babysit

Babysit a PR

Check mergeability first, then run one gh watcher and wait for its process to finish. Success means the intended checks completed for the current PR head, with any failures or missing checks accounted for. Babysitting does not authorize merging, bypassing gates, or posting review comments.

Check for conflicts before waiting

Use an explicit PR number and repository throughout:

gh pr view <n> -R <owner/repo> --json state,url,headRefOid,baseRefName,baseRefOid,mergeable,mergeStateStatus,isDraft,reviewDecision

Record the head SHA, base branch, and base SHA. Stop watching if the PR is closed or merged. Handle mergeable before starting a watcher:

• CONFLICTING: use resolve-pr-conflicts

against this PR's current base, then repeat this inspection after pushing.

• UNKNOWN: GitHub has not established mergeability. Recheck after a short

delay; if still unknown, inspect a fetched head/base locally or report the unresolved state. Do not assume it is conflict-free or loop indefinitely.

• MERGEABLE: proceed. Other merge gates can still block the PR;

mergeStateStatus: BLOCKED alone does not establish a textual conflict.

GitHub does not run workflows triggered by pull_request while merge conflicts exist. Resolving conflicts first prevents waiting for checks that cannot start. See GitHub's workflow troubleshooting.

Let gh watch the checks

gh pr checks <n> -R <owner/repo> --watch --interval 10

The PR checks command refreshes every ten seconds by default and waits until the reported checks finish. Use --fail-fast when a failure should return control immediately for repair. Add --required only when the task concerns required checks alone; it can exclude a review bot's optional check.

Run the command once as a long-running process. Use the execution tool's completion notification or wait on the same session with long practical waits. Do not alternate fresh gh pr checks calls with model-driven sleeps, restart the watcher on each tool yield, or narrate unchanged refreshes.

This is CLI-managed polling, not a server event subscription. The implementation refreshes while checks are pending and exits when none remain. It does not stay subscribed for checks created later or for future review comments.

Inspect the final output and exit status. Failed checks return nonzero; ordinary non-watch output uses exit code 8 for pending checks. Authentication, API, and missing-check errors also need diagnosis. Use a separate snapshot when structured details help:

gh pr checks <n> -R <owner/repo> --json name,bucket,state,workflow,link

Do not combine --json with --watch. JSON is a snapshot, and its successful export does not prove checks passed. Inspect bucket values; cancelled or skipped checks are not evidence that the intended tests ran.

Handle checks that cannot finish

If no checks are reported, the watcher exits with an error rather than waiting for their creation. If an expected check is missing, or a pending check has no progress beyond its normal runtime, inspect the workflow and run details. Check for conflicts again, trigger/path filters, skip annotations, workflow approval, and unavailable runners as the evidence warrants. Do not repeatedly restart a watcher for a workflow that cannot run. Resume once the blocker is resolved; otherwise report the blocker.

For a known GitHub Actions run, use its specific ID:

gh run watch <run-id> -R <owner/repo> --exit-status --compact --interval 10
gh run view <run-id> -R <owner/repo> --log-failed

The run watcher covers one Actions run, not every PR check. --exit-status makes run failure return nonzero. It does not support fine-grained PAT authentication. Inspect failure logs when needed; repair and push within the authorized task, then start with the conflict check again. Do not rerun failures repeatedly without a diagnosis.

A review service that publishes a check can be watched through gh pr checks. A service that only posts comments or reviews is outside these watchers. Use that service's completion mechanism if available; otherwise report that limitation rather than treating green CI as a completed review.

Verify the result

Re-read the PR state, head SHA, base branch and SHA, and mergeability after the watcher exits. Take a final checks snapshot and confirm the expected checks appeared. If the head or base changed, or conflicts appeared, repeat the preflight and watch the current checks before claiming success. A watcher exit alone does not prove merge readiness or review approval.

Report the PR, verified head SHA, check results, and any review or workflow blocker. Continue to merging only when the user's task authorizes it.