fix(spawn): verify preserved backlog state after interrupted spawn delivery - #3852
Merged
Merged
Conversation
… claims it The deferred-signal exit path asserted the paired task record and In-flight backlog state were preserved without reading either back, exactly when a reader is least able to check (fm-yi4j evidence, 2026-09-05). The commit's exit status alone has been observed to agree with a row that did not actually move. The exit path now re-reads the record and the row under the same per-task lock as the commit, repairs a row the commit believed it moved, and phrases the error as exactly what was verified or attempted - verified preserved, repaired and verified, or an explicit preservation-could-not-be-verified with the reason and hand-closeout instruction. Two behavior tests drive a lying tasks-axi start through a real interrupted spawn and assert the printed claim and the real backlog state agree.
…in backlog gate owner
Confidence Score: 5/5The PR appears safe to merge. No blocking failure remains. Reviews (4): Last reviewed commit: "no-mistakes(ci): Fixed the Greptile P1 o..." | Re-trigger Greptile |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
…signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change
…ib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean
…ib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean
Owner
|
Speaking as Kun's firstmate: this is merged. Thank you @RooseveltAdvisors — really appreciate you taking the time on this. Contract-class: restore — interrupted-spawn exit path already promised preserved backlog/record state; this makes the printed claim match a verified read-back (and honest attempted wording when unverifiable). Timeout bound applies only when VISION (per rule): One captain/interface — aligns (honest outcomes). Authority — aligns (no consent widen). Scripts/agents — aligns (deterministic verify). Restart non-event — aligns (durable backlog/record agree). Delegation spine — aligns (never intent-as-outcome). Fleet outlives vendor — aligns. Scope — aligns (command-layer correctness). CI all SUCCESS on tip; Greptile 5/5; attestation MATCH; MERGEABLE/CLEAN. |
RooseveltAdvisors
added a commit
to RooseveltAdvisors/firstmate
that referenced
this pull request
Sep 6, 2026
Second conflict pass: upstream main advanced another 10 commits (including kunchenguid#3852's interrupted-spawn verification tests and the fm_tasks_axi timeout bound, kunchenguid#3821 docs, kunchenguid#3732 Bash 3.2 work). - bin/fm-backlog-transition-lib.sh: keep this PR's collapsed gate (fm_backlog_source_present + one addressing decision) and route it through upstream's new fm_tasks_axi timeout bound. The addressing decision moved into fm_backlog_tasks_axi_addressing, which publishes FM_BACKLOG_AXI_ROOT/FM_BACKLOG_AXI_FILE for the caller's own single invocation subshell: fm_tasks_axi execs, so the gate's previous nested subshell would have added a process layer between tasks-axi and the lock-holding callers - exactly what kunchenguid#3852's test_dispatch_defers_interruption_across_backlog_commit counts on not existing (its stub kills its grandparent, the spawn). With the helper, the spawn-start topology matches upstream's one-subshell shape and the interruption tests pass. - AGENTS.md: keep this PR's configured-backend phrasing for config/backlog-backend; take upstream's refreshed backend-maturity text for config/backend. - docs/configuration.md: keep this PR's gate description, add upstream's interrupted-spawn verification sentence. Validation: bin/fm-lint.sh full run clean; ShellCheck clean on the resolved lib; tests/fm-backlog-atomicity 94 ok (including the new interruption and timeout cases), fm-lint 38 ok, fm-teardown 80 ok, fm-captain-hold-lifecycle 31 ok, fm-gotmp 3 ok, all rc=0.
RooseveltAdvisors
added a commit
to RooseveltAdvisors/firstmate
that referenced
this pull request
Sep 6, 2026
The PR had gone CONFLICTING against main. Merged forward rather than rebased, so 181bea4 and every published commit stay in this branch's history. One conflict, tests/lib.sh: main added a `umask 022` fixture pin in 85ad5e7 (kunchenguid#3852) after this branch's base, and a pipeline fix round had independently re-derived the same pin here with different comment wording. Resolved by taking main's version verbatim, so the file is byte-identical to main and carries no diff of its own.
eduardstan
added a commit
to eduardstan/firstmate
that referenced
this pull request
Sep 7, 2026
* fix(bin): wait out transient primary-checkout reads in the spawn worktree poll (#3834)
* fix(spawn): keep the worktree poll from adopting the repository primary
After `treehouse get` is sent, the worktree-discovery poll reads the pane's
foreground-process cwd. While treehouse is still fetching and checking a slot
out, the foreground process is treehouse itself and it reports the repository's
PRIMARY checkout as its cwd for several seconds. The poll accepted any path
that merely differed from the spawning project, so from a linked spawning home
- whose project is itself a worktree of that repository - it adopted the
primary, and the isolation guard then refused a launch whose slot treehouse
went on to create normally.
Screen every candidate with the isolation guard's own conditions, extracted as
spawn_worktree_isolated, so a read the guard would reject stays a transient the
poll keeps waiting through. The two-consecutive-reads rule and the guard as
final backstop are unchanged; a pane that never reaches an isolated worktree
still fails at the existing 60s deadline, now naming the last path it reported.
The already-settled timing assertion counted whole-spawn wall time against a
5s budget and failed on unmodified HEAD on slower machines; it now counts pane
reads, which is what "one confirming read, not an extra cycle" actually means.
* fix(spawn): say which path the worktree wait rejected, and why
Screening every discovery-poll candidate means a host that never reaches an
isolated worktree spends the whole 60s window before refusing. That wait is
deliberate - separating a transient from a terminal misconfiguration needs
machinery this path does not want - so the refusal explains itself instead:
the isolation check records why a candidate failed, and the deadline names the
last path seen together with that reason. Message and diagnostics only; the
poll's control flow is unchanged.
Two suites asserted the guard's wording on paths the poll now rejects rather
than adopts, so their refusal arrives from the deadline instead: realign
fm-tangle-guard's non-git and subdirectory-of-primary cases (each now also
asserting the stated reason, and the second the metadata absence it was
missing) and the herdr projection e2e's forced non-worktree cwd.
* no-mistakes(review): stub poll sleep in tangle-guard spawn isolation test
* no-mistakes(document): document spawn poll isolation screen in fm-spawn header
* test(spawn): make the non-git isolation case non-git anywhere
The refusal-reason assertion for a path outside any repository assumed TMPDIR
is not inside a git repository. Where it is, git walks up from the temporary
directory, finds that repository, and the spawn reports the subdirectory cause
instead - so the case passed or failed on a property of the host rather than on
the behaviour under test.
Build the path under a directory the test then names in GIT_CEILING_DIRECTORIES,
which git documents as not chdir-ing up into a listed directory while looking
for a repository. Git never excludes the directory being searched, so the
ceiling is the parent of the path handed to the spawn.
The assertions pin which cause fired rather than the sentence that explains it,
leaving the operator wording free to improve.
* no-mistakes(document): point spawn poll comment at the isolation screen's comparison
* no-mistakes(ci): Fixed the "Behavior portable serial 2" failure in tests/fm-tangle-guard.test.sh ("non-worktree spawn did not say why the path was rejected (missing: 'not inside a git worktree')"). Root cause, in this PR's code: bin/fm-spawn.sh's spawn_worktree_isolated resolved the git toplevel with `wt_top_real=$(cd "$SPAWN_WT_TOP" ...)`. For a path in no repository, `git rev-parse --show-toplevel` yields empty, and `cd ""` is a SUCCESSFUL no-op on bash before 5.3 (CI's ubuntu-latest ships bash 5.2). The empty toplevel therefore resolved to fm-spawn's own cwd — the CI checkout — so the poll reported "it is a subdirectory of worktree root '/home/runner/work/firstmate/firstmate'" instead of the correct "it is not inside a git worktree". Dev machines with bash 5.3 fail `cd ""`, which is why the suite passed locally and only failed on CI; it is a genuine shell-portability defect in the reason vocabulary this change added, not a test-environment artifact. Fix (smallest root-cause change, 1 line + comment, bin/fm-spawn.sh:2168-2173): guard the empty value so it never reaches `cd` — if [ -n "$SPAWN_WT_TOP" ] && ! wt_top_real=$(cd "$SPAWN_WT_TOP" 2>/dev/null && pwd -P); then No change to the poll's timing or deadline behavior (respecting the recorded refusal-latency and spawn-wt-reason-vocabulary decisions), no new tests, no other files touched. Verification: - Reproduced the exact CI failure locally by putting bash 3.2 (same `cd ""` semantics as CI's 5.2) first on PATH: fails before the fix with the identical message shape, passes after. - tests/fm-tangle-guard.test.sh passes under both bash 3.2 and bash 5.3. - tests/fm-spawn-worktree-settle.test.sh and tests/fm-spawn-pool-base-freshen.test.sh pass; shellcheck -x bin/fm-spawn.sh clean. - bin/fm-test-run.sh --changed: 46 suites completed, every FM_TEST_END exit=0, 1076 passing assertions, 0 "not ok" (including fm-tangle-guard, fm-control-relaunch, fm-lint). The run ended on my own 900s wall-clock cap (rc=124), not on any test failure
* fix: support stock macOS Bash 3.2 paths (#3732)
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): read orphaned green ci monitor and daemon-down failed record as not failed (#3846)
* fix(bin): read an orphaned green ci monitor as held-for-merge, not failed
A no-mistakes run held for a captain merge decision keeps its ci step
polling until merged or closed; when the shared daemon restarts under
that poll, the run is recorded failed although every substantive step
completed and GitHub reports the PR green. A monitor whose only
remaining job is to observe a human decision must not convert the
absence of that decision into a failure verdict.
fm-crew-state.sh now reclassifies a terminal failed run as done
(held-for-merge), surfacing the run's PR URL, when the steps table
shows every step completed except exactly ci failed and the ci log's
last recognized marker reads checks green. A genuinely red check, an
unreadable ci log, or a second failed step keeps the failure.
* no-mistakes(review): Read daemon-down coarse failed ledger as unknown, not failed
* no-mistakes(document): docs: align AGENTS.md failed-verdict guidance with crew-state reclassification
* fix(spawn): verify preserved backlog state after interrupted spawn delivery (#3852)
* fix(bin): read preserved spawn state back before the interrupted exit claims it
The deferred-signal exit path asserted the paired task record and
In-flight backlog state were preserved without reading either back,
exactly when a reader is least able to check (fm-yi4j evidence,
2026-09-05). The commit's exit status alone has been observed to agree
with a row that did not actually move.
The exit path now re-reads the record and the row under the same
per-task lock as the commit, repairs a row the commit believed it
moved, and phrases the error as exactly what was verified or attempted
- verified preserved, repaired and verified, or an explicit
preservation-could-not-be-verified with the reason and hand-closeout
instruction. Two behavior tests drive a lying tasks-axi start through a
real interrupted spawn and assert the printed claim and the real
backlog state agree.
* no-mistakes(test): Fix calm suite for Pi 0.85 and pin test umask
* no-mistakes(document): Document interrupted-spawn preservation claim in backlog gate owner
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-spawn.sh's deferred-signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-backlog-transition-lib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean
* no-mistakes(ci): Fixed the Greptile P1 on bin/fm-backlog-transition-lib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean
* docs: correct runtime-backend maturity labels for Herdr (#3821)
* docs: correct stale tmux/herdr backend maturity claims
Herdr now has 21 test files, its own required CI job (tests-herdr)
that installs a pinned build and hard-fails on "skip: herdr not
found", while tmux has 3 test files and is only required as a
dependency of the portable-serial e2e lane. zellij, orca, and cmux
still have no CI lane at all. AGENTS.md and docs/herdr-backend.md
still called Herdr merely "experimental" alongside those three,
misleading every session and reader about actual coverage.
Update AGENTS.md's config/backend entry, the opening lines of
docs/herdr-backend.md and docs/tmux-backend.md, the runtime-backend
section of docs/configuration.md, and the matching claims in
docs/architecture.md, CONTRIBUTING.md, and README.md so they agree
and distinguish tmux (default), herdr (own required CI lane, largest
suite, Windows still spike-only), and zellij/orca/cmux (still
experimental, no CI lane). No behavior, selection order, or
dispatch logic changes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GDYsWPEfwTuPNjQ2nGBCcj
* no-mistakes(review): docs: fix stale herdr label and CI-lane wording
* no-mistakes(document): docs: align tmux adapter label in scripts.md
* no-mistakes(review): docs: drop duplicated herdr CI claim from tmux page
* no-mistakes(review): docs: drop windows claim, align contributing backend wording
* no-mistakes(review): docs: trim duplicated CI claim from herdr opening line
* no-mistakes(review): docs: drop unguarded largest-test-suite superlative
* no-mistakes(review): docs: restore tmux verified label and README experimental scope
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: restore intent-targeted no-mistakes validation (#3865)
* fix: delete deterministic no-mistakes test baseline, restore intent-targeted Test
PR #3644 pinned commands.test to a fm-test-run.sh --changed walk of the
repository's 75-162 tests/*.test.sh scripts. no-mistakes runs commands.test
verbatim and unconditionally after every fix round, so that walk multiplied
by round count: measured at 32.7 minutes per validation versus 3.6 minutes
intent-targeted. Delete the pin and restore the 3.6-minute posture.
Add tests/fm-nm-test-contract.test.sh as a regression guard, parsing
.no-mistakes.yaml as YAML (ruby's bundled Psych, matching the parser
tests/fm-test-run.test.sh already uses for ci.yml) rather than grepping its
text, restoring in legal form what PR #823 added and PR #1282 removed.
Record the rule in docs/configuration.md's "Gate defaults" section (the
authoritative owner CONTRIBUTING.md already points at) and strengthen
CONTRIBUTING.md's existing local-Test guidance to state it plainly: never
configure commands.test to a deterministic test command, complete or partial.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSt9JvrQMVc4u3jPyUFCFC
* no-mistakes(review): Centralize no-mistakes test policy and narrow guard
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: verify Treehouse slot ownership before teardown (#3837)
* fix(bin): verify pool-slot ownership before returning a worktree slot
Workers were killed when cleanup returned a Treehouse pool slot that a
different, live task had already taken. Teardown now proves the slot is
genuinely this task's before releasing it: it refuses when another task
record claims the same live worktree path, or when the endpoint's working
directory contradicts the recorded slot, and that refusal holds under
--force. Slot allocation, metadata publication, ownership verification,
and slot return are serialized across linked firstmate homes, and forced
secondmate cleanup verifies descendant slot ownership before returning
any child worktree.
Regression coverage drives the scripts with two task records naming one
slot path and asserts the live worker survives and its slot is not reset.
* no-mistakes(review): Protect slots across cloned Firstmate homes
* no-mistakes(test): Gate teardown locking on genuine Treehouse slots
* no-mistakes(test): Clarify pooled descendant slot gating
* no-mistakes(test): Synchronize watcher re-arm test on process exit
* no-mistakes(test): Wait for watcher cleanup before timeout escalation
* no-mistakes(document): Document pool-slot ownership safeguards
* no-mistakes(ci): Fixed all reported CI issues: normalized bare local Git origins to the same Treehouse project-lock identity as absolute clone origins; resolved ShellCheck SC1091 with explicit conditional sourcing; and taught concurrent Herdr teardown coverage to retry expected Treehouse lock contention. Added behavioral regression coverage for bare/absolute origin lock identity. Verified endpoint-safety tests, watcher tests, full CI lint, and the previously failing Herdr teardown assertion
* fix(bin): resolve relative origins from repository root
* no-mistakes(ci): Fixed teardown so an exact recorded endpoint may change cwd without falsely vetoing cleanup. Removed cwd-based ownership refusal while preserving cross-home record exclusivity and project locking. Updated behavioral coverage for both foreign slot ownership refusal and moved-cwd teardown success. Endpoint-safety, backend, watcher, checkpoint, and targeted lint checks pass. Real Herdr presentation E2E progressed successfully but exceeded the 600s local timeout
* feat(bin): add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary (#3867)
* feat: add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary
Add omp as a verified harness: anchored process-name detection with a
Firstmate-owned FM_OMP_HARNESS launch marker that needs real omp ancestry,
the fm-spawn launch template with foreign-marker clearing, the tracked
.omp/fm-worker-overlay.yml posture overlay, --auto-approve, --cwd, and
pre-launch model validation scoped to providers 'omp models --json' lists.
Workers get a state-resident busy-state extension keyed on agent_end
without willContinue (omp has no agent_settled). The primary gets two
tracked .omp/extensions: a turn-end guard that answers omp's blocking
session_stop hook by compelling one continuation per turn, with the
pre-tool seatbelts and Run-tier session-start delivery, and a watcher
extension ported from the Pi one with fm_watch_arm_omp. Control tables,
composer busy footers, omp's status row as a bare-composer boundary, the
extension supervision model with an omp-keyed ownership proof, the
session-start diagnostic, and the supervision protocol snippet follow.
Verified live on omp 18.1.11 with openai-codex/gpt-6-astra: a Herdr scout
through spawn, busy state, steer, interrupt, exit, and teardown, and the
isolated rpc primary lab through extension auto-discovery, digest
delivery, lock identity, watcher arm, successor and wake delivery, and
the compelled guard continuation.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test: prove the omp guard continuation through a guard spy
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* fix(spawn): clear the gemini marker at the omp launch boundary
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): force the guard stage by freezing the watcher and clear lint findings
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): reap the live lab by path and record omp's rpc shutdown as a note
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): spawn a real secondmate for the discovery rule and classify the omp surfaces
Replace the template-extraction check with a genuine --secondmate launch
pinned to the fake tmux backend, assert the worker extension's handler set
through the executable rather than its bytes, classify the two new omp
surfaces in the documentation inventory, and record the Herdr worker
evidence in the runtime-backends verification doc.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* no-mistakes(review): omp: unverify remote routes, narrow busy regex, drop overlay approval pin
* no-mistakes(review): omp: validate config-pinned model, correct remote and marker docs
* no-mistakes(review): omp: pin config-model validation with a test, trim overlay
* no-mistakes(review): omp: sync guard evidence, drop dead param, map quota family
* no-mistakes(review): omp quota: refuse unmapped prefixes, match bare model scopes
* no-mistakes(document): docs: cover omp in cd-guard, quota, continuity, tmux
* no-mistakes(document): docs: add omp subagent-guard row, fix live test header
* no-mistakes(ci): Fixed both failing behavior shards and the Greptile P1 in bin/fm-composer-lib.sh. Root cause of "Behavior portable serial 1" and "Behavior portable parallel 2": the omp busy regex (FM_DELIVERY_OMP_BUSY_REGEX_DEFAULT) and omp status-row furniture regex (FM_COMPOSER_OMP_STATUS_RE_DEFAULT) used the bracket range [⠁-⣿]; BSD grep on macOS accepts it but GNU grep on Linux CI aborts with "Invalid collation character", failing every omp busy/furniture read (3 assertions across fm-omp-harness, fm-tmux-submit-busy, fm-composer-lib). Replaced the range with one shared explicit alternation FM_OMP_SPINNER_FRAMES_RE of omp 18.1.11's unicode-preset spinner frames (status set ⣾⣽⣻⢿⡿⣟⣯⣷ + activity set ⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏, read from the installed binary), the same pattern the Kimi busy regex already uses in CI. For Greptile's finding (the harness-agnostic furniture rule's first alternative matched any 1–4-byte token + ' · ', so wrapped typed input like 'fix · tests' with the cursor on it regressed from pending to unknown; reproduced locally vs base), pinned that alternative to omp's identity cell (π||pi, the icon.omp of each preset in the 18.1.11 binary). Tests: fm-composer-lib.test.sh asserts 'fix · tests' is not furniture, a status-set spinner row is furniture, and the wrapped composer screen reads pending under both locales (CAPS_TMUX cursor 3); fm-omp-harness.test.sh asserts a status-set frame reads busy. New negative cases fail against the pre-fix lib and pass after. Verified: fm-omp-harness, fm-tmux-submit-busy pass via bin/fm-test-run.sh; fm-composer-lib passes all cases except one pre-existing, unrelated local failure (Herdr half-block test uses printf '▀', unsupported by macOS bash 3.2; fails identically on a pristine HEAD export, passes on CI bash 5); shellcheck and bin/fm-lint.sh clean. Caveat: GNU grep is unavailable locally, so the Linux compile was not run directly; the fix uses only constructs already proven on CI's GNU grep (multibyte literal alternations, incl. under LC_ALL=C). Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh, tests/fm-omp-harness.test.sh. No docs needed changes (they describe the rule generically)
* no-mistakes(ci): Greptile Review: fixed. The omp status-row furniture regex FM_COMPOSER_OMP_STATUS_RE_DEFAULT in bin/fm-composer-lib.sh still accepted a literal `pi ·` opening, so wrapped composer input beginning with `pi ·` was truncated and misclassified. Read the installed omp 18.1.11 binary: the ascii preset's `icon.omp` is `pi` but its `sep.dot` separator is ` - ` (unicode/nerd use ` · `), so a real ascii status row never contains `pi ·` and that alternative could only ever match typed text. Removal-first fix: dropped `pi` from the identity alternation (now `(π|)`) and updated the comment to record why the ascii preset is excluded. Tests (tests/fm-composer-lib.test.sh): added a negative furniture case for 'pi · e · phi as the three constants' and a wrapped-screen assertion (CAPS_TMUX, cursor 3) that a continuation row opening `pi ·` reads pending in both locales; the new case fails against the unfixed lib and passes after. Verified: composer test with the half-block case skipped passes all 33 cases including the omp matrix; bin/fm-test-run.sh tests/fm-omp-harness.test.sh passes; shellcheck -x clean on both files; bin/fm-lint.sh clean. The full composer test via the runner fails locally only on the pre-existing half-block case (bash 3.2 printf cannot emit ▀; passes on CI bash 5), identical to before this change. Docs unchanged (they describe the rule generically and never mention the ascii identity cell). PR must be raised via no-mistakes: not caused by code. attestation.head_sha is cdddc60 while the PR head is cd51cf4 because the pipeline's ci-phase push moved the head; the outer executor's re-push will re-bind the attestation. No file change for that check. Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* fix(pi): invoke Bash helpers correctly on native Windows (#3843)
* Fix Pi shell invocation on native Windows
* no-mistakes(document): Document Pi Windows Bash transport
* no-mistakes(ci): Captain, staged a narrow fix: register the Pi Windows regression for both extension paths, make Windows mode emulation non-failing, and enforce LF shell checkouts. Mapping and coverage checks pass; CI/Require no-mistakes were approval-gated externally
* no-mistakes(review): Cover async Windows branch-outcome Bash invocation
* no-mistakes(review): Preserve Cygwin checks and refresh Windows timing
* no-mistakes(test): Invoke OpenCode operational-input owner through Bash on Windows
* validation-fixture
* no-mistakes(document): Document Windows Bash helper invocation
* no-mistakes(ci): Fixed PR-caused changed-selection failure by removing the malformed tracked evidence artifact and allowing deleted, unconsumed source paths to retire cleanly while preserving fail-closed behavior for live unmapped paths. Added regression coverage. Verified native-Windows Pi shell-seam test passes and --changed selects the Windows regression
---------
Co-authored-by: test <test@example.invalid>
* test(bin): pin teardown outcomes for squash-merged rebased branches (#3870)
* fix(bin): recognise squash-merged rebased work as landed at teardown
A pipeline rebase can leave the local worktree on pre-rebase commits while
GitHub squash-merges the rebased head. The landed-work test then compared
those stale commits against a squashed main and refused cleanup of work
that had already landed.
When the forge reports the recorded PR merged and its merge commit is on
the default branch, treat a local branch that only repeats paths from the
pipeline push as stale rather than unlanded. If the forge is unreachable,
the same coverage check runs against a PR head whose content is already
on default. Extra local paths still refuse.
* fix(bin): drop unprovable squash-rebase landed-work coverage
Path-set coverage treated a diverged local branch as landed whenever it
touched the same files as the squash merge. That accepts the reviewer's
failing sequence: same path, different content, work discarded.
git cherry and merge-tree containment were already too strict on the real
rebase-fold case. No remaining check is both safe and permissive enough
to recognise a stale pre-rebase copy without also accepting unlanded
edits, so that case still refuses.
Keep the proofs that hold: a merged PR head that contains local work, or
a clean content-in-default tree match. Tests now refuse same-path
different content and extra unlanded commits, and still allow a local
branch that followed the pipeline rebase.
* no-mistakes(review): drop recorded-pr-head fallback and reverted-design leftovers
* no-mistakes(review): silence squash-merge stdout corrupting test PR head
* no-mistakes(review): make unlanded follow-up commit sole cause of refusal
* no-mistakes(document): correct stale squash-rebase fixture comments in teardown tests
* no-mistakes(ci): Split the three reported checks: - CI (run 34061098467) and Require no-mistakes (run 34061098460) both concluded `action_required` — approval-gated workflow runs that never executed a step. Not caused by this PR's code; no change can clear them. - Greptile Review was a genuine defect in the new tests: the three new refusal cases (tests/fm-teardown.test.sh) asserted only exit status 1 and a REFUSED line, so a teardown regression that destroyed the worktree, branch, and task record before reporting refusal would still pass. Fix (tests only): added one `assert_refusal_retained_task_state` helper and called it from `test_squash_merged_same_file_different_content_refuses`, `test_squash_merged_rebased_local_with_unlanded_commit_refuses`, and `test_squash_merged_stale_local_refuses_when_forge_unreachable`, each capturing the worktree HEAD before `run_teardown`. It pins that the refusal left the isolated copy on disk, the task branch still checked out at the same unlanded commit, and state/task-x1.meta intact. Verification: the four squash tests pass; a sensitivity probe ran the ALLOW fixture (teardown completes) and pointed the same helper at the outcome — it fires, because a completed teardown detaches/deletes the branch and removes the task record, proving the assertions discriminate. Full tests/fm-teardown.test.sh: 83 passing. bin/fm-lint.sh clean with pinned ShellCheck 0.11.0 + actionlint 1.7.12 (plus an explicit --external-sources pass on the changed file). bin/fm-test-run.sh --check-coverage ok. Caveat: test_herdr_flat_teardown_preflight_refuses_before_changes (mode missing-adapter) fails on this machine. Verified it fails identically on base commit f91a950 via `git archive`, so it is a pre-existing local environment difference untouched by this diff; skipped to run the rest of the suite, not modified
---------
Co-authored-by: Morten Gad <mogad@itm8.com>
* fix(bin): keep supervision armed for registered custom checks (#3860)
* fix(bin): keep supervision armed for registered custom checks
A custom check bound by bin/fm-check-register.sh only ever runs inside the
watcher's check sweep, but fm_supervision_status counted in-flight tasks, the
relay poll shim, and process-event sources as supervision need, and not
registered checks. Tearing down the last task therefore stopped every
home-level check silently until the next spawn.
Count a state/<id>.check.sh that carries its state/<id>.check-trust binding as
supervision need. The relay shim keeps its own trust path and task PR polls
carry no such binding and are torn down with their task, so neither arms a home
by accident. Presence of the binding is the whole test: the sweep validates the
bytes at execution time and wakes firstmate when it rejects one, which is the
outcome an idle home needs.
Closes #3856
* no-mistakes(review): name registered checks in turn-end block banner and doc invariant
* no-mistakes(review): narrow PR poll predicate test to what it proves
* no-mistakes(document): point Grok re-arm step at supervision-need owner
* fix(bin): resolve Treehouse locks for remote secondmate homes (#3883)
* fix(bin): resolve the shared Treehouse project lock inside remote secondmate homes
Every spawn and teardown inside a remote-seeded secondmate home refused,
because the project lock's anchor could not be resolved there.
fm_firstmate_root_home walks a home's parent bindings upward to find the
anchor the lock lives in, and treated a remote parent binding as an error.
A remote-seeded home's parent is on another machine, so that walk can never
succeed from there - and neither can the home's own local descendants, whose
chain terminates at the same record. Both fail closed on every Treehouse-backed
spawn and every pool-slot teardown.
A remote parent now terminates the walk at the home holding it, which is the
correct anchor: a lock taken on this filesystem is neither held nor observable
across that boundary, and that home is already the top of the local tree
teardown's collect_local_firstmate_states enumerates, since that walk skips
remote registry entries for the same reason. Mutual exclusion is unchanged -
every home reachable through local parent links still derives one identical
lock file per project, and an unreadable binding, an unsupported route, an
unreachable local parent, a cycle, and an over-deep chain all still refuse.
Origin-less local-only projects keep resolving through their worktree top.
Regression coverage pins the anchor for the main-home layout, a local
secondmate, a remote-seeded home, and its local child; drives teardown
end-to-end in a remote-seeded home; keeps the cross-home slot-ownership
refusal across that boundary; and proves two homes still serialize on the
one shared lock file.
* no-mistakes(document): Clarify machine-local Treehouse lock ownership
* fix(bearings): repair board listening and decision reconciliation (#3872)
* fix(bearings): repair the board's listening, card hygiene, and reconcile path
Three defects made the fleet board go quiet and then lie about what still
needs the captain.
Never arm a poll on a session that is not live. `lavish-axi <file>` exits 0
even when it refuses to reopen a session the captain ended from the browser,
reporting `status: user-ended` with the same session id, so the build's
exit-status check accepted a dead session, printed `already-armed`, and left
the board reading "not listening". The build now proves the session is live
from a fresh authoritative listing immediately before arming - not from the
establish call's status alone, which is already stale by then - reopens once
when it finds the session ended, and refuses rather than arming when it stays
ended. A reopen also replaces the pre-reopen source generation before
reporting success, so a runner on its way out cannot be mistaken for a
listener, and a board whose source is registered but unowned gets a
replacement started before the build returns.
Let a dead generation's ownership actually move. Reclaiming a claim ran its
capture-reservation cleanup first, and that cleanup re-verifies the recorded
state-root identity, so a claim naming a pid and a process group that were
both provably gone could not be cleared: reconcile reported a start while
nothing attached, and retire refused with "cannot release source ownership".
Reservation records are keyed by claim token and every replacement claims a
fresh one, so they are hygiene, not an ownership invariant. Reclamation now
additionally requires the owning process group to be absent independently,
which keeps a reused pid whose poll child still runs from ever reading as a
gone generation. A live owner and a crashed leader whose owned group survives
are still never reclaimed.
Stop carding decisions whose subject already landed. The build drops a
decision card whose work item or PR appears in the payload's own landed rows,
and one whose task is no longer an open captain call, naming each drop on
stderr. A task whose state cannot be established is kept, because a call
wrongly hidden is worse than a card wrongly shown.
Add the reconcile choice, and make it structurally incapable of closing a
call. Every decision card carries a standard `reconcile` option, injected by
the build rather than left to the composer. The board now emits the picked
option and any freeform note as separate structured fields instead of fusing
them, so a reconcile selection is not expressible as an answer value at all -
the defect that let `reconcile - <note>` reach the intake as an ordinary
answer. The adapter routes selections from that structured field, creation of
a reconcile request is bound to a verified board source rather than the shared
keyed-answer intake, and the intake still refuses the reserved value on every
channel. Each authorization is bound to the captain-hold generation that
produced the card, so an obsolete card cannot close a later call, and both
terminal outcomes require a pending request: `reconcile close` records the
evidence under its own `reconciled` mode so it never reads as the captain's
words, and `reconcile note` leaves the call open. Anything unprovable -
an unversioned row, a missing generation, an unreadable state - refuses
rather than acting.
Regression coverage fails without each fix, and pins every leak path: a bare
reconcile, a standalone close or note with no pending request, an any-channel
reconcile, an annotated selection from a freeform card, and a
generation-skewed authorization. An opt-in guard re-proves the lavish-axi
shapes and the reopen against the installed tool.
* fix(bin): quote the done comparison in the reconcile intake
shellcheck SC1010 reads the bare word as the loop keyword. The failed run
never reached its lint step, so this shipped in the recovered content.
* no-mistakes(review): Publish reconciled parent resolution before request retirement
* no-mistakes(review): Clarify committed cleanup and reconcile reservation scope
* no-mistakes(review): Preserve remote cards and legacy answer compatibility
* no-mistakes(test): Separate live claim release from stale reclamation
* no-mistakes(test): Allow terminal self-retirement during active capture
* no-mistakes(document): Document Bearings repair contracts
* no-mistakes(ci): Stabilized the failing Herdr presentation E2E by serializing test-harness Treehouse allocator calls, preventing concurrent recovery spawns from claiming the same pool slot while preserving Herdr concurrency coverage. Verified with the full E2E suite on Herdr 0.8.2, bash syntax checks, ShellCheck, and git diff checks
* fix(bin): allow pooled spawns without a git origin (#3885)
* fix(bin): skip pooled-worktree freshness fetch when no origin is configured
An origin-less local-only project has nothing remote to be stale
against, so fm-spawn's freshen_spawn_worktree_base refused to launch
crews for it. Detect a missing origin remote and skip the fetch
freshness gate entirely; an existing-but-unreachable origin keeps
refusing as before.
* no-mistakes(review): Preserve pool safety for absent and unusable origins
* no-mistakes(review): Refuse empty origin configurations during pooled spawn
* no-mistakes(review): Detect empty origin sections across config includes
* no-mistakes(review): Honor globbed includes when detecting origin configuration
* no-mistakes(review): Document conservative conditional include handling
* no-mistakes(review): Use Git-resolved config files for origin detection
* no-mistakes(review): Document included empty-origin detection boundary
* no-mistakes(document): Document originless pooled spawn behavior
* feat(tests): run live harness guards by default when available (#3889)
* feat(tests): run live harness guards by default where the harness is installed
The 24 live-harness guards each opened with their own env check, so on the
machine that has every harness - the one the product and its validation
actually run on - all of them skipped and passed. Fourteen had never been run
by the pipeline at all.
tests/lib.sh gains fm_live_gate as the single owner of that decision: a guard
that spends no model tokens runs wherever its tools are installed, a guard that
submits prompts stays opt-in, an absent tool is a named capability skip, and a
guard's own variable or FM_LIVE forces it on (turning an absent tool into a
failure) or off. Every live guard now opens with it, which also carries the
test-suite gate-refusal bypass into the guards that never sourced the shared
helpers and were therefore refused whenever a gate agent ran them.
bin/fm-test-run.sh records what a skip means: the family's expected class is
live-capability rather than a bare env opt-in, and each gate skip's reason is
logged and written to the timing artifact, so a lane can say which tool this
host could not exercise.
Only the token-free guards flip to default-on: composer-matrix, the harness
liveness drift guard, and the Herdr version floor. cursor-primary submits three
prompts, so it stays opt-in.
Running the drift guard unasked immediately found a real defect it existed to
catch: it resolved the harness through a generic `command -v cursor`, which on
a machine that also has the Cursor editor finds the editor launcher rather than
cursor-agent. That binary exits at once, leaving a bare shell in the pane and a
liveness-drift failure no classifier change could fix. It now asks
fm_cursor_resolve_binary first, the same verified owner fm-spawn uses.
CI installs the public Pi package in the portable serial lane and fails on its
skip token, so the Pi extension tests stop passing silently against a package
that is not there. No secret is added.
Verified on macOS 26.5.2 arm64: the drift guard runs with no variable set and
classifies 8 installed harnesses alive; the Herdr version-floor guard runs by
default and checks 4 real releases; every live guard refuses together under
FM_LIVE=0.
* fix(tests): keep the composer-matrix guard opt-in
Running it unasked is red on a healthy machine for reasons no code change here
removes: a harness that has not trusted this checkout sits on its own trust
dialog, which the guard treats as an unreadable composer and correctly fails.
The opencode 1.18.29 and grok 1.0.13 composer drift it also surfaced reproduces
identically on main and is filed as separate work.
So this token-free guard stays opt-in with the reason stated in its header, and
the coding guidelines record the narrow exception: a guard whose verdict
depends on host state that installing its tools does not establish may stay
opt-in, because one that is permanently red is one the fleet learns to ignore.
The other two token-free guards keep running by default.
* no-mistakes(review): Wire bearings guard and remove composer exception policy
* no-mistakes(review): Run Pi responsiveness guard by default
* no-mistakes(review): Gate AFK Pi Herdr through authoritative family sweep
* no-mistakes(review): Sanitize live gate test environments
* no-mistakes(document): Document default-on live guard behavior
* no-mistakes(ci): Fixed CI by installing the Pi package in portable-parallel-1, where fm-pi-primary-types.test.sh runs, and enforcing its package-missing gate skip there. Verified with fm-lint.sh, workflow actionlint, coverage partition checks, lane membership, and git diff checks
* no-mistakes(ci): Fixed CI’s Pi typecheck skip enforcement by giving npm, tsc, and Pi-package capability skips a shared prefix and configuring both relevant CI lanes to fail on that prefix. Verified missing tsc emits the expected skip, missing Pi package becomes a runner failure, and actionlint, ShellCheck, and git diff checks pass
* fix(bin): refuse test runs in the primary checkout when a task marker is set (#3891)
* fix(bin): refuse the behavior suite in the repository primary checkout
A task worker's isolated worktree placement is verified exactly once, when
its task starts, and nothing re-checks it afterwards. A worker that later
changes directory into the repository's primary checkout runs its Git
commands, and this branch-switching suite, against the one checkout every
linked worktree resolves against and every landing merges into. A run that
dies mid-suite can leave that checkout on a stray branch.
bin/fm-test-run.sh now refuses that case. When FM_TASK_ID marks a task
worker and the runner resolves to the primary checkout, every executing mode
exits non-zero before selecting a suite, with one line naming the primary
path and pointing at the assigned task worktree. The predicate is the one
bin/fm-spawn.sh already uses for launch placement: the working tree's own
git dir is the repository's common git dir, which separates the primary from
every linked worktree even when their top levels differ. A run with no
FM_TASK_ID set is unchanged, and so are the inspection modes, which execute
nothing. When git resolves neither directory - a non-repository fixture, a
detached copy - nothing proves this is the primary, so the run proceeds.
bin/fm-spawn.sh sets the marker: ship and scout launches export FM_TASK_ID
into the pane shell on the same pre-launch channel as GOTMPDIR, and the name
joins the sanitized launch environment allowlist so an isolated launch keeps
it.
* no-mistakes(review): clear inherited task marker in test lib; name resolved ROOT
* no-mistakes(document): docs: record FM_TASK_ID marker and runner placement refusal
---------
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): bound stale alarms for backlog captain holds (#3842)
* fix(bin): bound a stale alarm with the backlog hold, not only the status line
A legitimate wait has two records and the stale alarm reads only one.
`status_is_paused_or_captain_held` takes a status line, so it sees a wait the
worker declared. It cannot see the wait firstmate records when it hands work to
the captain: `bin/fm-captain-hold.sh hold` writes that into the backlog and
leaves the status log alone, so a delivered task keeps `done: PR ...` as its
last line for the whole time the captain is deciding.
Both stale branches were blind to it, and each churned a new pane hash back into
its own alarm: a `done:` line is captain-relevant and reaches the terminal-stale
branch, while a held task whose last line is `working:` reaches
`surface_nonterminal_stale` and fails its declared-wait test.
Consult that second record where the watcher is about to alarm, through
`bin/fm-captain-hold.sh open`, which already owns the predicate's semantics, and
bound the alarm on the shared `.paused-resurfaced-<key>` marker and
`PAUSE_RESURFACE_SECS` window the declared-wait absorb already uses. The first
sight still alarms, the window's end alarms once more, and a held crew that goes
genuinely silent still escalates through the wedge timer.
Only an established open captain call bounds anything: an unreadable backlog, an
absent or incompatible tasks-axi, a row this home does not carry, and every task
with no hold keep alarming exactly as before. The backlog hold is deliberately
not recorded as a declared pause, because the loop-top reconciliation and
`pause_state_class` both read the status line and would clear a flag that line
does not support.
Extends the fix in #3443, which closed the forms of this loop that the status
line itself can express.
* fix(bin): identify the captain call a stale alarm is bounded by
Three gaps in the bound added by the previous commit, all in how the throttle is
scoped and where the backlog is consulted.
The scope carried only the status-log signature. A task can be held, answered
with `--release`, and re-held as a genuinely different captain call without any
status append, so the second call inherited the first one's marker and its first
sight was absorbed - the one thing this bound must never do. The task id is not
the call: `bin/fm-captain-hold.sh open` gains `--identity`, which reports the
call's own lifecycle - its hold-set stamp and the number of recorded answers -
on an exit 0 and only then, leaving the silent predicate every existing caller
reads unchanged. The throttle scope now carries that identity.
