Code Review

This skill is for reviewing existing changes. It routes the agent to the right checklist, review mode, and platform workflow without turning SKILL.md into a tool catalog.

Quick Reference

Task	Use	Primary Reference
General PR or diff review	Baseline review flow and severity rubric	references/operational-playbook.md
AI-generated or agent-created changes	Human-in-the-loop review rules and platform controls	references/automation-tools.md
Building or tuning an AI review tool/integration	Pre-review deterministic gate, size-gated planning pass, self-refutation filter pass	references/deterministic-vs-llm-routing.md
Backend/API review	Error handling, contracts, persistence, operability	assets/backend-api/api-review.md
Frontend review	Accessibility, responsive behavior, Core Web Vitals	assets/web-frontend/frontend-review.md
Mobile review	Platform patterns, lifecycle, permissions, UX	assets/mobile/mobile-review.md
Infrastructure review	CI/CD, IaC, secrets, deploy safety	assets/infrastructure/infrastructure-review.md
Smart contract review	Access control, reentrancy, unsafe assumptions	assets/blockchain/crypto-review.md
Data / ML review	Pipelines, experiments, models, deployment	assets/data-ml/ml-diff-review-overlay.md

When to Use This Skill

Use this skill when the primary task is to:

• review a PR, merge request, or diff
• find correctness, security, reliability, or maintainability issues
• assess test gaps and regression risk
• critique AI-generated or agent-created changes before merge
• recommend small, behavior-preserving refactors

When NOT to Use This Skill

• Greenfield architecture → software-architecture-design
• Deep AppSec design or formal threat modeling → software-security-appsec; this skill flags the taint path and the missing control, the sibling designs the fix
• Measured performance investigation (profiling, load tests, capacity) → software-performance for the optimization and qa-testing-performance for the test design; this skill flags hot-path regressions visible in the diff and hands off anything that needs a measurement
• Writing a new feature from scratch → use the stack-specific implementation skill
• Choosing or launching a team of review agents (review board, parallel reviewers, debate) → agents-subagents; this skill still supplies the review criteria each reviewer applies

Workflow

1. Confirm the review surface: diff, PR, merge request, generated code, or a focused file set.
2. Route architecture-only or formal security-design questions to the adjacent skill when review is not the primary task.
3. Trace the changed control flow and data contracts, then apply review modes in order: correctness, security, reliability, performance, and maintainability.
4. Pull in the stack overlay only when the code actually needs that domain-specific lens.
5. Return concrete findings with evidence, then verify any platform-specific claims via the navigation sources.

Review Routing

Apply review modes in this order unless the user asks for a narrower scope:

1. Correctness and edge cases
2. Security and data handling: trace controllable input to each sink in the diff (security-focused-review-guide.md); a finding that needs a threat model rather than a patch goes to software-security-appsec
3. Reliability and operability
4. Performance and cost: hot-path regressions visible in the diff only; measurement questions go to software-performance
5. Maintainability and test coverage

Stack overlays:

• Frontend: assets/web-frontend/frontend-review.md
• Backend/API: assets/backend-api/api-review.md
• Mobile: assets/mobile/mobile-review.md
• Infrastructure: assets/infrastructure/infrastructure-review.md
• Blockchain: assets/blockchain/crypto-review.md
• Data/ML: assets/data-ml/ml-diff-review-overlay.md

Platform overlays:

• GitHub: Copilot review, repository instructions, excluded files, Code Quality, merge queue
• GitLab: Duo review instructions, merge request approvals, approval rules
• Bitbucket: Code Insights, branch restrictions, required checks

If the user asks a platform-specific automation question, open the official entries in data/sources.json first and use references/automation-tools.md.

AI-Assisted Review Rules

• Human review is authoritative. AI findings are advisory until confirmed.
• Treat AI-generated code and agent-created PRs as higher-context review tasks, not lower-effort review tasks: the bottleneck moves from writing to verifying, so budget reviewer time accordingly rather than assuming automation shrinks the review workload.
• Judge AI-generated code against this codebase's actual conventions, not an abstract style guide — consistency with surrounding code is the more useful bar than "is this idiomatic in general."
• Prefer native platform controls before third-party bots:
  • repository or path-specific review instructions
  • excluded files / generated file filters
  • required checks, approval rules, and merge queues
• Do not accept benchmark, pricing, or feature-comparison claims from memory. Verify current tool status first — this space (GitHub Copilot code review, CodeRabbit, Qodo/PR-Agent, and adjacent tools) changes ownership, pricing, and capability frequently; see references/automation-tools.md and data/sources.json and re-verify before quoting specifics.
• Never pin or cite a specific AI model version/ID as a reviewer-facing fact; capability and behavior shift too fast for that to stay true.
• High agent PR volume: route by risk tier (auth, money, migrations, infra, public API escalate to CODEOWNERS for full human review), cap agent-PR intake to reviewer capacity, require stacked layers for large agent changes, name a human reviewer-of-record for every merge, and judge AI reviewers by acted-on rate and noise, not comment count. Details: AI-Volume Review Policy.

