Code Review

Code Review

Review a change for correctness first, quality second. A review is not a formatting pass — it's a search for the ways this diff is wrong or will become wrong. Approve when you'd be comfortable being paged for it at 3am.

Scope the review before reading line-by-line

Don't review the diff in isolation. A diff shows what changed, not whether the change is correct.

1. Read the description / linked issue. What is this change claiming to do? You're verifying

the claim, not just the syntax.

1. Read the surrounding code, not just the red/green lines. A removed null check looks fine in

the diff; whether it's a bug depends on the caller three functions up. Open the files.

1. Reproduce the claim mentally (or actually). Trace one real input through the new path. If the

PR says "fixes the retry loop," follow a failing request through the loop and confirm it now terminates. If you can run it, run it.

1. Check the tests changed with the code. New behavior with no new test is a finding. A test

that passes against both the old and new code tests nothing.

What to actually look for

Go in this order — correctness bugs are worth more than style nits.

| Category | Concretely | | --- | --- | | Logic bugs | Off-by-one, inverted conditions, wrong operator (&& vs \|\|), = vs ==, swapped args, fallthrough. | | Edge cases | Empty list/string, zero, negative, null/None/undefined, single element, max size, duplicate keys, unicode. | | Error handling | Swallowed exceptions, ignored return/error values, partial failure leaving inconsistent state, no cleanup on the error path, error message that loses the cause. | | Concurrency | Shared mutable state without a lock, check-then-act (TOCTOU), non-atomic read-modify-write, missing await, deadlock-prone lock ordering, assuming callback order. | | Resource safety | Leaked file/socket/connection, missing close/defer/finally/context manager, unbounded growth (caches, queues). | | API misuse | Calling a function with wrong assumptions about its contract, ignoring documented preconditions, depending on undefined behavior or implementation details. | | Boundaries / input | Untrusted input used unvalidated (defer real security to a dedicated security review, but flag the obvious). | | Tests | New branch with no test, happy-path-only coverage, assertions that can't fail, flaky timing-dependent tests. | | Performance | N+1 queries, work inside a hot loop that could hoist out, accidental O(n²), a query with no index, loading a whole collection to count it. Flag only when it matters at expected scale. | | Readability / reuse | Duplicated logic that already exists elsewhere, a 6-deep nesting that an early return flattens, a name that lies about what the thing does. |

Severity model

Tag every finding so the author knows what blocks merge and what doesn't.

| Severity | Meaning | Blocks merge? | | --- | --- | --- | | Blocker | Correctness bug, data loss, security hole, breaks the build/tests. | Yes | | Major | Likely-wrong under realistic conditions, missing error handling on a real failure path, missing test for new behavior. | Usually | | Minor | Works, but harder to maintain, slightly off-pattern, small inefficiency. | No | | Nit | Pure style/naming preference. Prefix with nit: so it's clearly optional. | No |

If you can't articulate the failure mode, it's at most a Minor — don't inflate taste into a Blocker.

How to phrase feedback

A useful comment points at a line, names the failure, and proposes a fix. Three parts:

> payment.go:142 (Blocker) — If charge() throws after markPaid() runs, the order is marked > paid but no charge exists; the next reconciliation will refund a payment that never happened. > Move markPaid() after the charge succeeds, or wrap both in a transaction.

• Point at file:line so it's locatable.
• Explain the failure mode ("if X then Y goes wrong"), not just "this is wrong."
• Suggest the fix — even a rough direction. A finding with no path forward stalls the PR.
• Ask, don't assert, when unsure. "Is items guaranteed non-empty here? If a caller can pass

[], this indexes out of bounds." — invites the author to confirm rather than starting a fight.

• Praise sparingly and specifically when a tricky thing is done well; it calibrates the rest.

Avoid: vague drive-bys ("this seems off"), restating the diff, bikeshedding naming on a Blocker-laden PR, and demanding rewrites that don't change behavior.

Approve vs request changes

• Approve — no Blockers or Majors; remaining items are Minors/Nits you're happy to see deferred.

Approving with a couple of nit:s the author can take or leave is normal and keeps things moving.

• Request changes — any Blocker, or a Major that materially risks correctness. Be explicit about

which findings gate the merge so the author isn't guessing.

• Comment (no verdict) — you have questions that change your assessment depending on the answer,

or you only skimmed and want to flag that.

Don't block a PR on preferences. Don't approve a PR you didn't actually understand — "LGTM" on a diff you skimmed is how bugs ship.

Review checklist

• [ ] I read the description and know what the change claims to do
• [ ] I read the surrounding code, not just the diff lines
• [ ] I traced at least one real input through the new path
• [ ] Edge cases covered: empty / null / zero / negative / max / duplicate
• [ ] Every error/failure path leaves state consistent and is observable
• [ ] No shared mutable state touched without synchronization
• [ ] Resources are released on every path (success and error)
• [ ] New/changed behavior has a test that fails against the old code
• [ ] No obvious N+1, accidental quadratic, or unindexed hot query
• [ ] No duplicated logic that already exists in the codebase
• [ ] Every finding is tagged with a severity and names a failure mode
• [ ] My verdict (approve / request changes) matches the findings

Source & license

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

• Author: 5dive-ai
• Source: 5dive-ai/skills
• License: MIT
• Homepage: https://5dive.ai

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