[SaulMoro]

Summary

#1761 dropped detached on Windows from the five spawns whose child must outlive its parent. Without it, libuv puts a Windows child in a kill-on-close job the parent holds, so the child ends when the parent exits: the workflow resume a session start's pull begins dies with the --bg-only pass, and its run stays running. This puts detached: true back at those five sites and keeps #1761's headless conhost for the WMI launch, which is what hides the background pass.

hook-dispatch-cli  background pass, plain fallback   detached: !win32 -> true
hook-handlers      pull after a git hook             detached: !win32 -> true
hook-handlers      workflow resume                   detached: !win32 -> true
local-agent        source reconcile-plugins          detached: !win32 -> true
skill-runner       workflow runner                   detached: !win32 -> true

It also closes #1763: a teamai/windows-hide rule fails a node:child_process call in production source (src/, outside src/__tests__/) whose options object literal does not set windowsHide: true, and the 56 calls it found on main now set it, whereOnWindows in src/utils/cli-path.ts among them. The rule follows the module through named, namespace and default imports, await import(), require and promisify.

Options built elsewhere are reported rather than trusted: the call site spreads them into a literal that sets the option ({ ...options, windowsHide: true }). A variable can be built far from the call, so only a literal keeps the option checkable at the call; a shared helper would be one more hop for the same line. No call needs a visible window, so nothing is suppressed. simple-git stays out of scope, as the issue says.

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Test Plan

New src/__tests__/e2e/background-lifetime-windows.test.ts, run as the last step of the windows CI job, drives the built CLI on the runner:

• (a0) control: a --bg-only pass started by a parent that stays alive reaches its prePull script under a temp path that is a file.
• (a) the same temp path fails the WMI payload write, so the dispatcher falls back to its own spawn; the pass still reaches prePull after the dispatcher exits, and logs its summary once released.
• (b) the WMI launch works; the teamai-finished-sessions run the --bg-only pass starts holds on a gate until every --bg-only process is gone, then succeeds.

On the Windows runner, the head before rebasing onto a docs-only commit (same code):

✓ (a0) control: with its parent alive, a --bg-only pass under the same unusable temp path reaches its prePull script  2385ms
✓ (a) WMI launch fails: the --bg-only pass the dispatcher starts finishes its handlers after the dispatcher exits  4450ms
✓ (b) WMI launch works: the workflow resume the --bg-only pass starts ends its run after that pass exits  8199ms
Tests  3 passed (3)

Without the fix, (b) failed on every run: the run stayed running and no workflow resume process was alive.

For #1763, after npm run build, a session start through the real CLI on macOS, with a user-scope install of a local team repo, in a git repository with no remote, reaches the fixed git remote get-url origin call in mr-hint.ts:

$ echo '{"cwd":"<tmp>/app","session_id":"wh-record"}' | HOME=<tmp>/home node dist/index.js hook-dispatch session-start --tool claude
dispatch exit=0
[DEBUG] mr-hint: no git remote, skipping
[DEBUG] [hook-dispatch] session-start/claude foreground: 89ms total, 7 handlers, ... mr-hint=64ms(ok) ...
[DEBUG] [hook-dispatch] session-start/claude background: 601ms total, 3 handlers, slowest pull=601ms | pull=601ms(ok) ...

whereOnWindows runs only on Windows; the windows job's providers-cli-cmd step calls it there.

• npm run typecheck and npm run lint pass
• npx vitest run passes (22720 tests), and npx vitest run oxlint-plugin-teamai passes (639, 18 of them the new rule's ruleHarness fixtures)
• Added/updated tests for the change

Related Issues

Fixes a regression from #1761. Closes #1763. Related to #1747. #1805 was fixed by #1845.

Notes for Reviewers

• On Windows detached also sets DETACHED_PROCESS, which cancels CREATE_NO_WINDOW, so a console program that the resume, the runner or the git-hook pull starts (git, an AI tool) may open a window again. The --bg-only pass itself still launches hidden through WMI. Keeping that work finishing came first; hiding its descendants is a follow-up.
• Wrapping these spawns in conhost --headless the way the WMI launch does was tried and does not work: a detached conhost never starts its child.
• Case (b) writes ~/.teamai in the runner's own profile, because a WMI-created process does not inherit the caller's environment. It runs only under GITHUB_ACTIONS, sets aside a ~/.teamai left by earlier steps and restores it, and stays the last step of the job.

[github-actions]

• [P2 non-blocking] oxlint-plugin-teamai/rules/windows-hide.js:38 — hidesWindow accepts any windowsHide: true property, even when a later spread overrides it. For example, spawn('git', [], { windowsHide: true, ...options }) passes lint when options.windowsHide is false, defeating the rule’s guarantee. Reject potentially overriding spreads after the property and add a fixture.

The PR description documents sufficient testing, including representative Windows real-CLI runs and a post-build macOS CLI run. The older-head Windows record is not blocking under the review rules.

Review was read-only; nothing from the PR was run, built, or installed.

[SaulMoro]

Fixed in b96d2ce. The last spread or windowsHide property in the object now decides, so { windowsHide: true, ...options } and { windowsHide: true, windowsHide: false } are rejected, with a fixture for both. The stricter check found two calls in src/scorers/scoring-evidence.ts that put windowsHide before ...timeLeft(gitIn). That spread only ever adds timeout, but both calls now put windowsHide: true last.

[github-actions]

• [P3 nit] oxlint-plugin-teamai/rules/windows-hide.js:45 — Computed properties are ignored when checking overrides. execSync('git status', { windowsHide: true, ['windowsHide']: false }) passes lint despite disabling window hiding. Recognize computed literal keys and add a fixture.