The terminal path recorded the throttle before publishing the durable wake. A
failed append exits the watcher with nothing queued, and the next sighting then
read that fresh marker and absorbed the retry, turning a delayed alarm into a
lost one. Recording moves behind the append, as the non-terminal path already
had it, and the comment claiming the marker could not outlive its wake is gone
because it was false.
The backlog was consulted only on a new terminal pane hash. A captain call can
open after a hash was absorbed as provably working, changing neither the pane nor
the status log, so nothing re-read the backlog and the wedge timer kept firing
possible-wedge alarms through a legitimate wait. That timer now consults the call
at its own alarm boundary and takes the same bounded cadence - and only at that
boundary, so an ordinary repeat poll under the bound stays the local-only read it
was.
Regression coverage for each, all driving churn through one watcher process
rather than relaunching per pane change: relaunch cost dominated the earlier
shape, and an absorbing watcher stays in its poll loop across churn in production
anyway. An unheld task still alarms on every new hash, and an elapsed wedge timer
with no open captain call still escalates as a possible wedge.
* fix(review): Compose stale throttles with captain-call lifecycle identity
* fix(review): Preserve bounded same-hash captain-call resurfacing
* revert(bin): narrow the captain-hold stale bound to its observed defect
Lifts the lifecycle-identity and cadence-ownership work back out, leaving the
change at the shape that matches the defect actually observed: the stale alarm
did not consult the backlog captain hold, on either stale branch.
Reviewing the wider version surfaced a series of adjacent gaps in the watcher's
alarm state machine - a call opening after the first alarm, marker invalidation
at the hold lifecycle boundary, and which deadline a terminal timer represents.
They are real, but fixing them turns a small extension into a state-machine
change to the alarm path, which is a different review on a subsystem that is
being actively reworked. They are named as known limitations rather than carried
here, and none of them is load-bearing for what remains: the bound does strictly
less than the reverted version, leaves the wedge path escalating on
STALE_ESCALATE_SECS exactly as before, and introduces no silence that the
existing terminal-alarm path did not already have.
Kept from the reverted work is the record-after-append ordering, because that is
a defect in the code being shipped rather than an adjacent one: recording the
cadence marker before publishing the durable wake let a failed append lose an
alarm outright instead of delaying it.
History is preserved: the earlier commits stay on the branch and this removal
sits on top of them.
* fix(review): Document secondmate captain-hold scope boundary
* fix(document): Document captain-hold stale alarm scope
* fix(bin): bind the stale throttle to the captain call, not the status log
The throttle this change introduces was scoped to the task's status-log
signature. Answering a call with `--release` and holding the task again creates a
genuinely different captain call without necessarily appending to that log, so
the second call inherited the first one's marker and its first sight was
absorbed.
That is the one alarm this bound must never swallow. A delivery announced twice
is noise; a decision waiting on the captain that is never surfaced is invisible,
because nobody asks for what they do not know to ask for.
Measured rather than assumed, on the same fixture - a delivered task held for the
captain, released, and re-held with no status append, driven through bin/fm-watch.sh:
base c499f84 call-1 first=ALARM call-1 churn=ALARM new call first sight=ALARM
before this fix call-1 first=ALARM call-1 churn=absorbed new call first sight=absorbed
after call-1 first=ALARM call-1 churn=absorbed new call first sight=ALARM
Base never suppresses the new call, so the suppression came from this change and
closing it completes the fix rather than widening it.
`bin/fm-captain-hold.sh open` gains `--identity`, printing the call's lifecycle -
its hold-set stamp and count of recorded answers - on an exit 0 and only then, so
the silent predicate bin/fm-teardown.sh reads is untouched. The throttle scope
carries that identity beside the status signature.
The sibling case was measured too and is NOT included: on the status-declared
path, where the last line is `captain-held:`, base already absorbs a re-held
call's first sight. That behaviour predates this change and stays documented as a
known limitation rather than repaired here.
* fix(document): Document captain-call throttle lifecycle scope
* fix(ci): isolate the Herdr restart fixtures from a claimed worktree
The Herdr behaviour test intermittently reused a local worktree still claimed
by an earlier fixture after a restart.
The restart scenarios now use an isolated Treehouse project.
The full Herdr test passes on Herdr 0.8.2; bash -n and git diff --check pass as
well.
* feat(pi): resolve extension-registered providers in the supervision branch (#3871)
* Let the supervision branch resolve extension-registered providers
The isolated branch ModelRuntime cannot see providers an extension
registered into main's runtime at run time, so a pin on pi-devin-auth's
devin/swe-1-7 (or an unpinned branch following a main session on devin)
failed with "unavailable to the isolated branch runtime".
Capture main's ModelRegistry alongside mainModel and copy each
extension-registered provider config into the branch runtime at
model-resolution time. The config carries the provider's own streamSimple
and oauth wiring by reference, so the custom gRPC transport reaches the
branch unchanged instead of being reimplemented. The /supervision-model
picker uses the same copy so those models are offered.
Update configuration.md and pi-supervision-branch.md, which previously
stated extension-registered providers were not offered.
* no-mistakes(document): docs: own devin provider carve-out in branch architecture doc
* no-mistakes(ci): Fixed the Greptile P1 finding: the /supervision-model picker copied extension-registered providers into the branch ModelRuntime but checked hasConfiguredAuth without refreshing them, so providers with provisional post-registration auth were omitted from the picker while the pin-resolution path (which did refresh) accepted them. Root-cause fix in .pi/extensions/fm-branch-supervision.ts: moved the `refresh({ providers, allowNetwork: false })` call into `copyExtensionProviders` (now async, refreshing every provider it copied) and removed the duplicate per-provider refresh from `resolveBranchModel`. Both the picker and the resolution path now share one copy-and-refresh step, so hasConfiguredAuth is real in both. Regression coverage in tests/fm-pi-branch-extension.test.sh: the stubbed ModelRuntime now mirrors the real runtime by leaving a registered provider's auth pending until `refresh()` runs for it. With that stub, the existing extension-registered-provider case fails against the pre-fix extension (picker offers only anthropic/main-model) and passes with the fix. Verification: tests/fm-pi-branch-extension.test.sh passes (42 ok, no failures); tests/fm-branch-supervision.test.sh passes; tests/fm-pi-primary-types.test.sh skips locally because tsc is not installed (the refresh signature reused is the one the existing code already called). Intent constraints preserved: isolation flags untouched, carve-out still scoped to provider registration, graceful fallthrough when no providers are registered
* ci: retrigger flaky Herdr/serial-1 lanes
* no-mistakes(document): docs already cover branch extension-provider copy
---------
Co-authored-by: Tiago <tiagop@hey.com>
Co-authored-by: 3264studios <3264studios@gmail.com>
Co-authored-by: Talon Stark <talonstark@gmail.com>
Co-authored-by: Jon Roosevelt <rooseveltadvisors@gmail.com>
Co-authored-by: kerry morrison <141675980+kmorebetter@users.noreply.github.com>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Co-authored-by: Kun Chen <3233006+kunchenguid@users.noreply.github.com>
Co-authored-by: Daniel Kuykendall IV <danielkuykendall23@gmail.com>
Co-authored-by: Cristian Rosescu <crosescu@gmail.com>
Co-authored-by: test <test@example.invalid>
Co-authored-by: MortenGad <55603022+MortenGad@users.noreply.github.com>
Co-authored-by: Morten Gad <mogad@itm8.com>
Co-authored-by: Gyute <9tempo@gmail.com>
Co-authored-by: Mickaël Rémond <mremond@process-one.net>
Co-authored-by: Pedro Guimarães <pedroguim@pm.me>
ysgdepaula
added a commit
to ysgdepaula/firstmate
that referenced
this pull request
Sep 8, 2026
* fix: support stock macOS Bash 3.2 paths (kunchenguid#3732) Co-authored-by: Talon Stark <talonstark@gmail.com> * fix(bin): read orphaned green ci monitor and daemon-down failed record as not failed (kunchenguid#3846) * fix(bin): read an orphaned green ci monitor as held-for-merge, not failed A no-mistakes run held for a captain merge decision keeps its ci step polling until merged or closed; when the shared daemon restarts under that poll, the run is recorded failed although every substantive step completed and GitHub reports the PR green. A monitor whose only remaining job is to observe a human decision must not convert the absence of that decision into a failure verdict. fm-crew-state.sh now reclassifies a terminal failed run as done (held-for-merge), surfacing the run's PR URL, when the steps table shows every step completed except exactly ci failed and the ci log's last recognized marker reads checks green. A genuinely red check, an unreadable ci log, or a second failed step keeps the failure. * no-mistakes(review): Read daemon-down coarse failed ledger as unknown, not failed * no-mistakes(document): docs: align AGENTS.md failed-verdict guidance with crew-state reclassification * fix(spawn): verify preserved backlog state after interrupted spawn delivery (kunchenguid#3852) * fix(bin): read preserved spawn state back before the interrupted exit claims it The deferred-signal exit path asserted the paired task record and In-flight backlog state were preserved without reading either back, exactly when a reader is least able to check (fm-yi4j evidence, 2026-09-05). The commit's exit status alone has been observed to agree with a row that did not actually move. The exit path now re-reads the record and the row under the same per-task lock as the commit, repairs a row the commit believed it moved, and phrases the error as exactly what was verified or attempted - verified preserved, repaired and verified, or an explicit preservation-could-not-be-verified with the reason and hand-closeout instruction. Two behavior tests drive a lying tasks-axi start through a real interrupted spawn and assert the printed claim and the real backlog state agree. * no-mistakes(test): Fix calm suite for Pi 0.85 and pin test umask * no-mistakes(document): Document interrupted-spawn preservation claim in backlog gate owner * no-mistakes(ci): Fixed the Greptile P1 in bin/fm-spawn.sh's deferred-signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change * no-mistakes(ci): Fixed the Greptile P1 in bin/fm-backlog-transition-lib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean * no-mistakes(ci): Fixed the Greptile P1 on bin/fm-backlog-transition-lib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean * docs: correct runtime-backend maturity labels for Herdr (kunchenguid#3821) * docs: correct stale tmux/herdr backend maturity claims Herdr now has 21 test files, its own required CI job (tests-herdr) that installs a pinned build and hard-fails on "skip: herdr not found", while tmux has 3 test files and is only required as a dependency of the portable-serial e2e lane. zellij, orca, and cmux still have no CI lane at all. AGENTS.md and docs/herdr-backend.md still called Herdr merely "experimental" alongside those three, misleading every session and reader about actual coverage. Update AGENTS.md's config/backend entry, the opening lines of docs/herdr-backend.md and docs/tmux-backend.md, the runtime-backend section of docs/configuration.md, and the matching claims in docs/architecture.md, CONTRIBUTING.md, and README.md so they agree and distinguish tmux (default), herdr (own required CI lane, largest suite, Windows still spike-only), and zellij/orca/cmux (still experimental, no CI lane). No behavior, selection order, or dispatch logic changes. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GDYsWPEfwTuPNjQ2nGBCcj * no-mistakes(review): docs: fix stale herdr label and CI-lane wording * no-mistakes(document): docs: align tmux adapter label in scripts.md * no-mistakes(review): docs: drop duplicated herdr CI claim from tmux page * no-mistakes(review): docs: drop windows claim, align contributing backend wording * no-mistakes(review): docs: trim duplicated CI claim from herdr opening line * no-mistakes(review): docs: drop unguarded largest-test-suite superlative * no-mistakes(review): docs: restore tmux verified label and README experimental scope --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> * fix: restore intent-targeted no-mistakes validation (kunchenguid#3865) * fix: delete deterministic no-mistakes test baseline, restore intent-targeted Test PR kunchenguid#3644 pinned commands.test to a fm-test-run.sh --changed walk of the repository's 75-162 tests/*.test.sh scripts. no-mistakes runs commands.test verbatim and unconditionally after every fix round, so that walk multiplied by round count: measured at 32.7 minutes per validation versus 3.6 minutes intent-targeted. Delete the pin and restore the 3.6-minute posture. Add tests/fm-nm-test-contract.test.sh as a regression guard, parsing .no-mistakes.yaml as YAML (ruby's bundled Psych, matching the parser tests/fm-test-run.test.sh already uses for ci.yml) rather than grepping its text, restoring in legal form what PR kunchenguid#823 added and PR kunchenguid#1282 removed. Record the rule in docs/configuration.md's "Gate defaults" section (the authoritative owner CONTRIBUTING.md already points at) and strengthen CONTRIBUTING.md's existing local-Test guidance to state it plainly: never configure commands.test to a deterministic test command, complete or partial. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CSt9JvrQMVc4u3jPyUFCFC * no-mistakes(review): Centralize no-mistakes test policy and narrow guard --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> * no-mistakes(review): Fix timeout portability and terminate hung task descendants * no-mistakes(review): Reuse shared GNU timeout cleanup and expand descendant regressions * no-mistakes(document): Correct stale backend and supervision documentation * no-mistakes: apply CI fixes * no-mistakes: apply CI fixes * no-mistakes(document): Clarify supervision and spawn safety comments * no-mistakes: apply CI fixes * no-mistakes(document): Clarify Calm working indicator documentation --------- Co-authored-by: 3264studios <3264studios@gmail.com> Co-authored-by: Talon Stark <talonstark@gmail.com> Co-authored-by: Jon Roosevelt <rooseveltadvisors@gmail.com> Co-authored-by: kerry morrison <141675980+kmorebetter@users.noreply.github.com> Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> Co-authored-by: Kun Chen <3233006+kunchenguid@users.noreply.github.com>
lytv
pushed a commit
to lytv/mymate
that referenced
this pull request
Sep 8, 2026
…livery (kunchenguid#3852) * fix(bin): read preserved spawn state back before the interrupted exit claims it The deferred-signal exit path asserted the paired task record and In-flight backlog state were preserved without reading either back, exactly when a reader is least able to check (fm-yi4j evidence, 2026-09-05). The commit's exit status alone has been observed to agree with a row that did not actually move. The exit path now re-reads the record and the row under the same per-task lock as the commit, repairs a row the commit believed it moved, and phrases the error as exactly what was verified or attempted - verified preserved, repaired and verified, or an explicit preservation-could-not-be-verified with the reason and hand-closeout instruction. Two behavior tests drive a lying tasks-axi start through a real interrupted spawn and assert the printed claim and the real backlog state agree. * no-mistakes(test): Fix calm suite for Pi 0.85 and pin test umask * no-mistakes(document): Document interrupted-spawn preservation claim in backlog gate owner * no-mistakes(ci): Fixed the Greptile P1 in bin/fm-spawn.sh's deferred-signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change * no-mistakes(ci): Fixed the Greptile P1 in bin/fm-backlog-transition-lib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean * no-mistakes(ci): Fixed the Greptile P1 on bin/fm-backlog-transition-lib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean
NewAiCoder-bot
added a commit
to NewAiCoder/firstmate
that referenced
this pull request
Sep 8, 2026
…rt (#28) * fix(bin): preserve subshell lock ownership on Bash 3.2 (#3789) * fix: distinguish subshell wake-lock owners on stock Bash Restore distinct process ownership for issue #3743 using the existing PID helper, consistently across lock publication, reclaim, release, role checks, and bounded handoff. The existing wake-queue regression fails on pristine upstream Bash 3.2 with rc=13. The complete suite now passes on Bash 3.2.57 and Bash 5.3.15, with added coverage for ownership when BASHPID is unset. Canonical lint and stock-Bash syntax checks pass. * no-mistakes(document): Correct lock grace-period documentation * no-mistakes(ci): Captain, fixed all 14 SC2031 false positives with nine ShellCheck source-boundary annotations across three tests. Full CI-mode lint and the complete wake-queue suite on stock Bash 3.2 passed. Runtime behavior is unchanged * fix(bin): resolve captain holds and legacy teardowns on non-markdown backends (#3782) * fix(bin): close legacy records on the Beads backend honestly Two pre-Beads reads blocked honest closure of leftover records: 1. fm-captain-hold.sh complete/verify resolved attested legacy hold ids only against the live backend and the pre-collapse derived identity, so a home whose holds fm-hold-migration rehomed under fm- ids failed with an empty-name absence message (the resolve failure was swallowed by the command substitution feeding verify_hold_durable). Resolution now falls back, on the Beads backend only, to the legacy id under the configured beads prefix and to the row whose notes carry the exact marker line 'migrated from data/backlog.md id <legacy id>'; every refusal names the id it could not resolve, and the markdown path is unchanged. 2. fm-teardown.sh refused any record without spawn_gen forever. A record that predates the field can now be torn down with an explicit --legacy-record flag once the recovery-grade endpoint classifier confirms the recorded endpoint dead or agent-less; the accepted incarnation is stamped into the record right before its close marker binds to it and named in the teardown line. Refusals leave the record byte-identical, the unlanded-work refusal is not relaxed, and a corrupt (multi-valued) spawn_gen is never accepted. The companion repair this branch carries (follow-up commit) is the backend-gated --file and markdown-file requirement in the mutate path and lifecycle gates: fm_backlog_mutate passed --file and required the markdown backlog file regardless of the resolved backend, and the transition gate plus row probe required that file before any backend work, so a home on a non-markdown backend could neither gate, probe, nor close its rows. Behavior tests: self-contained beads fixtures over a scratch bd graph (self-skipping on markdown-only tasks-axi installs), legacy meta fixtures for every teardown gate, and the relocated markdown backlog coverage stays green. * no-mistakes(review): fix(review): report migrated-hold scan refusals and guard legacy spawn_gen stamp against newline-less records * fix(backlog): address the configured backend for lifecycle writes Completes the fm-backlog-transition-lib repair the first commit's message claims: on this base fm_backlog_mutate passed --file and required the markdown backlog file regardless of the resolved backend, and fm_backlog_transition_applies plus fm_backlog_row_probe required that file before any backend work, so a home on a non-markdown backend could neither gate, probe, nor close its backlog rows. All three now gate the markdown file on the resolved tasks-axi backend: markdown keeps exactly its explicit <data>/backlog.md behavior, non-markdown homes address the backend their own configuration selects with no markdown file requirement. fm_backlog_row_show and fm_backlog_row_list already gated correctly and are unchanged. docs/configuration.md owns the contract line. Also extends the same backend gate to fm-captain-hold.sh's own mutation wrapper - hold/add/update/answer/done append the markdown --file only when the resolved backend is markdown, so a captain call on a Beads home reaches the Beads store end to end - and applies the review round's two direct remedies there: the [beads] graph path resolves against the backlog root when relative (never the process CWD), and a failed bd graph read reports bd's own trimmed stderr reason in the refusal. Coverage: tests/fm-backlog-atomicity.test.sh gains a stub-driven Beads completion case proving the transition gate applies, the row probe reads, and done runs without any markdown file or --file override; the relocated markdown backlog test stays green. * no-mistakes(review): Document root-tasks.toml-only beads settings for migrated-hold resolution * test(gotmp): stub fm_tasks_axi_backend so the fixture matches the backend-aware transition lib The legacy-records change made fm-backlog-transition-lib.sh resolve the configured backend via fm_tasks_axi_backend before the markdown-only skip. The gotmp fixture's fm-tasks-axi-lib stub lacked that function, so the markdown check fell through and teardown hit the incompatible-backend error with unbound FM_TASKS_AXI_MIN under set -u. Stub the backend as markdown and define the floor, restoring the intended no-backlog skip. * fix(teardown): roll the legacy stamp back when the close marker fails A legacy-record teardown stamps its accepted incarnation into the record right before the close marker binds to it; when that marker write then fails, the stamp survived, so a retried teardown sailed past the dead-or-agent-less endpoint gate the stamp now proved unnecessary. The failed marker write now truncates the record back to its exact pre-stamp bytes (verified by size), restoring the byte-identical-refusal invariant; when the rollback itself fails the operator is told to re-run with --legacy-record after reconciling the endpoint. Also completes the recorded review decision's coverage wording: the beads stub test now drives the answer close end to end (update and done through the gated wrapper), asserting no markdown file override reaches either verb. * fix(review): harden the legacy stamp rollback and resolve derived migrated ids The legacy-record stamp rollback now uses perl (already in the teardown curated PATH; truncate is not, and is absent on stock macOS), routes every failure branch inside the stamp block through the same size-verified rollback so the byte-identical-refusal invariant holds on those paths too, and gains behavior coverage: an unrecordable close (an invalid pr= link) fails the teardown, leaves the record byte-identical, keeps the backlog row in flight, and a flag-less retry still refuses. Migrated-hold resolution now probes the derived pre-collapse identity (<origin>-decision-<entry>) alongside the raw entry - fm-hold-migration recorded the DERIVED id in every migrated row's marker note - in both the prefix and the migration-note forms, with the ambiguity refusal naming every identity tried, plus behavior coverage for a bare decision key resolved through its derived identity's marker. Also aligns fm-backlog-transition-lib.sh's header ADDRESSING/SCOPE paragraphs with the backend-gated contract, drops an unreachable FORCE validity guard the parser rewrite left behind, and switches the new stub fixture to the portable sed -i.bak idiom. * no-mistakes(review): Name the configured backend in teardown's backlog reminder * no-mistakes(review): Scan migration markers before the prefix guess * no-mistakes(review): Document marker-first resolution and cover the prefix branch * no-mistakes(document): Record prefix-attestation audit and marker-line forms * no-mistakes(ci): Fixed the Greptile P1 on bin/fm-teardown.sh: a failed rollback of the synthetic legacy stamp let a retry bypass the dead-or-agent-less endpoint gate. Root cause: teardown minted `spawn_gen=legacy-<ts>-<pid>` into the task record before the close marker bound to it. When the close-marker write failed AND the rollback also failed, the record retained that token. On the next invocation `fm_backlog_meta_spawn_gen` succeeded, so `TEARDOWN_LEGACY_PENDING` stayed 0 and the endpoint gate was skipped entirely — even with `--legacy-record`. The script's own error text told the operator to "re-run teardown with --legacy-record", advice the code could not honor. Fix (bin/fm-teardown.sh): - A `legacy-*` spawn_gen is now recognized as a stamp this teardown path minted, never one a spawn published (fm-spawn.sh publishes `s<epoch>.<pid>.<random>`). Such a record still reads as the legacy record it is: it re-enters the endpoint gate, and a flag-less retry refuses naming `--legacy-record`. - Acceptance reuses the retained token instead of minting a second one; the append block is skipped when the record already carries it, so no duplicate spawn_gen is written. - The rollback attempt and its "could not be rolled back" message are guarded to runs that actually appended a stamp, so a run that appended nothing never claims a rollback it did not perform. - Usage header documents the retained-stamp rule. Test (tests/fm-teardown.test.sh): added `test_retained_legacy_stamp_still_faces_the_endpoint_gate`, an end-to-end reproduction — a `perl` stub that fails only the rollback's `truncate` (delegating every other perl call to the real interpreter) leaves the stamp behind, then the retry must still hit the gate, must not stamp a second incarnation, must not close the backlog row, and the flag-less retry must refuse. Verification: the new test fails against the pre-fix script on exactly the reported defect ("the retry skipped the dead-or-agent-less endpoint gate") and passes after. Full tests/fm-teardown.test.sh 80 ok / 0 failures / rc=0; tests/fm-backlog-atomicity.test.sh 80 ok / 0 failures / rc=0; bin/fm-lint.sh (pinned ShellCheck 0.11.0 + actionlint 1.7.12) clean * fix: reduce local ShellCheck source-analysis cost (#3778) * fix(lint): drop source following on the local changed-file gate The local lint step was inlining library closures through --external-sources and peaking above 8 GB on a single root. Keep full analysis in CI, on main, and without a merge-base; exclude the four cross-file codes from the local pass so those findings still land in CI. Co-authored-by: Cursor <cursoragent@cursor.com> * no-mistakes(review): Run local ShellCheck per root and document measurements * no-mistakes(review): Correct local source-following telemetry * no-mistakes(document): Clarify context-sensitive lint documentation --------- Co-authored-by: Cursor <cursoragent@cursor.com> * fix(pi): route decision-owned wake batches to main (#3776) * fix(pi): route needs-decision wakes and mixed batches wholly to main Skip the supervision branch for every needs-decision status append, the same way a check-kind wake already skips it. A coalesced signal/stale trigger batch containing any needs-decision row is delivered wholly to main, not split between the branch and a later main wake - the whole batch, including any co-present routine rows for a different task, travels together. Heartbeat and unread-status scans stay independent. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013LB2CeerSMCZN4oNsVfLeE * test(pi): cover distinct-file mixed batches and heartbeat independence Add a regression using two distinct files (not the same status file twice) in one coalesced trigger so a some-vs-every regression on the file-list cross-reference cannot hide behind a degenerate same-key case, and a heartbeat/needs-decision co-presence test proving a needs-decision row neither vetoes nor rides along with an otherwise eligible heartbeat scan. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013LB2CeerSMCZN4oNsVfLeE * no-mistakes(document): Clarify needs-decision and heartbeat routing * no-mistakes(ci): Fixed captain-held stale reminders so they bypass supervision and wake main, while unrelated unread rows and heartbeats remain independent. Added routing regressions and updated documentation. Pi watcher tests, strict TypeScript checks, lint, and diff checks pass. The no-mistakes attestation failure was pipeline-state related, not a source defect * no-mistakes(ci): Fixed CI lint by narrowly suppressing false-positive SC2031 diagnostics where background PIDs are captured immediately in the same shell. Verified with `CI=true bin/fm-lint.sh` and `git diff --check`. The no-mistakes attestation failure is pipeline-state-related (`test` was skipped), not a source defect * no-mistakes(review): Route stale open decisions directly to main * no-mistakes(review): Honor configured verbs in stale decision routing * no-mistakes(review): Route second-mate escalations and configured decisions to main * no-mistakes(review): Ignore trailing whitespace after captain holds * no-mistakes(review): Cache stale decision classification per status file * no-mistakes(review): Document unread decision precedence for later task wakes * no-mistakes(review): Cache unchanged stale decisions across scope scans * no-mistakes(review): Resolve decision aliases and reject symlinked statuses * no-mistakes(review): Route surfaced captain-held signals directly to main * no-mistakes(document): Document decision-owned main routing * no-mistakes(ci): Fixed captain-held spans to remain actionable while crew working evidence is positive, ensuring the watcher delivers their main-only marker. Updated the executable regression test to cover this case. Verified with the full fm-watch-triage suite, bash syntax checks, and git diff checks. Shellcheck reported only pre-existing test harness warnings (SC1091/SC2034) * no-mistakes(ci): Fixed the CI regression: captain-held transfers now retain their established non-actionable stale classification while the signal-routing side-band still surfaces them main-only. Verified with tests/fm-daemon.test.sh, tests/fm-watch-triage.test.sh, bash syntax checks, and git diff --check --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> * feat(bin): add opt-in worker launch environment allowlist (#3802) * feat(spawn): add an opt-in worker environment allowlist Honor a home-local launch-env-allowlist at the shared worker command boundary and inherit it into secondmate homes. Preserve the existing launch behavior when the file is absent. Keep the operational environment and explicit launch assignments, and account for filtered Muse credentials. Refs https://github.com/kunchenguid/firstmate/issues/3742 Verification: - Red on origin/main 1820316b66ac2c68e244dd04a02512859ee8c1f4: the new enabled-allowlist regression observed synthetic-unrelated in the worker; the absent-file control passed. - Green: fm-test-run.sh on fm-spawn-dispatch-profile, fm-muse-harness, and fm-trace-context-spawn; all three passed without skips. - Synthetic emitted-command probes ran through sh, stock Bash, and zsh. - Canonical lint, documentation audience checks, and stock Bash syntax checks passed. * no-mistakes(review): Reject inaccessible launch environment configuration * no-mistakes(review): Preserve inherited allowlists on source inspection errors * no-mistakes(document): Clarify worker environment grants and inheritance documentation * no-mistakes(lint): Fix inheritance test ShellCheck source boundary * feat(bin): add rovo crewmate/scout adapter with home-path file access (#3575) * feat(bin): add rovo as a verified crewmate/scout worker harness Wire the Atlassian Rovo CLI (202609.1.2) into the TUI-under-tmux/herdr adapter contract: detection with marker-precedence ordering, one-shot positional launch with --startup-receipt readiness polling instead of composer scraping, model/effort flags, a screen-scrape busy fallback scoped like grok's, and crew/scout-only lifecycle control that refuses secondmate launches. Ships with a portable regression suite, a live PTY guard against the real binary, a per-harness reference doc, and a dated verification record covering the silent OAuth refresh, the interrupt-ack divergence from the originating scout report, and the still-open composer-ghost and tmux/herdr pane-liveness gaps. * no-mistakes(review): revert rovo launch to positional brief, drop startup-receipt * no-mistakes(test): rewire rovo adapter to kimi-style launch-then-send shape * no-mistakes(document): add rovo to stale worker-harness enumerations in docs * docs(verification): close the rovo herdr-liveness gap with live isolated-lab evidence Placement, launch-then-send, and busy/idle rendering are now verified live in an isolated non-default Herdr lab session (bin/fm-herdr-lab.sh), driven directly through fm-spawn.sh's/fm-backend.sh's own shared primitives since the cross-session launcher-identity guard refuses this task's own ambient Herdr identity for a full fm-spawn.sh run. fm_backend_agent_state reported dead for a live, responding rovo pane at every point checked, because herdr's own agent-integration registry has no rovo entry (herdr integration status), so herdr agent get returns agent_not_found regardless of whether rovo is actually running. This is recorded as a Herdr-side integration gap rather than a firstmate bug, left unpatched to avoid a false-positive alive verdict for other idle shells. Updates docs/verification/rovo.md's backend-liveness section and its two cross-references (docs/verification/runtime-backends.md, docs/configuration.md) accordingly. * fix(bin): close rovo's failed-spawn leak and busy-scrape false idle Greptile P1s on PR #3575: a failed rovo readiness/submission/delivery gate exited without tearing down the just-created endpoint, leaving the launched --yolo rovo process running as an orphaned agent outside task control. Separately, the busy classifier's rendered-tail fallback returned definitive idle whenever the "Rovo is thinking" marker scrolled out of the last 12 nonblank lines of a long turn, which could make supervision wrongly conclude a still-working worker had gone idle. fm-spawn.sh: rovo_spawn_fail now calls rovo_endpoint_cleanup, which kills the created endpoint (tmux/herdr/zellij/cmux) via the same generic fm_backend_kill dispatch fm-spawn.sh's own orca-abort path already uses; orca's worktree and terminal remain owned by the separate ORCA_ABORT_CLEANUP trap. fm-busy-lib.sh: the rovo classifier arm now reports "unknown rovo-regex" instead of "idle rovo-regex" when the marker is absent, matching how muse and cursor already express "can't tell" for their own fallbacks. The positive busy match is unchanged. Extends tests/fm-rovo-harness.test.sh: the readiness and delivery failure tests now assert the endpoint is torn down (and the success test asserts it is not), and a new test drives the busy marker out of the tail window to confirm the verdict is unknown, never idle. bin/fm-lint.sh is clean on both changed files. * test(rovo): align spawn fixture with the launch-brief validation contract Upstream main now requires a brief's ## Captain's intent and ## Firstmate spec subsections (or a nonempty legacy # Task body) before spawn, and rewrites ship+no-mistakes briefs into launch-brief.md. Update the rovo harness fixture and pointer assertions to match, mirroring the kimi harness fixture. * no-mistakes(review): align rovo.md delivery-gate note with live herdr evidence * fix: prevent stale supervision wake loops (#3672) * fix(bin): stop the supervision branch's stale-ack and ghost-report loops Clean-slate implementation of the four authorized recommendations from the supervision-ghost-retrigger analysis (items 1, 2, 3, and 7), in their minimal form, superseding PR #3604: - fm_branch_report refuses a task the wake being handled never named. The extension fixes the reportable task set from the eligible rows before each prompt (signal and stale rows resolve to their tasks, a heartbeat allows any task with a live record, fleet is always allowed), so a report typed from memory about a task whose records teardown already removed is never stored or delivered. - An acknowledgement that consumes nothing says "nothing was acknowledged through N" and prints the exact --ack-through / --recovery-generation command for the current presented wake, instead of "re-run the drain", which re-fed the same stale acknowledgement in a loop. - bin/fm-guard.sh no longer tells the branch actor to drain queued wakes while it is handling them; it names the granted rows instead. - Teardown removes state/.<task>.branch-outcome-index for ordinary tasks and descendants; the index rebuild and the append-side index write both skip a task with neither a live record nor a status log, so the branch's report of a teardown it just performed is stored without recreating the index. No new locking, no spawn-generation binding, and no retired-task refusal: the branch can still report the outcome of a task it just tore down, and the teardown test now proves that path end to end. * fix(bin): narrow the branch report scope and guard silence to the minimal form Apply the four review decisions on the clean-slate branch: - A signal or stale prompt may report only the tasks its own rows resolve to; fleet is refused there too. A heartbeat review is not scoped by task at all, so the extension no longer tracks live task records and refuses nothing by task id during a fleet review. - The outcome-index rebuild no longer skips retired tasks; the append-side skip alone keeps a torn-down task's index from being recreated. - bin/fm-guard.sh keeps the queued-wakes warning silent for the branch actor instead of printing a replacement note. * no-mistakes(document): Align supervision docs with scoped wake handling * fix(bin): grant rovo the per-task home paths its standard crewmate flow needs rovo confines every file-tool operation to its worktree by default, and its bash tool independently refuses the same external paths regardless of any grant (confirmed live), so a rovo worker could not read its own brief or steering messages or write its status/report - all of which live in the firstmate home outside the worktree - without hand-feeding it. Grant toolPermissions.allowedExternalPaths for exactly the task's brief directory, steering inbox, and status file at launch time via --config-override, merged with agent.efficiencyLevel into one JSON object since that flag is single-value and silently discards a second occurrence. Extends the live PTY guard to prove, against the real binary, that the grant lets rovo read an external brief and append to an external status file, and that the same flow is blocked without the grant. * no-mistakes(document): align rovo reference Effort row with merged single --config-override * no-mistakes(document): document rovo file-access grant in harness reference --------- Co-authored-by: PUNEET PATWARI <ppatwari@atlassian.com> Co-authored-by: Kun Chen <3233006+kunchenguid@users.noreply.github.com> * fix(bin): prevent false pipeline blocks after drive timeouts (#3813) * fix(bin): read a crew's pipeline-death claim against the live run A crew's no-mistakes drive call blocks until the next gate or outcome, routinely far longer than its harness lets one command live, so the call gets killed or times out while the daemon runs the fix round on in the background. Crews read that as daemon death and block on it, and firstmate had nothing that contradicted them. Rule 7 of every generated brief now says a drive-call error or a harness command timeout is not a daemon error, requires `no-mistakes daemon status` plus `no-mistakes axi status` before a pipeline `blocked:`, and reserves that report for a refused socket or a run record failed with a daemon error. The no-mistakes definition of done adds the harness command limit and the background-and-poll shape that fits inside it. fm-crew-state gains one classification case: a `blocked:` line blaming the daemon, a timeout, or unreachability, while the run is running or fixing AND the pipeline reports fresh activity, now reads as superseded because the run is alive. Recency comes from the client's own `quiet` marker on active_steps.last_activity rather than a threshold invented here, and positive evidence is required, so a run record that outlives a genuinely dead daemon keeps the plain reading. stuck-crewmate-recovery gains the inverse-of-a-dead-endpoint playbook: firstmate reads both statuses itself, steers a reattach, never restarts the shared daemon on a crew's claim, and escalates only a refused socket. Nothing here depends on an unshipped no-mistakes capability. * no-mistakes(review): Prioritize daemon socket failure and narrow unreachable matching * no-mistakes(review): Honor socket refusal across coarse status and crew guidance * no-mistakes(test): Replace flaky settle timing assertion with pane-read count * no-mistakes(document): Document daemon timeout recovery contract * no-mistakes(ci): Fixed daemon socket failures being suppressed by terminal attributed runs. Positive refused/missing socket evidence now remains blocked regardless of run status. Added a behavioral regression test for terminal failed runs. Verified with fm-crew-state tests, project ShellCheck lint, and git diff checks * fix(bin): scope the worker role contract for ship and scout launches (#3797) * fix(brief): scope Firstmate workers to their launch contract * no-mistakes(document): Clarify supervisor scope and worker contract ownership * no-mistakes(ci): Removed the heading-based bypass so every ship/scout launch receives the current worker-role contract. Added a regression that failed before the fix and passes afterward. Dispatch, brief, and delivery suites, focused ShellCheck, and git diff --check all passed * no-mistakes(review): make launch overlay sole owner of worker role contract * no-mistakes(review): narrow heading test dimension and fix publish error wording * no-mistakes(review): gate role supersession, fix render guard, drop AGENTS twin * no-mistakes(document): align architecture AGENTS.md scope and spawn launch-brief header * fix(bin): stop ringing steering doorbells into dead panes (#3823) * fix(bin): stop ringing steering doorbells into dead panes The steering-inbox doorbell was a plain sentence plus Enter typed into a worker's pane, and the watcher re-rang it on the assumption that a ring is free. In a pane whose agent has exited that line is a shell command, and the re-ring ladder kept typing it into a shell that can never acknowledge it. - Prefix the doorbell with the shell no-op `: ` so a bare shell executes nothing while a live worker still reads the same self-describing line. `#` is not used because interactive zsh does not treat it as a comment by default and the claude harness binds it to memory mode. - fm_task_inbox_ring skips the pane (return 3) when the backend positively classifies the agent as dead; missing, ambiguous, unreadable, and unverified endpoints still ring so a blind classifier never starves a live worker. - The watcher caps the ladder for a dead pane: one stale wake for recovery, no ring, no ladder walk, and the durable record stays for stuck-crewmate-recovery. fm-send and the remote steer leg report the skip. Tests cover the no-op in real shells, the dead/live/unclassifiable ring verdicts, and the single-surfacing watcher path. * no-mistakes(review): Quote doorbell paths against shell injection * no-mistakes(review): Reject terminal-control paths before ringing * no-mistakes(review): Document accepted partial doorbell delivery race * no-mistakes(review): Skip unavailable endpoints before busy-state handling * no-mistakes(test): Respect shell startup PATH in environment allowlist test * no-mistakes(test): Fix doorbell test fixtures for endpoint liveness * no-mistakes(test): Prioritize confirmed restarts and clean shell test syntax * no-mistakes(document): Document dead and missing doorbell recovery * no-mistakes(ci): Fixed the persistence-reply timeout race by rechecking for a correlated reply immediately before falling back to a nudge. Added a deterministic regression covering replies arriving between the preliminary resolution pass and timeout handling. Verified with the targeted restart suite, project lint, coverage guard, bash syntax checks, and diff checks * fix(bin): wait out transient primary-checkout reads in the spawn worktree poll (#3834) * fix(spawn): keep the worktree poll from adopting the repository primary After `treehouse get` is sent, the worktree-discovery poll reads the pane's foreground-process cwd. While treehouse is still fetching and checking a slot out, the foreground process is treehouse itself and it reports the repository's PRIMARY checkout as its cwd for several seconds. The poll accepted any path that merely differed from the spawning project, so from a linked spawning home - whose project is itself a worktree of that repository - it adopted the primary, and the isolation guard then refused a launch whose slot treehouse went on to create normally. Screen every candidate with the isolation guard's own conditions, extracted as spawn_worktree_isolated, so a read the guard would reject stays a transient the poll keeps waiting through. The two-consecutive-reads rule and the guard as final backstop are unchanged; a pane that never reaches an isolated worktree still fails at the existing 60s deadline, now naming the last path it reported. The already-settled timing assertion counted whole-spawn wall time against a 5s budget and failed on unmodified HEAD on slower machines; it now counts pane reads, which is what "one confirming read, not an extra cycle" actually means. * fix(spawn): say which path the worktree wait rejected, and why Screening every discovery-poll candidate means a host that never reaches an isolated worktree spends the whole 60s window before refusing. That wait is deliberate - separating a transient from a terminal misconfiguration needs machinery this path does not want - so the refusal explains itself instead: the isolation check records why a candidate failed, and the deadline names the last path seen together with that reason. Message and diagnostics only; the poll's control flow is unchanged. Two suites asserted the guard's wording on paths the poll now rejects rather than adopts, so their refusal arrives from the deadline instead: realign fm-tangle-guard's non-git and subdirectory-of-primary cases (each now also asserting the stated reason, and the second the metadata absence it was missing) and the herdr projection e2e's forced non-worktree cwd. * no-mistakes(review): stub poll sleep in tangle-guard spawn isolation test * no-mistakes(document): document spawn poll isolation screen in fm-spawn header * test(spawn): make the non-git isolation case non-git anywhere The refusal-reason assertion for a path outside any repository assumed TMPDIR is not inside a git repository. Where it is, git walks up from the temporary directory, finds that repository, and the spawn reports the subdirectory cause instead - so the case passed or failed on a property of the host rather than on the behaviour under test. Build the path under a directory the test then names in GIT_CEILING_DIRECTORIES, which git documents as not chdir-ing up into a listed directory while looking for a repository. Git never excludes the directory being searched, so the ceiling is the parent of the path handed to the spawn. The assertions pin which cause fired rather than the sentence that explains it, leaving the operator wording free to improve. * no-mistakes(document): point spawn poll comment at the isolation screen's comparison * no-mistakes(ci): Fixed the "Behavior portable serial 2" failure in tests/fm-tangle-guard.test.sh ("non-worktree spawn did not say why the path was rejected (missing: 'not inside a git worktree')"). Root cause, in this PR's code: bin/fm-spawn.sh's spawn_worktree_isolated resolved the git toplevel with `wt_top_real=$(cd "$SPAWN_WT_TOP" ...)