watch-opencode

name

watch-opencode

description

Watch a PR's OpenCode review and notify the user when it finishes. Polls the OpenCode workflow run (opencode-review.yml, or the /oc-triggered opencode-comment.yml) until it settles - following supersede when a new push cancels an in-flight review - then reads OpenCode's comment, summarizes its findings, and pings the user via their preferred method (default: a completion chime + treemux attention + an in-chat summary). Use when the user wants to be told when the OpenCode bot is done reviewing a PR, e.g. 'watch the opencode review and let me know when it's done', 'ping me when opencode finishes on PR 1485'.

Watch OpenCode

Wait until the OpenCode bot has finished reviewing a pull request, then tell the user what it found. OpenCode runs as a GitHub Action (anomalyco/opencode/github) and posts its output as a top-level PR comment authored by github-actions[bot]. This skill blocks on the run in the background, so it costs almost nothing while waiting and wakes you exactly once - when OpenCode is done.

What "done" means here

OpenCode never literally writes "done". Each run posts one fresh issue comment with its findings (or a short "nothing to raise" note), so the authoritative completion signal is the workflow run status, not the comment text. Two workflows can produce that comment:

• opencode-review.yml ("PR Review") - the automatic reviewer. Runs on every push (synchronize) and uses cancel-in-progress concurrency, so a new push cancels the in-flight review and starts a new one. This skill follows that chain and only reports the review of the current tip.
• opencode-comment.yml ("OpenCode Comment Trigger") - fires when someone comments /oc or /opencode; this OpenCode can also push commits.

The helper script handles both and resolves the most recent run automatically.

Steps

1. Resolve the PR

• If the user gave a PR number or URL, use it.
• Otherwise use the current branch's PR: gh pr view --json number,url,headRefName,state.
• If there is no PR, tell the user and stop.

Store the PR number (call it PR).

Sanity-check that OpenCode will actually run: the review is skipped on draft PRs, bot-authored PRs, and fork PRs (see opencode-review.yml). If the PR is a draft, say so - OpenCode won't review it until it's marked ready.

2. Launch the watcher in the background

Run the helper in the background (run_in_background: true) so it survives past a single turn and re-invokes you when OpenCode settles:

.claude/skills/watch-opencode/scripts/watch-opencode.sh <PR>

It blocks until the newest OpenCode run on the PR's branch is completed and no superseding run is pending, then prints one line:

OPENCODE_RESULT {"runId":..,"status":"completed","conclusion":"success","headSha":..,"url":..,"name":"PR Review"}

Notable conclusion values:

• success / failure / cancelled - the run's real conclusion.
• none - no OpenCode run exists for this PR (draft / fork / bot PR, or nothing triggered it). Report that instead of waiting further.
• timeout - the safety backstop (~75 min) fired before it settled; re-launch or investigate.

Do not poll it yourself with sleep loops - the background process notifies you on exit. If you need a fallback heartbeat, ScheduleWakeup at a long interval (1200s+), but normally just wait for the completion notification.

3. Read OpenCode's comment

Once the watcher exits, fetch the newest github-actions[bot] issue comment on the PR - that's OpenCode's output for this run:

gh api "repos/{owner}/{repo}/issues/<PR>/comments" \
  --jq '[.[] | select(.user.login=="github-actions[bot]")] | last | {body, html_url, created_at}'

• If conclusion was success but the comment predates the run, the run may have posted nothing new - link the run (url) so the user can check logs.
• If conclusion was failure, OpenCode errored (often an OpenRouter/model hiccup) and likely posted no findings - report the failure with the run URL, don't invent findings.

Read the comment body and distill it: the real bugs/security/perf/UX items (lines prefixed [BUG], [SECURITY], [PERF], [UX], [CONVENTION], etc.), separating genuine findings from a "nothing to raise" all-clear.

4. Notify the user

Notify via the user's preferred method if they named one (e.g. they said "ping me on Slack"); otherwise use the default:

1. Completion sound (macOS): afplay /System/Library/Sounds/Glass.aiff
2. treemux attention so the sidebar flags this worktree: call needs_attention with a one-line summary (the user may be on another worktree).
3. In-chat summary: PR link, run conclusion, and OpenCode's key findings as a short bulleted list, with a link to the full comment (html_url). If OpenCode raised nothing, say so plainly.

Set treemux status to done (or error if the run failed).

Notes

• Everything here is read-only on GitHub - the skill never comments, pushes, or re-triggers anything. It only observes.
• OpenCode posts a new comment per run, so "the latest github-actions[bot] comment" is this run's output; older comments are prior reviews of earlier commits.
• If the user keeps pushing, the watcher keeps following the supersede chain and reports only once the tip's review settles - which is usually what "tell me when it's done" means.