AI-Generated Code Review Checklist

Apply these checks when reviewing AI-generated code or agent-created PRs:

Check	What to look for
Hallucinated imports	Packages or methods that don't exist; verify with npm info / pip index
Stale APIs	Deprecated methods, old signatures, removed features
Security gaps	Missing input validation, hardcoded secrets, SQL concatenation, unescaped output
Missing error handling	Happy-path only; no try/catch, no null checks, no timeout handling
Redundant abstractions	Unnecessary wrappers, premature generalization, over-engineered patterns
Copy-paste drift	Similar blocks with subtle inconsistencies across files
Test theater	Tests that assert implementation details, mock everything, or test the framework
Accessibility omissions	Missing alt text, broken ARIA, no keyboard handling, div soup
Design system violations	Components that ignore existing tokens, spacing, or component patterns
Confident but wrong comments	Docstrings that describe what the code should do, not what it actually does

Sources: Hallucination Detection, AI Design Antipatterns

Severity Rubric

Finding proof gate.

Before reporting a blocker or high-severity finding, provide the smallest proof appropriate to its class: a reachable execution or state-transition path, failing example or test, or concrete contract mismatch. Name the violated invariant and user impact, and identify any assumption you could not verify. Reserve controllable-input-to-sink tracing for security and tainted-data findings. If the necessary reachability or contract evidence is unknown, report a question or conditional risk rather than asserting a defect.

Priority	Label	Criteria	Review action
P0	BLOCKER	Data loss, security hole, correctness bug, crashes in prod path	Must fix before merge
P1	REQUIRED	Missing error handling, broken rollback, undefined behavior, SLA violation	Must fix before merge
P2	SUGGESTED	Test gap for non-critical path, minor inefficiency, readability issue	Fix in this PR or tracked issue
P3	OPTIONAL	Style preference, naming taste, future improvement	Author decides; no block

Mark each finding with its priority and label. Do not lump P0 and P3 findings in the same comment thread.

Calibration rules (apply in order; the first that fires wins):

1. Severity is impact if merged as-is, not effort to fix. A one-character off-by-one that corrupts money is P0; a 200-line ugly module that works is P3.
2. Unreachable in production (dead path, feature flag off, test-only code) drops one level and is reported as a question naming the assumption; a reviewer who cannot show reachability does not get to claim P0.
3. Anything on the auth, money-movement, migration, secrets, or public-API surface rises one level, because the fix after merge is a rollback or a data repair rather than a patch.
4. A missing test takes the severity of the behavior it would have caught: P1 for a bug-fix without a regression test on a risky path, P2 elsewhere, P3 when the code is low-change and the gap is tracked.
5. A finding you are unsure of is a question: at the severity it would have if true; never round uncertainty down to P3 to avoid friction, and never assert P0 on a hunch.
6. Performance findings without a measurement cap at P2 unless the diff introduces an obviously unbounded operation on a hot path (N+1 in a request loop, full-table scan on a new query, unbounded in-memory collection); a suspected regression that needs profiling to confirm is handed to software-performance.

What to skip so review time lands on risk: formatting a linter owns (turn the linter on instead), generated files and lockfiles once you have confirmed they carry no dependency or contract drift, pre-existing issues outside the diff's blast radius (file them), and restating the PR description as a comment.

When Not to Block a Merge

Blocking is a cost: it delays value delivery, encourages batching future changes to avoid another round, and burns reviewer credibility if used on low-stakes disagreements. Do not block on:

• P3-only findings (style, naming taste, a "nicer" abstraction) — leave as non-blocking suggestions and let the author decide.
• A correct approach you would have written differently, with no identified defect or maintainability cost.
• Missing test coverage for genuinely low-risk, low-change-frequency code, when the author explicitly acknowledges the gap and it is tracked.
• Pre-existing issues outside the diff's blast radius — file a follow-up instead of expanding this review's scope.
• Disagreements that are actually about product/requirements, not the code — escalate to the right owner rather than relitigating in review comments.

Do block on P0/P1 findings, missing tests for genuinely risky new behavior, and anything that would be expensive or unsafe to fix after merge (data migrations, public API shape, security boundaries). When in doubt, separate the blocking finding from the optional ones instead of letting one bleed into the other's priority.

