manual-pr-review
暂无简介
- License 许可
- MIT
- 最近更新
- 2026-11-10
- Stars
- ★ 0
- Forks
- 0
02 资源文件 1
- 参考文档agents/openai.yaml105 B
03 SKILL.md 正文
Manual PR review
Leave the reviewer able to explain the change and defend a decision about it. You are the guide, not the gate: the reviewer decides, you make the decision cheap.
1. Read the PR before you opine
Fetch, in order: the PR (title, body, changed files, base/head, additions/deletions), the full diff, and the issue(s) it closes plus any ADR or design note it cites. Read the whole diff before writing a single finding.
The PR body often already states the intent, the risks the author knows about, and the follow-ups. Quote it rather than inventing a frame: which PR is the base, and does the diff carry only this PR's commits or an ancestor's too.
Ground every "why" in the issue/ADR, never in a guess. If the PR closes #78 and an ADR exists, cite them.
2. Lead with an ELI5
The reviewer needs a model before details:
- one sentence: what changes, in plain words;
- the problem it solves: the failure or gap, cited to the issue;
- the mechanism: the smallest accurate account of how it works.
3. Show the process before and after
A file-by-file diff hides the shape of the change: what used to happen end to end, and what happens now. Walk the flow the diff touches twice, once before the change and once after, in the same plain vocabulary as the ELI5. Name the entry point, the steps in order, and where the new code takes over. Cover only the paths the diff changes, not the whole system. Two short paragraphs or a two-row before/after table both work. If the flow did not change, say that instead of inventing a difference.
4. Give a reading order, leaf → complex
Order files by dependency: types and leaves first, load-bearing logic last, so each file is understood on already-read ground. A table works: file, why here, and what to ask while reading it. The per-stop questions are the point ("is this injective?", "does the hash commit to what is enforced?", "any behaviour drift?").
5. Trace provenance of new inputs
For every new field, parameter, or config, say who authors it, where it comes from, and whether it is trustworthy. Distinguish values a witness/attestor signed from values the client merely asserts. "Where does this actually come from?" is the most common real question.
6. Findings: concrete, verified, cited
Reproduce before asserting. If you claim Array(1) collides with [], run it and paste the output. A finding without a demo is a guess.
Hunt specifically for:
- Doc drift: a comment that no longer matches the code (a schema that still says "no length check" after gaining one).
- Name collisions: the same identifier across entry points or subpaths meaning different things.
- Scope creep: unrelated work smuggled into the PR; recommend it move to its own.
- Asymmetry: one object under two rules (two fields length-checked, two not).
- Unchecked fields: signed but never verified; say what it buys and who must check it.
- Disclosure / privacy: a circuit
discloseor public field leaking more than intended; say whether that is accepted. - Removed guards: a deleted assert or validation; say what catches the case now.
- Frozen vectors moving: a changed hash, canonical form, or wire layout; name what else depends on the old one.
For each finding, say whether it is live or latent, and whether it is a bug, a doc fix, or a judgement call for the author. Cite file:line.
7. Say what's in and what's out
List the deferred work the PR names, plus anything you notice it does not cover. Reviewers need the boundary, not just the content.
8. Optional: a visual map
When the key file is dense (a hashing routine, a state machine, a wire layout), a standalone HTML map (purpose, pipeline, line ruler, review radar) compresses it better than prose. Build it only when it earns the effort; skip it for a small diff.
9. Answer questions the way they're asked
A yes/no question gets a yes or no first, then the reasoning. If the reviewer agrees with a finding, verify it before amplifying. If you disagree, say so and show why.
Output shape
Default to this order unless the reviewer asks otherwise:
- ELI5: sentence, problem, mechanism.
- The process before and after: the touched flow as it ran before and as it runs now, in the ELI5's vocabulary. Two short paragraphs or a two-row table.
- The changes: a short table: layer → before/after.
- Review order: the leaf→complex table with per-stop questions.
- Findings: numbered, each with a verdict (live/latent, bug/doc/judgement) and
file:line. - In / out: deliberate scope and known follow-ups.
- Offer: the visual map, or acting on a finding, when either helps.
Relationship to code-review
code-review produces a Standards/Spec compliance verdict via parallel sub-agents against a fixed point. This skill produces understanding and human-facing findings against a PR. They compose: run this to learn the change and gather findings, run that when a formal two-axis verdict is wanted. If the reviewer only wants the verdict, defer to code-review; if they want to understand and judge, stay here.
When asked to act, not just review
If the reviewer asks you to fix a finding: put it in the PR that owns the change (the one that introduced the code), not a drive-by in an unrelated one, and reply on the review thread with the commit. Watch scope: a review finding that belongs in a later PR should be filed, not smuggled.
04 同仓库更多 Skills
improve
暂无简介
create-verification-skill
Generate a project-local verification skill that drives your app the way a user does — any language, framework, or platform. Use for /create-verification-skill, "make a control skill for this repo", or when a project has no scripted way to prove UI/CLI/service behavior.
tdd
Test-driven development. Use when the user wants to build features or fix bugs test-first, mentions "red-green-refactor", or wants integration tests.
maintain-verification-skill
Periodic pass that keeps a project's verification skill and feature map honest: parallel source readers per feature, one live session driving every feature, at most one PR of proven corrections. Use for /maintain-verification-skill or "audit the verify skill".
verdict
The Reviewer's Verdict on a PR: re-run the repo's verification skill at the PR's head, hold the PR's Evidence against it, and record pass or fail with nico-agent.
research
Investigate a question against high-trust primary sources and capture the findings as a Markdown file in the repo. Use when the user wants a topic researched, docs or API facts gathered, or reading legwork delegated to a background agent.