Production Code Review

Purpose

Give a change a review with the rigor of a senior engineer who reads for consequences, not just syntax — and who is explicit about what's a proven fact versus a plausible risk versus a nice-to-have.

When to use

• A diff or pull request is ready (or claimed ready) to merge.
• Explicitly asked "review this," "is this safe to ship," or similar.
• After feature-to-production implements a change, as its review stage.

When NOT to use

• No specific change exists yet — reviewing an entire codebase's health is repo-architect's job, not this skill's.
• The only concern is naming/formatting with no bearing on correctness, security, or long-term maintainability — don't manufacture findings.

Required inputs

• The diff, PR, or set of changed files.
• Enough surrounding code (call sites, related tests, schema) to judge real impact rather than reviewing the diff in isolation.

Workflow

1. Read the full diff before forming any opinion — do not review hunk by hunk without the surrounding context of what the change is trying to do.
2. For each of the 12 dimensions below, actively check for issues — don't wait for something to jump out. Absence of a finding in a dimension is itself worth noting mentally, not proof there's nothing there.
3. For each candidate finding, gather the evidence (the actual line, the actual call site, the actual test — or lack of one) before writing it up.
4. Classify each finding: FACT (directly demonstrable from the code as written), LIKELY RISK (a plausible failure mode given the evidence, not certain), or OPTIONAL IMPROVEMENT (would be better, not a defect).
5. Assign severity per the rubric below.
6. Write up findings ranked most severe first, each with severity, file/location, evidence, impact, and a recommended fix.
7. Summarize: is this safe to merge as-is, safe with the P0/P1 items fixed, or not safe to merge.

Review dimensions

Correctness · Security · Reliability · Error handling · Concurrency · Performance · Data integrity · API contracts · Maintainability · Observability · Tests · Backward compatibility

• For security-sensitive changes (auth, input handling, secrets, injection surfaces), read references/security-checklist.md.
• For anything touching shared state, async code, hot paths, or resource usage, read references/concurrency-and-performance.md.
• For anything changing a public API, DB schema, or wire format, read references/api-contracts-and-compatibility.md.

Severity rubric

• P0 — critical: data loss/corruption, security vulnerability, or a correctness bug on a common path. Blocks merge.
• P1 — high: correctness bug on an edge case, missing error handling that causes a crash/hang, or a reliability gap in a critical path. Should block merge or ship with an explicit, informed exception.
• P2 — medium: real but bounded impact — a performance regression, a maintainability problem that will cost real time later, a test gap on a non-trivial path.
• P3 — improvement: worth doing, not blocking — clarity, minor duplication, a more idiomatic approach.

Tool & resource guidance

Read actual call sites and tests, don't assume behavior from function names. When a finding depends on how a function is used elsewhere, grep for its usages rather than asserting impact from the diff alone.

Output contract

A findings list, ranked P0 first, each with:

• Severity (P0-P3) and classification (FACT / LIKELY RISK / OPTIONAL IMPROVEMENT)
• File/location
• Evidence — the actual code/behavior that supports the finding
• Impact — what happens if this ships as-is
• Recommended fix

Followed by a one-paragraph merge recommendation.

Quality checks

• Every finding has file/location and evidence — no vague "this could be a problem somewhere."
• FACT findings are actually demonstrable, not classified as fact out of confidence alone.
• No P3 findings are presented as blocking; no real P0/P1 is downgraded to keep the review friendly.
• Style-only observations are excluded unless they materially affect maintainability (e.g. a pattern that will cause real bugs when copied).
• The merge recommendation matches the severities found — don't say "safe to merge" alongside an unresolved P0.

Edge cases

• Large diff spanning unrelated concerns: note that as a P2/P3 maintainability/process finding (should have been split), but still review the actual content fully.
• No tests exist for the changed path: this is a finding (severity depends on how critical the path is), not a blocker to reviewing the rest of the diff.
• Change is trivial (e.g. a config value bump): say so plainly and give a short review rather than manufacturing findings across all 12 dimensions.
• Reviewer disagrees with an existing architectural pattern the diff merely follows: don't flag pre-existing patterns as new findings unless the diff makes the specific instance worse.

References

See examples/idempotency-key-review.md for a full worked example of the output format, including severity classification and a merge recommendation.