[BoYanZh]

Problem

On Windows, normal dialogue occasionally fails with EPERM: operation not permitted, open '.../loops.json.lock'.

The cross-process lock in src/state.ts assumed POSIX semantics: any non-EEXIST error from link/readFile on the lock file was fatal. On Windows, benign handoffs (delete-pending handles, antivirus scans) surface as EPERM/EBUSY/EACCES/ENOTEMPTY on open, and every chat.message / system.transform / event plus the 30s lease heartbeat acquires the session lease — so the race is hit regularly. Only ENOENT was tolerated, so the transient error propagated into user-visible dialogue failures.

A second, latent bug: the V2 event consumer loop has no per-event handling, so a single failing event kills the whole consumer and silently disables all future event-driven scheduling until process restart.

Evidence from a real machine: orphaned loops.json.lock.owner.<pid>-* and loops.json.<pid>.ts.tmp files from dead processes, plus live multi-context contender files — confirming ongoing lock contention.

Approach

• src/state.ts: retry transient lock errors briefly (5s budget) on owner-create/link/read/release instead of throwing; genuine contention still waits on the EEXIST path with the 30s deadline. Lock release() is now best-effort so a release failure can never turn a successful mutation into a user-visible error. Opportunistically sweep dead-pid lock.owner.* orphans on acquisition (live contenders are never touched).
• src/server.ts: passive V1 hooks (chat.message, system.transform, compacting, event) and V2 hooks (context, event consumer) log and continue instead of breaking dialogue. Tool paths still surface errors so the agent can retry. The V2 consumer now catches per event.
• dist/server.js rebuilt per CONTRIBUTING step 5.

Verification

• bun test: 106 pass (102 baseline + 4 new), 0 fail.
• New regression tests: isTransientLockError classification table; V1 passive hooks + V2 context hook resolve (not throw) when state is unusable; V2 consumer keeps processing after a failing event. Confirmed the new tests fail on unfixed code (4 fail + 1 import error) and pass with the fix.
• bun run lint, bun run typecheck, bun run build, pack:dry-run all green.
• Also verified on Windows with a 50-way concurrent-mutation stress run (correct counts, no lock leftovers) and the original APPDATA failure case (state file valid, orphans cleaned).

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}