gh-ci-watch

GitHub CI Watch

Watch GitHub pull request CI until checks pass, fail with actionable logs, or reach a timeout. Keep noisy command output out of the final answer. Save detailed logs locally, then report the important failures clearly. If the logs show a trivial, high-confidence local fix, make it and stage only the files touched by that fix.

Workflow

1. Identify the PR:

   • If the user gives a PR number or URL, use it.
   • Otherwise use the repository at the current working directory.
   • Infer the PR associated with the current branch using gh pr view.
   • If there is no PR for the current branch, stop and report that blocker.
   • Do not switch branches or repositories unless the user explicitly asks.

2. Create a local temporary log directory:

   • Use a deterministic base under ${TMPDIR:-/tmp}/<owner>__<repo>/.
   • Prefer this shape: ${TMPDIR:-/tmp}/<owner>__<repo>/pr-<number>-checks/<utc-timestamp>/
   • Sanitize <owner>__<repo> and branch fallback names to lowercase letters, numbers, dots, underscores, and hyphens.
   • Use gh repo view --json nameWithOwner and gh pr view --json number when available.
   • If no PR number is available yet, use branch-<branch-slug> instead of pr-<number>-checks.
   • Create subfiles such as checks.txt, summary.md, and run-<id>-<job>.log as useful.
   • Tell the user the final log directory path.

3. Check status with short, visible poll turns:

   • Run gh pr checks and save output under the log directory.
   • Default poll interval: 20 seconds.
   • Default timeout: 30 minutes unless the user specifies otherwise.
   • Do not put the entire watch in one long shell loop with a huge block_until_ms / sleep budget. Cursor only surfaces that command's output when the shell finishes or is interrupted, so the chat looks idle while CI is still running.
   • Preferred pattern (multi-turn foreground):
     1. One Shell call: gh pr checks (and optional JSON rollup).
     2. If still pending: briefly tell the user which checks are pending vs done, then Shell sleep 20 (or AwaitShell with ~20s), then poll again in the next turn.
     3. Repeat until green, failed, or timeout. Track elapsed time across turns.
   • Optional background pattern (only if the user asks for background monitoring): start a poller with block_until_ms: 0 that appends to checks.txt and prints clear terminal lines such as CI_STATUS green, CI_STATUS failed, or CI_STATUS timeout. Use notify_on_output for those markers so the agent is woken when the watch ends. Still give an immediate first status in chat after starting it.
   • Between polls, keep chat updates short: pending count / names, any newly finished checks, elapsed time. Do not dump full check tables every turn unless something changed meaningfully.