The earlier spread-override finding is resolved: later spreads and non-computed overrides are rejected, with fixture coverage.

The PR description documents sufficient testing, including Windows real-CLI runs and a post-build macOS CLI run. The older-head Windows record is not blocking under the trusted review rules.

Review was read-only; nothing from the PR was run, built, or installed.

[SaulMoro]

Fixed in the latest commit. A computed key now counts like a spread: it may name windowsHide, so windowsHide: true must come after it. A computed key also never counts as windowsHide: true itself, so { [key]: true } is rejected. Both cases have a fixture.

[github-actions]

No new findings in the scoped diff.

• Both earlier findings are resolved: later spreads, explicit overrides, and computed keys are rejected, with fixture coverage.
• The PR description documents sufficient testing, including representative Windows real-CLI runs and a post-build macOS CLI run. The older-head Windows record is not blocking under the trusted review rules.

Review was read-only; nothing from the PR was run, built, or installed.

[github-actions]

No new findings in the scoped diff.

• Both earlier findings are resolved: later spreads, explicit overrides, and computed keys are rejected, with fixture coverage.
• The PR description documents sufficient testing, including representative Windows real-CLI runs and a post-build macOS CLI run. The older-head Windows record is not blocking under the trusted review rules.

Review was read-only; nothing from the PR was run, built, or installed.

[github-actions]

No new findings in the scoped diff.

• Both earlier findings are resolved: later spreads, explicit overrides, and computed keys are rejected, with fixture coverage.
• The PR description documents sufficient testing, including representative Windows real-CLI runs and a post-build macOS CLI run. The older-head Windows record is not blocking under the trusted review criteria.

Review was read-only; nothing from the PR was run, built, or installed.

[github-actions]

No new findings in the scoped diff.

• Both earlier findings are resolved: later spreads, explicit overrides, and computed keys are rejected, with fixture coverage.
• The PR description documents sufficient testing, including representative Windows real-CLI runs and a post-build macOS CLI run. The older-head Windows record is not blocking under the trusted review criteria.

Review was read-only; nothing from the PR was run, built, or installed.

[wstczyw]

Data point from a Windows 11 host for this PR (teamai-cli 0.29.0, Windows Terminal 1.24 as the default terminal), since the triage of #1747 noted that no Windows host was available:

• Each Claude Code session start produced about 14 visible Windows Terminal windows. Chain: hook-dispatch session-start --bg-only (WMI child) → score-sessions (spawned with detached: true, so it has no console) → whereOnWindows() → where.exe <tool> without windowsHide. Each where.exe was followed 40–100 ms later by OpenConsole.exe -Embedding and WindowsTerminal.exe -Embedding (parent svchost).
• whereOnWindows in src/utils/cli-path.ts on main still has no windowsHide, as Lint rule: require windowsHide: true on every child_process call in production source #1763 notes.
• Adding windowsHide: true to whereOnWindows and to 17 other console spawns reachable from score-sessions and the detached pull (patched in dist/cli.js) stopped it. In a repeat session start: 8 where.exe calls, 0 OpenConsole / WindowsTerminal -Embedding, 0 new visible top-level windows.
• The other sites: taskkill on judge abort, the sqlite3 CLI in the workbuddy and cursor stores, git in session-flow sync and in scoring-evidence, git --version and git cat-file --batch in src/utils/git.ts, git cat-file --batch-check in self-improvement, git ls-files in the repo and env resources, and the post-pull / team-scripts node spawns when stdio is not inherited. The lint rule from Lint rule: require windowsHide: true on every child_process call in production source #1763 should catch all of them.

[STiFLeR7]

@SaulMoro I resolved the three conflicts with main in merge commit 6656b913. Could you fast-forward test/windows-background-lifetime to that commit and push it to your fork?

git switch test/windows-background-lifetime
git fetch https://github.com/STiFLeR7/teamai-cli.git qa/1775-merge
git merge --ff-only FETCH_HEAD
git push origin test/windows-background-lifetime

Typecheck, Oxlint, 160 focused tests, and two isolated Windows subprocess end-to-end tests passed on the merge. The WMI-success case still needs the disposable CI runner.

[SaulMoro]

This was generated by AI during triage.

@STiFLeR7 thanks, I took your merge as is: 6656b913 is on the branch with your authorship. Main had moved since (#1827, #1831 and #1832), so f07cae0a merges 33fe3bbb on top. That merge had no conflicts. Typecheck and npm run lint pass, and so do 688 tests across the files the merge touches: hook-fast-path, mr-hint, pid-monitor, git/push, YAML merge, and the windows-hide rule.

@wstczyw thanks for the data point, it matches what this PR fixes. whereOnWindows in src/utils/cli-path.ts sets windowsHide: true on this branch. The other sites you list are covered too: the teamai/windows-hide lint rule fails any spawn/exec* call without it, including one where a later spread or a computed key could override it, and lint passes on the merged head.

[github-actions]

No new findings in the scoped diff.

• Both earlier findings are resolved: later spreads, explicit overrides, and computed keys are rejected, with fixture coverage.
• The PR description documents sufficient testing, including representative Windows real-CLI runs and a post-build macOS CLI run. The older-head Windows record is not blocking under the trusted review criteria.

Review was read-only; nothing from the PR was run, built, or installed.

[github-actions]

No new findings in the scoped diff.

• Both earlier findings are resolved: later spreads, explicit overrides, and computed keys are rejected, with fixture coverage.
• The PR description documents sufficient testing, including representative Windows real-CLI runs and a post-build macOS CLI run. The older-head Windows record is not blocking under the trusted review criteria.

Review was read-only; nothing from the PR was run, built, or installed.