Correctness Review

Correctness Review

Review whether the changed code actually does what it claims — the bugs that pass type-checking and linting but produce wrong results. This is the dimension the others miss: style, boundaries, coverage, and docs can all be clean while the logic is simply wrong.

Scope

1. Logic

• inverted or off-by-one conditions, wrong comparison or boolean operator
• wrong variable, field, or argument order used
• a calculation that produces the wrong result for a valid input

2. Edge cases

• unhandled empty, null, zero, negative, or maximum inputs
• a missing early return or default branch — a case the code silently skips
• a boundary where behavior flips (first/last, inclusive/exclusive)

3. Control flow and async

• missing await, unhandled rejection, or fire-and-forget that should block
• errors swallowed or caught too broadly; a failure path that returns success
• unreachable code, or a branch that can't be entered

4. Contract

• the function does not do what its name, signature, or doc claims
• a caller's assumption the change now breaks (return shape, nullability, ordering)
• shared state mutated without guarding a race, or written where callers don't expect it

Evidence threshold

Flag a bug only with a concrete trigger: name the input or sequence and the wrong result it produces. If you can't state the input that breaks it, it's not a finding — it's a worry. Don't re-flag what an existing test already covers.

Workflow

1. Read the diff and enough surrounding code to know what each change is supposed to do.
2. For each change, ask: what input or interleaving makes this do the wrong thing? Trace actual behavior against claimed behavior.
3. Check the tests — a green suite around wrong behavior means the test is wrong too; note both.
4. For large diffs (more than 3 files), fan out fast-tier readers — one per area — to surface candidate bugs. Verify each trigger yourself before reporting.

Output

For each finding: label (Critical / Fix / Consider / Nit — see review), file, the bug, triggering input → wrong result, fix direction.

• Bad: "This function looks fragile and could misbehave." (no trigger)
• Good: Critical — pricing.ts:40 applyDiscount uses > not >=, so an order exactly at the threshold gets nothing (total 100, threshold 100 → 0% instead of 10%).

Order Critical → Fix → Consider → Nit. If nothing breaks under a concrete input, report "No correctness findings" — don't invent worries.

See also

• review for severity framing and merge gating
• test-review for whether the bug's absence is guarded by a test
• debug for reproducing and root-causing a confirmed bug

Red flags

• Flagging style, naming, or structure as correctness — that's another dimension
• A finding with no triggering input — speculation, not a bug
• Re-deriving what a passing test already proves, with no new evidence
• "Could theoretically" / "might" with no concrete case

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: cniska
• Source: cniska/skills
• License: MIT

Install and usage instructions live in the source repository linked above.