// Use when evaluating review comments from AI reviewers (gemini-code-assist, copilot-pull-request-reviewer) on a PR. Default to assuming the reviewer is right; verify before declining. Includes reporting in chat and resolving threads after handling.
Use to monitor an open PR for AI/bot review comments, evaluate them critically, push fixes for real issues, and resolve threads — capped at a sensible duration so it self-terminates. Invoke as "babysit PR
Use when an extension change needs agentic verification in a real browser — content scripts, popup UI, network calls, fonts, console errors. Spawns the agent's own Chrome via the chrome-devtools-ext MCP. Complements (does not replace) human-eye verification via `pnpm dev:browser`.
Use when about to write or modify a Playwright e2e spec under `packages/danmaku-anywhere/e2e/`. Points at the canonical doctrine and surfaces the load-bearing rules so the spec doesn't get bounced in review.
Use when you need to load a specific published preview/nightly extension build into the agent's MCP browser by run-number, branch name, or tag. Faster than rebuilding locally when you only want to exercise an existing artifact.
Use when finishing a brainstorm, writing a design spec, or producing an implementation plan in this project. Directs the output to ClickUp as a doc on the DA-XXX task instead of saving to the repo or local disk.
| name | reviewing-ai-feedback |
| description | Use when evaluating review comments from AI reviewers (gemini-code-assist, copilot-pull-request-reviewer) on a PR. Default to assuming the reviewer is right; verify before declining. Includes reporting in chat and resolving threads after handling. |
AI reviewers catch real issues most of the time. Default to verifying the suggestion against the code, then accepting it. Decline only when you can name the specific rule (CLAUDE.md / AGENTS.md / skill convention) that makes the suggestion wrong here, and quote that rule in the reply.
If you're building an argument for why the reviewer is wrong, stop and re-read the comment. The cost of accepting an unnecessary nit is small; the cost of dismissing a real problem is large.
babysit-pr covers the GraphQL mechanics). Resolve applies to both accepted and declined.When you've worked a batch of threads, emit a structured summary to chat so the human can scan it without paging through the PR:
Reviewed N threads:
- accepted: <count> (commits: <sha-list>)
- declined: <count> (rules cited: <short list>)
- moot: <count> (lines/sections removed in earlier commits)
- handed off: <count> (links + reason)
Add one-line summaries for each batch worth calling out ("hardcoded IDs → placeholders"; "em dashes → hyphens per CLAUDE.md"). Anything handed off needs a one-line reason.
Narrow cases where declining is justified. Match exactly; don't generalize.
gh pr edit the bodyAccept: Fixed in <sha> by <thing>. Decline: cite the rule, e.g. Conflicts with CLAUDE.md <rule> or Single-use value; surrounding prose is the docs. Don't argue; don't relitigate. If the reviewer re-flags the same thing on a later push, back-reference the prior thread and resolve.