[hobostay]

Summary

hook-dispatch's Windows/WMI spawn path writes each hook payload to %TEMP%\teamai-hook-<pid>-<ms>.json, because WMI has no stdin pipe. Only readStdinFile() removed that file, so every early return before the read (the launcher gate, the another-host gate, or any error) left the full payload — prompts, tool input and output — on disk indefinitely. A parent killed before its detached child started leaked it the same way.

This makes cleanup independent of the read: hookDispatchCli now removes --stdin-file in a finally, and on Windows every --bg-only child sweeps teamai-hook-*.json files in the temp directory that are older than ten minutes, collecting files orphaned by a parent that died first. Fresh files are untouched, so a live sibling's payload is never swept. The sweep is Windows-only because the plain spawn paths pipe the payload and create no temp file. A CHANGELOG entry is included.

Fixes #1812.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature causing existing behavior to change)
• Documentation only
• Refactor / internal cleanup

Test Plan

• npx tsc --noEmit passes (npm run typecheck).
• npm run lint passes, including documentation checks.
• npx vitest run src/__tests__/hook-dispatch-cli.test.ts — 41 passed. New tests cover the early-return cleanup, that a cleanup failure cannot fail the dispatch, that sweepStaleHookPayloads removes only old teamai-hook-*.json files while keeping fresh and unrelated files, and that the sweep is skipped off Windows.
• Related hook suites pass: hook-dispatch-cli, hook-dispatch-scope, recall-attribution, dashboard-report-scope, score-drain-stop, votes-scope — 6 files, 311 tests.
• Real-CLI verification after npm run build (macOS, node dist/index.js):
  1. Created a payload file and ran TEAMAI_LAUNCHED_BY=judge node dist/index.js hook-dispatch stop --tool claude --stdin-file <file> → exit 0, file removed. This is the early-return path that previously left the file behind.
  2. Created a stale teamai-hook-*.json (mtime 11 minutes old) in os.tmpdir() and ran the same background pass with --bg-only --stdin-file <fresh> → exit 0, the fresh stdin file was removed, and the stale file remained, confirming the Windows-only gate on macOS.

I could not exercise the Windows WMI spawn itself on this macOS host. That spawn path is unchanged. The sweep's removal behavior is covered by the unit test that supplies a temp directory directly; the Windows-only integration is left to Windows CI. Windows CI does not currently run hook-dispatch-cli.test.ts.

Related Issues

Fixes #1812.

Notes for Reviewers

• The sweep only considers regular files whose name starts with teamai-hook- and ends with .json, and ignores every cleanup error, so it can never affect dispatch.
• A live child reads and removes its own payload immediately on entry; ten minutes is far beyond that window, so only abandoned files qualify.
• The finally also covers the normal path, where readStdinFile() already removed the file. fs.rmSync(..., { force: true }) makes the second removal a no-op.

🤖 Generated with Claude Code

[SaulMoro]

This was generated by AI during triage.

Thanks for this. #1813 fixes the same issue (#1812) in the same file. It was opened earlier and is already in review, so only one of the two should land. I compared them so the choice is quick:

	#1813	#1814
Removes the file on early returns	yes, finally in hookDispatchCli	yes, same finally
Removes it right after the read	no: the finally replaced the read-time removal, so the file stays for the whole background pass	yes: readStdinFile keeps its removal, and the finally is a no-op on that path
Sweep of stale teamai-hook-*.json	Windows only	every --bg-only pass on every platform, so a macOS or Linux hook lists the temp directory each time, though only the Windows path writes these files
CHANGELOG	entry for the fix and the sweep	none

This PR's read-time removal is the part I asked #1813 to add. Its sweep should be limited to Windows, as #1813's is, and it needs a CHANGELOG entry.

I'm leaving this PR open while the author of #1813 answers. @jeff-r2026 your call on this one: which of the two to keep, since they are the same fix for #1812.

[github-actions]

No findings in the specified PR diff.

The PR description includes representative real-CLI verification of both cleanup paths, satisfying the review rules’ testing-record requirement. Windows WMI execution is explicitly untested; that spawn path is unchanged.

I did not run, build, or install anything from the PR.

[SaulMoro]

This was generated by AI during triage.

#1813 is merged (531eb15f) and closes #1812, including the read-time removal this PR also had. @jeff-r2026 this one can be closed as a duplicate, unless you see something here #1813 does not cover.

[hobostay]

Applied both points at aec745b:

• The stale sweep now runs only when process.platform === 'win32'; macOS and Linux background passes do not list the temp directory.
• Added a CHANGELOG entry under Unreleased → Bug Fixes.

On payload lifetime: the read-time removal in readStdinFile() is untouched on this branch, so a payload the child actually reads is deleted immediately; the outer finally only covers exits that never read it (and swallows cleanup errors so a completed dispatch cannot fail on cleanup).

Tests at this head: hook-dispatch-cli.test.ts 41 passed, the six related hook suites 311 passed, npm run typecheck and npm run lint pass. Real CLI on macOS: the early-return path removes the file, and an old teamai-hook-*.json is left alone by a background pass, which verifies the Windows-only gate.

#1813 is the earlier PR for the same issue and now has the read-time removal too. I am happy to close this one in favour of #1813 if you prefer the first-opened branch; there is no need to land both.

[hobostay]

Superseded by #1813, which merged to main as 531eb15 while this was under review. The merged fix has the same read-time removal, outer finally for exits that never read, Windows-only ten-minute sweep and CHANGELOG entry, so there is nothing left to land here. Closing to avoid a duplicate; thanks to @Walternater and @SaulMoro for the triage and comparison.

[github-actions]

No findings in the specified PR diff. The earlier review also reported none, so there are no prior findings to mark resolved.

The PR description documents representative real-CLI verification, satisfying the testing-record requirement. Windows integration is explicitly deferred; the WMI spawn path is unchanged.

I did not run, build, or install anything from the PR.