Reviewing changes

Process

Always run the four default passes in order. Findings flow into one combined verdict. Pass 5 (AI-native practices) is opt-in: run it only when the user explicitly asks for it — some repos are deliberately not AI-native and should not be graded against that rubric.

1. Read the rules

Before any pass, internalise the language-rule skill that matches the diff (python-conventions, go-conventions, solidity-conventions, …) and engineering-philosophy. The rule skills are the source of truth — don't invent additional standards.

2. See the change

git diff <base>...HEAD
git log <base>..HEAD --oneline

For a GitHub PR:

gh pr view <N>
gh pr diff <N>

3. Pass 1 — Code quality

Check, in order:

• Philosophy violations — over-engineering (KISS, YAGNI), duplication (DRY), magic behaviour (No Magic), copy-paste-modified blocks.
• Redundant entity & local-pattern reuse: when the diff adds a function, accessor, or code path the repo already provides (an auto-generated getter, a wider getter that subsumes a narrower one, an idiom a sibling file implements), reuse the existing form instead of a second variant. DRY and YAGNI span the whole repo, not just this diff.
• SOLID violations — Single Responsibility first; flag classes/files that grew a second responsibility.
• Naming, readability, complexity — function lengths, parameter lists, deeply nested conditionals, clever one-liners that hide intent.
• Test coverage — was the change tested? If TDD discipline applied, was the failing test committed first?
• Tooling compliance — ruff/mypy strict for Python, golangci-lint for Go, solhint:all for Solidity, forge fmt --check for Solidity formatting.
• Configuration safety — production timeouts, connection pools, missing retries, missing rate limits.

4. Pass 2 — Security audit

Check, in order, against OWASP Top 10:

• Injection — SQL, command, LDAP, template, header injection via unsanitised input.
• Broken authentication — weak token handling, missing MFA, fragile session management.
• Broken access control — missing authz checks, privilege escalation, IDOR (insecure direct object reference).
• Sensitive data exposure — secrets in logs, error messages, or response bodies; missing TLS; weak ciphers.
• Misconfiguration — overly permissive CORS, missing security headers, debug endpoints exposed.
• Vulnerable components — pip-audit / npm audit / govulncheck / dep CVEs.
• XSS — unencoded output rendered as HTML/JS; missing CSP.
• Insecure deserialisation — pickle.loads on untrusted input, similar in JS/Java.
• Insufficient logging and monitoring — security-relevant events not logged, no alerting.
• Cryptographic issues — weak algorithms, hardcoded keys, missing key rotation, predictable IVs.
• Smart-contract specific (if Solidity) — reentrancy, integer over/underflow, unchecked external calls, access control on onlyOwner-style modifiers, front-running, MEV exposure, signature replay.

See reference/owasp-checklist.md for the canonical mapping with attack-vector notes.

5. Pass 3 — Architecture consistency

• Architecture map — does the diff respect the docs/architecture.md (or equivalent) responsibility split? See reference/architecture-map-pattern.md.
• Layer violations — dependencies pointing the wrong way (e.g., domain importing infrastructure).
• Boundary erosion — public methods sneaking into private packages; circular dependencies.
• Missing abstractions — same logic implemented twice with minor variations.
• Custom code where a library exists — presumptive Critical when the diff reinvents primitives the ecosystem already solves (cryptography, encoding, standard-format parsers, wire codecs, retry/rate-limiting, ORMs, validators). Major for general utility code with a battle-tested equivalent. Three sub-checks:
  • Already in tree — if the project's lockfile already pulls in a library that exports the function being hand-rolled, the hand-rolled version is Critical regardless of LoC. Don't import one symbol and reinvent the others.
  • Justification still valid — comments that justified hand-rolling earlier ("avoid coupling", "keep dep tree small", "minimise binary size") must still hold for this diff. Once the dep is in the tree, the original reason has expired.
  • What to grep for — custom encoders for standard formats, raw wire-protocol bytes as constants, hand-rolled crypto primitives, hand-written auth-token verification, custom retry-with-backoff loops.
• Pattern compliance — clean architecture / DDD bounded contexts, only when the project documents a pattern.

