CI Monitor

Watch GitHub Actions CI runs after a push. Auto-detects repo and branch, deduplicates concurrent monitors, reports results.

Interpreting skill arguments

Anything on the ARGUMENTS: line is a set of monitoring parameters, never a task description. Pick out the branch name and/or hex commit SHA and pass them as --branch / --sha; ignore filler words like "push", "commit", or "monitor". For example, ARGUMENTS: master push 0bed725ba7cb means:

uv run ~/.claude/skills/monitoring-ci/ci-monitor.py --branch master --sha 0bed725ba7cb

This skill is strictly read-only: run ci-monitor.py and report its result. Do NOT edit ci-monitor.py, this file, or any other file; do NOT commit, push, or change the script's interface — even if the arguments look like they don't match the script's flags. If you can't extract a branch or SHA from the arguments, run the script with no flags (it auto-detects) and say so in your report.

Usage

Run the script after pushing:

uv run ~/.claude/skills/monitoring-ci/ci-monitor.py

With explicit branch:

uv run ~/.claude/skills/monitoring-ci/ci-monitor.py --branch my-feature

How to Invoke from Claude

After a push, pass the branch and let the script resolve the SHA:

uv run ~/.claude/skills/monitoring-ci/ci-monitor.py --branch <branch-name> --watch-timeout 480

Run it in the FOREGROUND — do not pass runinbackground — and pass timeout: 600000 on the Bash call.

--watch-timeout 480 and timeout: 600000 are a matched pair, and both are required. The script's own default watch budget is 1800s, which is longer than the Bash tool can ever wait: when the tool timeout expires first, Bash moves the command to the background, this fork's turn ends immediately, and its completion notification carries "monitor running" instead of a verdict — the same verdict-reaches-nobody failure as backgrounding it deliberately, arrived at by a different route. Observed 2026-08-21: a green run was reported as merely "running" and had to be recovered by hand with gh run list.

Bounding the script at 480s keeps it inside the 600s tool ceiling, so it always exits with a real code of its own — 2 if CI genuinely outran the budget, which is honest and actionable, rather than silence.

If Bash reports the command was moved to the background anyway, do NOT end your turn. That message names an output file. Read it, and keep reading until the script's final verdict line appears, then report that. Ending the turn on "moved to the background" is the bug this section exists to prevent.

This skill is context: fork at the default background: true, so it is already detached from the user's conversation. The script blocking here costs them nothing, and the fork's own completion notification is what carries the verdict back.

Backgrounding the script inside the fork breaks exactly that. The fork's turn ends the instant the command is launched, so its notification fires seconds later carrying only "monitor running", while the real monitor keeps running inside a session that no longer exists — its output file is never read and the pass/fail reaches nobody. Observed 2026-08-17: a push was reported as monitored when nothing was watching it, and the status had to be recovered by hand with gh run list.

Report the script's actual verdict when it exits: 0 pass, 1 fail (include the failing logs it printed), 2 indeterminate — not a pass, so pass along the manual-check command from its output.

runinbackground: true is correct only when the monitor is launched from the main conversation instead of from this skill — committing-changes step 7 does this — because that session stays alive to receive the notification.

Do NOT pre-resolve the SHA yourself. The script already resolves it from the bookmark/branch (sha = args.sha or head_sha(branch)), which is correct under every workflow. Passing a hand-computed --sha only overrides that with something more likely to be wrong.

In particular, never derive it from @-. That is the parent of the working copy, which is only the pushed commit in the git-style flow where @ is a fresh empty change on top. In the jj describe @ + jj bookmark move --to @ flow the pushed commit is @, so @- resolves to the commit before it. The failure is silent and dangerous: the monitor watches the previous commit's already-finished run and reports its conclusion, so a green predecessor yields a false pass on unverified work.

Pass --sha only when a SHA was given explicitly on the ARGUMENTS: line.

Features

• Auto-detects branch: jj bookmarks first, falls back to git
• Deduplicates per push: sentinel at /tmp/{repo}-ci-monitor-{sha12} prevents double-watching the same push, while monitors for different pushes coexist; stale sentinels (dead monitor) are taken over automatically
• Polls for run: waits up to 180s for a CI run to appear on the branch (configurable via --timeout); sleeps 5s first so GitHub has time to register the run
• Watches until done: polls gh run view --json status,conclusion every 10s — a transient API error delays detection by one poll instead of crashing the watcher (which is what gh run watch does on a TCP reset)
• Reports failure logs: on failure, fetches gh run view --log-failed (last 3000 chars)
• Cleans up: always removes sentinel, even on error

Exit Codes

• 0 — CI passed (or another monitor already active for this push)
• 1 — CI failed (logs printed)
• 2 — indeterminate: no run found, watch timed out, or gh API kept erroring. NOT a pass — report the manual-check command from the output.

Requirements

• gh CLI authenticated
• GitHub Actions workflows in the repo