`. For a path in no repository, `git rev-parse --show-toplevel` yields empty, and `cd ""` is a SUCCESSFUL no-op on bash before 5.3 (CI's ubuntu-latest ships bash 5.2). The empty toplevel therefore resolved to fm-spawn's own cwd — the CI checkout — so the poll reported "it is a subdirectory of worktree root '/home/runner/work/firstmate/firstmate'" instead of the correct "it is not inside a git worktree". Dev machines with bash 5.3 fail `cd ""`, which is why the suite passed locally and only failed on CI; it is a genuine shell-portability defect in the reason vocabulary this change added, not a test-environment artifact. Fix (smallest root-cause change, 1 line + comment, bin/fm-spawn.sh:2168-2173): guard the empty value so it never reaches `cd` — if [ -n "$SPAWN_WT_TOP" ] && ! wt_top_real=$(cd "$SPAWN_WT_TOP" 2>/dev/null && pwd -P); then No change to the poll's timing or deadline behavior (respecting the recorded refusal-latency and spawn-wt-reason-vocabulary decisions), no new tests, no other files touched. Verification: - Reproduced the exact CI failure locally by putting bash 3.2 (same `cd ""` semantics as CI's 5.2) first on PATH: fails before the fix with the identical message shape, passes after. - tests/fm-tangle-guard.test.sh passes under both bash 3.2 and bash 5.3. - tests/fm-spawn-worktree-settle.test.sh and tests/fm-spawn-pool-base-freshen.test.sh pass; shellcheck -x bin/fm-spawn.sh clean. - bin/fm-test-run.sh --changed: 46 suites completed, every FM_TEST_END exit=0, 1076 passing assertions, 0 "not ok" (including fm-tangle-guard, fm-control-relaunch, fm-lint). The run ended on my own 900s wall-clock cap (rc=124), not on any test failure * fix: support stock macOS Bash 3.2 paths (#3732) Co-authored-by: Talon Stark <talonstark@gmail.com> * fix(bin): read orphaned green ci monitor and daemon-down failed record as not failed (#3846) * fix(bin): read an orphaned green ci monitor as held-for-merge, not failed A no-mistakes run held for a captain merge decision keeps its ci step polling until merged or closed; when the shared daemon restarts under that poll, the run is recorded failed although every substantive step completed and GitHub reports the PR green. A monitor whose only remaining job is to observe a human decision must not convert the absence of that decision into a failure verdict. fm-crew-state.sh now reclassifies a terminal failed run as done (held-for-merge), surfacing the run's PR URL, when the steps table shows every step completed except exactly ci failed and the ci log's last recognized marker reads checks green. A genuinely red check, an unreadable ci log, or a second failed step keeps the failure. * no-mistakes(review): Read daemon-down coarse failed ledger as unknown, not failed * no-mistakes(document): docs: align AGENTS.md failed-verdict guidance with crew-state reclassification * fix(spawn): verify preserved backlog state after interrupted spawn delivery (#3852) * fix(bin): read preserved spawn state back before the interrupted exit claims it The deferred-signal exit path asserted the paired task record and In-flight backlog state were preserved without reading either back, exactly when a reader is least able to check (fm-yi4j evidence, 2026-09-05). The commit's exit status alone has been observed to agree with a row that did not actually move. The exit path now re-reads the record and the row under the same per-task lock as the commit, repairs a row the commit believed it moved, and phrases the error as exactly what was verified or attempted - verified preserved, repaired and verified, or an explicit preservation-could-not-be-verified with the reason and hand-closeout instruction. Two behavior tests drive a lying tasks-axi start through a real interrupted spawn and assert the printed claim and the real backlog state agree. * no-mistakes(test): Fix calm suite for Pi 0.85 and pin test umask * no-mistakes(document): Document interrupted-spawn preservation claim in backlog gate owner * no-mistakes(ci): Fixed the Greptile P1 in bin/fm-spawn.sh's deferred-signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change * no-mistakes(ci): Fixed the Greptile P1 in bin/fm-backlog-transition-lib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean * no-mistakes(ci): Fixed the Greptile P1 on bin/fm-backlog-transition-lib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean * docs: correct runtime-backend maturity labels for Herdr (#3821) * docs: correct stale tmux/herdr backend maturity claims Herdr now has 21 test files, its own required CI job (tests-herdr) that installs a pinned build and hard-fails on "skip: herdr not found", while tmux has 3 test files and is only required as a dependency of the portable-serial e2e lane. zellij, orca, and cmux still have no CI lane at all. AGENTS.md and docs/herdr-backend.md still called Herdr merely "experimental" alongside those three, misleading every session and reader about actual coverage. Update AGENTS.md's config/backend entry, the opening lines of docs/herdr-backend.md and docs/tmux-backend.md, the runtime-backend section of docs/configuration.md, and the matching claims in docs/architecture.md, CONTRIBUTING.md, and README.md so they agree and distinguish tmux (default), herdr (own required CI lane, largest suite, Windows still spike-only), and zellij/orca/cmux (still experimental, no CI lane). No behavior, selection order, or dispatch logic changes. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GDYsWPEfwTuPNjQ2nGBCcj * no-mistakes(review): docs: fix stale herdr label and CI-lane wording * no-mistakes(document): docs: align tmux adapter label in scripts.md * no-mistakes(review): docs: drop duplicated herdr CI claim from tmux page * no-mistakes(review): docs: drop windows claim, align contributing backend wording * no-mistakes(review): docs: trim duplicated CI claim from herdr opening line * no-mistakes(review): docs: drop unguarded largest-test-suite superlative * no-mistakes(review): docs: restore tmux verified label and README experimental scope --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> * fix: restore intent-targeted no-mistakes validation (#3865) * fix: delete deterministic no-mistakes test baseline, restore intent-targeted Test PR #3644 pinned commands.test to a fm-test-run.sh --changed walk of the repository's 75-162 tests/*.test.sh scripts. no-mistakes runs commands.test verbatim and unconditionally after every fix round, so that walk multiplied by round count: measured at 32.7 minutes per validation versus 3.6 minutes intent-targeted. Delete the pin and restore the 3.6-minute posture. Add tests/fm-nm-test-contract.test.sh as a regression guard, parsing .no-mistakes.yaml as YAML (ruby's bundled Psych, matching the parser tests/fm-test-run.test.sh already uses for ci.yml) rather than grepping its text, restoring in legal form what PR #823 added and PR #1282 removed. Record the rule in docs/configuration.md's "Gate defaults" section (the authoritative owner CONTRIBUTING.md already points at) and strengthen CONTRIBUTING.md's existing local-Test guidance to state it plainly: never configure commands.test to a deterministic test command, complete or partial. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CSt9JvrQMVc4u3jPyUFCFC * no-mistakes(review): Centralize no-mistakes test policy and narrow guard --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> * fix: verify Treehouse slot ownership before teardown (#3837) * fix(bin): verify pool-slot ownership before returning a worktree slot Workers were killed when cleanup returned a Treehouse pool slot that a different, live task had already taken. Teardown now proves the slot is genuinely this task's before releasing it: it refuses when another task record claims the same live worktree path, or when the endpoint's working directory contradicts the recorded slot, and that refusal holds under --force. Slot allocation, metadata publication, ownership verification, and slot return are serialized across linked firstmate homes, and forced secondmate cleanup verifies descendant slot ownership before returning any child worktree. Regression coverage drives the scripts with two task records naming one slot path and asserts the live worker survives and its slot is not reset. * no-mistakes(review): Protect slots across cloned Firstmate homes * no-mistakes(test): Gate teardown locking on genuine Treehouse slots * no-mistakes(test): Clarify pooled descendant slot gating * no-mistakes(test): Synchronize watcher re-arm test on process exit * no-mistakes(test): Wait for watcher cleanup before timeout escalation * no-mistakes(document): Document pool-slot ownership safeguards * no-mistakes(ci): Fixed all reported CI issues: normalized bare local Git origins to the same Treehouse project-lock identity as absolute clone origins; resolved ShellCheck SC1091 with explicit conditional sourcing; and taught concurrent Herdr teardown coverage to retry expected Treehouse lock contention. Added behavioral regression coverage for bare/absolute origin lock identity. Verified endpoint-safety tests, watcher tests, full CI lint, and the previously failing Herdr teardown assertion * fix(bin): resolve relative origins from repository root * no-mistakes(ci): Fixed teardown so an exact recorded endpoint may change cwd without falsely vetoing cleanup. Removed cwd-based ownership refusal while preserving cross-home record exclusivity and project locking. Updated behavioral coverage for both foreign slot ownership refusal and moved-cwd teardown success. Endpoint-safety, backend, watcher, checkpoint, and targeted lint checks pass. Real Herdr presentation E2E progressed successfully but exceeded the 600s local timeout * feat(bin): add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary (#3867) * feat: add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary Add omp as a verified harness: anchored process-name detection with a Firstmate-owned FM_OMP_HARNESS launch marker that needs real omp ancestry, the fm-spawn launch template with foreign-marker clearing, the tracked .omp/fm-worker-overlay.yml posture overlay, --auto-approve, --cwd, and pre-launch model validation scoped to providers 'omp models --json' lists. Workers get a state-resident busy-state extension keyed on agent_end without willContinue (omp has no agent_settled). The primary gets two tracked .omp/extensions: a turn-end guard that answers omp's blocking session_stop hook by compelling one continuation per turn, with the pre-tool seatbelts and Run-tier session-start delivery, and a watcher extension ported from the Pi one with fm_watch_arm_omp. Control tables, composer busy footers, omp's status row as a bare-composer boundary, the extension supervision model with an omp-keyed ownership proof, the session-start diagnostic, and the supervision protocol snippet follow. Verified live on omp 18.1.11 with openai-codex/gpt-6-astra: a Herdr scout through spawn, busy state, steer, interrupt, exit, and teardown, and the isolated rpc primary lab through extension auto-discovery, digest delivery, lock identity, watcher arm, successor and wake delivery, and the compelled guard continuation. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh * test: prove the omp guard continuation through a guard spy Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh * fix(spawn): clear the gemini marker at the omp launch boundary Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh * test(omp): force the guard stage by freezing the watcher and clear lint findings Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh * test(omp): reap the live lab by path and record omp's rpc shutdown as a note Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh * test(omp): spawn a real secondmate for the discovery rule and classify the omp surfaces Replace the template-extraction check with a genuine --secondmate launch pinned to the fake tmux backend, assert the worker extension's handler set through the executable rather than its bytes, classify the two new omp surfaces in the documentation inventory, and record the Herdr worker evidence in the runtime-backends verification doc. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh * no-mistakes(review): omp: unverify remote routes, narrow busy regex, drop overlay approval pin * no-mistakes(review): omp: validate config-pinned model, correct remote and marker docs * no-mistakes(review): omp: pin config-model validation with a test, trim overlay * no-mistakes(review): omp: sync guard evidence, drop dead param, map quota family * no-mistakes(review): omp quota: refuse unmapped prefixes, match bare model scopes * no-mistakes(document): docs: cover omp in cd-guard, quota, continuity, tmux * no-mistakes(document): docs: add omp subagent-guard row, fix live test header * no-mistakes(ci): Fixed both failing behavior shards and the Greptile P1 in bin/fm-composer-lib.sh. Root cause of "Behavior portable serial 1" and "Behavior portable parallel 2": the omp busy regex (FM_DELIVERY_OMP_BUSY_REGEX_DEFAULT) and omp status-row furniture regex (FM_COMPOSER_OMP_STATUS_RE_DEFAULT) used the bracket range [⠁-⣿]; BSD grep on macOS accepts it but GNU grep on Linux CI aborts with "Invalid collation character", failing every omp busy/furniture read (3 assertions across fm-omp-harness, fm-tmux-submit-busy, fm-composer-lib). Replaced the range with one shared explicit alternation FM_OMP_SPINNER_FRAMES_RE of omp 18.1.11's unicode-preset spinner frames (status set ⣾⣽⣻⢿⡿⣟⣯⣷ + activity set ⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏, read from the installed binary), the same pattern the Kimi busy regex already uses in CI. For Greptile's finding (the harness-agnostic furniture rule's first alternative matched any 1–4-byte token + ' · ', so wrapped typed input like 'fix · tests' with the cursor on it regressed from pending to unknown; reproduced locally vs base), pinned that alternative to omp's identity cell (π||pi, the icon.omp of each preset in the 18.1.11 binary). Tests: fm-composer-lib.test.sh asserts 'fix · tests' is not furniture, a status-set spinner row is furniture, and the wrapped composer screen reads pending under both locales (CAPS_TMUX cursor 3); fm-omp-harness.test.sh asserts a status-set frame reads busy. New negative cases fail against the pre-fix lib and pass after. Verified: fm-omp-harness, fm-tmux-submit-busy pass via bin/fm-test-run.sh; fm-composer-lib passes all cases except one pre-existing, unrelated local failure (Herdr half-block test uses printf '▀', unsupported by macOS bash 3.2; fails identically on a pristine HEAD export, passes on CI bash 5); shellcheck and bin/fm-lint.sh clean. Caveat: GNU grep is unavailable locally, so the Linux compile was not run directly; the fix uses only constructs already proven on CI's GNU grep (multibyte literal alternations, incl. under LC_ALL=C). Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh, tests/fm-omp-harness.test.sh. No docs needed changes (they describe the rule generically) * no-mistakes(ci): Greptile Review: fixed. The omp status-row furniture regex FM_COMPOSER_OMP_STATUS_RE_DEFAULT in bin/fm-composer-lib.sh still accepted a literal `pi ·` opening, so wrapped composer input beginning with `pi ·` was truncated and misclassified. Read the installed omp 18.1.11 binary: the ascii preset's `icon.omp` is `pi` but its `sep.dot` separator is ` - ` (unicode/nerd use ` · `), so a real ascii status row never contains `pi ·` and that alternative could only ever match typed text. Removal-first fix: dropped `pi` from the identity alternation (now `(π|)`) and updated the comment to record why the ascii preset is excluded. Tests (tests/fm-composer-lib.test.sh): added a negative furniture case for 'pi · e · phi as the three constants' and a wrapped-screen assertion (CAPS_TMUX, cursor 3) that a continuation row opening `pi ·` reads pending in both locales; the new case fails against the unfixed lib and passes after. Verified: composer test with the half-block case skipped passes all 33 cases including the omp matrix; bin/fm-test-run.sh tests/fm-omp-harness.test.sh passes; shellcheck -x clean on both files; bin/fm-lint.sh clean. The full composer test via the runner fails locally only on the pre-existing half-block case (bash 3.2 printf cannot emit ▀; passes on CI bash 5), identical to before this change. Docs unchanged (they describe the rule generically and never mention the ascii identity cell). PR must be raised via no-mistakes: not caused by code. attestation.head_sha is cdddc60 while the PR head is cd51cf4 because the pipeline's ci-phase push moved the head; the outer executor's re-push will re-bind the attestation. No file change for that check. Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com> * fix(pi): invoke Bash helpers correctly on native Windows (#3843) * Fix Pi shell invocation on native Windows * no-mistakes(document): Document Pi Windows Bash transport * no-mistakes(ci): Captain, staged a narrow fix: register the Pi Windows regression for both extension paths, make Windows mode emulation non-failing, and enforce LF shell checkouts. Mapping and coverage checks pass; CI/Require no-mistakes were approval-gated externally * no-mistakes(review): Cover async Windows branch-outcome Bash invocation * no-mistakes(review): Preserve Cygwin checks and refresh Windows timing * no-mistakes(test): Invoke OpenCode operational-input owner through Bash on Windows * validation-fixture * no-mistakes(document): Document Windows Bash helper invocation * no-mistakes(ci): Fixed PR-caused changed-selection failure by removing the malformed tracked evidence artifact and allowing deleted, unconsumed source paths to retire cleanly while preserving fail-closed behavior for live unmapped paths. Added regression coverage. Verified native-Windows Pi shell-seam test passes and --changed selects the Windows regression --------- Co-authored-by: test <test@example.invalid> * test(bin): pin teardown outcomes for squash-merged rebased branches (#3870) * fix(bin): recognise squash-merged rebased work as landed at teardown A pipeline rebase can leave the local worktree on pre-rebase commits while GitHub squash-merges the rebased head. The landed-work test then compared those stale commits against a squashed main and refused cleanup of work that had already landed. When the forge reports the recorded PR merged and its merge commit is on the default branch, treat a local branch that only repeats paths from the pipeline push as stale rather than unlanded. If the forge is unreachable, the same coverage check runs against a PR head whose content is already on default. Extra local paths still refuse. * fix(bin): drop unprovable squash-rebase landed-work coverage Path-set coverage treated a diverged local branch as landed whenever it touched the same files as the squash merge. That accepts the reviewer's failing sequence: same path, different content, work discarded. git cherry and merge-tree containment were already too strict on the real rebase-fold case. No remaining check is both safe and permissive enough to recognise a stale pre-rebase copy without also accepting unlanded edits, so that case still refuses. Keep the proofs that hold: a merged PR head that contains local work, or a clean content-in-default tree match. Tests now refuse same-path different content and extra unlanded commits, and still allow a local branch that followed the pipeline rebase. * no-mistakes(review): drop recorded-pr-head fallback and reverted-design leftovers * no-mistakes(review): silence squash-merge stdout corrupting test PR head * no-mistakes(review): make unlanded follow-up commit sole cause of refusal * no-mistakes(document): correct stale squash-rebase fixture comments in teardown tests * no-mistakes(ci): Split the three reported checks: - CI (run 34061098467) and Require no-mistakes (run 34061098460) both concluded `action_required` — approval-gated workflow runs that never executed a step. Not caused by this PR's code; no change can clear them. - Greptile Review was a genuine defect in the new tests: the three new refusal cases (tests/fm-teardown.test.sh) asserted only exit status 1 and a REFUSED line, so a teardown regression that destroyed the worktree, branch, and task record before reporting refusal would still pass. Fix (tests only): added one `assert_refusal_retained_task_state` helper and called it from `test_squash_merged_same_file_different_content_refuses`, `test_squash_merged_rebased_local_with_unlanded_commit_refuses`, and `test_squash_merged_stale_local_refuses_when_forge_unreachable`, each capturing the worktree HEAD before `run_teardown`. It pins that the refusal left the isolated copy on disk, the task branch still checked out at the same unlanded commit, and state/task-x1.meta intact. Verification: the four squash tests pass; a sensitivity probe ran the ALLOW fixture (teardown completes) and pointed the same helper at the outcome — it fires, because a completed teardown detaches/deletes the branch and removes the task record, proving the assertions discriminate. Full tests/fm-teardown.test.sh: 83 passing. bin/fm-lint.sh clean with pinned ShellCheck 0.11.0 + actionlint 1.7.12 (plus an explicit --external-sources pass on the changed file). bin/fm-test-run.sh --check-coverage ok. Caveat: test_herdr_flat_teardown_preflight_refuses_before_changes (mode missing-adapter) fails on this machine. Verified it fails identically on base commit f91a950 via `git archive`, so it is a pre-existing local environment difference untouched by this diff; skipped to run the rest of the suite, not modified --------- Co-authored-by: Morten Gad <mogad@itm8.com> * fix(bin): keep supervision armed for registered custom checks (#3860) * fix(bin): keep supervision armed for registered custom checks A custom check bound by bin/fm-check-register.sh only ever runs inside the watcher's check sweep, but fm_supervision_status counted in-flight tasks, the relay poll shim, and process-event sources as supervision need, and not registered checks. Tearing down the last task therefore stopped every home-level check silently until the next spawn. Count a state/<id>.check.sh that carries its state/<id>.check-trust binding as supervision need. The relay shim keeps its own trust path and task PR polls carry no such binding and are torn down with their task, so neither arms a home by accident. Presence of the binding is the whole test: the sweep validates the bytes at execution time and wakes firstmate when it rejects one, which is the outcome an idle home needs. Closes #3856 * no-mistakes(review): name registered checks in turn-end block banner and doc invariant * no-mistakes(review): narrow PR poll predicate test to what it proves * no-mistakes(document): point Grok re-arm step at supervision-need owner * fix(bin): resolve Treehouse locks for remote secondmate homes (#3883) * fix(bin): resolve the shared Treehouse project lock inside remote secondmate homes Every spawn and teardown inside a remote-seeded secondmate home refused, because the project lock's anchor could not be resolved there. fm_firstmate_root_home walks a home's parent bindings upward to find the anchor the lock lives in, and treated a remote parent binding as an error. A remote-seeded home's parent is on another machine, so that walk can never succeed from there - and neither can the home's own local descendants, whose chain terminates at the same record. Both fail closed on every Treehouse-backed spawn and every pool-slot teardown. A remote parent now terminates the walk at the home holding it, which is the correct anchor: a lock taken on this filesystem is neither held nor observable across that boundary, and that home is already the top of the local tree teardown's collect_local_firstmate_states enumerates, since that walk skips remote registry entries for the same reason. Mutual exclusion is unchanged - every home reachable through local parent links still derives one identical lock file per project, and an unreadable binding, an unsupported route, an unreachable local parent, a cycle, and an over-deep chain all still refuse. Origin-less local-only projects keep resolving through their worktree top. Regression coverage pins the anchor for the main-home layout, a local secondmate, a remote-seeded home, and its local child; drives teardown end-to-end in a remote-seeded home; keeps the cross-home slot-ownership refusal across that boundary; and proves two homes still serialize on the one shared lock file. * no-mistakes(document): Clarify machine-local Treehouse lock ownership * fix(bearings): repair board listening and decision reconciliation (#3872) * fix(bearings): repair the board's listening, card hygiene, and reconcile path Three defects made the fleet board go quiet and then lie about what still needs the captain. Never arm a poll on a session that is not live. `lavish-axi <file>` exits 0 even when it refuses to reopen a session the captain ended from the browser, reporting `status: user-ended` with the same session id, so the build's exit-status check accepted a dead session, printed `already-armed`, and left the board reading "not listening". The build now proves the session is live from a fresh authoritative listing immediately before arming - not from the establish call's status alone, which is already stale by then - reopens once when it finds the session ended, and refuses rather than arming when it stays ended. A reopen also replaces the pre-reopen source generation before reporting success, so a runner on its way out cannot be mistaken for a listener, and a board whose source is registered but unowned gets a replacement started before the build returns. Let a dead generation's ownership actually move. Reclaiming a claim ran its capture-reservation cleanup first, and that cleanup re-verifies the recorded state-root identity, so a claim naming a pid and a process group that were both provably gone could not be cleared: reconcile reported a start while nothing attached, and retire refused with "cannot release source ownership". Reservation records are keyed by claim token and every replacement claims a fresh one, so they are hygiene, not an ownership invariant. Reclamation now additionally requires the owning process group to be absent independently, which keeps a reused pid whose poll child still runs from ever reading as a gone generation. A live owner and a crashed leader whose owned group survives are still never reclaimed. Stop carding decisions whose subject already landed. The build drops a decision card whose work item or PR appears in the payload's own landed rows, and one whose task is no longer an open captain call, naming each drop on stderr. A task whose state cannot be established is kept, because a call wrongly hidden is worse than a card wrongly shown. Add the reconcile choice, and make it structurally incapable of closing a call. Every decision card carries a standard `reconcile` option, injected by the build rather than left to the composer. The board now emits the picked option and any freeform note as separate structured fields instead of fusing them, so a reconcile selection is not expressible as an answer value at all - the defect that let `reconcile - <note>` reach the intake as an ordinary answer. The adapter routes selections from that structured field, creation of a reconcile request is bound to a verified board source rather than the shared keyed-answer intake, and the intake still refuses the reserved value on every channel. Each authorization is bound to the captain-hold generation that produced the card, so an obsolete card cannot close a later call, and both terminal outcomes require a pending request: `reconcile close` records the evidence under its own `reconciled` mode so it never reads as the captain's words, and `reconcile note` leaves the call open. Anything unprovable - an unversioned row, a missing generation, an unreadable…
Mauryanx
added a commit
to Mauryanx/firstmate
that referenced
this pull request
Sep 9, 2026
* fix(bin): scope the worker role contract for ship and scout launches (#3797)
* fix(brief): scope Firstmate workers to their launch contract
* no-mistakes(document): Clarify supervisor scope and worker contract ownership
* no-mistakes(ci): Removed the heading-based bypass so every ship/scout launch receives the current worker-role contract. Added a regression that failed before the fix and passes afterward. Dispatch, brief, and delivery suites, focused ShellCheck, and git diff --check all passed
* no-mistakes(review): make launch overlay sole owner of worker role contract
* no-mistakes(review): narrow heading test dimension and fix publish error wording
* no-mistakes(review): gate role supersession, fix render guard, drop AGENTS twin
* no-mistakes(document): align architecture AGENTS.md scope and spawn launch-brief header
* fix(bin): stop ringing steering doorbells into dead panes (#3823)
* fix(bin): stop ringing steering doorbells into dead panes
The steering-inbox doorbell was a plain sentence plus Enter typed into a
worker's pane, and the watcher re-rang it on the assumption that a ring is
free. In a pane whose agent has exited that line is a shell command, and the
re-ring ladder kept typing it into a shell that can never acknowledge it.
- Prefix the doorbell with the shell no-op `: ` so a bare shell executes
nothing while a live worker still reads the same self-describing line.
`#` is not used because interactive zsh does not treat it as a comment by
default and the claude harness binds it to memory mode.
- fm_task_inbox_ring skips the pane (return 3) when the backend positively
classifies the agent as dead; missing, ambiguous, unreadable, and unverified
endpoints still ring so a blind classifier never starves a live worker.
- The watcher caps the ladder for a dead pane: one stale wake for recovery,
no ring, no ladder walk, and the durable record stays for
stuck-crewmate-recovery. fm-send and the remote steer leg report the skip.
Tests cover the no-op in real shells, the dead/live/unclassifiable ring
verdicts, and the single-surfacing watcher path.
* no-mistakes(review): Quote doorbell paths against shell injection
* no-mistakes(review): Reject terminal-control paths before ringing
* no-mistakes(review): Document accepted partial doorbell delivery race
* no-mistakes(review): Skip unavailable endpoints before busy-state handling
* no-mistakes(test): Respect shell startup PATH in environment allowlist test
* no-mistakes(test): Fix doorbell test fixtures for endpoint liveness
* no-mistakes(test): Prioritize confirmed restarts and clean shell test syntax
* no-mistakes(document): Document dead and missing doorbell recovery
* no-mistakes(ci): Fixed the persistence-reply timeout race by rechecking for a correlated reply immediately before falling back to a nudge. Added a deterministic regression covering replies arriving between the preliminary resolution pass and timeout handling. Verified with the targeted restart suite, project lint, coverage guard, bash syntax checks, and diff checks
* fix(bin): wait out transient primary-checkout reads in the spawn worktree poll (#3834)
* fix(spawn): keep the worktree poll from adopting the repository primary
After `treehouse get` is sent, the worktree-discovery poll reads the pane's
foreground-process cwd. While treehouse is still fetching and checking a slot
out, the foreground process is treehouse itself and it reports the repository's
PRIMARY checkout as its cwd for several seconds. The poll accepted any path
that merely differed from the spawning project, so from a linked spawning home
- whose project is itself a worktree of that repository - it adopted the
primary, and the isolation guard then refused a launch whose slot treehouse
went on to create normally.
Screen every candidate with the isolation guard's own conditions, extracted as
spawn_worktree_isolated, so a read the guard would reject stays a transient the
poll keeps waiting through. The two-consecutive-reads rule and the guard as
final backstop are unchanged; a pane that never reaches an isolated worktree
still fails at the existing 60s deadline, now naming the last path it reported.
The already-settled timing assertion counted whole-spawn wall time against a
5s budget and failed on unmodified HEAD on slower machines; it now counts pane
reads, which is what "one confirming read, not an extra cycle" actually means.
* fix(spawn): say which path the worktree wait rejected, and why
Screening every discovery-poll candidate means a host that never reaches an
isolated worktree spends the whole 60s window before refusing. That wait is
deliberate - separating a transient from a terminal misconfiguration needs
machinery this path does not want - so the refusal explains itself instead:
the isolation check records why a candidate failed, and the deadline names the
last path seen together with that reason. Message and diagnostics only; the
poll's control flow is unchanged.
Two suites asserted the guard's wording on paths the poll now rejects rather
than adopts, so their refusal arrives from the deadline instead: realign
fm-tangle-guard's non-git and subdirectory-of-primary cases (each now also
asserting the stated reason, and the second the metadata absence it was
missing) and the herdr projection e2e's forced non-worktree cwd.
* no-mistakes(review): stub poll sleep in tangle-guard spawn isolation test
* no-mistakes(document): document spawn poll isolation screen in fm-spawn header
* test(spawn): make the non-git isolation case non-git anywhere
The refusal-reason assertion for a path outside any repository assumed TMPDIR
is not inside a git repository. Where it is, git walks up from the temporary
directory, finds that repository, and the spawn reports the subdirectory cause
instead - so the case passed or failed on a property of the host rather than on
the behaviour under test.
Build the path under a directory the test then names in GIT_CEILING_DIRECTORIES,
which git documents as not chdir-ing up into a listed directory while looking
for a repository. Git never excludes the directory being searched, so the
ceiling is the parent of the path handed to the spawn.
The assertions pin which cause fired rather than the sentence that explains it,
leaving the operator wording free to improve.
* no-mistakes(document): point spawn poll comment at the isolation screen's comparison
* no-mistakes(ci): Fixed the "Behavior portable serial 2" failure in tests/fm-tangle-guard.test.sh ("non-worktree spawn did not say why the path was rejected (missing: 'not inside a git worktree')"). Root cause, in this PR's code: bin/fm-spawn.sh's spawn_worktree_isolated resolved the git toplevel with `wt_top_real=$(cd "$SPAWN_WT_TOP" ...)`. For a path in no repository, `git rev-parse --show-toplevel` yields empty, and `cd ""` is a SUCCESSFUL no-op on bash before 5.3 (CI's ubuntu-latest ships bash 5.2). The empty toplevel therefore resolved to fm-spawn's own cwd — the CI checkout — so the poll reported "it is a subdirectory of worktree root '/home/runner/work/firstmate/firstmate'" instead of the correct "it is not inside a git worktree". Dev machines with bash 5.3 fail `cd ""`, which is why the suite passed locally and only failed on CI; it is a genuine shell-portability defect in the reason vocabulary this change added, not a test-environment artifact. Fix (smallest root-cause change, 1 line + comment, bin/fm-spawn.sh:2168-2173): guard the empty value so it never reaches `cd` — if [ -n "$SPAWN_WT_TOP" ] && ! wt_top_real=$(cd "$SPAWN_WT_TOP" 2>/dev/null && pwd -P); then No change to the poll's timing or deadline behavior (respecting the recorded refusal-latency and spawn-wt-reason-vocabulary decisions), no new tests, no other files touched. Verification: - Reproduced the exact CI failure locally by putting bash 3.2 (same `cd ""` semantics as CI's 5.2) first on PATH: fails before the fix with the identical message shape, passes after. - tests/fm-tangle-guard.test.sh passes under both bash 3.2 and bash 5.3. - tests/fm-spawn-worktree-settle.test.sh and tests/fm-spawn-pool-base-freshen.test.sh pass; shellcheck -x bin/fm-spawn.sh clean. - bin/fm-test-run.sh --changed: 46 suites completed, every FM_TEST_END exit=0, 1076 passing assertions, 0 "not ok" (including fm-tangle-guard, fm-control-relaunch, fm-lint). The run ended on my own 900s wall-clock cap (rc=124), not on any test failure
* fix: support stock macOS Bash 3.2 paths (#3732)
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): read orphaned green ci monitor and daemon-down failed record as not failed (#3846)
* fix(bin): read an orphaned green ci monitor as held-for-merge, not failed
A no-mistakes run held for a captain merge decision keeps its ci step
polling until merged or closed; when the shared daemon restarts under
that poll, the run is recorded failed although every substantive step
completed and GitHub reports the PR green. A monitor whose only
remaining job is to observe a human decision must not convert the
absence of that decision into a failure verdict.
fm-crew-state.sh now reclassifies a terminal failed run as done
(held-for-merge), surfacing the run's PR URL, when the steps table
shows every step completed except exactly ci failed and the ci log's
last recognized marker reads checks green. A genuinely red check, an
unreadable ci log, or a second failed step keeps the failure.
* no-mistakes(review): Read daemon-down coarse failed ledger as unknown, not failed
* no-mistakes(document): docs: align AGENTS.md failed-verdict guidance with crew-state reclassification
* fix(spawn): verify preserved backlog state after interrupted spawn delivery (#3852)
* fix(bin): read preserved spawn state back before the interrupted exit claims it
The deferred-signal exit path asserted the paired task record and
In-flight backlog state were preserved without reading either back,
exactly when a reader is least able to check (fm-yi4j evidence,
2026-09-05). The commit's exit status alone has been observed to agree
with a row that did not actually move.
The exit path now re-reads the record and the row under the same
per-task lock as the commit, repairs a row the commit believed it
moved, and phrases the error as exactly what was verified or attempted
- verified preserved, repaired and verified, or an explicit
preservation-could-not-be-verified with the reason and hand-closeout
instruction. Two behavior tests drive a lying tasks-axi start through a
real interrupted spawn and assert the printed claim and the real
backlog state agree.
* no-mistakes(test): Fix calm suite for Pi 0.85 and pin test umask
* no-mistakes(document): Document interrupted-spawn preservation claim in backlog gate owner
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-spawn.sh's deferred-signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-backlog-transition-lib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean
* no-mistakes(ci): Fixed the Greptile P1 on bin/fm-backlog-transition-lib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean
* docs: correct runtime-backend maturity labels for Herdr (#3821)
* docs: correct stale tmux/herdr backend maturity claims
Herdr now has 21 test files, its own required CI job (tests-herdr)
that installs a pinned build and hard-fails on "skip: herdr not
found", while tmux has 3 test files and is only required as a
dependency of the portable-serial e2e lane. zellij, orca, and cmux
still have no CI lane at all. AGENTS.md and docs/herdr-backend.md
still called Herdr merely "experimental" alongside those three,
misleading every session and reader about actual coverage.
Update AGENTS.md's config/backend entry, the opening lines of
docs/herdr-backend.md and docs/tmux-backend.md, the runtime-backend
section of docs/configuration.md, and the matching claims in
docs/architecture.md, CONTRIBUTING.md, and README.md so they agree
and distinguish tmux (default), herdr (own required CI lane, largest
suite, Windows still spike-only), and zellij/orca/cmux (still
experimental, no CI lane). No behavior, selection order, or
dispatch logic changes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GDYsWPEfwTuPNjQ2nGBCcj
* no-mistakes(review): docs: fix stale herdr label and CI-lane wording
* no-mistakes(document): docs: align tmux adapter label in scripts.md
* no-mistakes(review): docs: drop duplicated herdr CI claim from tmux page
* no-mistakes(review): docs: drop windows claim, align contributing backend wording
* no-mistakes(review): docs: trim duplicated CI claim from herdr opening line
* no-mistakes(review): docs: drop unguarded largest-test-suite superlative
* no-mistakes(review): docs: restore tmux verified label and README experimental scope
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: restore intent-targeted no-mistakes validation (#3865)
* fix: delete deterministic no-mistakes test baseline, restore intent-targeted Test
PR #3644 pinned commands.test to a fm-test-run.sh --changed walk of the
repository's 75-162 tests/*.test.sh scripts. no-mistakes runs commands.test
verbatim and unconditionally after every fix round, so that walk multiplied
by round count: measured at 32.7 minutes per validation versus 3.6 minutes
intent-targeted. Delete the pin and restore the 3.6-minute posture.
Add tests/fm-nm-test-contract.test.sh as a regression guard, parsing
.no-mistakes.yaml as YAML (ruby's bundled Psych, matching the parser
tests/fm-test-run.test.sh already uses for ci.yml) rather than grepping its
text, restoring in legal form what PR #823 added and PR #1282 removed.
Record the rule in docs/configuration.md's "Gate defaults" section (the
authoritative owner CONTRIBUTING.md already points at) and strengthen
CONTRIBUTING.md's existing local-Test guidance to state it plainly: never
configure commands.test to a deterministic test command, complete or partial.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSt9JvrQMVc4u3jPyUFCFC
* no-mistakes(review): Centralize no-mistakes test policy and narrow guard
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: verify Treehouse slot ownership before teardown (#3837)
* fix(bin): verify pool-slot ownership before returning a worktree slot
Workers were killed when cleanup returned a Treehouse pool slot that a
different, live task had already taken. Teardown now proves the slot is
genuinely this task's before releasing it: it refuses when another task
record claims the same live worktree path, or when the endpoint's working
directory contradicts the recorded slot, and that refusal holds under
--force. Slot allocation, metadata publication, ownership verification,
and slot return are serialized across linked firstmate homes, and forced
secondmate cleanup verifies descendant slot ownership before returning
any child worktree.
Regression coverage drives the scripts with two task records naming one
slot path and asserts the live worker survives and its slot is not reset.
* no-mistakes(review): Protect slots across cloned Firstmate homes
* no-mistakes(test): Gate teardown locking on genuine Treehouse slots
* no-mistakes(test): Clarify pooled descendant slot gating
* no-mistakes(test): Synchronize watcher re-arm test on process exit
* no-mistakes(test): Wait for watcher cleanup before timeout escalation
* no-mistakes(document): Document pool-slot ownership safeguards
* no-mistakes(ci): Fixed all reported CI issues: normalized bare local Git origins to the same Treehouse project-lock identity as absolute clone origins; resolved ShellCheck SC1091 with explicit conditional sourcing; and taught concurrent Herdr teardown coverage to retry expected Treehouse lock contention. Added behavioral regression coverage for bare/absolute origin lock identity. Verified endpoint-safety tests, watcher tests, full CI lint, and the previously failing Herdr teardown assertion
* fix(bin): resolve relative origins from repository root
* no-mistakes(ci): Fixed teardown so an exact recorded endpoint may change cwd without falsely vetoing cleanup. Removed cwd-based ownership refusal while preserving cross-home record exclusivity and project locking. Updated behavioral coverage for both foreign slot ownership refusal and moved-cwd teardown success. Endpoint-safety, backend, watcher, checkpoint, and targeted lint checks pass. Real Herdr presentation E2E progressed successfully but exceeded the 600s local timeout
* feat(bin): add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary (#3867)
* feat: add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary
Add omp as a verified harness: anchored process-name detection with a
Firstmate-owned FM_OMP_HARNESS launch marker that needs real omp ancestry,
the fm-spawn launch template with foreign-marker clearing, the tracked
.omp/fm-worker-overlay.yml posture overlay, --auto-approve, --cwd, and
pre-launch model validation scoped to providers 'omp models --json' lists.