Rubber-Stamp Detection

Fast, low-comment reviews are not automatically a problem — well-written, small, low-risk changes should review quickly. Treat these as a warning signal warranting a second look, not proof of bad review:

• Approval on a diff well over 400 LOC (Cisco/SmartBear: review under 200, never over 400) with review duration implausibly short for that size (see references/large-pr-review-strategies.md and references/code-review-metrics.md for the size/pace data this heuristic is based on).
• "LGTM" with zero substantive comments on a change touching auth, money movement, migrations, or public APIs.
• A team or individual with a defect-escape rate trending up while review turnaround trends down — the speed gain is likely coming from skipped scrutiny.
• Approvals that only restate the PR description back rather than referencing specific lines or behavior.

Use these as coaching signals (spot-check, pair on a review, ask what was actually read), not as a public leaderboard — see references/code-review-metrics.md for how to track this without creating gaming incentives.

Review the Tests, Not Just the Code

A diff with green tests and no test-quality review is only half-reviewed. Apply the same scrutiny to test code as to production code:

• Confirm new/changed behavior actually has a test that would fail without the fix, not just a test that happens to pass alongside it.
• Watch for tests that assert implementation details (mocking everything, checking internal call counts) instead of observable behavior — these pass trivially and catch nothing on refactor.
• Check that error paths and boundaries are tested, not just the happy path.
• Treat a bug fix with no regression test as incomplete, not as a style nit.

Known Traps

• Reviewing from the PR description first and the diff second. Treat descriptions as claims, not evidence.
• Spending most effort on style and almost none on behavioral risk, migration risk, or rollback safety.
• Treating generated files, snapshots, or lockfile churn as noise without checking whether they hide contract or dependency drift.
• Reviewing only changed lines when the bug is in the surrounding invariant, caller expectations, or teardown path.
• Calling out needs tests generically without naming the missing scenario, boundary, or regression.
• Accepting benchmark, security, or framework claims in the diff comments without live verification when the claim is version-sensitive.

Common Anti-Patterns

• Treating review as approval theater: light comments on naming while correctness and rollout risk remain unexamined.
• Rewriting the author's architecture in review when the real issue is a smaller bug, missing guardrail, or weak contract.
• Conflating preference with defect. Mark taste as optional and reserve blocking comments for correctness, safety, or maintainability risk.
• Collapsing multiple independent issues into one large comment thread instead of separating discrete findings with evidence.
• Using AI review output as authoritative instead of validating each finding against the actual diff and local context.

Default Review Output

Default to:

• a short summary of intent and risk
• findings grouped by P0 / P1 / P2 / P3
• REQUIRED vs OPTIONAL labeling
• concrete remediation: minimal diff, test case, or configuration fix
• follow-up questions only when requirements are genuinely unclear

Use assets/core/review-comment-guidelines.md for phrasing and assets/core/review-checklist-judgment.md for final pass judgment. Label comment intent using the Conventional Comments convention (suggestion:, issue:, question:, nitpick:, praise:) so blocking vs. non-blocking intent is unambiguous without relying on tone.

Navigation

Core references:

• references/operational-playbook.md
• references/automation-tools.md
• references/review-checklist-comprehensive.md — full checklist by concern plus the review-process practices from "Looks Good To Me" and "Implementing Effective Code Reviews"
• references/security-focused-review-guide.md
• references/large-pr-review-strategies.md
• references/code-review-metrics.md
• references/psychological-safety-guide.md
• references/dotnet-efcore-crypto-rules.md
• references/adversarial-review-protocol.md — stripped-context handoff and fixed-precedence finding reconciliation, two-reviewer gate, and batch sampling for pre-commit decision review
• references/complexity-only-review-pass.md — narrow, tagged, over-engineering-only review pass as a fast complement to the full checklist
• references/deterministic-vs-llm-routing.md — pre-review deterministic file gate, size-gated planning-pass threshold, and self-refutation filter pass for AI review tooling

Templates:

• assets/core/pull-request-description-template.md
• assets/core/review-comment-guidelines.md
• assets/backend-api/api-review.md
• assets/web-frontend/frontend-review.md
• assets/mobile/mobile-review.md
• assets/infrastructure/infrastructure-review.md
• assets/blockchain/crypto-review.md

Sources:

• data/sources.json - verified source set with platform status and replacements for retired links

Freshness Protocol

When the user asks for:

• the best or latest code review tool
• current PR automation practices
• GitHub Copilot vs CodeRabbit vs Qodo vs platform-native options
• whether a tool is still relevant

you must:

1. Use web search.
2. Prefer official docs and release/status pages over vendor comparison blogs.
3. Use data/sources.json as the starting source set.
4. Treat vendor metrics and case-study performance numbers as directional unless independently verified.

If web access is unavailable, say so and answer from data/sources.json, clearly marking time-sensitive advice as unverified.

Learnings Loop

When prior decisions or pitfalls are relevant, consult learnings.consolidated.md if present; use learnings.md only for needed history or as the available fallback. Otherwise skip both.

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.