fix(bin): sync upstream firstmate fixes for no-mistakes, watchers, and crew lifecycle

[danielkuykendall23-boop]

Intent

create a plan and use evwry tool and skill needed to fix bugs and everything that is broken in our envirment and fix it ... just fix every bug within our envirnmnet ... dont hallucinate no mistakes

Context: the environment is this Firstmate home (OMP primary on Herdr). A read-only audit (data/env-bug-audit/report.md in the live home ~/kun-agent-workspace) verified the defect below.

Defect (audit findings #10 and #4): this fork (origin danielkuykendall23-boop/firstmate) is 20 commits behind upstream kunchenguid/firstmate main. Upstream contains fixes relevant to this home, including c5131a3 (handle no-mistakes passed-with-override, required before upgrading no-mistakes past v1.77.1, which Main will do after this lands), 52fca51, dd9f2b4, 39f4c2a, 82dec2e, a8a2959 and 259a669.

What Changed

• Brings in 21 upstream kunchenguid/firstmate fixes, most of them in bin/. These include:
  • Treating the no-mistakes passed-with-override result as a pass.
  • Reporting green no-mistakes PRs that are waiting to be merged.
  • Requiring a non-draft PR before a PR-based done report.
  • Refusing a ship done: when the named head exists only in the worker copy.
  • Cleaning up workers after their PRs land, and letting windowless legacy task records be cleaned up too.
  • Supporting quota-axi schema 6 snapshots.
  • Getting the Lavish polling route from the board session.
  • Delivering failed public follow-ups with the updated AXI floors.
  • Recording away posture as soon as /afk runs.
  • Showing launches stuck on an interactive prompt as not-started.
• Makes watchers and wakes more reliable across fm-watch.sh, fm-wake-lib.sh, fm-busy-lib.sh, fm-task-inbox-lib.sh and the Pi/OMP watch extensions:
  • Watchers stop reliably when a poll is blocked. On bash 3.2, the watcher is woken so a single TERM stops it.
  • Stuck inbox doorbells are submitted instead of skipped.
  • A keyed answer no longer re-wakes the home.
  • A secondmate proven to be idle gets a ring before a wake-loop stall alarm is raised.
  • The Pi watcher's predecessor is kept, which stops false "down" alarms.
  • Secondmate relaunch no longer fails when watcher scratch files have disappeared.
• Updates docs, skills and AGENTS.md to match, and adds or extends the test suites. New suites are tests/fm-dod-lib.test.sh and tests/fm-launch-prompt-signals-live-e2e.test.sh. On the CI side, the no-mistakes required check is pinned to v1.80.1 and exempts kunchenguid.

🤖 Generated with Claude Code

Risk Assessment

✅ Low: The branch rebases the upstream fixes onto the fork faithfully. The final tree equals the earlier merge-based sync (upstream c576c2b merged into fork 4474a05, with only index/offset differences in the diffs) plus the fork's own #6 commit. That covers every intent-listed fix (c5131a3, 52fca51, dd9f2b4, 39f4c2a, 82dec2e, a8a2959, 259a669). The only non-upstream code is a small, well-scoped bash 3.2 signal ticker that watcher_cleanup cleans up on every normal stop path.

Testing

Ran the focused upstream regression cases under stock macOS bash 3.2, then ran the real watcher and crew-state scripts directly, before and after each fix. With fixture no-mistakes status, crew-state output went from unknown to done for passed-with-override. With a fake no-mistakes, teardown treats that outcome as terminal. Neither was driven against a live override run. The pre-fix watcher ignored one TERM for about 60 seconds while blocked. The fixed watcher stopped in 7–8 seconds with its lock released, and its helper also exits on SIGKILL. The upstream TERM test passed once on the pre-fix watcher only because its blocking holder ends after 30 seconds and the heavily loaded host (load ~117) let the test's 100-tick wait outlast that. The direct drive with a 60-second holder shows the difference clearly. Herdr lab lifecycle scenarios were not run. Transient test drivers were removed from the worktree.

• Live validation: ✅ go - 2 of 5 scenarios driven live against the product

Scenario	Result	Live	Evidence
Crew-state reports a worker whose no-mistakes run ended passed-with-override as done (not unknown)	⏸️ untested	no	The prior payload did not establish a live result. The real bin/fm-crew-state.sh was run, but the no-mistakes axi status output was a fixture. A live check needs a real no-mistakes run finished with a…
Teardown treats a run that lands on passed-with-override after abort as terminal (no REFUSED)	⏸️ untested	no	The prior payload did not establish a live result. Only the upstream test with a fake no-mistakes covered this. A live check needs a real no-mistakes run with an approved override exception.
On stock macOS bash 3.2, one SIGTERM stops a watcher blocked inside a pane capture and its cleanup releases the lock	✅ pass	live	live-watcher-signal-drive-head.txt: 3/3 runs exited 7–8s after one TERM with lock-present=no; live-watcher-signal-drive-prefix.txt: the pre-fix watcher took about 60–64s, ending only when the FIFO hol…
Adversarial: the SIGCHLD helper does not survive a SIGKILLed watcher	✅ pass	live	live-watcher-signal-drive-head.txt [kill-1]: 3 watcher processes before the kill; afterwards only the capture subshell still blocked on the FIFO remains, so the helper exited
Herdr lab lifecycle (spawn / watcher in a real fm-lab-* Herdr session) still works after the sync	⏸️ untested	no	Not run. A full bin/fm-herdr-lab.sh prepare/provision/run/teardown was not attempted on this host at load ~117. The required Herdr CI lane covers it, or run bin/fm-herdr-lab.sh with an fm-lab-* sessio…