Workers get a state-resident busy-state extension keyed on agent_end
without willContinue (omp has no agent_settled). The primary gets two
tracked .omp/extensions: a turn-end guard that answers omp's blocking
session_stop hook by compelling one continuation per turn, with the
pre-tool seatbelts and Run-tier session-start delivery, and a watcher
extension ported from the Pi one with fm_watch_arm_omp. Control tables,
composer busy footers, omp's status row as a bare-composer boundary, the
extension supervision model with an omp-keyed ownership proof, the
session-start diagnostic, and the supervision protocol snippet follow.
Verified live on omp 18.1.11 with openai-codex/gpt-6-astra: a Herdr scout
through spawn, busy state, steer, interrupt, exit, and teardown, and the
isolated rpc primary lab through extension auto-discovery, digest
delivery, lock identity, watcher arm, successor and wake delivery, and
the compelled guard continuation.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test: prove the omp guard continuation through a guard spy
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* fix(spawn): clear the gemini marker at the omp launch boundary
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): force the guard stage by freezing the watcher and clear lint findings
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): reap the live lab by path and record omp's rpc shutdown as a note
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): spawn a real secondmate for the discovery rule and classify the omp surfaces
Replace the template-extraction check with a genuine --secondmate launch
pinned to the fake tmux backend, assert the worker extension's handler set
through the executable rather than its bytes, classify the two new omp
surfaces in the documentation inventory, and record the Herdr worker
evidence in the runtime-backends verification doc.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* no-mistakes(review): omp: unverify remote routes, narrow busy regex, drop overlay approval pin
* no-mistakes(review): omp: validate config-pinned model, correct remote and marker docs
* no-mistakes(review): omp: pin config-model validation with a test, trim overlay
* no-mistakes(review): omp: sync guard evidence, drop dead param, map quota family
* no-mistakes(review): omp quota: refuse unmapped prefixes, match bare model scopes
* no-mistakes(document): docs: cover omp in cd-guard, quota, continuity, tmux
* no-mistakes(document): docs: add omp subagent-guard row, fix live test header
* no-mistakes(ci): Fixed both failing behavior shards and the Greptile P1 in bin/fm-composer-lib.sh. Root cause of "Behavior portable serial 1" and "Behavior portable parallel 2": the omp busy regex (FM_DELIVERY_OMP_BUSY_REGEX_DEFAULT) and omp status-row furniture regex (FM_COMPOSER_OMP_STATUS_RE_DEFAULT) used the bracket range [⠁-⣿]; BSD grep on macOS accepts it but GNU grep on Linux CI aborts with "Invalid collation character", failing every omp busy/furniture read (3 assertions across fm-omp-harness, fm-tmux-submit-busy, fm-composer-lib). Replaced the range with one shared explicit alternation FM_OMP_SPINNER_FRAMES_RE of omp 18.1.11's unicode-preset spinner frames (status set ⣾⣽⣻⢿⡿⣟⣯⣷ + activity set ⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏, read from the installed binary), the same pattern the Kimi busy regex already uses in CI. For Greptile's finding (the harness-agnostic furniture rule's first alternative matched any 1–4-byte token + ' · ', so wrapped typed input like 'fix · tests' with the cursor on it regressed from pending to unknown; reproduced locally vs base), pinned that alternative to omp's identity cell (π||pi, the icon.omp of each preset in the 18.1.11 binary). Tests: fm-composer-lib.test.sh asserts 'fix · tests' is not furniture, a status-set spinner row is furniture, and the wrapped composer screen reads pending under both locales (CAPS_TMUX cursor 3); fm-omp-harness.test.sh asserts a status-set frame reads busy. New negative cases fail against the pre-fix lib and pass after. Verified: fm-omp-harness, fm-tmux-submit-busy pass via bin/fm-test-run.sh; fm-composer-lib passes all cases except one pre-existing, unrelated local failure (Herdr half-block test uses printf '▀', unsupported by macOS bash 3.2; fails identically on a pristine HEAD export, passes on CI bash 5); shellcheck and bin/fm-lint.sh clean. Caveat: GNU grep is unavailable locally, so the Linux compile was not run directly; the fix uses only constructs already proven on CI's GNU grep (multibyte literal alternations, incl. under LC_ALL=C). Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh, tests/fm-omp-harness.test.sh. No docs needed changes (they describe the rule generically)
* no-mistakes(ci): Greptile Review: fixed. The omp status-row furniture regex FM_COMPOSER_OMP_STATUS_RE_DEFAULT in bin/fm-composer-lib.sh still accepted a literal `pi ·` opening, so wrapped composer input beginning with `pi ·` was truncated and misclassified. Read the installed omp 18.1.11 binary: the ascii preset's `icon.omp` is `pi` but its `sep.dot` separator is ` - ` (unicode/nerd use ` · `), so a real ascii status row never contains `pi ·` and that alternative could only ever match typed text. Removal-first fix: dropped `pi` from the identity alternation (now `(π|)`) and updated the comment to record why the ascii preset is excluded. Tests (tests/fm-composer-lib.test.sh): added a negative furniture case for 'pi · e · phi as the three constants' and a wrapped-screen assertion (CAPS_TMUX, cursor 3) that a continuation row opening `pi ·` reads pending in both locales; the new case fails against the unfixed lib and passes after. Verified: composer test with the half-block case skipped passes all 33 cases including the omp matrix; bin/fm-test-run.sh tests/fm-omp-harness.test.sh passes; shellcheck -x clean on both files; bin/fm-lint.sh clean. The full composer test via the runner fails locally only on the pre-existing half-block case (bash 3.2 printf cannot emit ▀; passes on CI bash 5), identical to before this change. Docs unchanged (they describe the rule generically and never mention the ascii identity cell). PR must be raised via no-mistakes: not caused by code. attestation.head_sha is cdddc60 while the PR head is cd51cf4 because the pipeline's ci-phase push moved the head; the outer executor's re-push will re-bind the attestation. No file change for that check. Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* fix(pi): invoke Bash helpers correctly on native Windows (#3843)
* Fix Pi shell invocation on native Windows
* no-mistakes(document): Document Pi Windows Bash transport
* no-mistakes(ci): Captain, staged a narrow fix: register the Pi Windows regression for both extension paths, make Windows mode emulation non-failing, and enforce LF shell checkouts. Mapping and coverage checks pass; CI/Require no-mistakes were approval-gated externally
* no-mistakes(review): Cover async Windows branch-outcome Bash invocation
* no-mistakes(review): Preserve Cygwin checks and refresh Windows timing
* no-mistakes(test): Invoke OpenCode operational-input owner through Bash on Windows
* validation-fixture
* no-mistakes(document): Document Windows Bash helper invocation
* no-mistakes(ci): Fixed PR-caused changed-selection failure by removing the malformed tracked evidence artifact and allowing deleted, unconsumed source paths to retire cleanly while preserving fail-closed behavior for live unmapped paths. Added regression coverage. Verified native-Windows Pi shell-seam test passes and --changed selects the Windows regression
---------
Co-authored-by: test <test@example.invalid>
* test(bin): pin teardown outcomes for squash-merged rebased branches (#3870)
* fix(bin): recognise squash-merged rebased work as landed at teardown
A pipeline rebase can leave the local worktree on pre-rebase commits while
GitHub squash-merges the rebased head. The landed-work test then compared
those stale commits against a squashed main and refused cleanup of work
that had already landed.
When the forge reports the recorded PR merged and its merge commit is on
the default branch, treat a local branch that only repeats paths from the
pipeline push as stale rather than unlanded. If the forge is unreachable,
the same coverage check runs against a PR head whose content is already
on default. Extra local paths still refuse.
* fix(bin): drop unprovable squash-rebase landed-work coverage
Path-set coverage treated a diverged local branch as landed whenever it
touched the same files as the squash merge. That accepts the reviewer's
failing sequence: same path, different content, work discarded.
git cherry and merge-tree containment were already too strict on the real
rebase-fold case. No remaining check is both safe and permissive enough
to recognise a stale pre-rebase copy without also accepting unlanded
edits, so that case still refuses.
Keep the proofs that hold: a merged PR head that contains local work, or
a clean content-in-default tree match. Tests now refuse same-path
different content and extra unlanded commits, and still allow a local
branch that followed the pipeline rebase.
* no-mistakes(review): drop recorded-pr-head fallback and reverted-design leftovers
* no-mistakes(review): silence squash-merge stdout corrupting test PR head
* no-mistakes(review): make unlanded follow-up commit sole cause of refusal
* no-mistakes(document): correct stale squash-rebase fixture comments in teardown tests
* no-mistakes(ci): Split the three reported checks: - CI (run 34061098467) and Require no-mistakes (run 34061098460) both concluded `action_required` — approval-gated workflow runs that never executed a step. Not caused by this PR's code; no change can clear them. - Greptile Review was a genuine defect in the new tests: the three new refusal cases (tests/fm-teardown.test.sh) asserted only exit status 1 and a REFUSED line, so a teardown regression that destroyed the worktree, branch, and task record before reporting refusal would still pass. Fix (tests only): added one `assert_refusal_retained_task_state` helper and called it from `test_squash_merged_same_file_different_content_refuses`, `test_squash_merged_rebased_local_with_unlanded_commit_refuses`, and `test_squash_merged_stale_local_refuses_when_forge_unreachable`, each capturing the worktree HEAD before `run_teardown`. It pins that the refusal left the isolated copy on disk, the task branch still checked out at the same unlanded commit, and state/task-x1.meta intact. Verification: the four squash tests pass; a sensitivity probe ran the ALLOW fixture (teardown completes) and pointed the same helper at the outcome — it fires, because a completed teardown detaches/deletes the branch and removes the task record, proving the assertions discriminate. Full tests/fm-teardown.test.sh: 83 passing. bin/fm-lint.sh clean with pinned ShellCheck 0.11.0 + actionlint 1.7.12 (plus an explicit --external-sources pass on the changed file). bin/fm-test-run.sh --check-coverage ok. Caveat: test_herdr_flat_teardown_preflight_refuses_before_changes (mode missing-adapter) fails on this machine. Verified it fails identically on base commit f91a950 via `git archive`, so it is a pre-existing local environment difference untouched by this diff; skipped to run the rest of the suite, not modified
---------
Co-authored-by: Morten Gad <mogad@itm8.com>
* fix(bin): keep supervision armed for registered custom checks (#3860)
* fix(bin): keep supervision armed for registered custom checks
A custom check bound by bin/fm-check-register.sh only ever runs inside the
watcher's check sweep, but fm_supervision_status counted in-flight tasks, the
relay poll shim, and process-event sources as supervision need, and not
registered checks. Tearing down the last task therefore stopped every
home-level check silently until the next spawn.
Count a state/<id>.check.sh that carries its state/<id>.check-trust binding as
supervision need. The relay shim keeps its own trust path and task PR polls
carry no such binding and are torn down with their task, so neither arms a home
by accident. Presence of the binding is the whole test: the sweep validates the
bytes at execution time and wakes firstmate when it rejects one, which is the
outcome an idle home needs.
Closes #3856
* no-mistakes(review): name registered checks in turn-end block banner and doc invariant
* no-mistakes(review): narrow PR poll predicate test to what it proves
* no-mistakes(document): point Grok re-arm step at supervision-need owner
* fix(bin): resolve Treehouse locks for remote secondmate homes (#3883)
* fix(bin): resolve the shared Treehouse project lock inside remote secondmate homes
Every spawn and teardown inside a remote-seeded secondmate home refused,
because the project lock's anchor could not be resolved there.
fm_firstmate_root_home walks a home's parent bindings upward to find the
anchor the lock lives in, and treated a remote parent binding as an error.
A remote-seeded home's parent is on another machine, so that walk can never
succeed from there - and neither can the home's own local descendants, whose
chain terminates at the same record. Both fail closed on every Treehouse-backed
spawn and every pool-slot teardown.
A remote parent now terminates the walk at the home holding it, which is the
correct anchor: a lock taken on this filesystem is neither held nor observable
across that boundary, and that home is already the top of the local tree
teardown's collect_local_firstmate_states enumerates, since that walk skips
remote registry entries for the same reason. Mutual exclusion is unchanged -
every home reachable through local parent links still derives one identical
lock file per project, and an unreadable binding, an unsupported route, an
unreachable local parent, a cycle, and an over-deep chain all still refuse.
Origin-less local-only projects keep resolving through their worktree top.
Regression coverage pins the anchor for the main-home layout, a local
secondmate, a remote-seeded home, and its local child; drives teardown
end-to-end in a remote-seeded home; keeps the cross-home slot-ownership
refusal across that boundary; and proves two homes still serialize on the
one shared lock file.
* no-mistakes(document): Clarify machine-local Treehouse lock ownership
* fix(bearings): repair board listening and decision reconciliation (#3872)
* fix(bearings): repair the board's listening, card hygiene, and reconcile path
Three defects made the fleet board go quiet and then lie about what still
needs the captain.
Never arm a poll on a session that is not live. `lavish-axi <file>` exits 0
even when it refuses to reopen a session the captain ended from the browser,
reporting `status: user-ended` with the same session id, so the build's
exit-status check accepted a dead session, printed `already-armed`, and left
the board reading "not listening". The build now proves the session is live
from a fresh authoritative listing immediately before arming - not from the
establish call's status alone, which is already stale by then - reopens once
when it finds the session ended, and refuses rather than arming when it stays
ended. A reopen also replaces the pre-reopen source generation before
reporting success, so a runner on its way out cannot be mistaken for a
listener, and a board whose source is registered but unowned gets a
replacement started before the build returns.
Let a dead generation's ownership actually move. Reclaiming a claim ran its
capture-reservation cleanup first, and that cleanup re-verifies the recorded
state-root identity, so a claim naming a pid and a process group that were
both provably gone could not be cleared: reconcile reported a start while
nothing attached, and retire refused with "cannot release source ownership".
Reservation records are keyed by claim token and every replacement claims a
fresh one, so they are hygiene, not an ownership invariant. Reclamation now
additionally requires the owning process group to be absent independently,
which keeps a reused pid whose poll child still runs from ever reading as a
gone generation. A live owner and a crashed leader whose owned group survives
are still never reclaimed.
Stop carding decisions whose subject already landed. The build drops a
decision card whose work item or PR appears in the payload's own landed rows,
and one whose task is no longer an open captain call, naming each drop on
stderr. A task whose state cannot be established is kept, because a call
wrongly hidden is worse than a card wrongly shown.
Add the reconcile choice, and make it structurally incapable of closing a
call. Every decision card carries a standard `reconcile` option, injected by
the build rather than left to the composer. The board now emits the picked
option and any freeform note as separate structured fields instead of fusing
them, so a reconcile selection is not expressible as an answer value at all -
the defect that let `reconcile - <note>` reach the intake as an ordinary
answer. The adapter routes selections from that structured field, creation of
a reconcile request is bound to a verified board source rather than the shared
keyed-answer intake, and the intake still refuses the reserved value on every
channel. Each authorization is bound to the captain-hold generation that
produced the card, so an obsolete card cannot close a later call, and both
terminal outcomes require a pending request: `reconcile close` records the
evidence under its own `reconciled` mode so it never reads as the captain's
words, and `reconcile note` leaves the call open. Anything unprovable -
an unversioned row, a missing generation, an unreadable state - refuses
rather than acting.
Regression coverage fails without each fix, and pins every leak path: a bare
reconcile, a standalone close or note with no pending request, an any-channel
reconcile, an annotated selection from a freeform card, and a
generation-skewed authorization. An opt-in guard re-proves the lavish-axi
shapes and the reopen against the installed tool.
* fix(bin): quote the done comparison in the reconcile intake
shellcheck SC1010 reads the bare word as the loop keyword. The failed run
never reached its lint step, so this shipped in the recovered content.
* no-mistakes(review): Publish reconciled parent resolution before request retirement
* no-mistakes(review): Clarify committed cleanup and reconcile reservation scope
* no-mistakes(review): Preserve remote cards and legacy answer compatibility
* no-mistakes(test): Separate live claim release from stale reclamation
* no-mistakes(test): Allow terminal self-retirement during active capture
* no-mistakes(document): Document Bearings repair contracts
* no-mistakes(ci): Stabilized the failing Herdr presentation E2E by serializing test-harness Treehouse allocator calls, preventing concurrent recovery spawns from claiming the same pool slot while preserving Herdr concurrency coverage. Verified with the full E2E suite on Herdr 0.8.2, bash syntax checks, ShellCheck, and git diff checks
* fix(bin): allow pooled spawns without a git origin (#3885)
* fix(bin): skip pooled-worktree freshness fetch when no origin is configured
An origin-less local-only project has nothing remote to be stale
against, so fm-spawn's freshen_spawn_worktree_base refused to launch
crews for it. Detect a missing origin remote and skip the fetch
freshness gate entirely; an existing-but-unreachable origin keeps
refusing as before.
* no-mistakes(review): Preserve pool safety for absent and unusable origins
* no-mistakes(review): Refuse empty origin configurations during pooled spawn
* no-mistakes(review): Detect empty origin sections across config includes
* no-mistakes(review): Honor globbed includes when detecting origin configuration
* no-mistakes(review): Document conservative conditional include handling
* no-mistakes(review): Use Git-resolved config files for origin detection
* no-mistakes(review): Document included empty-origin detection boundary
* no-mistakes(document): Document originless pooled spawn behavior
* feat(tests): run live harness guards by default when available (#3889)
* feat(tests): run live harness guards by default where the harness is installed
The 24 live-harness guards each opened with their own env check, so on the
machine that has every harness - the one the product and its validation
actually run on - all of them skipped and passed. Fourteen had never been run
by the pipeline at all.
tests/lib.sh gains fm_live_gate as the single owner of that decision: a guard
that spends no model tokens runs wherever its tools are installed, a guard that
submits prompts stays opt-in, an absent tool is a named capability skip, and a
guard's own variable or FM_LIVE forces it on (turning an absent tool into a
failure) or off. Every live guard now opens with it, which also carries the
test-suite gate-refusal bypass into the guards that never sourced the shared
helpers and were therefore refused whenever a gate agent ran them.
bin/fm-test-run.sh records what a skip means: the family's expected class is
live-capability rather than a bare env opt-in, and each gate skip's reason is
logged and written to the timing artifact, so a lane can say which tool this
host could not exercise.
Only the token-free guards flip to default-on: composer-matrix, the harness
liveness drift guard, and the Herdr version floor. cursor-primary submits three
prompts, so it stays opt-in.
Running the drift guard unasked immediately found a real defect it existed to
catch: it resolved the harness through a generic `command -v cursor`, which on
a machine that also has the Cursor editor finds the editor launcher rather than
cursor-agent. That binary exits at once, leaving a bare shell in the pane and a
liveness-drift failure no classifier change could fix. It now asks
fm_cursor_resolve_binary first, the same verified owner fm-spawn uses.
CI installs the public Pi package in the portable serial lane and fails on its
skip token, so the Pi extension tests stop passing silently against a package
that is not there. No secret is added.
Verified on macOS 26.5.2 arm64: the drift guard runs with no variable set and
classifies 8 installed harnesses alive; the Herdr version-floor guard runs by
default and checks 4 real releases; every live guard refuses together under
FM_LIVE=0.
* fix(tests): keep the composer-matrix guard opt-in
Running it unasked is red on a healthy machine for reasons no code change here
removes: a harness that has not trusted this checkout sits on its own trust
dialog, which the guard treats as an unreadable composer and correctly fails.
The opencode 1.18.29 and grok 1.0.13 composer drift it also surfaced reproduces
identically on main and is filed as separate work.
So this token-free guard stays opt-in with the reason stated in its header, and
the coding guidelines record the narrow exception: a guard whose verdict
depends on host state that installing its tools does not establish may stay
opt-in, because one that is permanently red is one the fleet learns to ignore.
The other two token-free guards keep running by default.
* no-mistakes(review): Wire bearings guard and remove composer exception policy
* no-mistakes(review): Run Pi responsiveness guard by default
* no-mistakes(review): Gate AFK Pi Herdr through authoritative family sweep
* no-mistakes(review): Sanitize live gate test environments
* no-mistakes(document): Document default-on live guard behavior
* no-mistakes(ci): Fixed CI by installing the Pi package in portable-parallel-1, where fm-pi-primary-types.test.sh runs, and enforcing its package-missing gate skip there. Verified with fm-lint.sh, workflow actionlint, coverage partition checks, lane membership, and git diff checks
* no-mistakes(ci): Fixed CI’s Pi typecheck skip enforcement by giving npm, tsc, and Pi-package capability skips a shared prefix and configuring both relevant CI lanes to fail on that prefix. Verified missing tsc emits the expected skip, missing Pi package becomes a runner failure, and actionlint, ShellCheck, and git diff checks pass
* fix(bin): refuse test runs in the primary checkout when a task marker is set (#3891)
* fix(bin): refuse the behavior suite in the repository primary checkout
A task worker's isolated worktree placement is verified exactly once, when
its task starts, and nothing re-checks it afterwards. A worker that later
changes directory into the repository's primary checkout runs its Git
commands, and this branch-switching suite, against the one checkout every
linked worktree resolves against and every landing merges into. A run that
dies mid-suite can leave that checkout on a stray branch.
bin/fm-test-run.sh now refuses that case. When FM_TASK_ID marks a task
worker and the runner resolves to the primary checkout, every executing mode
exits non-zero before selecting a suite, with one line naming the primary
path and pointing at the assigned task worktree. The predicate is the one
bin/fm-spawn.sh already uses for launch placement: the working tree's own
git dir is the repository's common git dir, which separates the primary from
every linked worktree even when their top levels differ. A run with no
FM_TASK_ID set is unchanged, and so are the inspection modes, which execute
nothing. When git resolves neither directory - a non-repository fixture, a
detached copy - nothing proves this is the primary, so the run proceeds.
bin/fm-spawn.sh sets the marker: ship and scout launches export FM_TASK_ID
into the pane shell on the same pre-launch channel as GOTMPDIR, and the name
joins the sanitized launch environment allowlist so an isolated launch keeps
it.
* no-mistakes(review): clear inherited task marker in test lib; name resolved ROOT
* no-mistakes(document): docs: record FM_TASK_ID marker and runner placement refusal
---------
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): bound stale alarms for backlog captain holds (#3842)
* fix(bin): bound a stale alarm with the backlog hold, not only the status line
A legitimate wait has two records and the stale alarm reads only one.
`status_is_paused_or_captain_held` takes a status line, so it sees a wait the
worker declared. It cannot see the wait firstmate records when it hands work to
the captain: `bin/fm-captain-hold.sh hold` writes that into the backlog and
leaves the status log alone, so a delivered task keeps `done: PR ...` as its
last line for the whole time the captain is deciding.
Both stale branches were blind to it, and each churned a new pane hash back into
its own alarm: a `done:` line is captain-relevant and reaches the terminal-stale
branch, while a held task whose last line is `working:` reaches
`surface_nonterminal_stale` and fails its declared-wait test.
Consult that second record where the watcher is about to alarm, through
`bin/fm-captain-hold.sh open`, which already owns the predicate's semantics, and
bound the alarm on the shared `.paused-resurfaced-<key>` marker and
`PAUSE_RESURFACE_SECS` window the declared-wait absorb already uses. The first
sight still alarms, the window's end alarms once more, and a held crew that goes
genuinely silent still escalates through the wedge timer.
Only an established open captain call bounds anything: an unreadable backlog, an
absent or incompatible tasks-axi, a row this home does not carry, and every task
with no hold keep alarming exactly as before. The backlog hold is deliberately
not recorded as a declared pause, because the loop-top reconciliation and
`pause_state_class` both read the status line and would clear a flag that line
does not support.
Extends the fix in #3443, which closed the forms of this loop that the status
line itself can express.
* fix(bin): identify the captain call a stale alarm is bounded by
Three gaps in the bound added by the previous commit, all in how the throttle is
scoped and where the backlog is consulted.
The scope carried only the status-log signature. A task can be held, answered
with `--release`, and re-held as a genuinely different captain call without any
status append, so the second call inherited the first one's marker and its first
sight was absorbed - the one thing this bound must never do. The task id is not
the call: `bin/fm-captain-hold.sh open` gains `--identity`, which reports the
call's own lifecycle - its hold-set stamp and the number of recorded answers -
on an exit 0 and only then, leaving the silent predicate every existing caller
reads unchanged. The throttle scope now carries that identity.
The terminal path recorded the throttle before publishing the durable wake. A
failed append exits the watcher with nothing queued, and the next sighting then
read that fresh marker and absorbed the retry, turning a delayed alarm into a
lost one. Recording moves behind the append, as the non-terminal path already
had it, and the comment claiming the marker could not outlive its wake is gone
because it was false.
The backlog was consulted only on a new terminal pane hash. A captain call can
open after a hash was absorbed as provably working, changing neither the pane nor
the status log, so nothing re-read the backlog and the wedge timer kept firing
possible-wedge alarms through a legitimate wait. That timer now consults the call
at its own alarm boundary and takes the same bounded cadence - and only at that
boundary, so an ordinary repeat poll under the bound stays the local-only read it
was.
Regression coverage for each, all driving churn through one watcher process
rather than relaunching per pane change: relaunch cost dominated the earlier
shape, and an absorbing watcher stays in its poll loop across churn in production
anyway. An unheld task still alarms on every new hash, and an elapsed wedge timer
with no open captain call still escalates as a possible wedge.
* fix(review): Compose stale throttles with captain-call lifecycle identity
* fix(review): Preserve bounded same-hash captain-call resurfacing
* revert(bin): narrow the captain-hold stale bound to its observed defect
Lifts the lifecycle-identity and cadence-ownership work back out, leaving the
change at the shape that matches the defect actually observed: the stale alarm
did not consult the backlog captain hold, on either stale branch.
Reviewing the wider version surfaced a series of adjacent gaps in the watcher's
alarm state machine - a call opening after the first alarm, marker invalidation
at the hold lifecycle boundary, and which deadline a terminal timer represents.
They are real, but fixing them turns a small extension into a state-machine
change to the alarm path, which is a different review on a subsystem that is
being actively reworked. They are named as known limitations rather than carried
here, and none of them is load-bearing for what remains: the bound does strictly
less than the reverted version, leaves the wedge path escalating on
STALE_ESCALATE_SECS exactly as before, and introduces no silence that the
existing terminal-alarm path did not already have.
Kept from the reverted work is the record-after-append ordering, because that is
a defect in the code being shipped rather than an adjacent one: recording the
cadence marker before publishing the durable wake let a failed append lose an
alarm outright instead of delaying it.
History is preserved: the earlier commits stay on the branch and this removal
sits on top of them.
* fix(review): Document secondmate captain-hold scope boundary
* fix(document): Document captain-hold stale alarm scope
* fix(bin): bind the stale throttle to the captain call, not the status log
The throttle this change introduces was scoped to the task's status-log
signature. Answering a call with `--release` and holding the task again creates a
genuinely different captain call without necessarily appending to that log, so
the second call inherited the first one's marker and its first sight was
absorbed.
That is the one alarm this bound must never swallow. A delivery announced twice
is noise; a decision waiting on the captain that is never surfaced is invisible,
because nobody asks for what they do not know to ask for.
Measured rather than assumed, on the same fixture - a delivered task held for the
captain, released, and re-held with no status append, driven through bin/fm-watch.sh:
base c499f84 call-1 first=ALARM call-1 churn=ALARM new call first sight=ALARM
before this fix call-1 first=ALARM call-1 churn=absorbed new call first sight=absorbed
after call-1 first=ALARM call-1 churn=absorbed new call first sight=ALARM
Base never suppresses the new call, so the suppression came from this change and
closing it completes the fix rather than widening it.
`bin/fm-captain-hold.sh open` gains `--identity`, printing the call's lifecycle -
its hold-set stamp and count of recorded answers - on an exit 0 and only then, so
the silent predicate bin/fm-teardown.sh reads is untouched. The throttle scope
carries that identity beside the status signature.
The sibling case was measured too and is NOT included: on the status-declared
path, where the last line is `captain-held:`, base already absorbs a re-held
call's first sight. That behaviour predates this change and stays documented as a
known limitation rather than repaired here.
* fix(document): Document captain-call throttle lifecycle scope
* fix(ci): isolate the Herdr restart fixtures from a claimed worktree
The Herdr behaviour test intermittently reused a local worktree still claimed
by an earlier fixture after a restart.
The restart scenarios now use an isolated Treehouse project.
The full Herdr test passes on Herdr 0.8.2; bash -n and git diff --check pass as
well.
* feat(pi): resolve extension-registered providers in the supervision branch (#3871)
* Let the supervision branch resolve extension-registered providers
The isolated branch ModelRuntime cannot see providers an extension
registered into main's runtime at run time, so a pin on pi-devin-auth's
devin/swe-1-7 (or an unpinned branch following a main session on devin)
failed with "unavailable to the isolated branch runtime".
Capture main's ModelRegistry alongside mainModel and copy each
extension-registered provider config into the branch runtime at
model-resolution time. The config carries the provider's own streamSimple
and oauth wiring by reference, so the custom gRPC transport reaches the
branch unchanged instead of being reimplemented. The /supervision-model
picker uses the same copy so those models are offered.
Update configuration.md and pi-supervision-branch.md, which previously
stated extension-registered providers were not offered.
* no-mistakes(document): docs: own devin provider carve-out in branch architecture doc
* no-mistakes(ci): Fixed the Greptile P1 finding: the /supervision-model picker copied extension-registered providers into the branch ModelRuntime but checked hasConfiguredAuth without refreshing them, so providers with provisional post-registration auth were omitted from the picker while the pin-resolution path (which did refresh) accepted them. Root-cause fix in .pi/extensions/fm-branch-supervision.ts: moved the `refresh({ providers, allowNetwork: false })` call into `copyExtensionProviders` (now async, refreshing every provider it copied) and removed the duplicate per-provider refresh from `resolveBranchModel`. Both the picker and the resolution path now share one copy-and-refresh step, so hasConfiguredAuth is real in both. Regression coverage in tests/fm-pi-branch-extension.test.sh: the stubbed ModelRuntime now mirrors the real runtime by leaving a registered provider's auth pending until `refresh()` runs for it. With that stub, the existing extension-registered-provider case fails against the pre-fix extension (picker offers only anthropic/main-model) and passes with the fix. Verification: tests/fm-pi-branch-extension.test.sh passes (42 ok, no failures); tests/fm-branch-supervision.test.sh passes; tests/fm-pi-primary-types.test.sh skips locally because tsc is not installed (the refresh signature reused is the one the existing code already called). Intent constraints preserved: isolation flags untouched, carve-out still scoped to provider registration, graceful fallthrough when no providers are registered
* ci: retrigger flaky Herdr/serial-1 lanes
* no-mistakes(document): docs already cover branch extension-provider copy
* fix(bin): gate secondmate wake-loop stall alerts on real queue no-progress (#3943)
* fix(watch): detect stalled secondmate queue progress
* no-mistakes(review): gate secondmate stall on active turns and progress episodes
* no-mistakes(document): align secondmate wake-stall docs with progress-episode detector
* no-mistakes(document): clarify active-turn gate in wake-stall config docs
* no-mistakes(ci): Fixed the one real defect behind the failing checks. ROOT CAUSE (Greptile P1, real code defect in this PR): `secondmate_wake_stall_tick` in bin/fm-watch.sh reset the no-progress timer only when the oldest actionable queue sequence INCREASED (`[ "$seq" -gt "$observed_seq" ]`). When a secondmate is retired and reprovisioned under the same task ID, its fresh home's queue sequence restarts BELOW the recorded position, so the comparison is false, no reset happens, and the new queue inherits the retired generation's already-expired idle interval — emitting a false `secondmate wake-loop stalled` on its very first observation. That is precisely the false-alarm class the user intent requires this PR to remove. FIX (smallest, removal-first): bin/fm-watch.sh:749 now resets when the drain position MOVES AT ALL (`-ne` instead of `-gt`). Draining moves it up, reprovisioning moves it down; neither is a continued no-progress episode. The asymmetric `-gt` branch is removed rather than special-cased or hardened. Updated the function header comment plus the two doc sentences in docs/architecture.md and docs/configuration.md that stated the old advance-only semantics. REGRESSION TEST: added `test_secondmate_reprovisioned_queue_starts_a_fresh_interval` to tests/fm-wake-queue.test.sh (registered in the invocation list). It drives the real watcher through retired generation (seq 9) -> reprovision (seq 3, later clock) -> freeze, asserting observable wake-queue output, no source-text inspection. VERIFICATION: - Fails before / passes after: with the fix reverted the suite aborts on `not ok - a reprovisioned queue generation inherited the retired generation's idle interval and alerted`; with the fix it passes, and its third leg confirms the restarted generation still escalates on a genuine freeze (row=3 idle=2s), so the fix does not merely mute the alarm. - `bash tests/fm-wake-queue.test.sh`: exit 0, 38/38 pass, all five secondmate cases green. - `bin/fm-lint.sh`: clean (ShellCheck 0.11.0, actionlint 1.7.12). CHECKS NOT CAUSED BY THE CODE: the CI and "Require no-mistakes" runs (34152740610, 34152740604) both ended with conclusion `action_required` — workflow approval pending, not a test/build failure. Separately, `bin/fm-test-run.sh --check-coverage` exits 1 in this environment, but I confirmed by stashing my changes that it fails identically on the unmodified base tree (locale-related `comm: input is not in sorted order`); it is pre-existing and this change adds no new test file for the partition to account for. Changes are left uncommitted in the worktree
* no-mistakes(ci): Fixed the one real code defect behind the failing checks. ROOT CAUSE (Greptile P1, second round, on the head commit e1304e6): `secondmate_wake_stall_tick` in bin/fm-watch.sh identified the queue's drain position by the sequence number ALONE. The previous round changed the comparison from `-gt` to `-ne`, which handles a reprovisioned queue that restarts BELOW the recorded position, but not one that restarts ON it. A mate retired and reprovisioned under the same task id gets a fresh home whose wake-queue sequence counter restarts at 1 — and the retained parent progress marker very plausibly holds a low sequence too (a queue frozen on its first row records seq 1). Equal sequence ⇒ no reset ⇒ the brand-new queue inherits the retired generation's long-expired idle interval and emits a false `secondmate wake-loop stalled` on its very first observation. That is exactly the false-alarm class this PR exists to remove. FIX (smallest, removal-first): the file already defines the identity of a queue row once, as `row_key="$epoch-$seq"` (used for stall receipts, the stall marker, and the notify key). The progress marker's separate, weaker seq-only identity is removed: `row_key` is now computed once right after the row is parsed, stored in the progress marker, and compared with `!=`. Across generations the epoch differs (the new generation's rows are appended later), so no sequence collision can carry a stale interval; within a generation the key is stable exactly while the position does not move. bin/fm-wake-lib.sh's `fm_wake_secondmate_progress_marker_write` now takes `<oldest-row-key>` and validates it the same way the two neighbouring row-key writers do. Updated the function header comment and the two doc sentences (docs/architecture.md, docs/configuration.md) that described the old sequence-only semantics. REGRESSION TEST: `test_secondmate_reprovisioned_queue_starts_a_fresh_interval` in tests/fm-wake-queue.test.sh now drives the reported case — the reprovisioned generation restarts on the SAME sequence 9 (epoch 200) that the retired generation recorded (epoch 100), at a later clock — and asserts observable watcher output only. Its third leg still confirms the restarted generation escalates on a genuine freeze (row=9 idle=2s), so the fix does not merely mute the alarm. Three seeded progress markers in the symlink, crash-window and prefix-receipt tests were updated to the epoch-sequence form. VERIFICATION: - Fails before / passes after: with bin/fm-watch.sh and bin/fm-wake-lib.sh reverted to HEAD and the new test in place, the suite aborts on `not ok - a reprovisioned queue generation inherited the retired generation's idle interval and alerted` (exit 1); with the fix, `bash tests/fm-wake-queue.test.sh` exits 0 with 38/38 pass, all six secondmate cases green. - `bin/fm-lint.sh`: clean (ShellCheck 0.11.0, actionlint 1.7.12). - I also started `bin/fm-test-run.sh tests/fm-watch-checkpoint.test.sh tests/fm-watch-triage.test.sh tests/fm-watch-recovery-loop.test.sh` as a blast-radius check; it was still running when this phase had to return, so its result is not included. No other suite references the stall detector or the progress marker (grep over tests/ for `wake-loop stall|SECONDMATE_WAKE_STALL|secondmate-wake-progress` matches only fm-wake-queue.test.sh), and the changed lib function has exactly one caller. CHECKS NOT CAUSED BY THE CODE: the CI and "Require no-mistakes" runs on the head commit (34154091191, 34154091236, 34154091945) all ended with conclusion `action_required` — pending workflow approval, not a test/…
tiago-peixoto
added a commit
to tiago-peixoto/firstmate
that referenced
this pull request
Sep 10, 2026
* fix(bin): scope the worker role contract for ship and scout launches (#3797)
* fix(brief): scope Firstmate workers to their launch contract
* no-mistakes(document): Clarify supervisor scope and worker contract ownership
* no-mistakes(ci): Removed the heading-based bypass so every ship/scout launch receives the current worker-role contract. Added a regression that failed before the fix and passes afterward. Dispatch, brief, and delivery suites, focused ShellCheck, and git diff --check all passed
* no-mistakes(review): make launch overlay sole owner of worker role contract
* no-mistakes(review): narrow heading test dimension and fix publish error wording
* no-mistakes(review): gate role supersession, fix render guard, drop AGENTS twin
* no-mistakes(document): align architecture AGENTS.md scope and spawn launch-brief header
* fix(bin): stop ringing steering doorbells into dead panes (#3823)
* fix(bin): stop ringing steering doorbells into dead panes
The steering-inbox doorbell was a plain sentence plus Enter typed into a
worker's pane, and the watcher re-rang it on the assumption that a ring is
free. In a pane whose agent has exited that line is a shell command, and the
re-ring ladder kept typing it into a shell that can never acknowledge it.
- Prefix the doorbell with the shell no-op `: ` so a bare shell executes
nothing while a live worker still reads the same self-describing line.
`#` is not used because interactive zsh does not treat it as a comment by
default and the claude harness binds it to memory mode.
- fm_task_inbox_ring skips the pane (return 3) when the backend positively
classifies the agent as dead; missing, ambiguous, unreadable, and unverified
endpoints still ring so a blind classifier never starves a live worker.
- The watcher caps the ladder for a dead pane: one stale wake for recovery,
no ring, no ladder walk, and the durable record stays for
stuck-crewmate-recovery. fm-send and the remote steer leg report the skip.
Tests cover the no-op in real shells, the dead/live/unclassifiable ring
verdicts, and the single-surfacing watcher path.
* no-mistakes(review): Quote doorbell paths against shell injection
* no-mistakes(review): Reject terminal-control paths before ringing
* no-mistakes(review): Document accepted partial doorbell delivery race
* no-mistakes(review): Skip unavailable endpoints before busy-state handling
* no-mistakes(test): Respect shell startup PATH in environment allowlist test
* no-mistakes(test): Fix doorbell test fixtures for endpoint liveness
* no-mistakes(test): Prioritize confirmed restarts and clean shell test syntax
* no-mistakes(document): Document dead and missing doorbell recovery
* no-mistakes(ci): Fixed the persistence-reply timeout race by rechecking for a correlated reply immediately before falling back to a nudge. Added a deterministic regression covering replies arriving between the preliminary resolution pass and timeout handling. Verified with the targeted restart suite, project lint, coverage guard, bash syntax checks, and diff checks
* fix(bin): wait out transient primary-checkout reads in the spawn worktree poll (#3834)
* fix(spawn): keep the worktree poll from adopting the repository primary
After `treehouse get` is sent, the worktree-discovery poll reads the pane's
foreground-process cwd. While treehouse is still fetching and checking a slot
out, the foreground process is treehouse itself and it reports the repository's
PRIMARY checkout as its cwd for several seconds. The poll accepted any path
that merely differed from the spawning project, so from a linked spawning home
- whose project is itself a worktree of that repository - it adopted the
primary, and the isolation guard then refused a launch whose slot treehouse
went on to create normally.
