
| --- | --- | --- |
| `over` | `command.mode` and `command.axes` (every flag that changes what a figure means; paths omitted); `tasks.count` and `tasks.digest` (sha256 over sorted, NUL-separated task ids — the corpus identity, never the task text); `filters`; `denominator` (`unit`, `definition`, `value`) | what the figures are over |
| `producer` | `commit` (or an explicit `unknown`, never a guess), `commit_source`, `dirty`, `model`, `provider`, `seed`, `started_at`, `finished_at`, `harness_version`, `go_version` | which run produced them |
| `trust` | `results_sha256`, `claimed` (the headline figures), `ledger`, `retention` | why they can still be checked |
The ledger accounts for every planned run, and the audit checks that the counts
form a funnel (`scored ≤ completed ≤ attempted`, `failed ≤ scored`,
`scored = pass1 + failed`, `rows = attempted + skipped`). A row count is never a
denominator: `denominator.value` is a count of scored tasks, and `rows` is
published only to show what was retained.
| Count | Meaning |
| --- | --- |
| `planned` | graded runs the batch set out to attempt: filtered tasks × trials |
| `attempted` | retained rows for runs that started (skipped rows are not attempts) |
| `completed` | attempts that ended and left a row; this harness writes a row only after an attempt ends, so it equals `attempted` and the equality is checked, not assumed |
| `scored` | completed first attempts of solvable, non-skipped tasks — the pass@1 denominator |
| `failed` | scored attempts the grader did not pass |
| `unmeasured` | attempts whose cost and tokens are unknown because no metrics landed; kept out of spend totals, never counted as zero |
| `excluded` | rows kept out of the accuracy denominator: skipped, `no_solution`, and retry entries |
Spend and cache fields are omitted from `claimed` when nothing was measured, so
an absent figure reads as unknown rather than as a measured zero. `seed` is
omitted for the same reason: the harness does not set a sampling seed.
`results_sha256` covers the values and the presence of every row; line endings,
indentation and key order are normalised away because none of them is evidence,
while number spelling is kept (a rewritten `1` is not `1.0`).
A successful audit records the verifier version (`audit.verifier_version`) and
the digest it checked in `batch.json`, so a re-verification can tell which code
checked the batch.
`results.jsonl` is retained unredacted beside the manifest — task prompts and
agent notes included — and is not public. The manifest itself carries a task-id
digest, no prompt text, no absolute paths, no endpoints or keys and no model
output. A claim whose rows are gone cannot be audited: there is nothing to
recompute from, and this tool will not certify it.
## Dataset retention
Keep every `-json` report and`-trajectories` directory from real runs: they
are the accumulating corpus — per-task contracts-to-be, full event
Keep every `-json` report,`-trajectories` directory and `-batch` directory from
real runs: they are the accumulating corpus — per-task contracts-to-be, full event
trajectories, checkpoint oracle verdicts, stop curves and phase traces — that
any future offline learning (routing, stop policies, budgets) would train
and evaluate on. The control plane stays deterministic and interpretable
go run ./cmd/e2ebench -suite benchmarks/realrepo -bin "$PWD/bin/reasonix" -task rt-stale-skills
```
The agent reads its provider config from `REASONIX_HOME` (default
`~/.reasonix`); point it at a disposable one to keep benchmark sessions out of
your own.
The default `-budget` (800k tokens) can stop the suite partway through, hence
`-budget 0`. Under `-trials`, each trial keeps its own
`<task>.trial-N.trajectory.jsonl`.
## The snapshot
Every `workdir/` is the same Go module (`module reasonix`, stdlib only, no
network needed to build or test):
- **Pinned code.** `git archive` of commit `42636ddcb425` for
`internal/contract/{pricing,ablation,agentgraph}` and
`internal/base/{retrieval,rulelist,tokencount,neterr,nilutil}` — about 5k
lines of real kernel code with its real tests and doc comments.
- **Fixture-authored files** in `base/`: `go.mod`, a trimmed `REASONIX.md` (the
project's own conventions, so the agent loads standing instructions the way
it does here), a `Makefile` with `check`, a module `README.md`, the
`cmd/benchctl` command that consumes `pricing` and `ablation` from outside
their package, and one project skill.
- **One pinned test anchored to the snapshot day.**
`TestCatalogPricesHaveBeenCheckedRecently` compares each row's `CheckedOn`
with the wall clock, which would fail every workdir 90 days after the oldest
pinned date (2026-11-20). `base/` replaces that file with one that reads
staleness at 2026-10-19 instead; nothing else in the pin is edited.
- **Per-task overlay** in `overlay/<task-id>/`, applied last. Only
`rt-stale-skills` has one.
`snapshot.sh` rebuilds every workdir from those three sources and
`snapshot.sh --check` fails if a committed workdir has drifted from them. Edit
`base/` or `overlay/`, never a workdir, then rerun the script. Moving the pin is
a new suite version: results from two pins do not compare.