Evidence: crew-state passed-with-override before/after

Source: crew-state passed-with-override before/after

HEAD: state: done · source: run-step · run passed: PR merged PRE-FIX: state: unknown · source: run-step · outcome: passed-with-override

\### HEAD
--- no-mistakes axi status fixture:
run:
  id: "01RUN"
  branch: fm/feat-override
  status: completed
  head: "86e74e2be5bb6bac01003ef285cadf90ad9fa568"
  pr: "https://github.com/o/r/pull/1"
  findings: none
outcome: passed-with-override
ci_override_reason: "live checks not all passed: Lint (fail)"
--- bin/fm-crew-state.sh feat-override output:
state: done · source: run-step · run passed: PR merged

\### PRE-FIX 12acf58^
--- no-mistakes axi status fixture:
run:
  id: "01RUN"
  branch: fm/feat-override
  status: completed
  head: "9f48d2ca9232d9c2bb7e791df4569bc3af2c3bf1"
  pr: "https://github.com/o/r/pull/1"
  findings: none
outcome: passed-with-override
ci_override_reason: "live checks not all passed: Lint (fail)"
--- bin/fm-crew-state.sh feat-override output:
state: unknown · source: run-step · outcome: passed-with-override

Evidence: watcher TERM/KILL drive at HEAD (bash 3.2)

Source: watcher TERM/KILL drive at HEAD (bash 3.2)

\### HEAD beaabe5 (with bash 3.2 SIGCHLD ticker)
[term-1] bash=3.2.57(1)-release watcher=34209 blocked=yes fm-watch-procs-before: 34209 40674 52710 
[term-1] SIGTERM: watcher exited after ~7s; lock-present=no
[term-1] fm-watch-procs-after (ticker survivors): [52710 ]
[term-2] bash=3.2.57(1)-release watcher=56797 blocked=yes fm-watch-procs-before: 56797 61901 78288 
[term-2] SIGTERM: watcher exited after ~7s; lock-present=no
[term-2] fm-watch-procs-after (ticker survivors): [78288 ]
[term-3] bash=3.2.57(1)-release watcher=81200 blocked=yes fm-watch-procs-before: 81200 85542 98345 
[term-3] SIGTERM: watcher exited after ~8s; lock-present=no
[term-3] fm-watch-procs-after (ticker survivors): [98345 ]
[kill-1] bash=3.2.57(1)-release watcher=298 blocked=yes fm-watch-procs-before: 298 8174 18544 
tests/.nmfocus-fm-watch-triage.test.sh: line 6070:   298 Killed: 9               PATH="$fakebin:$PATH" FM_FAKE_TMUX_WINDOW="$window" FM_FAKE_TMUX_CAPTURE="$fifo" FM_STATE_OVERRIDE="$state" FM_POLL=1 FM_SIGNAL_GRACE=1 FM_CHECK_INTERVAL=999999 FM_HEARTBEAT=999999 "$WATCH" > "$out" 2> "$dir/watch.err"
[kill-1] SIGKILL: watcher exited after ~1s; lock-present=yes
[kill-1] fm-watch-procs-after (ticker survivors): [18544 ]
FOCUS-DONE

Evidence: watcher TERM drive pre-fix (bash 3.2)

Source: watcher TERM drive pre-fix (bash 3.2)

\### PRE-FIX 687ab2e (no SIGCHLD ticker)
[term-1] bash=3.2.57(1)-release watcher=26805 blocked=yes fm-watch-procs-before: 26805 44951 
[term-1] SIGTERM: watcher exited after ~62s; lock-present=no
[term-1] fm-watch-procs-after (ticker survivors): []
[term-2] bash=3.2.57(1)-release watcher=64581 blocked=yes fm-watch-procs-before: 64581 75341 
[term-2] SIGTERM: watcher exited after ~60s; lock-present=no
[term-2] fm-watch-procs-after (ticker survivors): []
[term-3] bash=3.2.57(1)-release watcher=95097 blocked=yes fm-watch-procs-before: 10372 95097 
[term-3] SIGTERM: watcher exited after ~64s; lock-present=no
[term-3] fm-watch-procs-after (ticker survivors): []
FOCUS-DONE

