菜单
Use when the user asks to address PR review comments — from human reviewers or automated bots (Gemini, Coderabbit, Copilot, SonarCloud, etc.) — to triage and respond to feedback on a pull request
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | address-review |
| description | Use when the user asks to address PR review comments — from human reviewers or automated bots (Gemini, Coderabbit, Copilot, SonarCloud, etc.) — to triage and respond to feedback on a pull request |
Fetch, validate, triage, and address PR review comments from any reviewer — humans or automated bots. Core principle: skeptical triage, not blind compliance. Reviewers (especially bots) often lack cross-file context and project history — treat their suggestions as hints, not instructions.
Read references/github-api-patterns.md for exact gh api commands, field names, and reply formats.
PR: Use $ARGUMENTS if provided (e.g., /address-review 82 or /address-review 82 coderabbitai[bot] or /address-review 82 alice). Otherwise auto-detect:
gh pr view --json number --jq '.number'
Fail clearly if no PR is found.
Author filter: If a login is passed as the second argument, filter to that login (works for bots like coderabbitai[bot] or humans like alice). Otherwise default to all reviewers on the PR — the triage table makes it easy to scope down after the fact.
Run both calls from the reference file (inline comments + formal reviews). Filter by the login(s) from step 1. Skip comments where in_reply_to_id is set (those are replies, not original comments).
If zero comments found, say so and stop. No further action needed.
For each comment, extract: id, path, line, start_line, body, user.login (who said it), severity (if encoded), and any suggestion block. Group by file path.
Severity conventions vary — see the reference file. Bots typically encode badges/emoji; humans use cues like nit:, question:, blocking:, or plain prose. When nothing is encoded, treat as medium.
For each comment, read the actual file at the referenced lines. Then evaluate:
| Question | If No |
|---|---|
| Is the comment technically correct about the current code? | DISMISS |
| Is the suggestion syntactically and semantically correct? | DISMISS the suggestion (may still FIX differently) |
| Is this worth the churn? (severity vs effort vs risk) | DISMISS |
| Does it conflict with existing architecture or project conventions? | DISMISS |
| Does it require human judgment (trade-off, product decision)? | DISCUSS |
Apply receiving-code-review skepticism:
Weight human feedback more heavily than bot feedback by default — humans usually have context bots lack — but apply the same triage rubric to both.
Assign each comment: FIX / DISMISS / DISCUSS
Present a table:
| # | Author | File | Line | Severity | Verdict | Reason |
|---|--------------------|------------|------|----------|---------|--------|
| 1 | gemini-code-assist | src/foo.ts | 42 | medium | FIX | Redundant sort — suggestion is correct |
| 2 | alice | src/bar.ts | 18 | blocking | FIX | Missing null check on user-supplied input |
| 3 | coderabbitai | src/baz.ts | 7 | nitpick | DISMISS | Already handled by upstream validator |
Then for each comment, show the comment body (truncated), your analysis, and verdict rationale.
ASK THE USER: "These are my proposed verdicts. Approve, or tell me which to change."
DO NOT implement anything until the user explicitly approves. This gate is non-negotiable.
For approved FIX items only:
Reply to every comment — including dismissed ones. The path differs by author type:
Bot comments (*[bot] logins): post directly using the reply endpoint from the reference file. The step 5 triage gate is sufficient approval.
Human comments: confirm each reply with the user before posting. For every human comment, call AskUserQuestion:
question: include the author, path:line, a short snippet of their comment, and your proposed reply textheader: e.g. Reply 1/3 (keep under 12 chars)options:
{label: "Post this reply", description: "Send the proposed text as shown", preview: "<full proposed reply text>"}{label: "Skip this comment", description: "Don't reply to this thread"}The tool auto-adds an "Other" option so the user can type a custom reply. Send whichever text they pick via the endpoint from the reference file; if they choose Skip, leave the thread untouched.
Reply templates (starting point for both bot and human; the user may override for humans):
"Fixed in {sha}. {what changed}.""Not applicable — {technical reason}.""Flagged for discussion — {why}."No performative language. No "Great catch!" No "Thanks for the suggestion!" Just the technical substance — same standard for humans and bots.
Single commit for all fixes: fix: address review on PR #{number}
If all comments were dismissed (no code changes), skip the commit — replies are sufficient.
| Mistake | Fix |
|---|---|
| Implementing before triage approval | The hard gate at step 5 exists for a reason — wait |
| Blindly applying suggestion blocks | Suggestions can be wrong. Validate against the actual code |
| Creating new top-level comments instead of thread replies | Use the /replies endpoint, not /comments |
| Hardcoding owner/repo | Always detect dynamically via gh repo view |
| Hardcoding a single author filter | Default to all reviewers; let the user narrow via argument |
| Performative reply language | Direct and technical only — same standard for humans and bots |
| Skipping replies for dismissed comments | Every comment gets a reply — paper trail matters |
| Treating all comments equally | Severity + reviewer context + your own judgment determine priority |
| Trusting bot suggestions over human review | Humans usually have project context bots lack — weight accordingly |
| Auto-posting replies to humans | Always confirm human replies via AskUserQuestion first — bots can post directly after step 5 |
Orchestrates Android development tasks including project creation, deployment, SDK management, and environment diagnostics using the `android` command-line tool.
Use to find stuck coroutines, leaked jobs, and suspended awaits with no resumer — the concurrency bugs that don't crash but cause hangs, missing UI updates, or memory creep. Ephemerally add `kotlinx-coroutines-debug`, install `DebugProbes` in `Application.onCreate`, run the suspect flow, dump active coroutines with stack traces, fix, then remove the dependency and the install/dump calls. Reach for this when `Log.d` shows a `launch` ran but `collect` never received, or a screen leaves and something keeps running.
Use to flush out intermittent crashes and ANRs ("it crashes sometimes," "fails on slow devices," "happens after rotation a few times") by scripting the trigger sequence with `adb shell input` and running it in a shell loop against a cleared logcat, stopping on first match for `FATAL EXCEPTION` or `ANR in`. Inject deterministic stress (rotation, slow network, low memory) to surface latent races. The empirical answer to "is this still a bug, and under what conditions?"
Use to extract concrete numbers and slice-level evidence from a `.perfetto-trace` captured by `android-perfetto-capture`. Run SQL queries against the trace via the `trace_processor` binary, then hand the (small) result file to a Sonnet sub-agent for verdict — never read raw trace data inline. This is where "the app feels slow" becomes "frame 42 missed budget by 8.3 ms because `MainActivity.onResume` ran 18 ms on `main` while the IO dispatcher was idle." The analysis half of the perfetto loop.
Use to capture Perfetto system traces from a connected Android device — the input for any timing, threading, jank, or "is this on the main thread?" investigation. Writes a self-contained `.perfetto-trace` file to `/tmp` that downstream skills (`android-perfetto-analyze`, `android-trace-sections`) read with the `trace_processor` SQL backend. Pick the strategy that matches the question (one-shot, repeated, on-the-fly), drive the suspect flow during the recording window, and verify the trace contains the expected slices before declaring capture done.
Use when you need empirical proof a code path actually executed — branch entries, callback fires, coroutine continuations, suspect early returns. Insert temporary `Log.d` calls with a unique sentinel tag, redeploy, drive the app, read filtered logcat, then remove every probe before declaring done. The default investigation skill when the question is "did this run, in what order, with what values?"