vendor: actions/checkout@v6.0.2 with timeout-minutes input

[tenstorrent-github-bot]

Summary

Vendors actions/checkout@de0fac2e (v6.0.2) under .github/actions/checkout/ and adds a timeout-minutes input (default: 20 min) that hard-kills hung git subprocesses.

Why vendor?

timeout-minutes on steps inside a composite action is silently ignored by the GitHub runner (actions/runner#1979). The only way to enforce a wall-clock limit on git fetch/clone calls is to do it inside the action's own TypeScript code.

Why not Promise.race?

Promise.race rejects the outer promise when the timer fires but the underlying child_process spawned by @actions/exec keeps running in the OS — useless as a hung-fetch guard. This implementation uses child_process.spawn directly, sends SIGTERM on timeout, and escalates to SIGKILL after 5 s.

Commits

• Commit 1: verbatim copy of actions/checkout@de0fac2e — no modifications, easy to diff against upstream
• Commit 2: adds timeout-minutes input wired through IGitSourceSettings → GitCommandManager.execGitWithTimeout

Usage

- uses: tenstorrent/tt-github-actions/.github/actions/checkout@<sha>
  with:
    timeout-minutes: 30   # default: 20

[tenstorrent-github-bot]

Review of commit 2: timeout-minutes input

Overall this is a solid feature addition. The core approach — bypass @actions/exec and use child_process.spawn directly to get timeout control — is reasonable since @actions/exec doesn't expose timeout support.

However, there are two real bugs that should be fixed before merging, plus a couple of robustness improvements worth considering.

Bugs (must fix)

1. Inner SIGKILL timer leaks — the 5-second escalation timer inside the timeout callback is never cleared when the process exits. This keeps the Node event loop alive and fires kill() on a dead PID.
2. Infinity bypasses validation — parseFloat('Infinity') passes the > 0 check, and setTimeout(fn, Infinity) coerces to 0 in Node.js, making every git command time out immediately.

Robustness (nice to have)

3. Partial-line delivery for stdline/errline — splitting on \n per data event doesn't guarantee line-aligned chunks. The original @actions/exec uses internal line buffering. Acceptable for now since stdline/errline are only used for progress output, but worth a comment.
4. Windows: SIGTERM is a no-op for non-Node processes — child.kill('SIGTERM') does nothing to git.exe on Windows. The process only dies when the SIGKILL fires 5s later. Minor since GHA runners are mostly Linux, but the 5s delay is unnecessary on Windows.

[tenstorrent-github-bot]

Re-review of commit 3 (2b94e280) — Follow-up Pass

All items from the first review have been addressed. The killTimer cleanup, !isFinite() guard, improved error message, and partial-line comment are all good.

After a thorough second pass, I found one real bug and one minor inconsistency. Everything else checks out.

Verified Correct

• Timer race conditions: killTimer is assigned inside the outer timer callback. JavaScript's single-threaded event loop guarantees the entire callback (including the killTimer = setTimeout(...) assignment) runs to completion before any I/O event (like close) can fire. So killTimer is always assigned before close can check it. ✅
• timedOut flag race: Same event loop guarantee — either clearTimeout(timer) in close cancels the outer timer before it fires, or the outer timer fires first and sets timedOut = true. Both paths are deterministic. No race. ✅
• !isFinite(timeoutMinutes) validation: Correctly blocks Infinity and -Infinity (which pass !isNaN and > 0 checks). ✅
• dist/index.js sync: Verified the compiled JS matches the TS source for all changes (killTimer, isFinite, error message, comment). ✅
• env type cast: env as NodeJS.ProcessEnv is correct — both are {[key: string]: string | undefined}. ✅
• action.yml default vs input-helper.ts fallback: Both use '20'. The || '20' in TS is redundant since GHA injects the YAML default, but it's a fine defensive pattern and not worth removing. ✅
• spawn stdio: cp.spawn defaults to stdio: 'pipe' for all three streams, so child.stdout and child.stderr are always Readable streams. No null-check needed. ✅

Summary

One real issue (child.killed semantics), one nit (missing errline comment). Neither is a merge blocker — the child.killed issue is pre-existing from upstream actions/checkout, and git processes do respond to SIGTERM on Linux. But worth noting for awareness.

LGTM — good to merge. 🧠

[vmilosevic]

What is the purpose of this change?