Evidence: focused regression tests on bash 3.2

Source: focused regression tests on bash 3.2

=== fm-crew-state (bash: 3.2.57(1)-release)
ok - terminal passed run is authoritative
ok - terminal passed-with-override run reads done like a clean pass
FOCUS-DONE
/bin/bash tests/.nmfocus-$f.test.sh  0.47s user 0.59s system 1% cpu 1:18.66 total
=== fm-teardown (bash: 3.2.57(1)-release)
ok - a run that lands on passed-with-override after abort is still recognized as terminal
FOCUS-DONE
/bin/bash tests/.nmfocus-$f.test.sh  1.28s user 1.86s system 1% cpu 4:40.07 total
=== fm-watch-triage (bash: 3.2.57(1)-release)
~/.no-mistakes/worktrees/6ada68ffd386/01M36SFN68CEPZK8T5BMYZJ3BT/bin/fm-wake-lib.sh: line 554: 97923 Terminated: 15          ( while kill -CHLD "$WATCHER_PID" 2> /dev/null; do
    sleep 1;
done ) < /dev/null > /dev/null 2>&1
ok - TERM stops a watcher blocked inside a poll and still runs its cleanup
FOCUS-DONE
/bin/bash tests/.nmfocus-$f.test.sh  1.31s user 2.04s system 1% cpu 4:13.14 total

Evidence: upstream TERM test against pre-fix watcher

Source: upstream TERM test against pre-fix watcher

ok - TERM stops a watcher blocked inside a poll and still runs its cleanup
FOCUS-DONE
/bin/bash tests/.nmfocus-fm-watch-triage.test.sh  1.34s user 2.06s system 1% cpu 3:58.84 total

- Outcome: ⚠️ 1 info across 1 run (34m1s)

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

⚠️ **Review** - 1 info

• ℹ️ bin/fm-watch.sh:2334 - The new bash 3.2 SIGCHLD ticker (while kill -CHLD "$WATCHER_PID"; do sleep 1; done) only stops when watcher_cleanup kills it or when the watcher's PID stops existing. That's fine for normal HUP/TERM/INT stops because the EXIT trap kills the ticker. But if the watcher is SIGKILLed (OOM or a crash), the orphaned ticker keeps running. If the OS then reuses the watcher's PID, the ticker keeps sending SIGCHLD to that unrelated process for as long as it lives. SIGCHLD is ignored by default, so the practical impact is small: an orphaned bash/sleep 1 pair. If hardening is wanted, the loop could also exit when its parent changes. No action required for this sync.

⚠️ **Test** - 1 info

• ℹ️ bin/fm-watch.sh:2314 - When watcher cleanup stops the new bash 3.2 SIGCHLD helper, bash writes a 'Terminated: 15' job notice (attributed to fm-wake-lib.sh line 554) to the watcher's stderr. It is cosmetic only: the watcher still exits cleanly and releases its lock.
• Live validation: ✅ go - 2 of 5 scenarios driven live against the product

Scenario	Result	Live	Evidence
Crew-state reports a worker whose no-mistakes run ended passed-with-override as done (not unknown)	⏸️ untested	no	The prior payload did not establish a live result. The real bin/fm-crew-state.sh was run, but the no-mistakes axi status output was a fixture. A live check needs a real no-mistakes run finished with a…
Teardown treats a run that lands on passed-with-override after abort as terminal (no REFUSED)	⏸️ untested	no	The prior payload did not establish a live result. Only the upstream test with a fake no-mistakes covered this. A live check needs a real no-mistakes run with an approved override exception.
On stock macOS bash 3.2, one SIGTERM stops a watcher blocked inside a pane capture and its cleanup releases the lock	✅ pass	live	live-watcher-signal-drive-head.txt: 3/3 runs exited 7–8s after one TERM with lock-present=no; live-watcher-signal-drive-prefix.txt: the pre-fix watcher took about 60–64s, ending only when the FIFO hol…
Adversarial: the SIGCHLD helper does not survive a SIGKILLed watcher	✅ pass	live	live-watcher-signal-drive-head.txt [kill-1]: 3 watcher processes before the kill; afterwards only the capture subshell still blocked on the FIFO remains, so the helper exited
Herdr lab lifecycle (spawn / watcher in a real fm-lab-* Herdr session) still works after the sync	⏸️ untested	no	Not run. A full bin/fm-herdr-lab.sh prepare/provision/run/teardown was not attempted on this host at load ~117. The required Herdr CI lane covers it, or run bin/fm-herdr-lab.sh with an fm-lab-* sessio…

• /bin/bash (3.2.57) focused copy of tests/fm-crew-state.test.sh: test_terminal_passed, test_terminal_passed_with_override
• /bin/bash focused copy of tests/fm-teardown.test.sh: test_parked_own_run_concludes_on_passed_with_override_after_abort
• /bin/bash focused copy of tests/fm-watch-triage.test.sh: test_term_stops_a_watcher_blocked_inside_a_poll, on HEAD and with 687ab2e's bin/fm-watch.sh swapped in
• Custom driver: started the real bin/fm-watch.sh blocked on a FIFO pane capture, sent one TERM (3 runs), timed the exit, checked the lock was released and which processes survived; repeated on the pre-fix fm-watch.sh
• Custom driver: sent SIGKILL to a blocked watcher and checked that the SIGCHLD helper exits
• Before/after transcript of bin/fm-crew-state.sh feat-override for a passed-with-override status, on HEAD and on 12acf58^'s fm-crew-state.sh

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.