[Hyper-AI-Lab]

What

Release 3.5.1: fixes for the problems that reviewers found in the 3.5.0 runtime while checking a team project's upgrade to 3.5.0. No change to the method. Nothing to configure when upgrading.

Fixes

Each fix has a test in test/review-fixes.test.mjs that failed before it: 12 tests failed on 3.5.0 (a3a7fa9), and 6 more failed before the fixes from two fresh-context reviews of this branch. One more test guards behaviour that a fix keeps (19 tests).

Severity	Problem in 3.5.0	Fix
high	An invalid .agent/paacm.json crashed the stop hook (loadConfig outside any try). The hook is not fail-closed, so the turn ended without the gate (3.4.1 blocked). One project's error also discarded another project's block	The configuration error is reported like the gate's own (could not verify this task). Any other error blocks that stop once and names it; on the continuation the stop is let through as a hook error (stderr, exit 1), never silently. The other projects are still judged
medium	A gate.log that could not be written made paacm gate crash after recording the result	Warning; the result and the paacm gate: exit line stand. An untracked file that cannot be inspected (a symbolic link loop) no longer breaks the fingerprint
medium	Every exit 127 counted as "a missing tool", including a script the change removed or renamed. The agent was told not to fix it	Only the failing command's own not-found message counts. A command named by a path inside the project is an ordinary failure, and the verify instructions say so. A missing program stays a missing tool, also one named by a path outside the project (/opt/tools/bin/x). The hint and stop message keep 3.5.0's wording
medium	When git could not run in the hook (not on PATH, safe.directory), an edit in a sibling worktree was allowed	sameRepository(…, { strict: true }) throws with git's own message and the folder, and the fail-closed protect hook refuses the edit
medium	After the finish nudge, a dirty tree re-ran the whole gate once 15 s had passed (about 6 min for a DEEP task in the reporting project)	A continuation reuses a passing gate for the same fingerprint, however old it is; a failing one is run again. The fingerprint now includes untracked files' size and modification time (lstat)
medium	review-input included the content of untracked files, keys included	Untracked files named like secrets (.env, .env.*, *.pem, *.key, *.p12, *.pfx, id_rsa* and the other SSH keys), anywhere in the repository, are listed by repository-relative name only
medium	A blocked task in another worktree held every worktree whose commit had a change to its frozen files, even for test files added after its approval	Only files that existed at the approval count
low	The finish nudge came back at every turn while the agent waited for the human (15 s dedup)	Once per working tree and phase
low	An unreadable state.json silently hid a blocked task from the check	A warning on standard error
low	The skip-git-hooks rule backtracked exponentially on many -c options (about 40 hit the hook timeout)	Each option has one reading. 60 options take under 5 ms; the same commands still match

Deliberately not changed

• A change from the base branch still holds other worktrees over a blocked task. I tried exempting commits that a remote-tracking branch has, and reverted it. 3.5.0's audit rejected trusting remote-tracking refs for the audit, and a writer can create one without any approval (git fetch from a local folder). The new troubleshooting entry says how to settle it.
• A non-git session folder still runs each edited checkout's postEdit and checks. Making that opt-in is a change to the method. It only applies when PAACM is installed as a plugin and the session folder is not a repository.
• The broad catch in prune (session start) stays. Cleaning up old baselines must never fail a session start.
• 3.4.x baselines in the shared temporary folder are not removed by the upgrade. The changelog says when to delete them.

Docs

• CHANGELOG.md 3.5.1, docs/project/requirements.md release 3.5.1
• docs/guide/troubleshooting.md: the nudge, exit 127 for a path, and a new entry for "task … (in the worktree at …) is blocked"
• ALGORITHM.md §7 (exit 127)
• 07-verify.md and paacm-verify (exit 127 for a path)
• Version 3.5.1 (package, kit, marketplaces, README, setup guide)

Checks

• node bin/paacm.mjs check --tier quick → exit 0 (validate, build-drift, diff-whitespace, node --test 179 pass / 0 fail)
• Two fresh-context reviews of this branch (correctness, error handling) found six more problems, each fixed with a test that failed before it, and one wording problem (a sentence added to the missing-tool messages, now removed)
• npx markdownlint-cli2@0.23.3 "**/*.md" → 0 issues
• Two Windows-specific paths are skipped on Windows in the new tests (POSIX shell messages, emptying PATH); CI covers the rest there