Screen every candidate with the isolation guard's own conditions, extracted as
spawn_worktree_isolated, so a read the guard would reject stays a transient the
poll keeps waiting through. The two-consecutive-reads rule and the guard as
final backstop are unchanged; a pane that never reaches an isolated worktree
still fails at the existing 60s deadline, now naming the last path it reported.
The already-settled timing assertion counted whole-spawn wall time against a
5s budget and failed on unmodified HEAD on slower machines; it now counts pane
reads, which is what "one confirming read, not an extra cycle" actually means.
* fix(spawn): say which path the worktree wait rejected, and why
Screening every discovery-poll candidate means a host that never reaches an
isolated worktree spends the whole 60s window before refusing. That wait is
deliberate - separating a transient from a terminal misconfiguration needs
machinery this path does not want - so the refusal explains itself instead:
the isolation check records why a candidate failed, and the deadline names the
last path seen together with that reason. Message and diagnostics only; the
poll's control flow is unchanged.
Two suites asserted the guard's wording on paths the poll now rejects rather
than adopts, so their refusal arrives from the deadline instead: realign
fm-tangle-guard's non-git and subdirectory-of-primary cases (each now also
asserting the stated reason, and the second the metadata absence it was
missing) and the herdr projection e2e's forced non-worktree cwd.
* no-mistakes(review): stub poll sleep in tangle-guard spawn isolation test
* no-mistakes(document): document spawn poll isolation screen in fm-spawn header
* test(spawn): make the non-git isolation case non-git anywhere
The refusal-reason assertion for a path outside any repository assumed TMPDIR
is not inside a git repository. Where it is, git walks up from the temporary
directory, finds that repository, and the spawn reports the subdirectory cause
instead - so the case passed or failed on a property of the host rather than on
the behaviour under test.
Build the path under a directory the test then names in GIT_CEILING_DIRECTORIES,
which git documents as not chdir-ing up into a listed directory while looking
for a repository. Git never excludes the directory being searched, so the
ceiling is the parent of the path handed to the spawn.
The assertions pin which cause fired rather than the sentence that explains it,
leaving the operator wording free to improve.
* no-mistakes(document): point spawn poll comment at the isolation screen's comparison
* no-mistakes(ci): Fixed the "Behavior portable serial 2" failure in tests/fm-tangle-guard.test.sh ("non-worktree spawn did not say why the path was rejected (missing: 'not inside a git worktree')"). Root cause, in this PR's code: bin/fm-spawn.sh's spawn_worktree_isolated resolved the git toplevel with `wt_top_real=$(cd "$SPAWN_WT_TOP" ...)`. For a path in no repository, `git rev-parse --show-toplevel` yields empty, and `cd ""` is a SUCCESSFUL no-op on bash before 5.3 (CI's ubuntu-latest ships bash 5.2). The empty toplevel therefore resolved to fm-spawn's own cwd — the CI checkout — so the poll reported "it is a subdirectory of worktree root '/home/runner/work/firstmate/firstmate'" instead of the correct "it is not inside a git worktree". Dev machines with bash 5.3 fail `cd ""`, which is why the suite passed locally and only failed on CI; it is a genuine shell-portability defect in the reason vocabulary this change added, not a test-environment artifact. Fix (smallest root-cause change, 1 line + comment, bin/fm-spawn.sh:2168-2173): guard the empty value so it never reaches `cd` — if [ -n "$SPAWN_WT_TOP" ] && ! wt_top_real=$(cd "$SPAWN_WT_TOP" 2>/dev/null && pwd -P); then No change to the poll's timing or deadline behavior (respecting the recorded refusal-latency and spawn-wt-reason-vocabulary decisions), no new tests, no other files touched. Verification: - Reproduced the exact CI failure locally by putting bash 3.2 (same `cd ""` semantics as CI's 5.2) first on PATH: fails before the fix with the identical message shape, passes after. - tests/fm-tangle-guard.test.sh passes under both bash 3.2 and bash 5.3. - tests/fm-spawn-worktree-settle.test.sh and tests/fm-spawn-pool-base-freshen.test.sh pass; shellcheck -x bin/fm-spawn.sh clean. - bin/fm-test-run.sh --changed: 46 suites completed, every FM_TEST_END exit=0, 1076 passing assertions, 0 "not ok" (including fm-tangle-guard, fm-control-relaunch, fm-lint). The run ended on my own 900s wall-clock cap (rc=124), not on any test failure
* fix: support stock macOS Bash 3.2 paths (#3732)
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): read orphaned green ci monitor and daemon-down failed record as not failed (#3846)
* fix(bin): read an orphaned green ci monitor as held-for-merge, not failed
A no-mistakes run held for a captain merge decision keeps its ci step
polling until merged or closed; when the shared daemon restarts under
that poll, the run is recorded failed although every substantive step
completed and GitHub reports the PR green. A monitor whose only
remaining job is to observe a human decision must not convert the
absence of that decision into a failure verdict.
fm-crew-state.sh now reclassifies a terminal failed run as done
(held-for-merge), surfacing the run's PR URL, when the steps table
shows every step completed except exactly ci failed and the ci log's
last recognized marker reads checks green. A genuinely red check, an
unreadable ci log, or a second failed step keeps the failure.
* no-mistakes(review): Read daemon-down coarse failed ledger as unknown, not failed
* no-mistakes(document): docs: align AGENTS.md failed-verdict guidance with crew-state reclassification
* fix(spawn): verify preserved backlog state after interrupted spawn delivery (#3852)
* fix(bin): read preserved spawn state back before the interrupted exit claims it
The deferred-signal exit path asserted the paired task record and
In-flight backlog state were preserved without reading either back,
exactly when a reader is least able to check (fm-yi4j evidence,
2026-09-05). The commit's exit status alone has been observed to agree
with a row that did not actually move.
The exit path now re-reads the record and the row under the same
per-task lock as the commit, repairs a row the commit believed it
moved, and phrases the error as exactly what was verified or attempted
- verified preserved, repaired and verified, or an explicit
preservation-could-not-be-verified with the reason and hand-closeout
instruction. Two behavior tests drive a lying tasks-axi start through a
real interrupted spawn and assert the printed claim and the real
backlog state agree.
* no-mistakes(test): Fix calm suite for Pi 0.85 and pin test umask
* no-mistakes(document): Document interrupted-spawn preservation claim in backlog gate owner
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-spawn.sh's deferred-signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-backlog-transition-lib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean
* no-mistakes(ci): Fixed the Greptile P1 on bin/fm-backlog-transition-lib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean
* docs: correct runtime-backend maturity labels for Herdr (#3821)
* docs: correct stale tmux/herdr backend maturity claims
Herdr now has 21 test files, its own required CI job (tests-herdr)
that installs a pinned build and hard-fails on "skip: herdr not
found", while tmux has 3 test files and is only required as a
dependency of the portable-serial e2e lane. zellij, orca, and cmux
still have no CI lane at all. AGENTS.md and docs/herdr-backend.md
still called Herdr merely "experimental" alongside those three,
misleading every session and reader about actual coverage.
Update AGENTS.md's config/backend entry, the opening lines of
docs/herdr-backend.md and docs/tmux-backend.md, the runtime-backend
section of docs/configuration.md, and the matching claims in
docs/architecture.md, CONTRIBUTING.md, and README.md so they agree
and distinguish tmux (default), herdr (own required CI lane, largest
suite, Windows still spike-only), and zellij/orca/cmux (still
experimental, no CI lane). No behavior, selection order, or
dispatch logic changes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GDYsWPEfwTuPNjQ2nGBCcj
* no-mistakes(review): docs: fix stale herdr label and CI-lane wording
* no-mistakes(document): docs: align tmux adapter label in scripts.md
* no-mistakes(review): docs: drop duplicated herdr CI claim from tmux page
* no-mistakes(review): docs: drop windows claim, align contributing backend wording
* no-mistakes(review): docs: trim duplicated CI claim from herdr opening line
* no-mistakes(review): docs: drop unguarded largest-test-suite superlative
* no-mistakes(review): docs: restore tmux verified label and README experimental scope
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: restore intent-targeted no-mistakes validation (#3865)
* fix: delete deterministic no-mistakes test baseline, restore intent-targeted Test
PR #3644 pinned commands.test to a fm-test-run.sh --changed walk of the
repository's 75-162 tests/*.test.sh scripts. no-mistakes runs commands.test
verbatim and unconditionally after every fix round, so that walk multiplied
by round count: measured at 32.7 minutes per validation versus 3.6 minutes
intent-targeted. Delete the pin and restore the 3.6-minute posture.
Add tests/fm-nm-test-contract.test.sh as a regression guard, parsing
.no-mistakes.yaml as YAML (ruby's bundled Psych, matching the parser
tests/fm-test-run.test.sh already uses for ci.yml) rather than grepping its
text, restoring in legal form what PR #823 added and PR #1282 removed.
Record the rule in docs/configuration.md's "Gate defaults" section (the
authoritative owner CONTRIBUTING.md already points at) and strengthen
CONTRIBUTING.md's existing local-Test guidance to state it plainly: never
configure commands.test to a deterministic test command, complete or partial.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSt9JvrQMVc4u3jPyUFCFC
* no-mistakes(review): Centralize no-mistakes test policy and narrow guard
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: verify Treehouse slot ownership before teardown (#3837)
* fix(bin): verify pool-slot ownership before returning a worktree slot
Workers were killed when cleanup returned a Treehouse pool slot that a
different, live task had already taken. Teardown now proves the slot is
genuinely this task's before releasing it: it refuses when another task
record claims the same live worktree path, or when the endpoint's working
directory contradicts the recorded slot, and that refusal holds under
--force. Slot allocation, metadata publication, ownership verification,
and slot return are serialized across linked firstmate homes, and forced
secondmate cleanup verifies descendant slot ownership before returning
any child worktree.
Regression coverage drives the scripts with two task records naming one
slot path and asserts the live worker survives and its slot is not reset.
* no-mistakes(review): Protect slots across cloned Firstmate homes
* no-mistakes(test): Gate teardown locking on genuine Treehouse slots
* no-mistakes(test): Clarify pooled descendant slot gating
* no-mistakes(test): Synchronize watcher re-arm test on process exit
* no-mistakes(test): Wait for watcher cleanup before timeout escalation
* no-mistakes(document): Document pool-slot ownership safeguards
* no-mistakes(ci): Fixed all reported CI issues: normalized bare local Git origins to the same Treehouse project-lock identity as absolute clone origins; resolved ShellCheck SC1091 with explicit conditional sourcing; and taught concurrent Herdr teardown coverage to retry expected Treehouse lock contention. Added behavioral regression coverage for bare/absolute origin lock identity. Verified endpoint-safety tests, watcher tests, full CI lint, and the previously failing Herdr teardown assertion
* fix(bin): resolve relative origins from repository root
* no-mistakes(ci): Fixed teardown so an exact recorded endpoint may change cwd without falsely vetoing cleanup. Removed cwd-based ownership refusal while preserving cross-home record exclusivity and project locking. Updated behavioral coverage for both foreign slot ownership refusal and moved-cwd teardown success. Endpoint-safety, backend, watcher, checkpoint, and targeted lint checks pass. Real Herdr presentation E2E progressed successfully but exceeded the 600s local timeout
* feat(bin): add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary (#3867)
* feat: add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary
Add omp as a verified harness: anchored process-name detection with a
Firstmate-owned FM_OMP_HARNESS launch marker that needs real omp ancestry,
the fm-spawn launch template with foreign-marker clearing, the tracked
.omp/fm-worker-overlay.yml posture overlay, --auto-approve, --cwd, and
pre-launch model validation scoped to providers 'omp models --json' lists.
Workers get a state-resident busy-state extension keyed on agent_end
without willContinue (omp has no agent_settled). The primary gets two
tracked .omp/extensions: a turn-end guard that answers omp's blocking
session_stop hook by compelling one continuation per turn, with the
pre-tool seatbelts and Run-tier session-start delivery, and a watcher
extension ported from the Pi one with fm_watch_arm_omp. Control tables,
composer busy footers, omp's status row as a bare-composer boundary, the
extension supervision model with an omp-keyed ownership proof, the
session-start diagnostic, and the supervision protocol snippet follow.
Verified live on omp 18.1.11 with openai-codex/gpt-6-astra: a Herdr scout
through spawn, busy state, steer, interrupt, exit, and teardown, and the
isolated rpc primary lab through extension auto-discovery, digest
delivery, lock identity, watcher arm, successor and wake delivery, and
the compelled guard continuation.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test: prove the omp guard continuation through a guard spy
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* fix(spawn): clear the gemini marker at the omp launch boundary
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): force the guard stage by freezing the watcher and clear lint findings
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): reap the live lab by path and record omp's rpc shutdown as a note
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): spawn a real secondmate for the discovery rule and classify the omp surfaces
Replace the template-extraction check with a genuine --secondmate launch
pinned to the fake tmux backend, assert the worker extension's handler set
through the executable rather than its bytes, classify the two new omp
surfaces in the documentation inventory, and record the Herdr worker
evidence in the runtime-backends verification doc.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* no-mistakes(review): omp: unverify remote routes, narrow busy regex, drop overlay approval pin
* no-mistakes(review): omp: validate config-pinned model, correct remote and marker docs
* no-mistakes(review): omp: pin config-model validation with a test, trim overlay
* no-mistakes(review): omp: sync guard evidence, drop dead param, map quota family
* no-mistakes(review): omp quota: refuse unmapped prefixes, match bare model scopes
* no-mistakes(document): docs: cover omp in cd-guard, quota, continuity, tmux
* no-mistakes(document): docs: add omp subagent-guard row, fix live test header
* no-mistakes(ci): Fixed both failing behavior shards and the Greptile P1 in bin/fm-composer-lib.sh. Root cause of "Behavior portable serial 1" and "Behavior portable parallel 2": the omp busy regex (FM_DELIVERY_OMP_BUSY_REGEX_DEFAULT) and omp status-row furniture regex (FM_COMPOSER_OMP_STATUS_RE_DEFAULT) used the bracket range [⠁-⣿]; BSD grep on macOS accepts it but GNU grep on Linux CI aborts with "Invalid collation character", failing every omp busy/furniture read (3 assertions across fm-omp-harness, fm-tmux-submit-busy, fm-composer-lib). Replaced the range with one shared explicit alternation FM_OMP_SPINNER_FRAMES_RE of omp 18.1.11's unicode-preset spinner frames (status set ⣾⣽⣻⢿⡿⣟⣯⣷ + activity set ⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏, read from the installed binary), the same pattern the Kimi busy regex already uses in CI. For Greptile's finding (the harness-agnostic furniture rule's first alternative matched any 1–4-byte token + ' · ', so wrapped typed input like 'fix · tests' with the cursor on it regressed from pending to unknown; reproduced locally vs base), pinned that alternative to omp's identity cell (π||pi, the icon.omp of each preset in the 18.1.11 binary). Tests: fm-composer-lib.test.sh asserts 'fix · tests' is not furniture, a status-set spinner row is furniture, and the wrapped composer screen reads pending under both locales (CAPS_TMUX cursor 3); fm-omp-harness.test.sh asserts a status-set frame reads busy. New negative cases fail against the pre-fix lib and pass after. Verified: fm-omp-harness, fm-tmux-submit-busy pass via bin/fm-test-run.sh; fm-composer-lib passes all cases except one pre-existing, unrelated local failure (Herdr half-block test uses printf '▀', unsupported by macOS bash 3.2; fails identically on a pristine HEAD export, passes on CI bash 5); shellcheck and bin/fm-lint.sh clean. Caveat: GNU grep is unavailable locally, so the Linux compile was not run directly; the fix uses only constructs already proven on CI's GNU grep (multibyte literal alternations, incl. under LC_ALL=C). Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh, tests/fm-omp-harness.test.sh. No docs needed changes (they describe the rule generically)
* no-mistakes(ci): Greptile Review: fixed. The omp status-row furniture regex FM_COMPOSER_OMP_STATUS_RE_DEFAULT in bin/fm-composer-lib.sh still accepted a literal `pi ·` opening, so wrapped composer input beginning with `pi ·` was truncated and misclassified. Read the installed omp 18.1.11 binary: the ascii preset's `icon.omp` is `pi` but its `sep.dot` separator is ` - ` (unicode/nerd use ` · `), so a real ascii status row never contains `pi ·` and that alternative could only ever match typed text. Removal-first fix: dropped `pi` from the identity alternation (now `(π|)`) and updated the comment to record why the ascii preset is excluded. Tests (tests/fm-composer-lib.test.sh): added a negative furniture case for 'pi · e · phi as the three constants' and a wrapped-screen assertion (CAPS_TMUX, cursor 3) that a continuation row opening `pi ·` reads pending in both locales; the new case fails against the unfixed lib and passes after. Verified: composer test with the half-block case skipped passes all 33 cases including the omp matrix; bin/fm-test-run.sh tests/fm-omp-harness.test.sh passes; shellcheck -x clean on both files; bin/fm-lint.sh clean. The full composer test via the runner fails locally only on the pre-existing half-block case (bash 3.2 printf cannot emit ▀; passes on CI bash 5), identical to before this change. Docs unchanged (they describe the rule generically and never mention the ascii identity cell). PR must be raised via no-mistakes: not caused by code. attestation.head_sha is cdddc60 while the PR head is cd51cf4 because the pipeline's ci-phase push moved the head; the outer executor's re-push will re-bind the attestation. No file change for that check. Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* fix(pi): invoke Bash helpers correctly on native Windows (#3843)
* Fix Pi shell invocation on native Windows
* no-mistakes(document): Document Pi Windows Bash transport
* no-mistakes(ci): Captain, staged a narrow fix: register the Pi Windows regression for both extension paths, make Windows mode emulation non-failing, and enforce LF shell checkouts. Mapping and coverage checks pass; CI/Require no-mistakes were approval-gated externally
* no-mistakes(review): Cover async Windows branch-outcome Bash invocation
* no-mistakes(review): Preserve Cygwin checks and refresh Windows timing
* no-mistakes(test): Invoke OpenCode operational-input owner through Bash on Windows
* validation-fixture
* no-mistakes(document): Document Windows Bash helper invocation
* no-mistakes(ci): Fixed PR-caused changed-selection failure by removing the malformed tracked evidence artifact and allowing deleted, unconsumed source paths to retire cleanly while preserving fail-closed behavior for live unmapped paths. Added regression coverage. Verified native-Windows Pi shell-seam test passes and --changed selects the Windows regression
---------
Co-authored-by: test <test@example.invalid>
* test(bin): pin teardown outcomes for squash-merged rebased branches (#3870)
* fix(bin): recognise squash-merged rebased work as landed at teardown
A pipeline rebase can leave the local worktree on pre-rebase commits while
GitHub squash-merges the rebased head. The landed-work test then compared
those stale commits against a squashed main and refused cleanup of work
that had already landed.
When the forge reports the recorded PR merged and its merge commit is on
the default branch, treat a local branch that only repeats paths from the
pipeline push as stale rather than unlanded. If the forge is unreachable,
the same coverage check runs against a PR head whose content is already
on default. Extra local paths still refuse.
* fix(bin): drop unprovable squash-rebase landed-work coverage
Path-set coverage treated a diverged local branch as landed whenever it
touched the same files as the squash merge. That accepts the reviewer's
failing sequence: same path, different content, work discarded.
git cherry and merge-tree containment were already too strict on the real
rebase-fold case. No remaining check is both safe and permissive enough
to recognise a stale pre-rebase copy without also accepting unlanded
edits, so that case still refuses.
Keep the proofs that hold: a merged PR head that contains local work, or
a clean content-in-default tree match. Tests now refuse same-path
different content and extra unlanded commits, and still allow a local
branch that followed the pipeline rebase.
* no-mistakes(review): drop recorded-pr-head fallback and reverted-design leftovers
* no-mistakes(review): silence squash-merge stdout corrupting test PR head
* no-mistakes(review): make unlanded follow-up commit sole cause of refusal
* no-mistakes(document): correct stale squash-rebase fixture comments in teardown tests
* no-mistakes(ci): Split the three reported checks: - CI (run 34061098467) and Require no-mistakes (run 34061098460) both concluded `action_required` — approval-gated workflow runs that never executed a step. Not caused by this PR's code; no change can clear them. - Greptile Review was a genuine defect in the new tests: the three new refusal cases (tests/fm-teardown.test.sh) asserted only exit status 1 and a REFUSED line, so a teardown regression that destroyed the worktree, branch, and task record before reporting refusal would still pass. Fix (tests only): added one `assert_refusal_retained_task_state` helper and called it from `test_squash_merged_same_file_different_content_refuses`, `test_squash_merged_rebased_local_with_unlanded_commit_refuses`, and `test_squash_merged_stale_local_refuses_when_forge_unreachable`, each capturing the worktree HEAD before `run_teardown`. It pins that the refusal left the isolated copy on disk, the task branch still checked out at the same unlanded commit, and state/task-x1.meta intact. Verification: the four squash tests pass; a sensitivity probe ran the ALLOW fixture (teardown completes) and pointed the same helper at the outcome — it fires, because a completed teardown detaches/deletes the branch and removes the task record, proving the assertions discriminate. Full tests/fm-teardown.test.sh: 83 passing. bin/fm-lint.sh clean with pinned ShellCheck 0.11.0 + actionlint 1.7.12 (plus an explicit --external-sources pass on the changed file). bin/fm-test-run.sh --check-coverage ok. Caveat: test_herdr_flat_teardown_preflight_refuses_before_changes (mode missing-adapter) fails on this machine. Verified it fails identically on base commit f91a950 via `git archive`, so it is a pre-existing local environment difference untouched by this diff; skipped to run the rest of the suite, not modified
---------
Co-authored-by: Morten Gad <mogad@itm8.com>
* fix(bin): keep supervision armed for registered custom checks (#3860)
* fix(bin): keep supervision armed for registered custom checks
A custom check bound by bin/fm-check-register.sh only ever runs inside the
watcher's check sweep, but fm_supervision_status counted in-flight tasks, the
relay poll shim, and process-event sources as supervision need, and not
registered checks. Tearing down the last task therefore stopped every
home-level check silently until the next spawn.
Count a state/<id>.check.sh that carries its state/<id>.check-trust binding as
supervision need. The relay shim keeps its own trust path and task PR polls
carry no such binding and are torn down with their task, so neither arms a home
by accident. Presence of the binding is the whole test: the sweep validates the
bytes at execution time and wakes firstmate when it rejects one, which is the
outcome an idle home needs.
Closes #3856
* no-mistakes(review): name registered checks in turn-end block banner and doc invariant
* no-mistakes(review): narrow PR poll predicate test to what it proves
* no-mistakes(document): point Grok re-arm step at supervision-need owner
* fix(bin): resolve Treehouse locks for remote secondmate homes (#3883)
* fix(bin): resolve the shared Treehouse project lock inside remote secondmate homes
Every spawn and teardown inside a remote-seeded secondmate home refused,
because the project lock's anchor could not be resolved there.
fm_firstmate_root_home walks a home's parent bindings upward to find the
anchor the lock lives in, and treated a remote parent binding as an error.
A remote-seeded home's parent is on another machine, so that walk can never
succeed from there - and neither can the home's own local descendants, whose
chain terminates at the same record. Both fail closed on every Treehouse-backed
spawn and every pool-slot teardown.
A remote parent now terminates the walk at the home holding it, which is the
correct anchor: a lock taken on this filesystem is neither held nor observable
across that boundary, and that home is already the top of the local tree
teardown's collect_local_firstmate_states enumerates, since that walk skips
remote registry entries for the same reason. Mutual exclusion is unchanged -
every home reachable through local parent links still derives one identical
lock file per project, and an unreadable binding, an unsupported route, an
unreachable local parent, a cycle, and an over-deep chain all still refuse.
Origin-less local-only projects keep resolving through their worktree top.
Regression coverage pins the anchor for the main-home layout, a local
secondmate, a remote-seeded home, and its local child; drives teardown
end-to-end in a remote-seeded home; keeps the cross-home slot-ownership
refusal across that boundary; and proves two homes still serialize on the
one shared lock file.
* no-mistakes(document): Clarify machine-local Treehouse lock ownership
* fix(bearings): repair board listening and decision reconciliation (#3872)
* fix(bearings): repair the board's listening, card hygiene, and reconcile path
Three defects made the fleet board go quiet and then lie about what still
needs the captain.
Never arm a poll on a session that is not live. `lavish-axi <file>` exits 0
even when it refuses to reopen a session the captain ended from the browser,
reporting `status: user-ended` with the same session id, so the build's
exit-status check accepted a dead session, printed `already-armed`, and left
the board reading "not listening". The build now proves the session is live
from a fresh authoritative listing immediately before arming - not from the
establish call's status alone, which is already stale by then - reopens once
when it finds the session ended, and refuses rather than arming when it stays
ended. A reopen also replaces the pre-reopen source generation before
reporting success, so a runner on its way out cannot be mistaken for a
listener, and a board whose source is registered but unowned gets a
replacement started before the build returns.
Let a dead generation's ownership actually move. Reclaiming a claim ran its
capture-reservation cleanup first, and that cleanup re-verifies the recorded
state-root identity, so a claim naming a pid and a process group that were
both provably gone could not be cleared: reconcile reported a start while
nothing attached, and retire refused with "cannot release source ownership".
Reservation records are keyed by claim token and every replacement claims a
fresh one, so they are hygiene, not an ownership invariant. Reclamation now
additionally requires the owning process group to be absent independently,
which keeps a reused pid whose poll child still runs from ever reading as a
gone generation. A live owner and a crashed leader whose owned group survives
are still never reclaimed.
Stop carding decisions whose subject already landed. The build drops a
decision card whose work item or PR appears in the payload's own landed rows,
and one whose task is no longer an open captain call, naming each drop on
stderr. A task whose state cannot be established is kept, because a call
wrongly hidden is worse than a card wrongly shown.
Add the reconcile choice, and make it structurally incapable of closing a
call. Every decision card carries a standard `reconcile` option, injected by
the build rather than left to the composer. The board now emits the picked
option and any freeform note as separate structured fields instead of fusing
them, so a reconcile selection is not expressible as an answer value at all -
the defect that let `reconcile - <note>` reach the intake as an ordinary
answer. The adapter routes selections from that structured field, creation of
a reconcile request is bound to a verified board source rather than the shared
keyed-answer intake, and the intake still refuses the reserved value on every
channel. Each authorization is bound to the captain-hold generation that
produced the card, so an obsolete card cannot close a later call, and both
terminal outcomes require a pending request: `reconcile close` records the
evidence under its own `reconciled` mode so it never reads as the captain's
words, and `reconcile note` leaves the call open. Anything unprovable -
an unversioned row, a missing generation, an unreadable state - refuses
rather than acting.
Regression coverage fails without each fix, and pins every leak path: a bare
reconcile, a standalone close or note with no pending request, an any-channel
reconcile, an annotated selection from a freeform card, and a
generation-skewed authorization. An opt-in guard re-proves the lavish-axi
shapes and the reopen against the installed tool.
* fix(bin): quote the done comparison in the reconcile intake
shellcheck SC1010 reads the bare word as the loop keyword. The failed run
never reached its lint step, so this shipped in the recovered content.
* no-mistakes(review): Publish reconciled parent resolution before request retirement
* no-mistakes(review): Clarify committed cleanup and reconcile reservation scope
* no-mistakes(review): Preserve remote cards and legacy answer compatibility
* no-mistakes(test): Separate live claim release from stale reclamation
* no-mistakes(test): Allow terminal self-retirement during active capture
* no-mistakes(document): Document Bearings repair contracts
* no-mistakes(ci): Stabilized the failing Herdr presentation E2E by serializing test-harness Treehouse allocator calls, preventing concurrent recovery spawns from claiming the same pool slot while preserving Herdr concurrency coverage. Verified with the full E2E suite on Herdr 0.8.2, bash syntax checks, ShellCheck, and git diff checks
* fix(bin): allow pooled spawns without a git origin (#3885)
* fix(bin): skip pooled-worktree freshness fetch when no origin is configured
An origin-less local-only project has nothing remote to be stale
against, so fm-spawn's freshen_spawn_worktree_base refused to launch
crews for it. Detect a missing origin remote and skip the fetch
freshness gate entirely; an existing-but-unreachable origin keeps
refusing as before.
* no-mistakes(review): Preserve pool safety for absent and unusable origins
* no-mistakes(review): Refuse empty origin configurations during pooled spawn
* no-mistakes(review): Detect empty origin sections across config includes
* no-mistakes(review): Honor globbed includes when detecting origin configuration
* no-mistakes(review): Document conservative conditional include handling
* no-mistakes(review): Use Git-resolved config files for origin detection
* no-mistakes(review): Document included empty-origin detection boundary
* no-mistakes(document): Document originless pooled spawn behavior
* feat(tests): run live harness guards by default when available (#3889)
* feat(tests): run live harness guards by default where the harness is installed
The 24 live-harness guards each opened with their own env check, so on the
machine that has every harness - the one the product and its validation
actually run on - all of them skipped and passed. Fourteen had never been run
by the pipeline at all.
tests/lib.sh gains fm_live_gate as the single owner of that decision: a guard
that spends no model tokens runs wherever its tools are installed, a guard that
submits prompts stays opt-in, an absent tool is a named capability skip, and a
guard's own variable or FM_LIVE forces it on (turning an absent tool into a
failure) or off. Every live guard now opens with it, which also carries the
test-suite gate-refusal bypass into the guards that never sourced the shared
helpers and were therefore refused whenever a gate agent ran them.
bin/fm-test-run.sh records what a skip means: the family's expected class is
live-capability rather than a bare env opt-in, and each gate skip's reason is
logged and written to the timing artifact, so a lane can say which tool this
host could not exercise.
Only the token-free guards flip to default-on: composer-matrix, the harness
liveness drift guard, and the Herdr version floor. cursor-primary submits three
prompts, so it stays opt-in.
Running the drift guard unasked immediately found a real defect it existed to
catch: it resolved the harness through a generic `command -v cursor`, which on
a machine that also has the Cursor editor finds the editor launcher rather than
cursor-agent. That binary exits at once, leaving a bare shell in the pane and a
liveness-drift failure no classifier change could fix. It now asks
fm_cursor_resolve_binary first, the same verified owner fm-spawn uses.
CI installs the public Pi package in the portable serial lane and fails on its
skip token, so the Pi extension tests stop passing silently against a package
that is not there. No secret is added.
Verified on macOS 26.5.2 arm64: the drift guard runs with no variable set and
classifies 8 installed harnesses alive; the Herdr version-floor guard runs by
default and checks 4 real releases; every live guard refuses together under
FM_LIVE=0.
* fix(tests): keep the composer-matrix guard opt-in
Running it unasked is red on a healthy machine for reasons no code change here
removes: a harness that has not trusted this checkout sits on its own trust
dialog, which the guard treats as an unreadable composer and correctly fails.
The opencode 1.18.29 and grok 1.0.13 composer drift it also surfaced reproduces
identically on main and is filed as separate work.
So this token-free guard stays opt-in with the reason stated in its header, and
the coding guidelines record the narrow exception: a guard whose verdict
depends on host state that installing its tools does not establish may stay
opt-in, because one that is permanently red is one the fleet learns to ignore.
The other two token-free guards keep running by default.
* no-mistakes(review): Wire bearings guard and remove composer exception policy
* no-mistakes(review): Run Pi responsiveness guard by default
* no-mistakes(review): Gate AFK Pi Herdr through authoritative family sweep
* no-mistakes(review): Sanitize live gate test environments
* no-mistakes(document): Document default-on live guard behavior
* no-mistakes(ci): Fixed CI by installing the Pi package in portable-parallel-1, where fm-pi-primary-types.test.sh runs, and enforcing its package-missing gate skip there. Verified with fm-lint.sh, workflow actionlint, coverage partition checks, lane membership, and git diff checks
* no-mistakes(ci): Fixed CI’s Pi typecheck skip enforcement by giving npm, tsc, and Pi-package capability skips a shared prefix and configuring both relevant CI lanes to fail on that prefix. Verified missing tsc emits the expected skip, missing Pi package becomes a runner failure, and actionlint, ShellCheck, and git diff checks pass
* fix(bin): refuse test runs in the primary checkout when a task marker is set (#3891)
* fix(bin): refuse the behavior suite in the repository primary checkout
A task worker's isolated worktree placement is verified exactly once, when
its task starts, and nothing re-checks it afterwards. A worker that later
changes directory into the repository's primary checkout runs its Git
commands, and this branch-switching suite, against the one checkout every
linked worktree resolves against and every landing merges into. A run that
dies mid-suite can leave that checkout on a stray branch.
bin/fm-test-run.sh now refuses that case. When FM_TASK_ID marks a task
worker and the runner resolves to the primary checkout, every executing mode
exits non-zero before selecting a suite, with one line naming the primary
path and pointing at the assigned task worktree. The predicate is the one
bin/fm-spawn.sh already uses for launch placement: the working tree's own
git dir is the repository's common git dir, which separates the primary from
every linked worktree even when their top levels differ. A run with no
FM_TASK_ID set is unchanged, and so are the inspection modes, which execute
nothing. When git resolves neither directory - a non-repository fixture, a
detached copy - nothing proves this is the primary, so the run proceeds.
bin/fm-spawn.sh sets the marker: ship and scout launches export FM_TASK_ID
into the pane shell on the same pre-launch channel as GOTMPDIR, and the name
joins the sanitized launch environment allowlist so an isolated launch keeps
it.
* no-mistakes(review): clear inherited task marker in test lib; name resolved ROOT
* no-mistakes(document): docs: record FM_TASK_ID marker and runner placement refusal
---------
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): bound stale alarms for backlog captain holds (#3842)
* fix(bin): bound a stale alarm with the backlog hold, not only the status line
A legitimate wait has two records and the stale alarm reads only one.
`status_is_paused_or_captain_held` takes a status line, so it sees a wait the
worker declared. It cannot see the wait firstmate records when it hands work to
the captain: `bin/fm-captain-hold.sh hold` writes that into the backlog and
leaves the status log alone, so a delivered task keeps `done: PR ...` as its
last line for the whole time the captain is deciding.
Both stale branches were blind to it, and each churned a new pane hash back into
its own alarm: a `done:` line is captain-relevant and reaches the terminal-stale
branch, while a held task whose last line is `working:` reaches
`surface_nonterminal_stale` and fails its declared-wait test.
Consult that second record where the watcher is about to alarm, through
`bin/fm-captain-hold.sh open`, which already owns the predicate's semantics, and
bound the alarm on the shared `.paused-resurfaced-<key>` marker and
`PAUSE_RESURFACE_SECS` window the declared-wait absorb already uses. The first
sight still alarms, the window's end alarms once more, and a held crew that goes
genuinely silent still escalates through the wedge timer.
Only an established open captain call bounds anything: an unreadable backlog, an
absent or incompatible tasks-axi, a row this home does not carry, and every task
with no hold keep alarming exactly as before. The backlog hold is deliberately
not recorded as a declared pause, because the loop-top reconciliation and
`pause_state_class` both read the status line and would clear a flag that line
does not support.
Extends the fix in #3443, which closed the forms of this loop that the status
line itself can express.
* fix(bin): identify the captain call a stale alarm is bounded by
Three gaps in the bound added by the previous commit, all in how the throttle is
scoped and where the backlog is consulted.
The scope carried only the status-log signature. A task can be held, answered
with `--release`, and re-held as a genuinely different captain call without any
status append, so the second call inherited the first one's marker and its first
sight was absorbed - the one thing this bound must never do. The task id is not
the call: `bin/fm-captain-hold.sh open` gains `--identity`, which reports the
call's own lifecycle - its hold-set stamp and the number of recorded answers -
on an exit 0 and only then, leaving the silent predicate every existing caller
reads unchanged. The throttle scope now carries that identity.
The terminal path recorded the throttle before publishing the durable wake. A
failed append exits the watcher with nothing queued, and the next sighting then
read that fresh marker and absorbed the retry, turning a delayed alarm into a
lost one. Recording moves behind the append, as the non-terminal path already
had it, and the comment claiming the marker could not outlive its wake is gone
because it was false.
The backlog was consulted only on a new terminal pane hash. A captain call can
open after a hash was absorbed as provably working, changing neither the pane nor
the status log, so nothing re-read the backlog and the wedge timer kept firing
possible-wedge alarms through a legitimate wait. That timer now consults the call
at its own alarm boundary and takes the same bounded cadence - and only at that
boundary, so an ordinary repeat poll under the bound stays the local-only read it
was.
Regression coverage for each, all driving churn through one watcher process
rather than relaunching per pane change: relaunch cost dominated the earlier
shape, and an absorbing watcher stays in its poll loop across churn in production
anyway. An unheld task still alarms on every new hash, and an elapsed wedge timer
with no open captain call still escalates as a possible wedge.
* fix(review): Compose stale throttles with captain-call lifecycle identity
* fix(review): Preserve bounded same-hash captain-call resurfacing
* revert(bin): narrow the captain-hold stale bound to its observed defect
Lifts the lifecycle-identity and cadence-ownership work back out, leaving the
change at the shape that matches the defect actually observed: the stale alarm
did not consult the backlog captain hold, on either stale branch.
Reviewing the wider version surfaced a series of adjacent gaps in the watcher's
alarm state machine - a call opening after the first alarm, marker invalidation
at the hold lifecycle boundary, and which deadline a terminal timer represents.
They are real, but fixing them turns a small extension into a state-machine
change to the alarm path, which is a different review on a subsystem that is
being actively reworked. They are named as known limitations rather than carried
here, and none of them is load-bearing for what remains: the bound does strictly
less than the reverted version, leaves the wedge path escalating on
STALE_ESCALATE_SECS exactly as before, and introduces no silence that the
existing terminal-alarm path did not already have.
Kept from the reverted work is the record-after-append ordering, because that is
a defect in the code being shipped rather than an adjacent one: recording the
cadence marker before publishing the durable wake let a failed append lose an
alarm outright instead of delaying it.
History is preserved: the earlier commits stay on the branch and this removal
sits on top of them.
* fix(review): Document secondmate captain-hold scope boundary
* fix(document): Document captain-hold stale alarm scope
* fix(bin): bind the stale throttle to the captain call, not the status log
The throttle this change introduces was scoped to the task's status-log
signature. Answering a call with `--release` and holding the task again creates a
genuinely different captain call without necessarily appending to that log, so
the second call inherited the first one's marker and its first sight was
absorbed.
That is the one alarm this bound must never swallow. A delivery announced twice
is noise; a decision waiting on the captain that is never surfaced is invisible,
because nobody asks for what they do not know to ask for.
Measured rather than assumed, on the same fixture - a delivered task held for the
captain, released, and re-held with no status append, driven through bin/fm-watch.sh:
base c499f84 call-1 first=ALARM call-1 churn=ALARM new call first sight=ALARM
before this fix call-1 first=ALARM call-1 churn=absorbed new call first sight=absorbed
after call-1 first=ALARM call-1 churn=absorbed new call first sight=ALARM
Base never suppresses the new call, so the suppression came from this change and
closing it completes the fix rather than widening it.
`bin/fm-captain-hold.sh open` gains `--identity`, printing the call's lifecycle -
its hold-set stamp and count of recorded answers - on an exit 0 and only then, so
the silent predicate bin/fm-teardown.sh reads is untouched. The throttle scope
carries that identity beside the status signature.
The sibling case was measured too and is NOT included: on the status-declared
path, where the last line is `captain-held:`, base already absorbs a re-held
call's first sight. That behaviour predates this change and stays documented as a
known limitation rather than repaired here.
* fix(document): Document captain-call throttle lifecycle scope
* fix(ci): isolate the Herdr restart fixtures from a claimed worktree
The Herdr behaviour test intermittently reused a local worktree still claimed
by an earlier fixture after a restart.
The restart scenarios now use an isolated Treehouse project.
The full Herdr test passes on Herdr 0.8.2; bash -n and git diff --check pass as
well.
* feat(pi): resolve extension-registered providers in the supervision branch (#3871)
* Let the supervision branch resolve extension-registered providers
The isolated branch ModelRuntime cannot see providers an extension
registered into main's runtime at run time, so a pin on pi-devin-auth's
devin/swe-1-7 (or an unpinned branch following a main session on devin)
failed with "unavailable to the isolated branch runtime".