## Tasks
| Task | Kind | What it exercises |
| --- | --- | --- |
| `rt-cache-write-rate` | multi-file Go change, test, README | Adds a `CacheWrite` rate to `RateCard` and threads it through cost, fingerprint, card identity and the catalog, then prices Anthropic's `claude-sonnet-4-5` with it: a new `officialRates` row, documentation and checked-on constants, an endpoint host, and `benchctl rate`/`cost` output. The fingerprint has to stay byte-identical for every existing card, so the change to `PricingFingerprint` cannot be the obvious one. |
| `rt-stale-skills` | documentation only, no recognised verifier | Three project skills name paths, functions, `make` targets and `benchctl` subcommands and flags that do not exist, and one contradicts the append-only rule in `REASONIX.md`. Fixing them takes reading `rates.go`, `ablation.go`, both `benchctl` files and the `Makefile`, then editing only Markdown — the shape behind #13293, where `stale_verification` stays owed after a skill edit because no check covers it. |
| `rt-rate-lookup-signatures` | API migration with a likely compile-error loop | Changes `CurrentRate` to `(RateCard, bool)`, `LookupCatalog` to a `CatalogKey` struct, and `IsOfficialCard`, `MatchesCatalog` and `LiveTariff` to a `ModelRef`: about 20 call sites across 11 files. Most are in `_test.go` files and dereference the old pointer, so `go build ./...` can pass while `go vet` and `go test` fail. |
| `rt-plan-ablation-module` | plan-mode with `complete_step` | Five numbered steps adding a `memory` ablation module, summaries, `ModuleHelp` and `benchctl arm -list`. The prompt opens with the host's own plan-first marker (`control.PlannerRouteMarker`), which routes the turn through the planner, and asks for each step to be signed off before the next. An existing test pins the `all` arm name and has to be updated, which adds a red-test step. |
Every grader stages hidden tests and checks after the run: `gofmt -l`, `go vet
./...` and `go test ./...` for the Go tasks; for `rt-stale-skills`, every
backticked path, identifier, `make` target and `benchctl` command line in each
skill must resolve against the tree, and a SHA-256 manifest proves no Go source, `go.mod`,
`Makefile` or `REASONIX.md` changed. Each task fails its pristine seed and
passes its `solution/`, both checked by `cmd/e2ebench`'s corpus tests.
## Reading the numbers
```text
Metric: requests per task
Anchor: the trajectory digest's requests_by_source, summed over sources
Population: every graded trial of a task that ran to its own stop
Excludes: tasks the token budget skipped, and runs that hit timeout_sec
```
Requests is the headline because round trips are what these tasks exist to
move. Read it beside the solve rate and the report's round outcomes: a change
that drops requests by giving up on verification has not cut a round trip, it
has cut a check. `rt-stale-skills` is read for whether the run finishes at all
and how many rounds go to re-verifying after its last write.
## Baseline
`deepseek-flash`, 2 trials per task, at studio `0b8de9e63`. Requests are the
trajectory's executor requests; all 8 runs ran to their own stop.
prompt="""Some providers bill tokens written into their prompt cache at a rate of their own. Add that to the pricing contract in this repository:
- `RateCard` gains a `CacheWrite` rate, per 1M tokens like the others. Zero means the card has no separate cache-write price, and cache-written tokens keep being charged at `Input` exactly as they are today.
- When a card sets `CacheWrite`, `OriginalCostAmount` charges the billed cache-write units (`CacheWriteBilledTokens` when set, otherwise `CacheWriteTokens`) at `CacheWrite` instead of `Input`.
- `PricingFingerprint` must tell apart two cards that differ only in `CacheWrite`, but a card without one must keep exactly the fingerprint it has today: stored ledgers are keyed by it.
- Two cards that differ only in `CacheWrite` are different cards everywhere cards are compared, including official-card recognition and catalog matching.
- `CatalogEntry` carries the rate too, and `RateCardFromCatalog` and `CatalogEntry.RatesAt` hand it back.
- Add Anthropic to the official price book: `claude-sonnet-4-5`, pay-as-you-go in USD, cache hit 0.30, input 3, output 15 and cache write 3.75 per 1M tokens, documented at https://www.anthropic.com/pricing and checked 2026-10-19. `api.anthropic.com` is Anthropic's own endpoint.
- `benchctl cost` gains a `-write` flag for cache-written tokens, and `benchctl rate` shows a card's cache-write rate when it has one.
- Add tests for the new behaviour and update the README's pricing section."""
{ID:"titled",Fields:map[string]string{"title":"Package manager","body":"We migrated to pnpm."}},
{ID:"mention",Fields:map[string]string{"title":"Sprint notes","body":"Someone asked about the package manager once; also weather, lunch, and a package of stickers."}},
}
hits:=RankV2("package manager",docs)
iflen(hits)!=2||hits[0].ID!="titled"{
t.Fatalf("field-weighted doc must outrank body mentions: %+v",hits)
// Five bytes is more than one token's worth; truncating would let a long
// run of short strings sum to nothing.
ifgot:=Text("abcde");got!=2{
t.Fatalf("Text(\"abcde\") = %d, want 2",got)
}
}
Loaded 100 of 2842 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
esengine