Crash/reboot resilience + atomic O_EXCL lock (fix WSL2 mkdir non-atomicity)#48
Open
owenzhang26-sys wants to merge 2 commits into
Open
Crash/reboot resilience + atomic O_EXCL lock (fix WSL2 mkdir non-atomicity)#48owenzhang26-sys wants to merge 2 commits into
owenzhang26-sys wants to merge 2 commits into
Conversation
added 2 commits
June 23, 2026 10:34
…ocal.json in dirty check fm-spawn waits specifically for a */.treehouse/* cwd rather than any change off the project dir, so a transient default-cwd reading is never misrecorded as the worktree. fm-teardown's dirty check ignores the tracked .claude/settings.local.json that the turn-end hook modifies, so teardown is not blocked by firstmate's own hook.
Crash/reboot resilience: a WSL VM teardown (host sleep, idle timeout, Windows Update reboot) kills tmux and every crewmate at once and nothing relaunched firstmate afterward. systemd/firstmate.service runs bin/fm-resume.sh as a watchdog that recreates the persistent firstmate tmux session on boot and self-heals if it dies (KillMode=process, so stopping the unit never kills a live firstmate). bin/fm-install-autostart.sh installs/removes it. Pair with ~/.wslconfig vmIdleTimeout=-1 to prevent the idle teardown in the first place. Lock correctness: the wake-queue/singleton lock used mkdir as its atomic primitive, but mkdir is NOT atomic on WSL2's filesystem (verified: 4 concurrent mkdir calls succeeded on one path in a barrier race), so the lock double-granted under contention - duplicate watchers, raced wake-queue drains, flaky tests. Replaced with an O_EXCL (noclobber) create, atomic on Linux, WSL2, and macOS, which also writes the holder pid in the same step (no empty-pid window). Dead holders are reclaimed in one call; a live holder's lock is never stolen. Tests: fm-lock-exclusivity (canonical mutex probe + reclaim) and fm-resume (base-index-safe create + idempotency) guard both fixes; the singleton-start test now polls for the invariant instead of a fixed sleep.
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
1 participant
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.
Why
A WSL VM teardown (host sleep, idle timeout, or a Windows Update reboot) kills tmux and every crewmate at once, and nothing relaunched firstmate after the VM came back — the fleet stayed dead until the captain reconnected. Separately, the test suite was intermittently flaky.
Crash/reboot resilience
systemd/firstmate.service+bin/fm-resume.sh— a watchdog recreates the persistentfirstmatetmux session on boot and self-heals if it dies.KillMode=processmeans stopping the unit never kills a live firstmate. Verified: restarts ~10s after a kill; survivor session untouched.bin/fm-install-autostart.sh—install/status/uninstall; fully reversible, never touches a running session.~/.wslconfigwithvmIdleTimeout=-1to prevent the idle teardown in the first place (needswsl --shutdownto apply).Lock correctness (root cause of the flakiness)
The wake-queue / watcher-singleton lock used
mkdiras its atomic primitive, butmkdiris not atomic on WSL2's filesystem — a 20-way barrier race produced up to 4 concurrentmkdirsuccesses on one path. The lock therefore double-granted under contention (duplicate watchers, raced wake-queue drains, flaky tests). A canonical mutex probe showed 4–8 double-grants per run.Replaced the primitive with an
O_EXCL(noclobber) create — atomic on Linux, WSL2, and macOS — which also writes the holder pid in the same step, eliminating the empty-pid window. Dead holders are reclaimed in a singletry_acquire(sofm-watch.shtakes over a crashed watcher); a live holder's lock is never stolen. Probe now reports 0 double-grants.Tests
tests/fm-lock-exclusivity.test.sh— canonical mutex probe + dead/live reclaim (guards the lock fix).tests/fm-resume.test.sh— base-index-safe session creation + idempotency (runs against an isolated tmux server).tests/fm-wake-queue.test.sh— singleton-start now polls for the invariant instead of a fixed sleep.Local CI green:
shellcheck bin/*.sh tests/*.sh, all 5 test files, and the repo invariants.Also included
fix(spawn,teardown): fm-spawn waits for a*/.treehouse/*cwd before recording the worktree; fm-teardown ignores the tracked.claude/settings.local.json(turn-end hook) in its dirty check.🤖 Generated with Claude Code