Capture main's ModelRegistry alongside mainModel and copy each
extension-registered provider config into the branch runtime at
model-resolution time. The config carries the provider's own streamSimple
and oauth wiring by reference, so the custom gRPC transport reaches the
branch unchanged instead of being reimplemented. The /supervision-model
picker uses the same copy so those models are offered.
Update configuration.md and pi-supervision-branch.md, which previously
stated extension-registered providers were not offered.
* no-mistakes(document): docs: own devin provider carve-out in branch architecture doc
* no-mistakes(ci): Fixed the Greptile P1 finding: the /supervision-model picker copied extension-registered providers into the branch ModelRuntime but checked hasConfiguredAuth without refreshing them, so providers with provisional post-registration auth were omitted from the picker while the pin-resolution path (which did refresh) accepted them. Root-cause fix in .pi/extensions/fm-branch-supervision.ts: moved the `refresh({ providers, allowNetwork: false })` call into `copyExtensionProviders` (now async, refreshing every provider it copied) and removed the duplicate per-provider refresh from `resolveBranchModel`. Both the picker and the resolution path now share one copy-and-refresh step, so hasConfiguredAuth is real in both. Regression coverage in tests/fm-pi-branch-extension.test.sh: the stubbed ModelRuntime now mirrors the real runtime by leaving a registered provider's auth pending until `refresh()` runs for it. With that stub, the existing extension-registered-provider case fails against the pre-fix extension (picker offers only anthropic/main-model) and passes with the fix. Verification: tests/fm-pi-branch-extension.test.sh passes (42 ok, no failures); tests/fm-branch-supervision.test.sh passes; tests/fm-pi-primary-types.test.sh skips locally because tsc is not installed (the refresh signature reused is the one the existing code already called). Intent constraints preserved: isolation flags untouched, carve-out still scoped to provider registration, graceful fallthrough when no providers are registered
* ci: retrigger flaky Herdr/serial-1 lanes
* no-mistakes(document): docs already cover branch extension-provider copy
* fix(bin): gate secondmate wake-loop stall alerts on real queue no-progress (#3943)
* fix(watch): detect stalled secondmate queue progress
* no-mistakes(review): gate secondmate stall on active turns and progress episodes
* no-mistakes(document): align secondmate wake-stall docs with progress-episode detector
* no-mistakes(document): clarify active-turn gate in wake-stall config docs
* no-mistakes(ci): Fixed the one real defect behind the failing checks. ROOT CAUSE (Greptile P1, real code defect in this PR): `secondmate_wake_stall_tick` in bin/fm-watch.sh reset the no-progress timer only when the oldest actionable queue sequence INCREASED (`[ "$seq" -gt "$observed_seq" ]`). When a secondmate is retired and reprovisioned under the same task ID, its fresh home's queue sequence restarts BELOW the recorded position, so the comparison is false, no reset happens, and the new queue inherits the retired generation's already-expired idle interval — emitting a false `secondmate wake-loop stalled` on its very first observation. That is precisely the false-alarm class the user intent requires this PR to remove. FIX (smallest, removal-first): bin/fm-watch.sh:749 now resets when the drain position MOVES AT ALL (`-ne` instead of `-gt`). Draining moves it up, reprovisioning moves it down; neither is a continued no-progress episode. The asymmetric `-gt` branch is removed rather than special-cased or hardened. Updated the function header comment plus the two doc sentences in docs/architecture.md and docs/configuration.md that stated the old advance-only semantics. REGRESSION TEST: added `test_secondmate_reprovisioned_queue_starts_a_fresh_interval` to tests/fm-wake-queue.test.sh (registered in the invocation list). It drives the real watcher through retired generation (seq 9) -> reprovision (seq 3, later clock) -> freeze, asserting observable wake-queue output, no source-text inspection. VERIFICATION: - Fails before / passes after: with the fix reverted the suite aborts on `not ok - a reprovisioned queue generation inherited the retired generation's idle interval and alerted`; with the fix it passes, and its third leg confirms the restarted generation still escalates on a genuine freeze (row=3 idle=2s), so the fix does not merely mute the alarm. - `bash tests/fm-wake-queue.test.sh`: exit 0, 38/38 pass, all five secondmate cases green. - `bin/fm-lint.sh`: clean (ShellCheck 0.11.0, actionlint 1.7.12). CHECKS NOT CAUSED BY THE CODE: the CI and "Require no-mistakes" runs (34152740610, 34152740604) both ended with conclusion `action_required` — workflow approval pending, not a test/build failure. Separately, `bin/fm-test-run.sh --check-coverage` exits 1 in this environment, but I confirmed by stashing my changes that it fails identically on the unmodified base tree (locale-related `comm: input is not in sorted order`); it is pre-existing and this change adds no new test file for the partition to account for. Changes are left uncommitted in the worktree
* no-mistakes(ci): Fixed the one real code defect behind the failing checks. ROOT CAUSE (Greptile P1, second round, on the head commit e1304e6): `secondmate_wake_stall_tick` in bin/fm-watch.sh identified the queue's drain position by the sequence number ALONE. The previous round changed the comparison from `-gt` to `-ne`, which handles a reprovisioned queue that restarts BELOW the recorded position, but not one that restarts ON it. A mate retired and reprovisioned under the same task id gets a fresh home whose wake-queue sequence counter restarts at 1 — and the retained parent progress marker very plausibly holds a low sequence too (a queue frozen on its first row records seq 1). Equal sequence ⇒ no reset ⇒ the brand-new queue inherits the retired generation's long-expired idle interval and emits a false `secondmate wake-loop stalled` on its very first observation. That is exactly the false-alarm class this PR exists to remove. FIX (smallest, removal-first): the file already defines the identity of a queue row once, as `row_key="$epoch-$seq"` (used for stall receipts, the stall marker, and the notify key). The progress marker's separate, weaker seq-only identity is removed: `row_key` is now computed once right after the row is parsed, stored in the progress marker, and compared with `!=`. Across generations the epoch differs (the new generation's rows are appended later), so no sequence collision can carry a stale interval; within a generation the key is stable exactly while the position does not move. bin/fm-wake-lib.sh's `fm_wake_secondmate_progress_marker_write` now takes `<oldest-row-key>` and validates it the same way the two neighbouring row-key writers do. Updated the function header comment and the two doc sentences (docs/architecture.md, docs/configuration.md) that described the old sequence-only semantics. REGRESSION TEST: `test_secondmate_reprovisioned_queue_starts_a_fresh_interval` in tests/fm-wake-queue.test.sh now drives the reported case — the reprovisioned generation restarts on the SAME sequence 9 (epoch 200) that the retired generation recorded (epoch 100), at a later clock — and asserts observable watcher output only. Its third leg still confirms the restarted generation escalates on a genuine freeze (row=9 idle=2s), so the fix does not merely mute the alarm. Three seeded progress markers in the symlink, crash-window and prefix-receipt tests were updated to the epoch-sequence form. VERIFICATION: - Fails before / passes after: with bin/fm-watch.sh and bin/fm-wake-lib.sh reverted to HEAD and the new test in place, the suite aborts on `not ok - a reprovisioned queue generation inherited the retired generation's idle interval and alerted` (exit 1); with the fix, `bash tests/fm-wake-queue.test.sh` exits 0 with 38/38 pass, all six secondmate cases green. - `bin/fm-lint.sh`: clean (ShellCheck 0.11.0, actionlint 1.7.12). - I also started `bin/fm-test-run.sh tests/fm-watch-checkpoint.test.sh tests/fm-watch-triage.test.sh tests/fm-watch-recovery-loop.test.sh` as a blast-radius check; it was still running when this phase had to return, so its result is not included. No other suite references the stall detector or the progress marker (grep over tests/ for `wake-loop stall|SECONDMATE_WAKE_STALL|secondmate-wake-progress` matches only fm-wake-queue.test.sh), and the changed lib function has exactly one caller. CHECKS NOT CAUSED BY THE CODE: the CI and "Require no-mistakes" runs on the head commit (34154091191, 34154091236, 34154091945) all ended with conclusion `action_required` — pending workflow approval, not a test/bui…
tiago-peixoto
added a commit
to tiago-peixoto/firstmate
that referenced
this pull request
Sep 12, 2026
* fix(bin): scope the worker role contract for ship and scout launches (#3797)
* fix(brief): scope Firstmate workers to their launch contract
* no-mistakes(document): Clarify supervisor scope and worker contract ownership
* no-mistakes(ci): Removed the heading-based bypass so every ship/scout launch receives the current worker-role contract. Added a regression that failed before the fix and passes afterward. Dispatch, brief, and delivery suites, focused ShellCheck, and git diff --check all passed
* no-mistakes(review): make launch overlay sole owner of worker role contract
* no-mistakes(review): narrow heading test dimension and fix publish error wording
* no-mistakes(review): gate role supersession, fix render guard, drop AGENTS twin
* no-mistakes(document): align architecture AGENTS.md scope and spawn launch-brief header
* fix(bin): stop ringing steering doorbells into dead panes (#3823)
* fix(bin): stop ringing steering doorbells into dead panes
The steering-inbox doorbell was a plain sentence plus Enter typed into a
worker's pane, and the watcher re-rang it on the assumption that a ring is
free. In a pane whose agent has exited that line is a shell command, and the
re-ring ladder kept typing it into a shell that can never acknowledge it.
- Prefix the doorbell with the shell no-op `: ` so a bare shell executes
nothing while a live worker still reads the same self-describing line.
`#` is not used because interactive zsh does not treat it as a comment by
default and the claude harness binds it to memory mode.
- fm_task_inbox_ring skips the pane (return 3) when the backend positively
classifies the agent as dead; missing, ambiguous, unreadable, and unverified
endpoints still ring so a blind classifier never starves a live worker.
- The watcher caps the ladder for a dead pane: one stale wake for recovery,
no ring, no ladder walk, and the durable record stays for
stuck-crewmate-recovery. fm-send and the remote steer leg report the skip.
Tests cover the no-op in real shells, the dead/live/unclassifiable ring
verdicts, and the single-surfacing watcher path.
* no-mistakes(review): Quote doorbell paths against shell injection
* no-mistakes(review): Reject terminal-control paths before ringing
* no-mistakes(review): Document accepted partial doorbell delivery race
* no-mistakes(review): Skip unavailable endpoints before busy-state handling
* no-mistakes(test): Respect shell startup PATH in environment allowlist test
* no-mistakes(test): Fix doorbell test fixtures for endpoint liveness
* no-mistakes(test): Prioritize confirmed restarts and clean shell test syntax
* no-mistakes(document): Document dead and missing doorbell recovery
* no-mistakes(ci): Fixed the persistence-reply timeout race by rechecking for a correlated reply immediately before falling back to a nudge. Added a deterministic regression covering replies arriving between the preliminary resolution pass and timeout handling. Verified with the targeted restart suite, project lint, coverage guard, bash syntax checks, and diff checks
* fix(bin): wait out transient primary-checkout reads in the spawn worktree poll (#3834)
* fix(spawn): keep the worktree poll from adopting the repository primary
After `treehouse get` is sent, the worktree-discovery poll reads the pane's
foreground-process cwd. While treehouse is still fetching and checking a slot
out, the foreground process is treehouse itself and it reports the repository's
PRIMARY checkout as its cwd for several seconds. The poll accepted any path
that merely differed from the spawning project, so from a linked spawning home
- whose project is itself a worktree of that repository - it adopted the
primary, and the isolation guard then refused a launch whose slot treehouse
went on to create normally.
Screen every candidate with the isolation guard's own conditions, extracted as
spawn_worktree_isolated, so a read the guard would reject stays a transient the
poll keeps waiting through. The two-consecutive-reads rule and the guard as
final backstop are unchanged; a pane that never reaches an isolated worktree
still fails at the existing 60s deadline, now naming the last path it reported.
The already-settled timing assertion counted whole-spawn wall time against a
5s budget and failed on unmodified HEAD on slower machines; it now counts pane
reads, which is what "one confirming read, not an extra cycle" actually means.
* fix(spawn): say which path the worktree wait rejected, and why
Screening every discovery-poll candidate means a host that never reaches an
isolated worktree spends the whole 60s window before refusing. That wait is
deliberate - separating a transient from a terminal misconfiguration needs
machinery this path does not want - so the refusal explains itself instead:
the isolation check records why a candidate failed, and the deadline names the
last path seen together with that reason. Message and diagnostics only; the
poll's control flow is unchanged.
Two suites asserted the guard's wording on paths the poll now rejects rather
than adopts, so their refusal arrives from the deadline instead: realign
fm-tangle-guard's non-git and subdirectory-of-primary cases (each now also
asserting the stated reason, and the second the metadata absence it was
missing) and the herdr projection e2e's forced non-worktree cwd.
* no-mistakes(review): stub poll sleep in tangle-guard spawn isolation test
* no-mistakes(document): document spawn poll isolation screen in fm-spawn header
* test(spawn): make the non-git isolation case non-git anywhere
The refusal-reason assertion for a path outside any repository assumed TMPDIR
is not inside a git repository. Where it is, git walks up from the temporary
directory, finds that repository, and the spawn reports the subdirectory cause
instead - so the case passed or failed on a property of the host rather than on
the behaviour under test.
Build the path under a directory the test then names in GIT_CEILING_DIRECTORIES,
which git documents as not chdir-ing up into a listed directory while looking
for a repository. Git never excludes the directory being searched, so the
ceiling is the parent of the path handed to the spawn.
The assertions pin which cause fired rather than the sentence that explains it,
leaving the operator wording free to improve.
* no-mistakes(document): point spawn poll comment at the isolation screen's comparison
* no-mistakes(ci): Fixed the "Behavior portable serial 2" failure in tests/fm-tangle-guard.test.sh ("non-worktree spawn did not say why the path was rejected (missing: 'not inside a git worktree')"). Root cause, in this PR's code: bin/fm-spawn.sh's spawn_worktree_isolated resolved the git toplevel with `wt_top_real=$(cd "$SPAWN_WT_TOP" ...)`. For a path in no repository, `git rev-parse --show-toplevel` yields empty, and `cd ""` is a SUCCESSFUL no-op on bash before 5.3 (CI's ubuntu-latest ships bash 5.2). The empty toplevel therefore resolved to fm-spawn's own cwd — the CI checkout — so the poll reported "it is a subdirectory of worktree root '/home/runner/work/firstmate/firstmate'" instead of the correct "it is not inside a git worktree". Dev machines with bash 5.3 fail `cd ""`, which is why the suite passed locally and only failed on CI; it is a genuine shell-portability defect in the reason vocabulary this change added, not a test-environment artifact. Fix (smallest root-cause change, 1 line + comment, bin/fm-spawn.sh:2168-2173): guard the empty value so it never reaches `cd` — if [ -n "$SPAWN_WT_TOP" ] && ! wt_top_real=$(cd "$SPAWN_WT_TOP" 2>/dev/null && pwd -P); then No change to the poll's timing or deadline behavior (respecting the recorded refusal-latency and spawn-wt-reason-vocabulary decisions), no new tests, no other files touched. Verification: - Reproduced the exact CI failure locally by putting bash 3.2 (same `cd ""` semantics as CI's 5.2) first on PATH: fails before the fix with the identical message shape, passes after. - tests/fm-tangle-guard.test.sh passes under both bash 3.2 and bash 5.3. - tests/fm-spawn-worktree-settle.test.sh and tests/fm-spawn-pool-base-freshen.test.sh pass; shellcheck -x bin/fm-spawn.sh clean. - bin/fm-test-run.sh --changed: 46 suites completed, every FM_TEST_END exit=0, 1076 passing assertions, 0 "not ok" (including fm-tangle-guard, fm-control-relaunch, fm-lint). The run ended on my own 900s wall-clock cap (rc=124), not on any test failure
* fix: support stock macOS Bash 3.2 paths (#3732)
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): read orphaned green ci monitor and daemon-down failed record as not failed (#3846)
* fix(bin): read an orphaned green ci monitor as held-for-merge, not failed
A no-mistakes run held for a captain merge decision keeps its ci step
polling until merged or closed; when the shared daemon restarts under
that poll, the run is recorded failed although every substantive step
completed and GitHub reports the PR green. A monitor whose only
remaining job is to observe a human decision must not convert the
absence of that decision into a failure verdict.
fm-crew-state.sh now reclassifies a terminal failed run as done
(held-for-merge), surfacing the run's PR URL, when the steps table
shows every step completed except exactly ci failed and the ci log's
last recognized marker reads checks green. A genuinely red check, an
unreadable ci log, or a second failed step keeps the failure.
* no-mistakes(review): Read daemon-down coarse failed ledger as unknown, not failed
* no-mistakes(document): docs: align AGENTS.md failed-verdict guidance with crew-state reclassification
* fix(spawn): verify preserved backlog state after interrupted spawn delivery (#3852)
* fix(bin): read preserved spawn state back before the interrupted exit claims it
The deferred-signal exit path asserted the paired task record and
In-flight backlog state were preserved without reading either back,
exactly when a reader is least able to check (fm-yi4j evidence,
2026-09-05). The commit's exit status alone has been observed to agree
with a row that did not actually move.
The exit path now re-reads the record and the row under the same
per-task lock as the commit, repairs a row the commit believed it
moved, and phrases the error as exactly what was verified or attempted
- verified preserved, repaired and verified, or an explicit
preservation-could-not-be-verified with the reason and hand-closeout
instruction. Two behavior tests drive a lying tasks-axi start through a
real interrupted spawn and assert the printed claim and the real
backlog state agree.
* no-mistakes(test): Fix calm suite for Pi 0.85 and pin test umask
* no-mistakes(document): Document interrupted-spawn preservation claim in backlog gate owner
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-spawn.sh's deferred-signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-backlog-transition-lib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean
* no-mistakes(ci): Fixed the Greptile P1 on bin/fm-backlog-transition-lib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean
* docs: correct runtime-backend maturity labels for Herdr (#3821)
* docs: correct stale tmux/herdr backend maturity claims
Herdr now has 21 test files, its own required CI job (tests-herdr)
that installs a pinned build and hard-fails on "skip: herdr not
found", while tmux has 3 test files and is only required as a
dependency of the portable-serial e2e lane. zellij, orca, and cmux
still have no CI lane at all. AGENTS.md and docs/herdr-backend.md
still called Herdr merely "experimental" alongside those three,
misleading every session and reader about actual coverage.
Update AGENTS.md's config/backend entry, the opening lines of
docs/herdr-backend.md and docs/tmux-backend.md, the runtime-backend
section of docs/configuration.md, and the matching claims in
docs/architecture.md, CONTRIBUTING.md, and README.md so they agree
and distinguish tmux (default), herdr (own required CI lane, largest
suite, Windows still spike-only), and zellij/orca/cmux (still
experimental, no CI lane). No behavior, selection order, or
dispatch logic changes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GDYsWPEfwTuPNjQ2nGBCcj
* no-mistakes(review): docs: fix stale herdr label and CI-lane wording
* no-mistakes(document): docs: align tmux adapter label in scripts.md
* no-mistakes(review): docs: drop duplicated herdr CI claim from tmux page
* no-mistakes(review): docs: drop windows claim, align contributing backend wording
* no-mistakes(review): docs: trim duplicated CI claim from herdr opening line
* no-mistakes(review): docs: drop unguarded largest-test-suite superlative
* no-mistakes(review): docs: restore tmux verified label and README experimental scope
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: restore intent-targeted no-mistakes validation (#3865)
* fix: delete deterministic no-mistakes test baseline, restore intent-targeted Test
PR #3644 pinned commands.test to a fm-test-run.sh --changed walk of the
repository's 75-162 tests/*.test.sh scripts. no-mistakes runs commands.test
verbatim and unconditionally after every fix round, so that walk multiplied
by round count: measured at 32.7 minutes per validation versus 3.6 minutes
intent-targeted. Delete the pin and restore the 3.6-minute posture.
Add tests/fm-nm-test-contract.test.sh as a regression guard, parsing
.no-mistakes.yaml as YAML (ruby's bundled Psych, matching the parser
tests/fm-test-run.test.sh already uses for ci.yml) rather than grepping its
text, restoring in legal form what PR #823 added and PR #1282 removed.
Record the rule in docs/configuration.md's "Gate defaults" section (the
authoritative owner CONTRIBUTING.md already points at) and strengthen
CONTRIBUTING.md's existing local-Test guidance to state it plainly: never
configure commands.test to a deterministic test command, complete or partial.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSt9JvrQMVc4u3jPyUFCFC
* no-mistakes(review): Centralize no-mistakes test policy and narrow guard
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: verify Treehouse slot ownership before teardown (#3837)
* fix(bin): verify pool-slot ownership before returning a worktree slot
Workers were killed when cleanup returned a Treehouse pool slot that a
different, live task had already taken. Teardown now proves the slot is
genuinely this task's before releasing it: it refuses when another task
record claims the same live worktree path, or when the endpoint's working
directory contradicts the recorded slot, and that refusal holds under
--force. Slot allocation, metadata publication, ownership verification,
and slot return are serialized across linked firstmate homes, and forced
secondmate cleanup verifies descendant slot ownership before returning
any child worktree.
Regression coverage drives the scripts with two task records naming one
slot path and asserts the live worker survives and its slot is not reset.
* no-mistakes(review): Protect slots across cloned Firstmate homes
* no-mistakes(test): Gate teardown locking on genuine Treehouse slots
* no-mistakes(test): Clarify pooled descendant slot gating
* no-mistakes(test): Synchronize watcher re-arm test on process exit
* no-mistakes(test): Wait for watcher cleanup before timeout escalation
* no-mistakes(document): Document pool-slot ownership safeguards
* no-mistakes(ci): Fixed all reported CI issues: normalized bare local Git origins to the same Treehouse project-lock identity as absolute clone origins; resolved ShellCheck SC1091 with explicit conditional sourcing; and taught concurrent Herdr teardown coverage to retry expected Treehouse lock contention. Added behavioral regression coverage for bare/absolute origin lock identity. Verified endpoint-safety tests, watcher tests, full CI lint, and the previously failing Herdr teardown assertion
* fix(bin): resolve relative origins from repository root
* no-mistakes(ci): Fixed teardown so an exact recorded endpoint may change cwd without falsely vetoing cleanup. Removed cwd-based ownership refusal while preserving cross-home record exclusivity and project locking. Updated behavioral coverage for both foreign slot ownership refusal and moved-cwd teardown success. Endpoint-safety, backend, watcher, checkpoint, and targeted lint checks pass. Real Herdr presentation E2E progressed successfully but exceeded the 600s local timeout
* feat(bin): add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary (#3867)
* feat: add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary
Add omp as a verified harness: anchored process-name detection with a
Firstmate-owned FM_OMP_HARNESS launch marker that needs real omp ancestry,
the fm-spawn launch template with foreign-marker clearing, the tracked
.omp/fm-worker-overlay.yml posture overlay, --auto-approve, --cwd, and
pre-launch model validation scoped to providers 'omp models --json' lists.
Workers get a state-resident busy-state extension keyed on agent_end
without willContinue (omp has no agent_settled). The primary gets two
tracked .omp/extensions: a turn-end guard that answers omp's blocking
session_stop hook by compelling one continuation per turn, with the
pre-tool seatbelts and Run-tier session-start delivery, and a watcher
extension ported from the Pi one with fm_watch_arm_omp. Control tables,
composer busy footers, omp's status row as a bare-composer boundary, the
extension supervision model with an omp-keyed ownership proof, the
session-start diagnostic, and the supervision protocol snippet follow.
Verified live on omp 18.1.11 with openai-codex/gpt-6-astra: a Herdr scout
through spawn, busy state, steer, interrupt, exit, and teardown, and the
isolated rpc primary lab through extension auto-discovery, digest
delivery, lock identity, watcher arm, successor and wake delivery, and
the compelled guard continuation.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test: prove the omp guard continuation through a guard spy
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* fix(spawn): clear the gemini marker at the omp launch boundary
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): force the guard stage by freezing the watcher and clear lint findings
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): reap the live lab by path and record omp's rpc shutdown as a note
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): spawn a real secondmate for the discovery rule and classify the omp surfaces
Replace the template-extraction check with a genuine --secondmate launch
pinned to the fake tmux backend, assert the worker extension's handler set
through the executable rather than its bytes, classify the two new omp
surfaces in the documentation inventory, and record the Herdr worker
evidence in the runtime-backends verification doc.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* no-mistakes(review): omp: unverify remote routes, narrow busy regex, drop overlay approval pin
* no-mistakes(review): omp: validate config-pinned model, correct remote and marker docs
* no-mistakes(review): omp: pin config-model validation with a test, trim overlay
* no-mistakes(review): omp: sync guard evidence, drop dead param, map quota family
* no-mistakes(review): omp quota: refuse unmapped prefixes, match bare model scopes
* no-mistakes(document): docs: cover omp in cd-guard, quota, continuity, tmux
* no-mistakes(document): docs: add omp subagent-guard row, fix live test header
* no-mistakes(ci): Fixed both failing behavior shards and the Greptile P1 in bin/fm-composer-lib.sh. Root cause of "Behavior portable serial 1" and "Behavior portable parallel 2": the omp busy regex (FM_DELIVERY_OMP_BUSY_REGEX_DEFAULT) and omp status-row furniture regex (FM_COMPOSER_OMP_STATUS_RE_DEFAULT) used the bracket range [⠁-⣿]; BSD grep on macOS accepts it but GNU grep on Linux CI aborts with "Invalid collation character", failing every omp busy/furniture read (3 assertions across fm-omp-harness, fm-tmux-submit-busy, fm-composer-lib). Replaced the range with one shared explicit alternation FM_OMP_SPINNER_FRAMES_RE of omp 18.1.11's unicode-preset spinner frames (status set ⣾⣽⣻⢿⡿⣟⣯⣷ + activity set ⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏, read from the installed binary), the same pattern the Kimi busy regex already uses in CI. For Greptile's finding (the harness-agnostic furniture rule's first alternative matched any 1–4-byte token + ' · ', so wrapped typed input like 'fix · tests' with the cursor on it regressed from pending to unknown; reproduced locally vs base), pinned that alternative to omp's identity cell (π||pi, the icon.omp of each preset in the 18.1.11 binary). Tests: fm-composer-lib.test.sh asserts 'fix · tests' is not furniture, a status-set spinner row is furniture, and the wrapped composer screen reads pending under both locales (CAPS_TMUX cursor 3); fm-omp-harness.test.sh asserts a status-set frame reads busy. New negative cases fail against the pre-fix lib and pass after. Verified: fm-omp-harness, fm-tmux-submit-busy pass via bin/fm-test-run.sh; fm-composer-lib passes all cases except one pre-existing, unrelated local failure (Herdr half-block test uses printf '▀', unsupported by macOS bash 3.2; fails identically on a pristine HEAD export, passes on CI bash 5); shellcheck and bin/fm-lint.sh clean. Caveat: GNU grep is unavailable locally, so the Linux compile was not run directly; the fix uses only constructs already proven on CI's GNU grep (multibyte literal alternations, incl. under LC_ALL=C). Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh, tests/fm-omp-harness.test.sh. No docs needed changes (they describe the rule generically)
* no-mistakes(ci): Greptile Review: fixed. The omp status-row furniture regex FM_COMPOSER_OMP_STATUS_RE_DEFAULT in bin/fm-composer-lib.sh still accepted a literal `pi ·` opening, so wrapped composer input beginning with `pi ·` was truncated and misclassified. Read the installed omp 18.1.11 binary: the ascii preset's `icon.omp` is `pi` but its `sep.dot` separator is ` - ` (unicode/nerd use ` · `), so a real ascii status row never contains `pi ·` and that alternative could only ever match typed text. Removal-first fix: dropped `pi` from the identity alternation (now `(π|)`) and updated the comment to record why the ascii preset is excluded. Tests (tests/fm-composer-lib.test.sh): added a negative furniture case for 'pi · e · phi as the three constants' and a wrapped-screen assertion (CAPS_TMUX, cursor 3) that a continuation row opening `pi ·` reads pending in both locales; the new case fails against the unfixed lib and passes after. Verified: composer test with the half-block case skipped passes all 33 cases including the omp matrix; bin/fm-test-run.sh tests/fm-omp-harness.test.sh passes; shellcheck -x clean on both files; bin/fm-lint.sh clean. The full composer test via the runner fails locally only on the pre-existing half-block case (bash 3.2 printf cannot emit ▀; passes on CI bash 5), identical to before this change. Docs unchanged (they describe the rule generically and never mention the ascii identity cell). PR must be raised via no-mistakes: not caused by code. attestation.head_sha is cdddc60 while the PR head is cd51cf4 because the pipeline's ci-phase push moved the head; the outer executor's re-push will re-bind the attestation. No file change for that check. Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* fix(pi): invoke Bash helpers correctly on native Windows (#3843)
* Fix Pi shell invocation on native Windows
* no-mistakes(document): Document Pi Windows Bash transport
* no-mistakes(ci): Captain, staged a narrow fix: register the Pi Windows regression for both extension paths, make Windows mode emulation non-failing, and enforce LF shell checkouts. Mapping and coverage checks pass; CI/Require no-mistakes were approval-gated externally
* no-mistakes(review): Cover async Windows branch-outcome Bash invocation
* no-mistakes(review): Preserve Cygwin checks and refresh Windows timing
* no-mistakes(test): Invoke OpenCode operational-input owner through Bash on Windows
* validation-fixture
* no-mistakes(document): Document Windows Bash helper invocation
* no-mistakes(ci): Fixed PR-caused changed-selection failure by removing the malformed tracked evidence artifact and allowing deleted, unconsumed source paths to retire cleanly while preserving fail-closed behavior for live unmapped paths. Added regression coverage. Verified native-Windows Pi shell-seam test passes and --changed selects the Windows regression
---------
Co-authored-by: test <test@example.invalid>
* test(bin): pin teardown outcomes for squash-merged rebased branches (#3870)
* fix(bin): recognise squash-merged rebased work as landed at teardown
A pipeline rebase can leave the local worktree on pre-rebase commits while
GitHub squash-merges the rebased head. The landed-work test then compared
those stale commits against a squashed main and refused cleanup of work
that had already landed.
When the forge reports the recorded PR merged and its merge commit is on
the default branch, treat a local branch that only repeats paths from the
pipeline push as stale rather than unlanded. If the forge is unreachable,
the same coverage check runs against a PR head whose content is already
on default. Extra local paths still refuse.
* fix(bin): drop unprovable squash-rebase landed-work coverage
Path-set coverage treated a diverged local branch as landed whenever it
touched the same files as the squash merge. That accepts the reviewer's
failing sequence: same path, different content, work discarded.
git cherry and merge-tree containment were already too strict on the real
rebase-fold case. No remaining check is both safe and permissive enough
to recognise a stale pre-rebase copy without also accepting unlanded
edits, so that case still refuses.
Keep the proofs that hold: a merged PR head that contains local work, or
a clean content-in-default tree match. Tests now refuse same-path
different content and extra unlanded commits, and still allow a local
branch that followed the pipeline rebase.
* no-mistakes(review): drop recorded-pr-head fallback and reverted-design leftovers
* no-mistakes(review): silence squash-merge stdout corrupting test PR head
* no-mistakes(review): make unlanded follow-up commit sole cause of refusal
* no-mistakes(document): correct stale squash-rebase fixture comments in teardown tests
* no-mistakes(ci): Split the three reported checks: - CI (run 34061098467) and Require no-mistakes (run 34061098460) both concluded `action_required` — approval-gated workflow runs that never executed a step. Not caused by this PR's code; no change can clear them. - Greptile Review was a genuine defect in the new tests: the three new refusal cases (tests/fm-teardown.test.sh) asserted only exit status 1 and a REFUSED line, so a teardown regression that destroyed the worktree, branch, and task record before reporting refusal would still pass. Fix (tests only): added one `assert_refusal_retained_task_state` helper and called it from `test_squash_merged_same_file_different_content_refuses`, `test_squash_merged_rebased_local_with_unlanded_commit_refuses`, and `test_squash_merged_stale_local_refuses_when_forge_unreachable`, each capturing the worktree HEAD before `run_teardown`. It pins that the refusal left the isolated copy on disk, the task branch still checked out at the same unlanded commit, and state/task-x1.meta intact. Verification: the four squash tests pass; a sensitivity probe ran the ALLOW fixture (teardown completes) and pointed the same helper at the outcome — it fires, because a completed teardown detaches/deletes the branch and removes the task record, proving the assertions discriminate. Full tests/fm-teardown.test.sh: 83 passing. bin/fm-lint.sh clean with pinned ShellCheck 0.11.0 + actionlint 1.7.12 (plus an explicit --external-sources pass on the changed file). bin/fm-test-run.sh --check-coverage ok. Caveat: test_herdr_flat_teardown_preflight_refuses_before_changes (mode missing-adapter) fails on this machine. Verified it fails identically on base commit f91a950 via `git archive`, so it is a pre-existing local environment difference untouched by this diff; skipped to run the rest of the suite, not modified
---------
Co-authored-by: Morten Gad <mogad@itm8.com>
* fix(bin): keep supervision armed for registered custom checks (#3860)
* fix(bin): keep supervision armed for registered custom checks
A custom check bound by bin/fm-check-register.sh only ever runs inside the
watcher's check sweep, but fm_supervision_status counted in-flight tasks, the
relay poll shim, and process-event sources as supervision need, and not
registered checks. Tearing down the last task therefore stopped every
home-level check silently until the next spawn.
Count a state/<id>.check.sh that carries its state/<id>.check-trust binding as
supervision need. The relay shim keeps its own trust path and task PR polls
carry no such binding and are torn down with their task, so neither arms a home
by accident. Presence of the binding is the whole test: the sweep validates the
bytes at execution time and wakes firstmate when it rejects one, which is the
outcome an idle home needs.
Closes #3856
* no-mistakes(review): name registered checks in turn-end block banner and doc invariant
* no-mistakes(review): narrow PR poll predicate test to what it proves
* no-mistakes(document): point Grok re-arm step at supervision-need owner
* fix(bin): resolve Treehouse locks for remote secondmate homes (#3883)
* fix(bin): resolve the shared Treehouse project lock inside remote secondmate homes
Every spawn and teardown inside a remote-seeded secondmate home refused,
because the project lock's anchor could not be resolved there.
fm_firstmate_root_home walks a home's parent bindings upward to find the
anchor the lock lives in, and treated a remote parent binding as an error.
A remote-seeded home's parent is on another machine, so that walk can never
succeed from there - and neither can the home's own local descendants, whose
chain terminates at the same record. Both fail closed on every Treehouse-backed
spawn and every pool-slot teardown.
A remote parent now terminates the walk at the home holding it, which is the
correct anchor: a lock taken on this filesystem is neither held nor observable
across that boundary, and that home is already the top of the local tree
teardown's collect_local_firstmate_states enumerates, since that walk skips
remote registry entries for the same reason. Mutual exclusion is unchanged -
every home reachable through local parent links still derives one identical
lock file per project, and an unreadable binding, an unsupported route, an
unreachable local parent, a cycle, and an over-deep chain all still refuse.
Origin-less local-only projects keep resolving through their worktree top.
Regression coverage pins the anchor for the main-home layout, a local
secondmate, a remote-seeded home, and its local child; drives teardown
end-to-end in a remote-seeded home; keeps the cross-home slot-ownership
refusal across that boundary; and proves two homes still serialize on the
one shared lock file.
* no-mistakes(document): Clarify machine-local Treehouse lock ownership
* fix(bearings): repair board listening and decision reconciliation (#3872)
* fix(bearings): repair the board's listening, card hygiene, and reconcile path
Three defects made the fleet board go quiet and then lie about what still
needs the captain.
Never arm a poll on a session that is not live. `lavish-axi <file>` exits 0
even when it refuses to reopen a session the captain ended from the browser,
reporting `status: user-ended` with the same session id, so the build's
exit-status check accepted a dead session, printed `already-armed`, and left
the board reading "not listening". The build now proves the session is live
from a fresh authoritative listing immediately before arming - not from the
establish call's status alone, which is already stale by then - reopens once
when it finds the session ended, and refuses rather than arming when it stays
ended. A reopen also replaces the pre-reopen source generation before
reporting success, so a runner on its way out cannot be mistaken for a
listener, and a board whose source is registered but unowned gets a
replacement started before the build returns.
Let a dead generation's ownership actually move. Reclaiming a claim ran its
capture-reservation cleanup first, and that cleanup re-verifies the recorded
state-root identity, so a claim naming a pid and a process group that were
both provably gone could not be cleared: reconcile reported a start while
nothing attached, and retire refused with "cannot release source ownership".
Reservation records are keyed by claim token and every replacement claims a
fresh one, so they are hygiene, not an ownership invariant. Reclamation now
additionally requires the owning process group to be absent independently,
which keeps a reused pid whose poll child still runs from ever reading as a
gone generation. A live owner and a crashed leader whose owned group survives
are still never reclaimed.
Stop carding decisions whose subject already landed. The build drops a
decision card whose work item or PR appears in the payload's own landed rows,
and one whose task is no longer an open captain call, naming each drop on
stderr. A task whose state cannot be established is kept, because a call
wrongly hidden is worse than a card wrongly shown.
Add the reconcile choice, and make it structurally incapable of closing a
call. Every decision card carries a standard `reconcile` option, injected by
the build rather than left to the composer. The board now emits the picked
option and any freeform note as separate structured fields instead of fusing
them, so a reconcile selection is not expressible as an answer value at all -
the defect that let `reconcile - <note>` reach the intake as an ordinary
answer. The adapter routes selections from that structured field, creation of
a reconcile request is bound to a verified board source rather than the shared
keyed-answer intake, and the intake still refuses the reserved value on every
channel. Each authorization is bound to the captain-hold generation that
produced the card, so an obsolete card cannot close a later call, and both
terminal outcomes require a pending request: `reconcile close` records the
evidence under its own `reconciled` mode so it never reads as the captain's
words, and `reconcile note` leaves the call open. Anything unprovable -
an unversioned row, a missing generation, an unreadable state - refuses
rather than acting.
Regression coverage fails without each fix, and pins every leak path: a bare
reconcile, a standalone close or note with no pending request, an any-channel
reconcile, an annotated selection from a freeform card, and a
generation-skewed authorization. An opt-in guard re-proves the lavish-axi
shapes and the reopen against the installed tool.
* fix(bin): quote the done comparison in the reconcile intake
shellcheck SC1010 reads the bare word as the loop keyword. The failed run
never reached its lint step, so this shipped in the recovered content.
* no-mistakes(review): Publish reconciled parent resolution before request retirement
* no-mistakes(review): Clarify committed cleanup and reconcile reservation scope
* no-mistakes(review): Preserve remote cards and legacy answer compatibility
* no-mistakes(test): Separate live claim release from stale reclamation
* no-mistakes(test): Allow terminal self-retirement during active capture
* no-mistakes(document): Document Bearings repair contracts
* no-mistakes(ci): Stabilized the failing Herdr presentation E2E by serializing test-harness Treehouse allocator calls, preventing concurrent recovery spawns from claiming the same pool slot while preserving Herdr concurrency coverage. Verified with the full E2E suite on Herdr 0.8.2, bash syntax checks, ShellCheck, and git diff checks
* fix(bin): allow pooled spawns without a git origin (#3885)
* fix(bin): skip pooled-worktree freshness fetch when no origin is configured
An origin-less local-only project has nothing remote to be stale
against, so fm-spawn's freshen_spawn_worktree_base refused to launch
crews for it. Detect a missing origin remote and skip the fetch
freshness gate entirely; an existing-but-unreachable origin keeps
refusing as before.
