Fix a Helion pull request

Goal: bring a PR green by fixing CI failures and addressing unresolved review comments. Leave the fixes uncommitted and unstaged — the user handles committing and updating the PR.

This skill fixes the single PR for the HEAD commit and is for pull requests you do not own, or one-off branches not managed with stack-pr. For a stack-pr stack of your own (even one commit), use the fix-pr-stack skill: it amends the fixes, resubmits and watches CI. If GitHub is unreachable from your session, step 0 of fix-pr-stack lists the fallback routes.

1. Identify the target PR

Run git log -n1 to see the current commit. The local commit must correspond to the PR being fixed.

Resolve the PR number using whichever sources are available:

• User-provided URL (e.g. https://github.com/pytorch/helion/pull/1234): extract the PR number directly.
• Commit message stack-info line: look for a line of the form

  stack-info: PR: https://github.com/pytorch/helion/pull/<number>, branch: ...

  Use the URL on that line. Ignore the branch: value — do not run any checkout, branch, or switch commands.
• Both present: if the user-provided URL and the stack-info URL disagree, abort with an error.
• Neither present: abort with an error explaining that the PR could not be identified.

Then fetch the PR metadata with gh pr view <number> --repo pytorch/helion --json title,body,.... Verify the PR title matches the local commit's subject line. If they differ, abort with an error — the local commit does not match the PR.

2. Fix CI failures

Use gh to list the PR's check runs and pull failure logs:

gh pr checks <number> --repo pytorch/helion
gh run view <run-id> --log-failed --repo pytorch/helion

For each failing check:

• Infra failures (most commonly CUDA Compute Check, runner provisioning errors, transient network issues): do not attempt to fix. Note them and report at the end.
• Real failures (lint, type-check, test failures, build errors): read the logs, locate the root cause, and fix it in the working tree. Re-run the relevant check locally when feasible (e.g. ./lint.sh, pytest <file>::<test>) to confirm the fix.

3. Address unresolved review comments

Fetch inline review threads with GraphQL — the REST endpoints don't report whether a thread is resolved:

gh api graphql -F n=<number> -f query='query($n:Int!){repository(owner:"pytorch",name:"helion"){pullRequest(number:$n){reviewThreads(first:100){nodes{isResolved isOutdated path comments(first:20){nodes{author{login} body diffHunk}}}}}}}'

Review summaries and the top-level conversation come from:

gh api repos/pytorch/helion/pulls/<number>/reviews
gh pr view <number> --repo pytorch/helion --comments

For each thread with isResolved: false, and each unaddressed review summary or top-level comment, apply the requested change in the working tree. If isOutdated is true, the code changed after the comment: find the spot via path + diffHunk and check whether it's already addressed. Skip comments that are non-actionable (praise, questions answered in thread) or that the author explicitly waved off.

4. Wrap up

• Leave all fixes uncommitted and unstaged. Do not run git add, git commit, git push, or any rebase/checkout commands.
• Report a short summary covering:
  • PR identified (number + title)
  • CI failures fixed (with file:line where useful)
  • Review comments addressed
  • Any infra failures skipped (so the user knows to retry them)
  • Anything that could not be fixed and why