/watch-ci

name: watch-ci
description: "Monitor GitHub Actions CI runs until completion. Use when: watching CI after push, checking build status, monitoring PR checks, waiting for CI completion, user says 'watch CI', 'check CI', 'CI status', 'monitor build', or /watch-ci. Not for: pushing code (use push-ci), creating PRs (use create-pr). Output: per-run verdict (pass/fail/timeout)."
allowed-tools: Bash(gh:*), Bash(git:*), Read, Monitor

Watch CI

Monitor GitHub Actions CI runs for the current HEAD (or a specified SHA) until completion, then report verdict.

Trigger

• Keywords: watch CI, check CI, CI status, monitor build, build status, is CI passing, watch actions, CI result

When NOT to Use

• Pushing code to remote (use `/push-ci`)
• Creating pull requests (use `/create-pr`)
• Running local tests (use `/verify` or `/precommit`)

Workflow

Auto-detect (branch + SHA) → Find matching runs → Quick-check status → Watch or Report → Verdict

Step 1: Resolve Target

Determine which CI runs to monitor. Use arguments if provided, otherwise auto-detect.

BRANCH=${ARG_BRANCH:-$(git rev-parse --abbrev-ref HEAD)}
HEAD_SHA=${ARG_SHA:-$(git rev-parse HEAD)}
TIMEOUT=${ARG_TIMEOUT:-10}
INTERVAL=${ARG_INTERVAL:-30}

If `--run-id <id>` is specified, skip run discovery and monitor that specific run directly.

Step 2: Find CI Runs

Find runs matching the target SHA on the target branch:

gh run list --branch "$BRANCH" --limit 30 \
  --json databaseId,headSha,status,name,url

> **Note**: Use `--limit 30` (not 10) to avoid missing target SHA runs on busy branches. Filter results client-side by `HEAD_SHA`.

Filter results to those matching `HEAD_SHA`.

**Retry logic**: If no matching runs found, retry up to 3 times by re-running the `gh run list` command. The natural processing delay between retries provides sufficient wait time — **do not use `sleep N` (N ≥ 2) as the first command**, the harness will block it. All retry commands must start with `gh` or `git` to match `allowed-tools`. CI workflows may take a few seconds to trigger after push.

If still no runs found after retries:

⚠️ No CI run detected for SHA <sha>. Possible causes:
- No workflow configured for this branch
- Path-filtered workflow didn't trigger
- Check: gh run list --branch <branch>

Step 3a: Quick Status Check

Before starting a long-running watch, check if runs are already completed:

gh run view <run-id> --json status,conclusion,name,url

| Result | Action | |--------|--------| | All runs completed | Skip to Step 4 (Verdict) immediately — no watching needed | | Some completed, some in progress | Report completed verdicts, watch remaining (Step 3b) | | All in progress | Proceed to Step 3b |

Step 3b: Watch Runs

For each in-progress run, monitor with `gh run watch`:

gh run watch <run-id> --exit-status -i "$INTERVAL"

**Poll interval**: `$INTERVAL` defaults to 30 seconds (configurable via `--interval`). `gh`'s own default is 3 seconds; at that rate, Monitor streaming surfaces ~20 notifications per minute, which has been reported as noisy. 30 seconds reduces poll noise by ~90% at the cost of ≤ 27 s additional completion-detection lag, which is negligible for typical multi-minute CI runs. Pass `--interval 3` to restore the old cadence when near-real-time feedback matters.

**Execution mode**: Monitor streaming is the default — non-blocking, reliable notifications for each status line.

| Mode | When | Behavior | |------|------|----------| | Monitor (default) | No mode flag | Stream `gh run watch` via Monitor tool. Each stdout line arrives as a notification. Claude processes verdict on completion. Non-blocking. | | Foreground (`--blocking`) | `--blocking` flag passed | Execute `gh run watch` inline (blocking). Claude waits for completion, then reports verdict. Use when Monitor is unavailable or for simple single-run cases. | | Background (`--background`) | `--background` flag passed | Launch with `Bash(run_in_background: true)`. Legacy fallback only — `run_in_background` delivers a single completion event, not streaming progress, so Monitor is preferred for rich updates. Provide a manual check command. |

**Monitor mode (default) — behavior**: 1. Launch `gh run watch <run-id> --exit-status -i "$INTERVAL"` via Monitor tool with `description: "CI run <run-id> (<name>)"` and `timeout_ms: TIMEOUT * 60 * 1000` 2. Each stdout line (status update) arrives as a streaming notification — `$INTERVAL` controls how often those lines fire 3. On exit (run completes or fails), parse final output for pass/fail status 4. Report verdict

**Foreground mode (`--blocking`) — behavior**: 1. Execute `gh run watch <run-id> --exit-status -i "$INTERVAL"` inline via Bash 2. Wait for completion (blocking) — `$INTERVAL` only affects how often `gh` polls the API, not wall-clock completion 3. Parse output for pass/fail status 4. Report verdict

**Background mode (`--background`) — legacy fallback only**: 1. Quick-check (Step 3a) first — if already completed, report immediately and skip background 2. If still running, launch `gh run watch <run-id> --exit-status -i "$INTERVAL"` with `Bash(run_in_background: true)`; `$INTERVAL` still applies (same `gh` call), but because background mode only surfaces a single completion event, poll cadence has no user-visible effect here 3. Inform the user honestly: "CI monitoring launched in background for run `<id>`. Background notifications may not auto-report reliably. To check manually: `gh run view <id>` or re-run `/watch-ci`" 4. **Do NOT promise streaming progress updates** — `Bash(run_in_background: true)` only delivers a single completion event, not per-status-line streaming; for rich updates, use Monitor mode

**Multiple runs**: If multiple workflow runs match (e.g. CI + Auto Release), launch parallel Monitor instances — one per run. Each Monitor reports its own per-run verdict via notifications. Overall verdict = worst individual result (any fa