* no-mistakes(review): Preserve pool safety for absent and unusable origins
* no-mistakes(review): Refuse empty origin configurations during pooled spawn
* no-mistakes(review): Detect empty origin sections across config includes
* no-mistakes(review): Honor globbed includes when detecting origin configuration
* no-mistakes(review): Document conservative conditional include handling
* no-mistakes(review): Use Git-resolved config files for origin detection
* no-mistakes(review): Document included empty-origin detection boundary
* no-mistakes(document): Document originless pooled spawn behavior
* feat(tests): run live harness guards by default when available (#3889)
* feat(tests): run live harness guards by default where the harness is installed
The 24 live-harness guards each opened with their own env check, so on the
machine that has every harness - the one the product and its validation
actually run on - all of them skipped and passed. Fourteen had never been run
by the pipeline at all.
tests/lib.sh gains fm_live_gate as the single owner of that decision: a guard
that spends no model tokens runs wherever its tools are installed, a guard that
submits prompts stays opt-in, an absent tool is a named capability skip, and a
guard's own variable or FM_LIVE forces it on (turning an absent tool into a
failure) or off. Every live guard now opens with it, which also carries the
test-suite gate-refusal bypass into the guards that never sourced the shared
helpers and were therefore refused whenever a gate agent ran them.
bin/fm-test-run.sh records what a skip means: the family's expected class is
live-capability rather than a bare env opt-in, and each gate skip's reason is
logged and written to the timing artifact, so a lane can say which tool this
host could not exercise.
Only the token-free guards flip to default-on: composer-matrix, the harness
liveness drift guard, and the Herdr version floor. cursor-primary submits three
prompts, so it stays opt-in.
Running the drift guard unasked immediately found a real defect it existed to
catch: it resolved the harness through a generic `command -v cursor`, which on
a machine that also has the Cursor editor finds the editor launcher rather than
cursor-agent. That binary exits at once, leaving a bare shell in the pane and a
liveness-drift failure no classifier change could fix. It now asks
fm_cursor_resolve_binary first, the same verified owner fm-spawn uses.
CI installs the public Pi package in the portable serial lane and fails on its
skip token, so the Pi extension tests stop passing silently against a package
that is not there. No secret is added.
Verified on macOS 26.5.2 arm64: the drift guard runs with no variable set and
classifies 8 installed harnesses alive; the Herdr version-floor guard runs by
default and checks 4 real releases; every live guard refuses together under
FM_LIVE=0.
* fix(tests): keep the composer-matrix guard opt-in
Running it unasked is red on a healthy machine for reasons no code change here
removes: a harness that has not trusted this checkout sits on its own trust
dialog, which the guard treats as an unreadable composer and correctly fails.
The opencode 1.18.29 and grok 1.0.13 composer drift it also surfaced reproduces
identically on main and is filed as separate work.
So this token-free guard stays opt-in with the reason stated in its header, and
the coding guidelines record the narrow exception: a guard whose verdict
depends on host state that installing its tools does not establish may stay
opt-in, because one that is permanently red is one the fleet learns to ignore.
The other two token-free guards keep running by default.
* no-mistakes(review): Wire bearings guard and remove composer exception policy
* no-mistakes(review): Run Pi responsiveness guard by default
* no-mistakes(review): Gate AFK Pi Herdr through authoritative family sweep
* no-mistakes(review): Sanitize live gate test environments
* no-mistakes(document): Document default-on live guard behavior
* no-mistakes(ci): Fixed CI by installing the Pi package in portable-parallel-1, where fm-pi-primary-types.test.sh runs, and enforcing its package-missing gate skip there. Verified with fm-lint.sh, workflow actionlint, coverage partition checks, lane membership, and git diff checks
* no-mistakes(ci): Fixed CI’s Pi typecheck skip enforcement by giving npm, tsc, and Pi-package capability skips a shared prefix and configuring both relevant CI lanes to fail on that prefix. Verified missing tsc emits the expected skip, missing Pi package becomes a runner failure, and actionlint, ShellCheck, and git diff checks pass
* fix(bin): refuse test runs in the primary checkout when a task marker is set (#3891)
* fix(bin): refuse the behavior suite in the repository primary checkout
A task worker's isolated worktree placement is verified exactly once, when
its task starts, and nothing re-checks it afterwards. A worker that later
changes directory into the repository's primary checkout runs its Git
commands, and this branch-switching suite, against the one checkout every
linked worktree resolves against and every landing merges into. A run that
dies mid-suite can leave that checkout on a stray branch.
bin/fm-test-run.sh now refuses that case. When FM_TASK_ID marks a task
worker and the runner resolves to the primary checkout, every executing mode
exits non-zero before selecting a suite, with one line naming the primary
path and pointing at the assigned task worktree. The predicate is the one
bin/fm-spawn.sh already uses for launch placement: the working tree's own
git dir is the repository's common git dir, which separates the primary from
every linked worktree even when their top levels differ. A run with no
FM_TASK_ID set is unchanged, and so are the inspection modes, which execute
nothing. When git resolves neither directory - a non-repository fixture, a
detached copy - nothing proves this is the primary, so the run proceeds.
bin/fm-spawn.sh sets the marker: ship and scout launches export FM_TASK_ID
into the pane shell on the same pre-launch channel as GOTMPDIR, and the name
joins the sanitized launch environment allowlist so an isolated launch keeps
it.
* no-mistakes(review): clear inherited task marker in test lib; name resolved ROOT
* no-mistakes(document): docs: record FM_TASK_ID marker and runner placement refusal
---------
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): bound stale alarms for backlog captain holds (#3842)
* fix(bin): bound a stale alarm with the backlog hold, not only the status line
A legitimate wait has two records and the stale alarm reads only one.
`status_is_paused_or_captain_held` takes a status line, so it sees a wait the
worker declared. It cannot see the wait firstmate records when it hands work to
the captain: `bin/fm-captain-hold.sh hold` writes that into the backlog and
leaves the status log alone, so a delivered task keeps `done: PR ...` as its
last line for the whole time the captain is deciding.
Both stale branches were blind to it, and each churned a new pane hash back into
its own alarm: a `done:` line is captain-relevant and reaches the terminal-stale
branch, while a held task whose last line is `working:` reaches
`surface_nonterminal_stale` and fails its declared-wait test.
Consult that second record where the watcher is about to alarm, through
`bin/fm-captain-hold.sh open`, which already owns the predicate's semantics, and
bound the alarm on the shared `.paused-resurfaced-<key>` marker and
`PAUSE_RESURFACE_SECS` window the declared-wait absorb already uses. The first
sight still alarms, the window's end alarms once more, and a held crew that goes
genuinely silent still escalates through the wedge timer.
Only an established open captain call bounds anything: an unreadable backlog, an
absent or incompatible tasks-axi, a row this home does not carry, and every task
with no hold keep alarming exactly as before. The backlog hold is deliberately
not recorded as a declared pause, because the loop-top reconciliation and
`pause_state_class` both read the status line and would clear a flag that line
does not support.
Extends the fix in #3443, which closed the forms of this loop that the status
line itself can express.
* fix(bin): identify the captain call a stale alarm is bounded by
Three gaps in the bound added by the previous commit, all in how the throttle is
scoped and where the backlog is consulted.
The scope carried only the status-log signature. A task can be held, answered
with `--release`, and re-held as a genuinely different captain call without any
status append, so the second call inherited the first one's marker and its first
sight was absorbed - the one thing this bound must never do. The task id is not
the call: `bin/fm-captain-hold.sh open` gains `--identity`, which reports the
call's own lifecycle - its hold-set stamp and the number of recorded answers -
on an exit 0 and only then, leaving the silent predicate every existing caller
reads unchanged. The throttle scope now carries that identity.
The terminal path recorded the throttle before publishing the durable wake. A
failed append exits the watcher with nothing queued, and the next sighting then
read that fresh marker and absorbed the retry, turning a delayed alarm into a
lost one. Recording moves behind the append, as the non-terminal path already
had it, and the comment claiming the marker could not outlive its wake is gone
because it was false.
The backlog was consulted only on a new terminal pane hash. A captain call can
open after a hash was absorbed as provably working, changing neither the pane nor
the status log, so nothing re-read the backlog and the wedge timer kept firing
possible-wedge alarms through a legitimate wait. That timer now consults the call
at its own alarm boundary and takes the same bounded cadence - and only at that
boundary, so an ordinary repeat poll under the bound stays the local-only read it
was.
Regression coverage for each, all driving churn through one watcher process
rather than relaunching per pane change: relaunch cost dominated the earlier
shape, and an absorbing watcher stays in its poll loop across churn in production
anyway. An unheld task still alarms on every new hash, and an elapsed wedge timer
with no open captain call still escalates as a possible wedge.
* fix(review): Compose stale throttles with captain-call lifecycle identity
* fix(review): Preserve bounded same-hash captain-call resurfacing
* revert(bin): narrow the captain-hold stale bound to its observed defect
Lifts the lifecycle-identity and cadence-ownership work back out, leaving the
change at the shape that matches the defect actually observed: the stale alarm
did not consult the backlog captain hold, on either stale branch.
Reviewing the wider version surfaced a series of adjacent gaps in the watcher's
alarm state machine - a call opening after the first alarm, marker invalidation
at the hold lifecycle boundary, and which deadline a terminal timer represents.
They are real, but fixing them turns a small extension into a state-machine
change to the alarm path, which is a different review on a subsystem that is
being actively reworked. They are named as known limitations rather than carried
here, and none of them is load-bearing for what remains: the bound does strictly
less than the reverted version, leaves the wedge path escalating on
STALE_ESCALATE_SECS exactly as before, and introduces no silence that the
existing terminal-alarm path did not already have.
Kept from the reverted work is the record-after-append ordering, because that is
a defect in the code being shipped rather than an adjacent one: recording the
cadence marker before publishing the durable wake let a failed append lose an
alarm outright instead of delaying it.
History is preserved: the earlier commits stay on the branch and this removal
sits on top of them.
* fix(review): Document secondmate captain-hold scope boundary
* fix(document): Document captain-hold stale alarm scope
* fix(bin): bind the stale throttle to the captain call, not the status log
The throttle this change introduces was scoped to the task's status-log
signature. Answering a call with `--release` and holding the task again creates a
genuinely different captain call without necessarily appending to that log, so
the second call inherited the first one's marker and its first sight was
absorbed.
That is the one alarm this bound must never swallow. A delivery announced twice
is noise; a decision waiting on the captain that is never surfaced is invisible,
because nobody asks for what they do not know to ask for.
Measured rather than assumed, on the same fixture - a delivered task held for the
captain, released, and re-held with no status append, driven through bin/fm-watch.sh:
base c499f84 call-1 first=ALARM call-1 churn=ALARM new call first sight=ALARM
before this fix call-1 first=ALARM call-1 churn=absorbed new call first sight=absorbed
after call-1 first=ALARM call-1 churn=absorbed new call first sight=ALARM
Base never suppresses the new call, so the suppression came from this change and
closing it completes the fix rather than widening it.
`bin/fm-captain-hold.sh open` gains `--identity`, printing the call's lifecycle -
its hold-set stamp and count of recorded answers - on an exit 0 and only then, so
the silent predicate bin/fm-teardown.sh reads is untouched. The throttle scope
carries that identity beside the status signature.
The sibling case was measured too and is NOT included: on the status-declared
path, where the last line is `captain-held:`, base already absorbs a re-held
call's first sight. That behaviour predates this change and stays documented as a
known limitation rather than repaired here.
* fix(document): Document captain-call throttle lifecycle scope
* fix(ci): isolate the Herdr restart fixtures from a claimed worktree
The Herdr behaviour test intermittently reused a local worktree still claimed
by an earlier fixture after a restart.
The restart scenarios now use an isolated Treehouse project.
The full Herdr test passes on Herdr 0.8.2; bash -n and git diff --check pass as
well.
* feat(pi): resolve extension-registered providers in the supervision branch (#3871)
* Let the supervision branch resolve extension-registered providers
The isolated branch ModelRuntime cannot see providers an extension
registered into main's runtime at run time, so a pin on pi-devin-auth's
devin/swe-1-7 (or an unpinned branch following a main session on devin)
failed with "unavailable to the isolated branch runtime".
Capture main's ModelRegistry alongside mainModel and copy each
extension-registered provider config into the branch runtime at
model-resolution time. The config carries the provider's own streamSimple
and oauth wiring by reference, so the custom gRPC transport reaches the
branch unchanged instead of being reimplemented. The /supervision-model
picker uses the same copy so those models are offered.
Update configuration.md and pi-supervision-branch.md, which previously
stated extension-registered providers were not offered.
* no-mistakes(document): docs: own devin provider carve-out in branch architecture doc
* no-mistakes(ci): Fixed the Greptile P1 finding: the /supervision-model picker copied extension-registered providers into the branch ModelRuntime but checked hasConfiguredAuth without refreshing them, so providers with provisional post-registration auth were omitted from the picker while the pin-resolution path (which did refresh) accepted them. Root-cause fix in .pi/extensions/fm-branch-supervision.ts: moved the `refresh({ providers, allowNetwork: false })` call into `copyExtensionProviders` (now async, refreshing every provider it copied) and removed the duplicate per-provider refresh from `resolveBranchModel`. Both the picker and the resolution path now share one copy-and-refresh step, so hasConfiguredAuth is real in both. Regression coverage in tests/fm-pi-branch-extension.test.sh: the stubbed ModelRuntime now mirrors the real runtime by leaving a registered provider's auth pending until `refresh()` runs for it. With that stub, the existing extension-registered-provider case fails against the pre-fix extension (picker offers only anthropic/main-model) and passes with the fix. Verification: tests/fm-pi-branch-extension.test.sh passes (42 ok, no failures); tests/fm-branch-supervision.test.sh passes; tests/fm-pi-primary-types.test.sh skips locally because tsc is not installed (the refresh signature reused is the one the existing code already called). Intent constraints preserved: isolation flags untouched, carve-out still scoped to provider registration, graceful fallthrough when no providers are registered
* ci: retrigger flaky Herdr/serial-1 lanes
* no-mistakes(document): docs already cover branch extension-provider copy
* fix(bin): gate secondmate wake-loop stall alerts on real queue no-progress (#3943)
* fix(watch): detect stalled secondmate queue progress
* no-mistakes(review): gate secondmate stall on active turns and progress episodes
* no-mistakes(document): align secondmate wake-stall docs with progress-episode detector
* no-mistakes(document): clarify active-turn gate in wake-stall config docs
* no-mistakes(ci): Fixed the one real defect behind the failing checks. ROOT CAUSE (Greptile P1, real code defect in this PR): `secondmate_wake_stall_tick` in bin/fm-watch.sh reset the no-progress timer only when the oldest actionable queue sequence INCREASED (`[ "$seq" -gt "$observed_seq" ]`). When a secondmate is retired and reprovisioned under the same task ID, its fresh home's queue sequence restarts BELOW the recorded position, so the comparison is false, no reset happens, and the new queue inherits the retired generation's already-expired idle interval — emitting a false `secondmate wake-loop stalled` on its very first observation. That is precisely the false-alarm class the user intent requires this PR to remove. FIX (smallest, removal-first): bin/fm-watch.sh:749 now resets when the drain position MOVES AT ALL (`-ne` instead of `-gt`). Draining moves it up, reprovisioning moves it down; neither is a continued no-progress episode. The asymmetric `-gt` branch is removed rather than special-cased or hardened. Updated the function header comment plus the two doc sentences in docs/architecture.md and docs/configuration.md that stated the old advance-only semantics. REGRESSION TEST: added `test_secondmate_reprovisioned_queue_starts_a_fresh_interval` to tests/fm-wake-queue.test.sh (registered in the invocation list). It drives the real watcher through retired generation (seq 9) -> reprovision (seq 3, later clock) -> freeze, asserting observable wake-queue output, no source-text inspection. VERIFICATION: - Fails before / passes after: with the fix reverted the suite aborts on `not ok - a reprovisioned queue generation inherited the retired generation's idle interval and alerted`; with the fix it passes, and its third leg confirms the restarted generation still escalates on a genuine freeze (row=3 idle=2s), so the fix does not merely mute the alarm. - `bash tests/fm-wake-queue.test.sh`: exit 0, 38/38 pass, all five secondmate cases green. - `bin/fm-lint.sh`: clean (ShellCheck 0.11.0, actionlint 1.7.12). CHECKS NOT CAUSED BY THE CODE: the CI and "Require no-mistakes" runs (34152740610, 34152740604) both ended with conclusion `action_required` — workflow approval pending, not a test/build failure. Separately, `bin/fm-test-run.sh --check-coverage` exits 1 in this environment, but I confirmed by stashing my changes that it fails identically on the unmodified base tree (locale-related `comm: input is not in sorted order`); it is pre-existing and this change adds no new test file for the partition to account for. Changes are left uncommitted in the worktree
* no-mistakes(ci): Fixed the one real code defect behind the failing checks. ROOT CAUSE (Greptile P1, second round, on the head commit e1304e6): `secondmate_wake_stall_tick` in bin/fm-watch.sh identified the queue's drain position by the sequence number ALONE. The previous round changed the comparison from `-gt` to `-ne`, which handles a reprovisioned queue that restarts BELOW the recorded position, but not one that restarts ON it. A mate retired and reprovisioned under the same task id gets a fresh home whose wake-queue sequence counter restarts at 1 — and the retained parent progress marker very plausibly holds a low sequence too (a queue frozen on its first row records seq 1). Equal sequence ⇒ no reset ⇒ the brand-new queue inherits the retired generation's long-expired idle interval and emits a false `secondmate wake-loop stalled` on its very first observation. That is exactly the false-alarm class this PR exists to remove. FIX (smallest, removal-first): the file already defines the identity of a queue row once, as `row_key="$epoch-$seq"` (used for stall receipts, the stall marker, and the notify key). The progress marker's separate, weaker seq-only identity is removed: `row_key` is now computed once right after the row is parsed, stored in the progress marker, and compared with `!=`. Across generations the epoch differs (the new generation's rows are appended later), so no sequence collision can carry a stale interval; within a generation the key is stable exactly while the position does not move. bin/fm-wake-lib.sh's `fm_wake_secondmate_progress_marker_write` now takes `<oldest-row-key>` and validates it the same way the two neighbouring row-key writers do. Updated the function header comment and the two doc sentences (docs/architecture.md, docs/configuration.md) that described the old sequence-only semantics. REGRESSION TEST: `test_secondmate_reprovisioned_queue_starts_a_fresh_interval` in tests/fm-wake-queue.test.sh now drives the reported case — the reprovisioned generation restarts on the SAME sequence 9 (epoch 200) that the retired generation recorded (epoch 100), at a later clock — and asserts observable watcher output only. Its third leg still confirms the restarted generation escalates on a genuine freeze (row=9 idle=2s), so the fix does not merely mute the alarm. Three seeded progress markers in the symlink, crash-window and prefix-receipt tests were updated to the epoch-sequence form. VERIFICATION: - Fails before / passes after: with bin/fm-watch.sh and bin/fm-wake-lib.sh reverted to HEAD and the new test in place, the suite aborts on `not ok - a reprovisioned queue generation inherited the retired generation's idle interval and alerted` (exit 1); with the fix, `bash tests/fm-wake-queue.test.sh` exits 0 with 38/38 pass, all six secondmate cases green. - `bin/fm-lint.sh`: clean (ShellCheck 0.11.0, actionlint 1.7.12). - I also started `bin/fm-test-run.sh tests/fm-watch-checkpoint.test.sh tests/fm-watch-triage.test.sh tests/fm-watch-recovery-loop.test.sh` as a blast-radius check; it was still running when this phase had to return, so its result is not included. No other suite references the stall detector or the progress marker (grep over tests/ for `wake-loop stall|SECONDMATE_WAKE_STALL|secondmate-wake-progress` matches only fm-wake-queue.test.sh), and the changed lib function has exactly one caller. CHECKS NOT CAUSED BY THE CODE: the CI and "Require no-mistakes" runs on the head commit (34154091191, 34154091236, 34154091945) all ended with conclusion `action_required` — pending workflow approval, not a test/bui…
BenWilcox8
pushed a commit
to BenWilcox8/firstmate
that referenced
this pull request
Sep 12, 2026
…livery (kunchenguid#3852) * fix(bin): read preserved spawn state back before the interrupted exit claims it The deferred-signal exit path asserted the paired task record and In-flight backlog state were preserved without reading either back, exactly when a reader is least able to check (fm-yi4j evidence, 2026-09-05). The commit's exit status alone has been observed to agree with a row that did not actually move. The exit path now re-reads the record and the row under the same per-task lock as the commit, repairs a row the commit believed it moved, and phrases the error as exactly what was verified or attempted - verified preserved, repaired and verified, or an explicit preservation-could-not-be-verified with the reason and hand-closeout instruction. Two behavior tests drive a lying tasks-axi start through a real interrupted spawn and assert the printed claim and the real backlog state agree. * no-mistakes(test): Fix calm suite for Pi 0.85 and pin test umask * no-mistakes(document): Document interrupted-spawn preservation claim in backlog gate owner * no-mistakes(ci): Fixed the Greptile P1 in bin/fm-spawn.sh's deferred-signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change * no-mistakes(ci): Fixed the Greptile P1 in bin/fm-backlog-transition-lib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean * no-mistakes(ci): Fixed the Greptile P1 on bin/fm-backlog-transition-lib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean
eduardstan
added a commit
to eduardstan/firstmate
that referenced
this pull request
Sep 13, 2026
* fix(bin): wait out transient primary-checkout reads in the spawn worktree poll (#3834)
* fix(spawn): keep the worktree poll from adopting the repository primary
After `treehouse get` is sent, the worktree-discovery poll reads the pane's
foreground-process cwd. While treehouse is still fetching and checking a slot
out, the foreground process is treehouse itself and it reports the repository's
PRIMARY checkout as its cwd for several seconds. The poll accepted any path
that merely differed from the spawning project, so from a linked spawning home
- whose project is itself a worktree of that repository - it adopted the
primary, and the isolation guard then refused a launch whose slot treehouse
went on to create normally.
Screen every candidate with the isolation guard's own conditions, extracted as
spawn_worktree_isolated, so a read the guard would reject stays a transient the
poll keeps waiting through. The two-consecutive-reads rule and the guard as
final backstop are unchanged; a pane that never reaches an isolated worktree
still fails at the existing 60s deadline, now naming the last path it reported.
The already-settled timing assertion counted whole-spawn wall time against a
5s budget and failed on unmodified HEAD on slower machines; it now counts pane
reads, which is what "one confirming read, not an extra cycle" actually means.
* fix(spawn): say which path the worktree wait rejected, and why
Screening every discovery-poll candidate means a host that never reaches an
isolated worktree spends the whole 60s window before refusing. That wait is
deliberate - separating a transient from a terminal misconfiguration needs
machinery this path does not want - so the refusal explains itself instead:
the isolation check records why a candidate failed, and the deadline names the
last path seen together with that reason. Message and diagnostics only; the
poll's control flow is unchanged.
Two suites asserted the guard's wording on paths the poll now rejects rather
than adopts, so their refusal arrives from the deadline instead: realign
fm-tangle-guard's non-git and subdirectory-of-primary cases (each now also
asserting the stated reason, and the second the metadata absence it was
missing) and the herdr projection e2e's forced non-worktree cwd.
* no-mistakes(review): stub poll sleep in tangle-guard spawn isolation test
* no-mistakes(document): document spawn poll isolation screen in fm-spawn header
* test(spawn): make the non-git isolation case non-git anywhere
The refusal-reason assertion for a path outside any repository assumed TMPDIR
is not inside a git repository. Where it is, git walks up from the temporary
directory, finds that repository, and the spawn reports the subdirectory cause
instead - so the case passed or failed on a property of the host rather than on
the behaviour under test.
Build the path under a directory the test then names in GIT_CEILING_DIRECTORIES,
which git documents as not chdir-ing up into a listed directory while looking
for a repository. Git never excludes the directory being searched, so the
ceiling is the parent of the path handed to the spawn.
The assertions pin which cause fired rather than the sentence that explains it,
leaving the operator wording free to improve.
* no-mistakes(document): point spawn poll comment at the isolation screen's comparison
* no-mistakes(ci): Fixed the "Behavior portable serial 2" failure in tests/fm-tangle-guard.test.sh ("non-worktree spawn did not say why the path was rejected (missing: 'not inside a git worktree')"). Root cause, in this PR's code: bin/fm-spawn.sh's spawn_worktree_isolated resolved the git toplevel with `wt_top_real=$(cd "$SPAWN_WT_TOP" ...)`. For a path in no repository, `git rev-parse --show-toplevel` yields empty, and `cd ""` is a SUCCESSFUL no-op on bash before 5.3 (CI's ubuntu-latest ships bash 5.2). The empty toplevel therefore resolved to fm-spawn's own cwd — the CI checkout — so the poll reported "it is a subdirectory of worktree root '/home/runner/work/firstmate/firstmate'" instead of the correct "it is not inside a git worktree". Dev machines with bash 5.3 fail `cd ""`, which is why the suite passed locally and only failed on CI; it is a genuine shell-portability defect in the reason vocabulary this change added, not a test-environment artifact. Fix (smallest root-cause change, 1 line + comment, bin/fm-spawn.sh:2168-2173): guard the empty value so it never reaches `cd` — if [ -n "$SPAWN_WT_TOP" ] && ! wt_top_real=$(cd "$SPAWN_WT_TOP" 2>/dev/null && pwd -P); then No change to the poll's timing or deadline behavior (respecting the recorded refusal-latency and spawn-wt-reason-vocabulary decisions), no new tests, no other files touched. Verification: - Reproduced the exact CI failure locally by putting bash 3.2 (same `cd ""` semantics as CI's 5.2) first on PATH: fails before the fix with the identical message shape, passes after. - tests/fm-tangle-guard.test.sh passes under both bash 3.2 and bash 5.3. - tests/fm-spawn-worktree-settle.test.sh and tests/fm-spawn-pool-base-freshen.test.sh pass; shellcheck -x bin/fm-spawn.sh clean. - bin/fm-test-run.sh --changed: 46 suites completed, every FM_TEST_END exit=0, 1076 passing assertions, 0 "not ok" (including fm-tangle-guard, fm-control-relaunch, fm-lint). The run ended on my own 900s wall-clock cap (rc=124), not on any test failure
* fix: support stock macOS Bash 3.2 paths (#3732)
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): read orphaned green ci monitor and daemon-down failed record as not failed (#3846)
* fix(bin): read an orphaned green ci monitor as held-for-merge, not failed
A no-mistakes run held for a captain merge decision keeps its ci step
polling until merged or closed; when the shared daemon restarts under
that poll, the run is recorded failed although every substantive step
completed and GitHub reports the PR green. A monitor whose only
remaining job is to observe a human decision must not convert the
absence of that decision into a failure verdict.
fm-crew-state.sh now reclassifies a terminal failed run as done
(held-for-merge), surfacing the run's PR URL, when the steps table
shows every step completed except exactly ci failed and the ci log's
last recognized marker reads checks green. A genuinely red check, an
unreadable ci log, or a second failed step keeps the failure.
* no-mistakes(review): Read daemon-down coarse failed ledger as unknown, not failed
* no-mistakes(document): docs: align AGENTS.md failed-verdict guidance with crew-state reclassification
* fix(spawn): verify preserved backlog state after interrupted spawn delivery (#3852)
* fix(bin): read preserved spawn state back before the interrupted exit claims it
The deferred-signal exit path asserted the paired task record and
In-flight backlog state were preserved without reading either back,
exactly when a reader is least able to check (fm-yi4j evidence,
2026-09-05). The commit's exit status alone has been observed to agree
with a row that did not actually move.
The exit path now re-reads the record and the row under the same
per-task lock as the commit, repairs a row the commit believed it
moved, and phrases the error as exactly what was verified or attempted
- verified preserved, repaired and verified, or an explicit
preservation-could-not-be-verified with the reason and hand-closeout
instruction. Two behavior tests drive a lying tasks-axi start through a
real interrupted spawn and assert the printed claim and the real
backlog state agree.
* no-mistakes(test): Fix calm suite for Pi 0.85 and pin test umask
* no-mistakes(document): Document interrupted-spawn preservation claim in backlog gate owner
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-spawn.sh's deferred-signal exit path: during preservation verification, the no-op HUP/INT/TERM re-trap combined with an unresponsive `tasks-axi show`/`start` (bash cannot run traps while a foreground child runs) held the per-task meta lock - and every lifecycle operation waiting on it - indefinitely. Root-cause fix: bound every tasks-axi invocation made under the lock. bin/fm-backlog-transition-lib.sh gains fm_tasks_axi, an exec-based wrapper (GNU timeout, gtimeout fallback) used by fm_backlog_row_show and fm_backlog_mutate that preserves the exact process placement of the plain tasks-axi call; bin/fm-spawn.sh sets FM_TASKS_AXI_TIMEOUT (default 30s) at the commit point so both the commit and the read-back verification are bounded. A timed-out call fails through the existing error plumbing and probe/mutate name the timeout as the reason, so the interrupted exit path prints honest 'preservation could not be verified ... (reason)' wording - never intent phrased as outcome, matching the author's intent. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives a real interrupted spawn through a lying tasks-axi whose repair start never answers; it asserts the spawn exits promptly (self-bounded by an outer timeout), the attempted wording names the timeout, and the printed claim agrees with the real record/backlog state. Confirmed the test fails on the unfixed tree and passes with the fix. Verified: fm-backlog-atomicity (83 ok), fm-transition-lib, fm-backlog-handoff, fm-captain-hold, fm-teardown, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-spawn-batch, fm-spawn-dispatch-profile, fm-task-delivery, fm-control-relaunch all pass; bin/fm-lint.sh clean. The fm-bootstrap 'unsplit run lost its local diagnostic' failure reproduces on the pristine base commit and is unrelated to this change
* no-mistakes(ci): Fixed the Greptile P1 in bin/fm-backlog-transition-lib.sh: fm_tasks_axi bounded tasks-axi only through GNU timeout/gtimeout and fell through to an unbounded exec on hosts with neither (stock macOS), so an unresponsive call could hold the per-task meta lock forever during interrupted-spawn verification. Root-cause fix: the bound now has no unbounded path. GNU timeout is preferred, gtimeout next, then a small perl watchdog (fork + waitpid WNOHANG polling at 50ms, TERM on expiry, one bound of grace, then KILL, exit 124 so the callers' existing timeout plumbing reports it; exit statuses and output pass through unchanged). Polling was chosen over alarm+die to avoid perl's platform-dependent syscall-restart semantics. When a bound is requested but no bounding mechanism exists, the call fails closed (exit 127 with a diagnostic) rather than running unbounded, so the interrupted exit path prints honest attempted wording, never intent as outcome. The unbounded exec remains only for the no-bound plain-call case. Added three behavior tests in tests/fm-backlog-atomicity.test.sh driving fm_tasks_axi through a PATH with no timeout binary: a hanging stub must exit 124 within the bound (verified to fail on the pre-fix code), a failing stub's status/output must pass through, and a tool-less PATH must fail closed with the diagnostic. Verified: fm-backlog-atomicity 86/86 ok, fm-transition-lib, fm-backlog-handoff, fm-teardown, fm-spawn-batch, fm-task-delivery, fm-fleet-snapshot-view, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile all pass; bin/fm-lint.sh clean
* no-mistakes(ci): Fixed the Greptile P1 on bin/fm-backlog-transition-lib.sh: fm_tasks_axi's GNU timeout and gtimeout paths sent TERM at the bound but had no kill-after, so a tasks-axi that ignores SIGTERM kept the bounded call - and the per-task meta lock - held indefinitely during interrupted-spawn verification. Root-cause fix: both GNU execs now carry -k "$bound" (TERM at the bound, KILL after one further bound of grace), giving every bounded path the same forced-termination contract the perl watchdog already had. Because GNU timeout exits 137 (128+SIGKILL) when the kill-after fires - versus 124 for a TERM expiry - the probe/mutate timeout detection now goes through a new fm_tasks_axi_timeout_expired helper that treats 124 and 137 alike, so the interrupted exit path still names the timeout as the reason; the helper keeps the bound check in one place. Added a behavior test in tests/fm-backlog-atomicity.test.sh that drives fm_tasks_axi through a real GNU timeout with a tasks-axi stub that traps and ignores TERM (the ignored disposition survives exec into sleep) and asserts a bound-expiry status plus completion within bound+grace; on the pre-fix code the suite hangs until killed, confirming the reproduction. Verified: fm-backlog-atomicity 87/87 ok, fm-transition-lib, fm-backlog-handoff, fm-spawn-batch, fm-task-delivery, fm-teardown, fm-secondmate-reconcile, fm-control-relaunch, fm-spawn-dispatch-profile, fm-captain-hold-lifecycle all pass; bin/fm-lint.sh clean
* docs: correct runtime-backend maturity labels for Herdr (#3821)
* docs: correct stale tmux/herdr backend maturity claims
Herdr now has 21 test files, its own required CI job (tests-herdr)
that installs a pinned build and hard-fails on "skip: herdr not
found", while tmux has 3 test files and is only required as a
dependency of the portable-serial e2e lane. zellij, orca, and cmux
still have no CI lane at all. AGENTS.md and docs/herdr-backend.md
still called Herdr merely "experimental" alongside those three,
misleading every session and reader about actual coverage.
Update AGENTS.md's config/backend entry, the opening lines of
docs/herdr-backend.md and docs/tmux-backend.md, the runtime-backend
section of docs/configuration.md, and the matching claims in
docs/architecture.md, CONTRIBUTING.md, and README.md so they agree
and distinguish tmux (default), herdr (own required CI lane, largest
suite, Windows still spike-only), and zellij/orca/cmux (still
experimental, no CI lane). No behavior, selection order, or
dispatch logic changes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GDYsWPEfwTuPNjQ2nGBCcj
* no-mistakes(review): docs: fix stale herdr label and CI-lane wording
* no-mistakes(document): docs: align tmux adapter label in scripts.md
* no-mistakes(review): docs: drop duplicated herdr CI claim from tmux page
* no-mistakes(review): docs: drop windows claim, align contributing backend wording
* no-mistakes(review): docs: trim duplicated CI claim from herdr opening line
* no-mistakes(review): docs: drop unguarded largest-test-suite superlative
* no-mistakes(review): docs: restore tmux verified label and README experimental scope
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: restore intent-targeted no-mistakes validation (#3865)
* fix: delete deterministic no-mistakes test baseline, restore intent-targeted Test
PR #3644 pinned commands.test to a fm-test-run.sh --changed walk of the
repository's 75-162 tests/*.test.sh scripts. no-mistakes runs commands.test
verbatim and unconditionally after every fix round, so that walk multiplied
by round count: measured at 32.7 minutes per validation versus 3.6 minutes
intent-targeted. Delete the pin and restore the 3.6-minute posture.
Add tests/fm-nm-test-contract.test.sh as a regression guard, parsing
.no-mistakes.yaml as YAML (ruby's bundled Psych, matching the parser
tests/fm-test-run.test.sh already uses for ci.yml) rather than grepping its
text, restoring in legal form what PR #823 added and PR #1282 removed.
Record the rule in docs/configuration.md's "Gate defaults" section (the
authoritative owner CONTRIBUTING.md already points at) and strengthen
CONTRIBUTING.md's existing local-Test guidance to state it plainly: never
configure commands.test to a deterministic test command, complete or partial.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSt9JvrQMVc4u3jPyUFCFC
* no-mistakes(review): Centralize no-mistakes test policy and narrow guard
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix: verify Treehouse slot ownership before teardown (#3837)
* fix(bin): verify pool-slot ownership before returning a worktree slot
Workers were killed when cleanup returned a Treehouse pool slot that a
different, live task had already taken. Teardown now proves the slot is
genuinely this task's before releasing it: it refuses when another task
record claims the same live worktree path, or when the endpoint's working
directory contradicts the recorded slot, and that refusal holds under
--force. Slot allocation, metadata publication, ownership verification,
and slot return are serialized across linked firstmate homes, and forced
secondmate cleanup verifies descendant slot ownership before returning
any child worktree.
Regression coverage drives the scripts with two task records naming one
slot path and asserts the live worker survives and its slot is not reset.
* no-mistakes(review): Protect slots across cloned Firstmate homes
* no-mistakes(test): Gate teardown locking on genuine Treehouse slots
* no-mistakes(test): Clarify pooled descendant slot gating
* no-mistakes(test): Synchronize watcher re-arm test on process exit
* no-mistakes(test): Wait for watcher cleanup before timeout escalation
* no-mistakes(document): Document pool-slot ownership safeguards
* no-mistakes(ci): Fixed all reported CI issues: normalized bare local Git origins to the same Treehouse project-lock identity as absolute clone origins; resolved ShellCheck SC1091 with explicit conditional sourcing; and taught concurrent Herdr teardown coverage to retry expected Treehouse lock contention. Added behavioral regression coverage for bare/absolute origin lock identity. Verified endpoint-safety tests, watcher tests, full CI lint, and the previously failing Herdr teardown assertion
* fix(bin): resolve relative origins from repository root
* no-mistakes(ci): Fixed teardown so an exact recorded endpoint may change cwd without falsely vetoing cleanup. Removed cwd-based ownership refusal while preserving cross-home record exclusivity and project locking. Updated behavioral coverage for both foreign slot ownership refusal and moved-cwd teardown success. Endpoint-safety, backend, watcher, checkpoint, and targeted lint checks pass. Real Herdr presentation E2E progressed successfully but exceeded the 600s local timeout
* feat(bin): add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary (#3867)
* feat: add verified omp (Oh My Pi) harness adapter for crew, secondmate, and primary
Add omp as a verified harness: anchored process-name detection with a
Firstmate-owned FM_OMP_HARNESS launch marker that needs real omp ancestry,
the fm-spawn launch template with foreign-marker clearing, the tracked
.omp/fm-worker-overlay.yml posture overlay, --auto-approve, --cwd, and
pre-launch model validation scoped to providers 'omp models --json' lists.
Workers get a state-resident busy-state extension keyed on agent_end
without willContinue (omp has no agent_settled). The primary gets two
tracked .omp/extensions: a turn-end guard that answers omp's blocking
session_stop hook by compelling one continuation per turn, with the
pre-tool seatbelts and Run-tier session-start delivery, and a watcher
extension ported from the Pi one with fm_watch_arm_omp. Control tables,
composer busy footers, omp's status row as a bare-composer boundary, the
extension supervision model with an omp-keyed ownership proof, the
session-start diagnostic, and the supervision protocol snippet follow.
