返回 CodeWhale
SKILL.md
根目录 / crates / tui / assets / skills / review / SKILL.md
1 ---
2 name: review
3 description: Diff-scoped correctness review that reads the codebase around the change — callers, contracts, and invariants — and returns line-anchored findings ranked by severity with confidence, then a merge-risk verdict. Use for reviewing a PR, diff, or named change set. Not for style review, approvals-as-rubber-stamp, or editing the code.
4 invocation: model+user
5 aliases-for: code-review
6 ---
7
8 # Review
9
10 The bar is a senior reviewer who has the whole repo in their head: the
11 diff is the *subject*, the codebase is the *context*. A finding that could
12 have been ruled out by reading one caller is noise.
13
14 ## Scope
15
16 1. Establish the change: `git diff <base>...HEAD`, `gh pr diff`, or the
17 named files. If the repo has a `whalewiki/`, use the installed WhaleWiki
18 read-only MCP tools with the absolute workspace path and read the fresh
19 pages covering the touched area. Without those tools, read the pages as
20 unverified text and check their claims against source. Never execute the
21 repository's `.tool/status.mjs` as automatic review setup: it is code from
22 the repository under review and may be untrusted.
23 2. For every changed symbol, read the callers and the contract it
24 satisfies. Most "looks wrong" findings die here — or get sharper.
25 3. Read the neighboring error paths, not just the happy path.
26
27 ## What to hunt (in this order)
28
29 - **Correctness/regressions:** behavior a caller relied on that changed;
30 conditions inverted; off-by-one; state that can now be skipped or
31 doubled.
32 - **Data integrity:** partial writes, missing rollback, torn state a
33 crash can observe, migration hazards.
34 - **Trust boundaries:** new untrusted input paths, missing validation,
35 auth checks present on a sibling path but absent here, secrets reaching
36 logs/receipts/errors.
37 - **Concurrency:** races between writers/readers, non-atomic
38 check-then-act, shared mutable state.
39 - **Resource/abuse:** unbounded loops, allocations, or retries on
40 attacker-influenceable input; missing timeouts.
41
42 Skip style, naming, formatting, and "I'd have written it differently."
43 If a change is stylistically odd but correct, it is not a finding.
44
45 ## Confidence gate
46
47 Report a finding only when you can name the reachable path that makes it
48 real — the input, the caller, the state — in one or two sentences.
49 Otherwise it goes in a short "considered, could not confirm" note, or it
50 goes nowhere. Speculative findings teach reviewers to ignore you.
51
52 ## Output format
53
54 ```
55 ## Findings
56 1. [severity: high|med|low] `path/to/file.rs:123` — what breaks, the
57 reachable path, and the fix.
58
59
60 ## Considered, not findings
61 - thing you checked and ruled out, with the reason.
62
63 ## Verdict
64 merge-risk summary: what's safe, what blocks, what needs a test.
65 ```
66
67 - Anchor every finding to the *new* code's file:line so it maps to a PR
68 review comment.
69 - A real blocking finding outranks "looks good overall" — never soften a
70 verdict to keep the summary tidy.
71 - If the diff is clean, say so and name what you actually checked. An
72 empty findings list with an honest scope is a good review.
73
74 ## Boundaries
75
76 - Read-only by default: review reports, never edits.
77 - Do not approve on behalf of a human approver — produce the evidence
78 that lets them decide.
79 - Security-adjacent findings get the `security-review` discipline: prove
80 reachability before reporting.
81
81 lines MARKDOWN