Verified live on omp 18.1.11 with openai-codex/gpt-6-astra: a Herdr scout
through spawn, busy state, steer, interrupt, exit, and teardown, and the
isolated rpc primary lab through extension auto-discovery, digest
delivery, lock identity, watcher arm, successor and wake delivery, and
the compelled guard continuation.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test: prove the omp guard continuation through a guard spy
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* fix(spawn): clear the gemini marker at the omp launch boundary
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): force the guard stage by freezing the watcher and clear lint findings
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): reap the live lab by path and record omp's rpc shutdown as a note
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* test(omp): spawn a real secondmate for the discovery rule and classify the omp surfaces
Replace the template-extraction check with a genuine --secondmate launch
pinned to the fake tmux backend, assert the worker extension's handler set
through the executable rather than its bytes, classify the two new omp
surfaces in the documentation inventory, and record the Herdr worker
evidence in the runtime-backends verification doc.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012AMyYaHU42Ltotn6fauPAh
* no-mistakes(review): omp: unverify remote routes, narrow busy regex, drop overlay approval pin
* no-mistakes(review): omp: validate config-pinned model, correct remote and marker docs
* no-mistakes(review): omp: pin config-model validation with a test, trim overlay
* no-mistakes(review): omp: sync guard evidence, drop dead param, map quota family
* no-mistakes(review): omp quota: refuse unmapped prefixes, match bare model scopes
* no-mistakes(document): docs: cover omp in cd-guard, quota, continuity, tmux
* no-mistakes(document): docs: add omp subagent-guard row, fix live test header
* no-mistakes(ci): Fixed both failing behavior shards and the Greptile P1 in bin/fm-composer-lib.sh. Root cause of "Behavior portable serial 1" and "Behavior portable parallel 2": the omp busy regex (FM_DELIVERY_OMP_BUSY_REGEX_DEFAULT) and omp status-row furniture regex (FM_COMPOSER_OMP_STATUS_RE_DEFAULT) used the bracket range [⠁-⣿]; BSD grep on macOS accepts it but GNU grep on Linux CI aborts with "Invalid collation character", failing every omp busy/furniture read (3 assertions across fm-omp-harness, fm-tmux-submit-busy, fm-composer-lib). Replaced the range with one shared explicit alternation FM_OMP_SPINNER_FRAMES_RE of omp 18.1.11's unicode-preset spinner frames (status set ⣾⣽⣻⢿⡿⣟⣯⣷ + activity set ⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏, read from the installed binary), the same pattern the Kimi busy regex already uses in CI. For Greptile's finding (the harness-agnostic furniture rule's first alternative matched any 1–4-byte token + ' · ', so wrapped typed input like 'fix · tests' with the cursor on it regressed from pending to unknown; reproduced locally vs base), pinned that alternative to omp's identity cell (π||pi, the icon.omp of each preset in the 18.1.11 binary). Tests: fm-composer-lib.test.sh asserts 'fix · tests' is not furniture, a status-set spinner row is furniture, and the wrapped composer screen reads pending under both locales (CAPS_TMUX cursor 3); fm-omp-harness.test.sh asserts a status-set frame reads busy. New negative cases fail against the pre-fix lib and pass after. Verified: fm-omp-harness, fm-tmux-submit-busy pass via bin/fm-test-run.sh; fm-composer-lib passes all cases except one pre-existing, unrelated local failure (Herdr half-block test uses printf '▀', unsupported by macOS bash 3.2; fails identically on a pristine HEAD export, passes on CI bash 5); shellcheck and bin/fm-lint.sh clean. Caveat: GNU grep is unavailable locally, so the Linux compile was not run directly; the fix uses only constructs already proven on CI's GNU grep (multibyte literal alternations, incl. under LC_ALL=C). Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh, tests/fm-omp-harness.test.sh. No docs needed changes (they describe the rule generically)
* no-mistakes(ci): Greptile Review: fixed. The omp status-row furniture regex FM_COMPOSER_OMP_STATUS_RE_DEFAULT in bin/fm-composer-lib.sh still accepted a literal `pi ·` opening, so wrapped composer input beginning with `pi ·` was truncated and misclassified. Read the installed omp 18.1.11 binary: the ascii preset's `icon.omp` is `pi` but its `sep.dot` separator is ` - ` (unicode/nerd use ` · `), so a real ascii status row never contains `pi ·` and that alternative could only ever match typed text. Removal-first fix: dropped `pi` from the identity alternation (now `(π|)`) and updated the comment to record why the ascii preset is excluded. Tests (tests/fm-composer-lib.test.sh): added a negative furniture case for 'pi · e · phi as the three constants' and a wrapped-screen assertion (CAPS_TMUX, cursor 3) that a continuation row opening `pi ·` reads pending in both locales; the new case fails against the unfixed lib and passes after. Verified: composer test with the half-block case skipped passes all 33 cases including the omp matrix; bin/fm-test-run.sh tests/fm-omp-harness.test.sh passes; shellcheck -x clean on both files; bin/fm-lint.sh clean. The full composer test via the runner fails locally only on the pre-existing half-block case (bash 3.2 printf cannot emit ▀; passes on CI bash 5), identical to before this change. Docs unchanged (they describe the rule generically and never mention the ascii identity cell). PR must be raised via no-mistakes: not caused by code. attestation.head_sha is cdddc60 while the PR head is cd51cf4 because the pipeline's ci-phase push moved the head; the outer executor's re-push will re-bind the attestation. No file change for that check. Files changed: bin/fm-composer-lib.sh, tests/fm-composer-lib.test.sh
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* fix(pi): invoke Bash helpers correctly on native Windows (#3843)
* Fix Pi shell invocation on native Windows
* no-mistakes(document): Document Pi Windows Bash transport
* no-mistakes(ci): Captain, staged a narrow fix: register the Pi Windows regression for both extension paths, make Windows mode emulation non-failing, and enforce LF shell checkouts. Mapping and coverage checks pass; CI/Require no-mistakes were approval-gated externally
* no-mistakes(review): Cover async Windows branch-outcome Bash invocation
* no-mistakes(review): Preserve Cygwin checks and refresh Windows timing
* no-mistakes(test): Invoke OpenCode operational-input owner through Bash on Windows
* validation-fixture
* no-mistakes(document): Document Windows Bash helper invocation
* no-mistakes(ci): Fixed PR-caused changed-selection failure by removing the malformed tracked evidence artifact and allowing deleted, unconsumed source paths to retire cleanly while preserving fail-closed behavior for live unmapped paths. Added regression coverage. Verified native-Windows Pi shell-seam test passes and --changed selects the Windows regression
---------
Co-authored-by: test <test@example.invalid>
* test(bin): pin teardown outcomes for squash-merged rebased branches (#3870)
* fix(bin): recognise squash-merged rebased work as landed at teardown
A pipeline rebase can leave the local worktree on pre-rebase commits while
GitHub squash-merges the rebased head. The landed-work test then compared
those stale commits against a squashed main and refused cleanup of work
that had already landed.
When the forge reports the recorded PR merged and its merge commit is on
the default branch, treat a local branch that only repeats paths from the
pipeline push as stale rather than unlanded. If the forge is unreachable,
the same coverage check runs against a PR head whose content is already
on default. Extra local paths still refuse.
* fix(bin): drop unprovable squash-rebase landed-work coverage
Path-set coverage treated a diverged local branch as landed whenever it
touched the same files as the squash merge. That accepts the reviewer's
failing sequence: same path, different content, work discarded.
git cherry and merge-tree containment were already too strict on the real
rebase-fold case. No remaining check is both safe and permissive enough
to recognise a stale pre-rebase copy without also accepting unlanded
edits, so that case still refuses.
Keep the proofs that hold: a merged PR head that contains local work, or
a clean content-in-default tree match. Tests now refuse same-path
different content and extra unlanded commits, and still allow a local
branch that followed the pipeline rebase.
* no-mistakes(review): drop recorded-pr-head fallback and reverted-design leftovers
* no-mistakes(review): silence squash-merge stdout corrupting test PR head
* no-mistakes(review): make unlanded follow-up commit sole cause of refusal
* no-mistakes(document): correct stale squash-rebase fixture comments in teardown tests
* no-mistakes(ci): Split the three reported checks: - CI (run 34061098467) and Require no-mistakes (run 34061098460) both concluded `action_required` — approval-gated workflow runs that never executed a step. Not caused by this PR's code; no change can clear them. - Greptile Review was a genuine defect in the new tests: the three new refusal cases (tests/fm-teardown.test.sh) asserted only exit status 1 and a REFUSED line, so a teardown regression that destroyed the worktree, branch, and task record before reporting refusal would still pass. Fix (tests only): added one `assert_refusal_retained_task_state` helper and called it from `test_squash_merged_same_file_different_content_refuses`, `test_squash_merged_rebased_local_with_unlanded_commit_refuses`, and `test_squash_merged_stale_local_refuses_when_forge_unreachable`, each capturing the worktree HEAD before `run_teardown`. It pins that the refusal left the isolated copy on disk, the task branch still checked out at the same unlanded commit, and state/task-x1.meta intact. Verification: the four squash tests pass; a sensitivity probe ran the ALLOW fixture (teardown completes) and pointed the same helper at the outcome — it fires, because a completed teardown detaches/deletes the branch and removes the task record, proving the assertions discriminate. Full tests/fm-teardown.test.sh: 83 passing. bin/fm-lint.sh clean with pinned ShellCheck 0.11.0 + actionlint 1.7.12 (plus an explicit --external-sources pass on the changed file). bin/fm-test-run.sh --check-coverage ok. Caveat: test_herdr_flat_teardown_preflight_refuses_before_changes (mode missing-adapter) fails on this machine. Verified it fails identically on base commit f91a950 via `git archive`, so it is a pre-existing local environment difference untouched by this diff; skipped to run the rest of the suite, not modified
---------
Co-authored-by: Morten Gad <mogad@itm8.com>
* fix(bin): keep supervision armed for registered custom checks (#3860)
* fix(bin): keep supervision armed for registered custom checks
A custom check bound by bin/fm-check-register.sh only ever runs inside the
watcher's check sweep, but fm_supervision_status counted in-flight tasks, the
relay poll shim, and process-event sources as supervision need, and not
registered checks. Tearing down the last task therefore stopped every
home-level check silently until the next spawn.
Count a state/<id>.check.sh that carries its state/<id>.check-trust binding as
supervision need. The relay shim keeps its own trust path and task PR polls
carry no such binding and are torn down with their task, so neither arms a home
by accident. Presence of the binding is the whole test: the sweep validates the
bytes at execution time and wakes firstmate when it rejects one, which is the
outcome an idle home needs.
Closes #3856
* no-mistakes(review): name registered checks in turn-end block banner and doc invariant
* no-mistakes(review): narrow PR poll predicate test to what it proves
* no-mistakes(document): point Grok re-arm step at supervision-need owner
* fix(bin): resolve Treehouse locks for remote secondmate homes (#3883)
* fix(bin): resolve the shared Treehouse project lock inside remote secondmate homes
Every spawn and teardown inside a remote-seeded secondmate home refused,
because the project lock's anchor could not be resolved there.
fm_firstmate_root_home walks a home's parent bindings upward to find the
anchor the lock lives in, and treated a remote parent binding as an error.
A remote-seeded home's parent is on another machine, so that walk can never
succeed from there - and neither can the home's own local descendants, whose
chain terminates at the same record. Both fail closed on every Treehouse-backed
spawn and every pool-slot teardown.
A remote parent now terminates the walk at the home holding it, which is the
correct anchor: a lock taken on this filesystem is neither held nor observable
across that boundary, and that home is already the top of the local tree
teardown's collect_local_firstmate_states enumerates, since that walk skips
remote registry entries for the same reason. Mutual exclusion is unchanged -
every home reachable through local parent links still derives one identical
lock file per project, and an unreadable binding, an unsupported route, an
unreachable local parent, a cycle, and an over-deep chain all still refuse.
Origin-less local-only projects keep resolving through their worktree top.
Regression coverage pins the anchor for the main-home layout, a local
secondmate, a remote-seeded home, and its local child; drives teardown
end-to-end in a remote-seeded home; keeps the cross-home slot-ownership
refusal across that boundary; and proves two homes still serialize on the
one shared lock file.
* no-mistakes(document): Clarify machine-local Treehouse lock ownership
* fix(bearings): repair board listening and decision reconciliation (#3872)
* fix(bearings): repair the board's listening, card hygiene, and reconcile path
Three defects made the fleet board go quiet and then lie about what still
needs the captain.
Never arm a poll on a session that is not live. `lavish-axi <file>` exits 0
even when it refuses to reopen a session the captain ended from the browser,
reporting `status: user-ended` with the same session id, so the build's
exit-status check accepted a dead session, printed `already-armed`, and left
the board reading "not listening". The build now proves the session is live
from a fresh authoritative listing immediately before arming - not from the
establish call's status alone, which is already stale by then - reopens once
when it finds the session ended, and refuses rather than arming when it stays
ended. A reopen also replaces the pre-reopen source generation before
reporting success, so a runner on its way out cannot be mistaken for a
listener, and a board whose source is registered but unowned gets a
replacement started before the build returns.
Let a dead generation's ownership actually move. Reclaiming a claim ran its
capture-reservation cleanup first, and that cleanup re-verifies the recorded
state-root identity, so a claim naming a pid and a process group that were
both provably gone could not be cleared: reconcile reported a start while
nothing attached, and retire refused with "cannot release source ownership".
Reservation records are keyed by claim token and every replacement claims a
fresh one, so they are hygiene, not an ownership invariant. Reclamation now
additionally requires the owning process group to be absent independently,
which keeps a reused pid whose poll child still runs from ever reading as a
gone generation. A live owner and a crashed leader whose owned group survives
are still never reclaimed.
Stop carding decisions whose subject already landed. The build drops a
decision card whose work item or PR appears in the payload's own landed rows,
and one whose task is no longer an open captain call, naming each drop on
stderr. A task whose state cannot be established is kept, because a call
wrongly hidden is worse than a card wrongly shown.
Add the reconcile choice, and make it structurally incapable of closing a
call. Every decision card carries a standard `reconcile` option, injected by
the build rather than left to the composer. The board now emits the picked
option and any freeform note as separate structured fields instead of fusing
them, so a reconcile selection is not expressible as an answer value at all -
the defect that let `reconcile - <note>` reach the intake as an ordinary
answer. The adapter routes selections from that structured field, creation of
a reconcile request is bound to a verified board source rather than the shared
keyed-answer intake, and the intake still refuses the reserved value on every
channel. Each authorization is bound to the captain-hold generation that
produced the card, so an obsolete card cannot close a later call, and both
terminal outcomes require a pending request: `reconcile close` records the
evidence under its own `reconciled` mode so it never reads as the captain's
words, and `reconcile note` leaves the call open. Anything unprovable -
an unversioned row, a missing generation, an unreadable state - refuses
rather than acting.
Regression coverage fails without each fix, and pins every leak path: a bare
reconcile, a standalone close or note with no pending request, an any-channel
reconcile, an annotated selection from a freeform card, and a
generation-skewed authorization. An opt-in guard re-proves the lavish-axi
shapes and the reopen against the installed tool.
* fix(bin): quote the done comparison in the reconcile intake
shellcheck SC1010 reads the bare word as the loop keyword. The failed run
never reached its lint step, so this shipped in the recovered content.
* no-mistakes(review): Publish reconciled parent resolution before request retirement
* no-mistakes(review): Clarify committed cleanup and reconcile reservation scope
* no-mistakes(review): Preserve remote cards and legacy answer compatibility
* no-mistakes(test): Separate live claim release from stale reclamation
* no-mistakes(test): Allow terminal self-retirement during active capture
* no-mistakes(document): Document Bearings repair contracts
* no-mistakes(ci): Stabilized the failing Herdr presentation E2E by serializing test-harness Treehouse allocator calls, preventing concurrent recovery spawns from claiming the same pool slot while preserving Herdr concurrency coverage. Verified with the full E2E suite on Herdr 0.8.2, bash syntax checks, ShellCheck, and git diff checks
* fix(bin): allow pooled spawns without a git origin (#3885)
* fix(bin): skip pooled-worktree freshness fetch when no origin is configured
An origin-less local-only project has nothing remote to be stale
against, so fm-spawn's freshen_spawn_worktree_base refused to launch
crews for it. Detect a missing origin remote and skip the fetch
freshness gate entirely; an existing-but-unreachable origin keeps
refusing as before.
* no-mistakes(review): Preserve pool safety for absent and unusable origins
* no-mistakes(review): Refuse empty origin configurations during pooled spawn
* no-mistakes(review): Detect empty origin sections across config includes
* no-mistakes(review): Honor globbed includes when detecting origin configuration
* no-mistakes(review): Document conservative conditional include handling
* no-mistakes(review): Use Git-resolved config files for origin detection
* no-mistakes(review): Document included empty-origin detection boundary
* no-mistakes(document): Document originless pooled spawn behavior
* feat(tests): run live harness guards by default when available (#3889)
* feat(tests): run live harness guards by default where the harness is installed
The 24 live-harness guards each opened with their own env check, so on the
machine that has every harness - the one the product and its validation
actually run on - all of them skipped and passed. Fourteen had never been run
by the pipeline at all.
tests/lib.sh gains fm_live_gate as the single owner of that decision: a guard
that spends no model tokens runs wherever its tools are installed, a guard that
submits prompts stays opt-in, an absent tool is a named capability skip, and a
guard's own variable or FM_LIVE forces it on (turning an absent tool into a
failure) or off. Every live guard now opens with it, which also carries the
test-suite gate-refusal bypass into the guards that never sourced the shared
helpers and were therefore refused whenever a gate agent ran them.
bin/fm-test-run.sh records what a skip means: the family's expected class is
live-capability rather than a bare env opt-in, and each gate skip's reason is
logged and written to the timing artifact, so a lane can say which tool this
host could not exercise.
Only the token-free guards flip to default-on: composer-matrix, the harness
liveness drift guard, and the Herdr version floor. cursor-primary submits three
prompts, so it stays opt-in.
Running the drift guard unasked immediately found a real defect it existed to
catch: it resolved the harness through a generic `command -v cursor`, which on
a machine that also has the Cursor editor finds the editor launcher rather than
cursor-agent. That binary exits at once, leaving a bare shell in the pane and a
liveness-drift failure no classifier change could fix. It now asks
fm_cursor_resolve_binary first, the same verified owner fm-spawn uses.
CI installs the public Pi package in the portable serial lane and fails on its
skip token, so the Pi extension tests stop passing silently against a package
that is not there. No secret is added.
Verified on macOS 26.5.2 arm64: the drift guard runs with no variable set and
classifies 8 installed harnesses alive; the Herdr version-floor guard runs by
default and checks 4 real releases; every live guard refuses together under
FM_LIVE=0.
* fix(tests): keep the composer-matrix guard opt-in
Running it unasked is red on a healthy machine for reasons no code change here
removes: a harness that has not trusted this checkout sits on its own trust
dialog, which the guard treats as an unreadable composer and correctly fails.
The opencode 1.18.29 and grok 1.0.13 composer drift it also surfaced reproduces
identically on main and is filed as separate work.
So this token-free guard stays opt-in with the reason stated in its header, and
the coding guidelines record the narrow exception: a guard whose verdict
depends on host state that installing its tools does not establish may stay
opt-in, because one that is permanently red is one the fleet learns to ignore.
The other two token-free guards keep running by default.
* no-mistakes(review): Wire bearings guard and remove composer exception policy
* no-mistakes(review): Run Pi responsiveness guard by default
* no-mistakes(review): Gate AFK Pi Herdr through authoritative family sweep
* no-mistakes(review): Sanitize live gate test environments
* no-mistakes(document): Document default-on live guard behavior
* no-mistakes(ci): Fixed CI by installing the Pi package in portable-parallel-1, where fm-pi-primary-types.test.sh runs, and enforcing its package-missing gate skip there. Verified with fm-lint.sh, workflow actionlint, coverage partition checks, lane membership, and git diff checks
* no-mistakes(ci): Fixed CI’s Pi typecheck skip enforcement by giving npm, tsc, and Pi-package capability skips a shared prefix and configuring both relevant CI lanes to fail on that prefix. Verified missing tsc emits the expected skip, missing Pi package becomes a runner failure, and actionlint, ShellCheck, and git diff checks pass
* fix(bin): refuse test runs in the primary checkout when a task marker is set (#3891)
* fix(bin): refuse the behavior suite in the repository primary checkout
A task worker's isolated worktree placement is verified exactly once, when
its task starts, and nothing re-checks it afterwards. A worker that later
changes directory into the repository's primary checkout runs its Git
commands, and this branch-switching suite, against the one checkout every
linked worktree resolves against and every landing merges into. A run that
dies mid-suite can leave that checkout on a stray branch.
bin/fm-test-run.sh now refuses that case. When FM_TASK_ID marks a task
worker and the runner resolves to the primary checkout, every executing mode
exits non-zero before selecting a suite, with one line naming the primary
path and pointing at the assigned task worktree. The predicate is the one
bin/fm-spawn.sh already uses for launch placement: the working tree's own
git dir is the repository's common git dir, which separates the primary from
every linked worktree even when their top levels differ. A run with no
FM_TASK_ID set is unchanged, and so are the inspection modes, which execute
nothing. When git resolves neither directory - a non-repository fixture, a
detached copy - nothing proves this is the primary, so the run proceeds.
bin/fm-spawn.sh sets the marker: ship and scout launches export FM_TASK_ID
into the pane shell on the same pre-launch channel as GOTMPDIR, and the name
joins the sanitized launch environment allowlist so an isolated launch keeps
it.
* no-mistakes(review): clear inherited task marker in test lib; name resolved ROOT
* no-mistakes(document): docs: record FM_TASK_ID marker and runner placement refusal
---------
Co-authored-by: Talon Stark <talonstark@gmail.com>
* fix(bin): bound stale alarms for backlog captain holds (#3842)
* fix(bin): bound a stale alarm with the backlog hold, not only the status line
A legitimate wait has two records and the stale alarm reads only one.
`status_is_paused_or_captain_held` takes a status line, so it sees a wait the
worker declared. It cannot see the wait firstmate records when it hands work to
the captain: `bin/fm-captain-hold.sh hold` writes that into the backlog and
leaves the status log alone, so a delivered task keeps `done: PR ...` as its
last line for the whole time the captain is deciding.
Both stale branches were blind to it, and each churned a new pane hash back into
its own alarm: a `done:` line is captain-relevant and reaches the terminal-stale
branch, while a held task whose last line is `working:` reaches
`surface_nonterminal_stale` and fails its declared-wait test.
Consult that second record where the watcher is about to alarm, through
`bin/fm-captain-hold.sh open`, which already owns the predicate's semantics, and
bound the alarm on the shared `.paused-resurfaced-<key>` marker and
`PAUSE_RESURFACE_SECS` window the declared-wait absorb already uses. The first
sight still alarms, the window's end alarms once more, and a held crew that goes
genuinely silent still escalates through the wedge timer.
Only an established open captain call bounds anything: an unreadable backlog, an
absent or incompatible tasks-axi, a row this home does not carry, and every task
with no hold keep alarming exactly as before. The backlog hold is deliberately
not recorded as a declared pause, because the loop-top reconciliation and
`pause_state_class` both read the status line and would clear a flag that line
does not support.
Extends the fix in #3443, which closed the forms of this loop that the status
line itself can express.
* fix(bin): identify the captain call a stale alarm is bounded by
Three gaps in the bound added by the previous commit, all in how the throttle is
scoped and where the backlog is consulted.
The scope carried only the status-log signature. A task can be held, answered
with `--release`, and re-held as a genuinely different captain call without any
status append, so the second call inherited the first one's marker and its first
sight was absorbed - the one thing this bound must never do. The task id is not
the call: `bin/fm-captain-hold.sh open` gains `--identity`, which reports the
call's own lifecycle - its hold-set stamp and the number of recorded answers -
on an exit 0 and only then, leaving the silent predicate every existing caller
reads unchanged. The throttle scope now carries that identity.
The terminal path recorded the throttle before publishing the durable wake. A
failed append exits the watcher with nothing queued, and the next sighting then
read that fresh marker and absorbed the retry, turning a delayed alarm into a
lost one. Recording moves behind the append, as the non-terminal path already
had it, and the comment claiming the marker could not outlive its wake is gone
because it was false.
The backlog was consulted only on a new terminal pane hash. A captain call can
open after a hash was absorbed as provably working, changing neither the pane nor
the status log, so nothing re-read the backlog and the wedge timer kept firing
possible-wedge alarms through a legitimate wait. That timer now consults the call
at its own alarm boundary and takes the same bounded cadence - and only at that
boundary, so an ordinary repeat poll under the bound stays the local-only read it
was.
Regression coverage for each, all driving churn through one watcher process
rather than relaunching per pane change: relaunch cost dominated the earlier
shape, and an absorbing watcher stays in its poll loop across churn in production
anyway. An unheld task still alarms on every new hash, and an elapsed wedge timer
with no open captain call still escalates as a possible wedge.
* fix(review): Compose stale throttles with captain-call lifecycle identity
* fix(review): Preserve bounded same-hash captain-call resurfacing
* revert(bin): narrow the captain-hold stale bound to its observed defect
Lifts the lifecycle-identity and cadence-ownership work back out, leaving the
change at the shape that matches the defect actually observed: the stale alarm
did not consult the backlog captain hold, on either stale branch.
Reviewing the wider version surfaced a series of adjacent gaps in the watcher's
alarm state machine - a call opening after the first alarm, marker invalidation
at the hold lifecycle boundary, and which deadline a terminal timer represents.
They are real, but fixing them turns a small extension into a state-machine
change to the alarm path, which is a different review on a subsystem that is
being actively reworked. They are named as known limitations rather than carried
here, and none of them is load-bearing for what remains: the bound does strictly
less than the reverted version, leaves the wedge path escalating on
STALE_ESCALATE_SECS exactly as before, and introduces no silence that the
existing terminal-alarm path did not already have.
Kept from the reverted work is the record-after-append ordering, because that is
a defect in the code being shipped rather than an adjacent one: recording the
cadence marker before publishing the durable wake let a failed append lose an
alarm outright instead of delaying it.
History is preserved: the earlier commits stay on the branch and this removal
sits on top of them.
* fix(review): Document secondmate captain-hold scope boundary
* fix(document): Document captain-hold stale alarm scope
* fix(bin): bind the stale throttle to the captain call, not the status log
The throttle this change introduces was scoped to the task's status-log
signature. Answering a call with `--release` and holding the task again creates a
genuinely different captain call without necessarily appending to that log, so
the second call inherited the first one's marker and its first sight was
absorbed.
That is the one alarm this bound must never swallow. A delivery announced twice
is noise; a decision waiting on the captain that is never surfaced is invisible,
because nobody asks for what they do not know to ask for.
Measured rather than assumed, on the same fixture - a delivered task held for the
captain, released, and re-held with no status append, driven through bin/fm-watch.sh:
base c499f84 call-1 first=ALARM call-1 churn=ALARM new call first sight=ALARM
before this fix call-1 first=ALARM call-1 churn=absorbed new call first sight=absorbed
after call-1 first=ALARM call-1 churn=absorbed new call first sight=ALARM
Base never suppresses the new call, so the suppression came from this change and
closing it completes the fix rather than widening it.
`bin/fm-captain-hold.sh open` gains `--identity`, printing the call's lifecycle -
its hold-set stamp and count of recorded answers - on an exit 0 and only then, so
the silent predicate bin/fm-teardown.sh reads is untouched. The throttle scope
carries that identity beside the status signature.
The sibling case was measured too and is NOT included: on the status-declared
path, where the last line is `captain-held:`, base already absorbs a re-held
call's first sight. That behaviour predates this change and stays documented as a
known limitation rather than repaired here.
* fix(document): Document captain-call throttle lifecycle scope
* fix(ci): isolate the Herdr restart fixtures from a claimed worktree
The Herdr behaviour test intermittently reused a local worktree still claimed
by an earlier fixture after a restart.
The restart scenarios now use an isolated Treehouse project.
The full Herdr test passes on Herdr 0.8.2; bash -n and git diff --check pass as
well.
* feat(pi): resolve extension-registered providers in the supervision branch (#3871)
* Let the supervision branch resolve extension-registered providers
The isolated branch ModelRuntime cannot see providers an extension
registered into main's runtime at run time, so a pin on pi-devin-auth's
devin/swe-1-7 (or an unpinned branch following a main session on devin)
failed with "unavailable to the isolated branch runtime".
Capture main's ModelRegistry alongside mainModel and copy each
extension-registered provider config into the branch runtime at
model-resolution time. The config carries the provider's own streamSimple
and oauth wiring by reference, so the custom gRPC transport reaches the
branch unchanged instead of being reimplemented. The /supervision-model
picker uses the same copy so those models are offered.
Update configuration.md and pi-supervision-branch.md, which previously
stated extension-registered providers were not offered.
* no-mistakes(document): docs: own devin provider carve-out in branch architecture doc
* no-mistakes(ci): Fixed the Greptile P1 finding: the /supervision-model picker copied extension-registered providers into the branch ModelRuntime but checked hasConfiguredAuth without refreshing them, so providers with provisional post-registration auth were omitted from the picker while the pin-resolution path (which did refresh) accepted them. Root-cause fix in .pi/extensions/fm-branch-supervision.ts: moved the `refresh({ providers, allowNetwork: false })` call into `copyExtensionProviders` (now async, refreshing every provider it copied) and removed the duplicate per-provider refresh from `resolveBranchModel`. Both the picker and the resolution path now share one copy-and-refresh step, so hasConfiguredAuth is real in both. Regression coverage in tests/fm-pi-branch-extension.test.sh: the stubbed ModelRuntime now mirrors the real runtime by leaving a registered provider's auth pending until `refresh()` runs for it. With that stub, the existing extension-registered-provider case fails against the pre-fix extension (picker offers only anthropic/main-model) and passes with the fix. Verification: tests/fm-pi-branch-extension.test.sh passes (42 ok, no failures); tests/fm-branch-supervision.test.sh passes; tests/fm-pi-primary-types.test.sh skips locally because tsc is not installed (the refresh signature reused is the one the existing code already called). Intent constraints preserved: isolation flags untouched, carve-out still scoped to provider registration, graceful fallthrough when no providers are registered
* ci: retrigger flaky Herdr/serial-1 lanes
* no-mistakes(document): docs already cover branch extension-provider copy
* fix(bin): gate secondmate wake-loop stall alerts on real queue no-progress (#3943)
* fix(watch): detect stalled secondmate queue progress
* no-mistakes(review): gate secondmate stall on active turns and progress episodes
* no-mistakes(document): align secondmate wake-stall docs with progress-episode detector
* no-mistakes(document): clarify active-turn gate in wake-stall config docs
* no-mistakes(ci): Fixed the one real defect behind the failing checks. ROOT CAUSE (Greptile P1, real code defect in this PR): `secondmate_wake_stall_tick` in bin/fm-watch.sh reset the no-progress timer only when the oldest actionable queue sequence INCREASED (`[ "$seq" -gt "$observed_seq" ]`). When a secondmate is retired and reprovisioned under the same task ID, its fresh home's queue sequence restarts BELOW the recorded position, so the comparison is false, no reset happens, and the new queue inherits the retired generation's already-expired idle interval — emitting a false `secondmate wake-loop stalled` on its very first observation. That is precisely the false-alarm class the user intent requires this PR to remove. FIX (smallest, removal-first): bin/fm-watch.sh:749 now resets when the drain position MOVES AT ALL (`-ne` instead of `-gt`). Draining moves it up, reprovisioning moves it down; neither is a continued no-progress episode. The asymmetric `-gt` branch is removed rather than special-cased or hardened. Updated the function header comment plus the two doc sentences in docs/architecture.md and docs/configuration.md that stated the old advance-only semantics. REGRESSION TEST: added `test_secondmate_reprovisioned_queue_starts_a_fresh_interval` to tests/fm-wake-queue.test.sh (registered in the invocation list). It drives the real watcher through retired generation (seq 9) -> reprovision (seq 3, later clock) -> freeze, asserting observable wake-queue output, no source-text inspection. VERIFICATION: - Fails before / passes after: with the fix reverted the suite aborts on `not ok - a reprovisioned queue generation inherited the retired generation's idle interval and alerted`; with the fix it passes, and its third leg confirms the restarted generation still escalates on a genuine freeze (row=3 idle=2s), so the fix does not merely mute the alarm. - `bash tests/fm-wake-queue.test.sh`: exit 0, 38/38 pass, all five secondmate cases green. - `bin/fm-lint.sh`: clean (ShellCheck 0.11.0, actionlint 1.7.12). CHECKS NOT CAUSED BY THE CODE: the CI and "Require no-mistakes" runs (34152740610, 34152740604) both ended with conclusion `action_required` — workflow approval pending, not a test/build failure. Separately, `bin/fm-test-run.sh --check-coverage` exits 1 in this environment, but I confirmed by stashing my changes that it fails identically on the unmodified base tree (locale-related `comm: input is not in sorted order`); it is pre-existing and this change adds no new test file for the partition to account for. Changes are left uncommitted in the worktree
* no-mistakes(ci): Fixed the one real code defect behind the failing checks. ROOT CAUSE (Greptile P1, second round, on the head commit e1304e6): `secondmate_wake_stall_tick` in bin/fm-watch.sh identified the queue's drain position by the sequence number ALONE. The previous round changed the comparison from `-gt` to `-ne`, which handles a reprovisioned queue that restarts BELOW the recorded position, but not one that restarts ON it. A mate retired and reprovisioned under the same task id gets a fresh home whose wake-queue sequence counter restarts at 1 — and the retained parent progress marker very plausibly holds a low sequence too (a queue frozen on its first row records seq 1). Equal sequence ⇒ no reset ⇒ the brand-new queue inherits the retired generation's long-expired idle interval and emits a false `secondmate wake-loop stalled` on its very first observation. That is exactly the false-alarm class this PR exists to remove. FIX (smallest, removal-first): the file already defines the identity of a queue row once, as `row_key="$epoch-$seq"` (used for stall receipts, the stall marker, and the notify key). The progress marker's separate, weaker seq-only identity is removed: `row_key` is now computed once right after the row is parsed, stored in the progress marker, and compared with `!=`. Across generations the epoch differs (the new generation's rows are appended later), so no sequence collision can carry a stale interval; within a generation the key is stable exactly while the position does not move. bin/fm-wake-lib.sh's `fm_wake_secondmate_progress_marker_write` now takes `<oldest-row-key>` and validates it the same way the two neighbouring row-key writers do. Updated the function header comment and the two doc sentences (docs/architecture.md, docs/configuration.md) that described the old sequence-only semantics. REGRESSION TEST: `test_secondmate_reprovisioned_queue_starts_a_fresh_interval` in tests/fm-wake-queue.test.sh now drives the reported case — the reprovisioned generation restarts on the SAME sequence 9 (epoch 200) that the retired generation recorded (epoch 100), at a later clock — and asserts observable watcher output only. Its third leg still confirms the restarted generation escalates on a genuine freeze (row=9 idle=2s), so the fix does not merely mute the alarm. Three seeded progress markers in the symlink, crash-window and prefix-receipt tests were updated to the epoch-sequence form. VERIFICATION: - Fails before / passes after: with bin/fm-watch.sh and bin/fm-wake-lib.sh reverted to HEAD and the new test in place, the suite aborts on `not ok - a reprovisioned queue generation inherited the retired generation's idle interval and alerted` (exit 1); with the fix, `bash tests/fm-wake-queue.test.sh` exits 0 with 38/38 pass, all six secondmate cases green. - `bin/fm-lint.sh`: clean (ShellCheck 0.11.0, actionlint 1.7.12). - I also started `bin/fm-test-run.sh tests/fm-watch-checkpoint.test.sh tests/fm-watch-triage.test.sh tests/fm-watch-recovery-loop.test.sh` as a blast-radius check; it was still running when this phase had to return, so its result is not included. No other suite references the stall detector or the progress marker (grep over tests/ for `wake-loop stall|SECONDMATE_WAKE_STALL|secondmate-wake-progress` matches only fm-wake-queue.test.sh), and the changed lib function has exactly one caller. CHECKS NOT CAUSED BY THE CODE: the CI and "Require no-mistakes" runs on the head commit (34154091191, 34154091236, 34154091945) all ended with conclusion `action_required` — pending workflow approval, not a test/build failure. `bin/fm-test-run.sh --check-coverage` still exits 1 in this environment for the pre-existing locale reason recorded in the previous phase (`comm: input is not in sorted order` on the unmodified base tree); this change adds no new test file. Changes are left uncommitted in the worktree: bin/fm-watch.sh, bin/fm-wake-lib.sh, docs/architecture.md, docs/configuration.md, tests/fm-wake-queue.test.sh
---------
Co-authored-by: Alex William <awilliam@v2202608403614505120.powersrv.de>
* test(calm): harden the export-DOM render step and record Pi 0.85.1 evidence (#3952)
* fix(tests): make the Calm export-DOM render step retry and report
The Calm suite's rendered-export-DOM assertion started breaking CI with a
bare "could not render calm-mode HTML export DOM", which read like a Pi
0.85 rendering change. It is not one. Calm's rendered rows are identical
across Pi 0.84.4, 0.85.0, and 0.85.1, and the CI break appeared in exactly
one of the thirteen most recent runs, all on the same Pi 0.85.1, with the
main runs immediately before and after it passing.
What actually failed is headless Chrome's start-up. The render step made a
single unattended attempt and discarded both Chrome's stderr and its exit
status, so the log held nothing to tell a Chrome crash apart from a real
change in Pi's export shape.
Rendering is a vendor-tool step; the DOM assertions that follow it are what
protect the Calm conversation boundary. So the step now retries a bounded
number of Chrome start-ups on a fresh profile, drops Chrome's background
network and /dev/shm dependencies without changing what a local file renders
to, and, when every attempt fails, reports the Chrome binary, its version,
the installed Pi version, each attempt's exit status, and Chrome's own
stderr. test_export_dom_render_guard pins that with real processes and no
browser: one clean render, one that only succeeds after a start-up failure,
and one that never renders and must report enough to diagnose itself.
The verification record adds the 0.85.1 evidence this contract is now
pinned to, the cross-version comparison run through isolated installs, and
the Pi 0.85.0 packaging gap - its dist/experimental/server.js statically
imports @earendil-works/pi-server, which 0.85.0 does not declare - that
made the contract look version-sensitive in the first place.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y7rr2DHf51MjMvy7sMRauo
* no-mistakes(review): docs: attribute Pi 0.85 calm contract adaptation to renderer change
* no-mistakes(review): tests: drop inert chrome flags, report render timeouts
* no-mistakes(document): docs: fix stale Pi version facts and doc-lint link
---------
Co-authored-by: Alex William <awilliam@v2202608403614505120.powersrv.de>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* fix(bin): make every counted wake queue row presentable or retired (#3950)
* fix(bin): make every counted wake queue row presentable or retired
A wake row could be counted as queued while no drain wo…
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Intent
When a firstmate spawn is interrupted after launch delivery began, its error path claimed the paired task record and In-flight backlog state were preserved without reading that state back, so the printed claim could disagree with the real backlog item (observed 2026-09-05, the beads failure pattern). The fix makes the interrupted-signal exit path in bin/fm-spawn.sh re-read the record and backlog row under the same per-task lock as the commit, repair the In-flight transition where the commit believed it succeeded but the row did not move, and print verified wording (verified preserved / moved to In flight now and verified) or honest attempted wording (preservation could not be verified, with reason and hand-closeout instruction) - never phrasing intent as outcome. Two new behavior tests in tests/fm-backlog-atomicity.test.sh drive a lying tasks-axi start through a real interrupted spawn and assert the printed claim and the real backlog state agree; the existing interruption test was updated to the verified wording. The change is surgical to the deferred-signal exit path; bin/fm-lint.sh is clean.
What Changed
bin/fm-spawn.shnow re-reads the paired task record and backlog row under the same per-task lock as the commit, repairs an In-flight transition the commit believed succeeded but that did not read back, and prints verified wording ("verified preserved" / "moved to In flight now and verified") or honest attempted wording ("preservation could not be verified" with a reason and hand-closeout instruction) — never intent phrased as outcome. The path also re-defers signals during the read-back so verification cannot itself be killed halfway.tests/fm-backlog-atomicity.test.shthat drive a lyingtasks-axi startthrough a real interrupted spawn and assert the printed claim agrees with the real backlog state (one where the repair succeeds, one where it fails), and updated the existing interruption test to the new verified wording.tests/lib.shpinsumask 022for fixtures, andtests/fm-calm-pi-extension.test.shadapts its Pi-equivalence baseline and waits to Pi >= 0.85 (explicit stock tool definitions instead of the removed implicit substitution, "Working" editor-border indicator matching, retry-tolerant Escape aborts).docs/configuration.md.Risk Assessment
✅ Low: The change is surgical to the deferred-signal exit path, every outcome branch of the new verification sets an honest claim before returning, the lock is provably held through the read-back and released on every exit via the existing EXIT trap, and the new tests assert behavior through the real spawn and real backlog state rather than source contents.
Testing
Completed 1 recorded test check.
Pipeline
Updates from git push no-mistakes
✅ **intent** - passed
✅ No issues found.
✅ **Rebase** - passed
✅ No issues found.
bin/fm-spawn.sh:3474- The read-back repair calls fm_backlog_start without first checking fm_backlog_row_dispatchable, unlike the sibling commit path (bin/fm-backlog-transition-lib.sh:557) which only starts a 'queued no no' row. For a non-dispatchable row (e.g. done or held) the repair attempts a start that can only fail, so the printed claim adds a 'moving it to In flight failed (...)' clause for an inherently unfixable state. The claim stays honest in every such trace (state is printed verbatim and the hand-closeout instruction is present), so this is noise, not a wrong claim; no action needed for this change.bin/fm-test-run.sh --changed --exclude-family real-herdr-gated🔧 Fix: Fix calm suite for Pi 0.85 and pin test umask
1 error still open:
bin/fm-test-run.sh --changed --exclude-family real-herdr-gateddocs/calm.md:6- docs/calm.md still names the hidden stock row as theWorking...transcript row, while the new test comments establish that Pi 0.85+ embeds the indicator in the editor border asWorking. This is pre-existing Pi-version drift the change only revealed (the calm extension behavior is untouched by this branch), so it was left alone per scope discipline.✅ **Lint** - passed
✅ No issues found.
✅ **Push** - passed
✅ No issues found.