6. Pass 4 — Acceptance / intent alignment

Does the diff actually solve the contract — linked GitHub issue, PR description, or active Spec Kit specs/<NNN>-<feature>/tasks.md Block? Cover the three axes:

• Drift — diff implements something related but not the asked feature.
• Partial — diff covers some required behaviours but misses others.
• Overreach — diff includes changes the issue did not request.

See agents/acceptance-auditor.md for the full procedure.

7. Pass 5 — AI-Native-Coding Practices (opt-in, not part of the default gate)

Run only on explicit request (e.g. /coding-skills:review ai-native or the user asking for an AI-native check). Validates the diff and the surrounding project against the empirically-grounded rubric for working with AI coding agents. Eight rules, citation-grounded:

• R1 — Comments WHY not WHAT (load-bearing for LLM coding; inaccurate comments harmful).
• R2 — Durable agent context belongs in instruction files. At least one of AGENTS.md / CLAUDE.md / .cursor/rules/ must exist at repo root; section structure is guidance, not a graded checklist.
• R3 — Tests prefer real objects over mocks; mock only at I/O boundaries.
• R4 — Architectural decisions need ADRs — only for projects using Spec Kit (detected via specs/ + plan.md/tasks.md or .specify/). Non-Spec-Kit projects skip R4.
• R5 — Code review and PR hygiene survive AI throughput (small commits, decomposed PRs).
• R6 — Conversational interaction is iterative, not one-shot.
• R7 — Minimize context: delete, don't tombstone. Every line in always-loaded files costs every turn — remove rather than mark removed.
• R8 — Mechanical-rubric subset belongs in CI; bundled templates ship with the plugin.

The rubric, with citations, lives at reference/ai-native-rubric.md. The mechanical-check templates ship at reference/ai-native-templates/. See agents/ai-native-reviewer.md for the full procedure.

Output

## Quality Gate Summary

| Review              | Verdict        | Critical | Major | Minor |
|---------------------|----------------|----------|-------|-------|
| Code                | pass/warn/fail | N        | N     | N     |
| Security            | pass/warn/fail | N        | N     | N     |
| Architecture        | pass/warn/fail | N        | N     | N     |
| Acceptance          | pass/warn/fail | N        | N     | N     |

**Overall**: PASS / NEEDS WORK / FAIL

### Action items
1. <Critical/Major items, ordered>

Append an AI-Native Practices row only when the opt-in Pass 5 ran.

For each individual finding:

• Rule — which rule was violated (with the language-rule skill or engineering-philosophy reference) or "best practice" if no codified rule.
• Severity — Critical / Major / Minor.
• Location — file:line.
• Issue — what's wrong and (for security) the attack vector.
• Fix — concrete suggestion, with a short code example when it clarifies the change.

Behavioural traits

• Constructive, educational tone. Teach; don't just flag.
• Specific, actionable feedback. "This is too complex" without a fix is useless.
• Severity matches reality. Critical for "this could ship a bug or a CVE today"; Major for "this will hurt within six months"; Minor for style and polish.
• Practical over theoretical security risks. If an attack requires three impossible preconditions, mark Minor.
• Defence in depth. Multiple weak controls beat one perfect control.
• Read-only. This skill never edits the diff itself; it reports.

Cross-references

• running-tdd-cycles — preceding workflow; review confirms TDD discipline.
• committing-changes — commit-message + branch hygiene checks fold into the code-quality pass.
• python-conventions / go-conventions / solidity-conventions — the language rule the diff is being checked against.
• engineering-philosophy — KISS, YAGNI, DRY, SOLID weights for code-quality and architecture passes.

Reference

• reference/owasp-checklist.md — canonical OWASP Top 10 mapping with attack vectors and fix patterns.
• reference/architecture-map-pattern.md — the optional docs/architecture.md convention this skill expects.

The live subagent shims that wrap this skill for parallel execution live one level up:

• agents/code-reviewer.md — Pass 1 (code quality)
• agents/security-auditor.md — Pass 2 (security)
• agents/architect-review.md — Pass 3 (architecture)
• agents/acceptance-auditor.md — Pass 4 (acceptance / intent)
• agents/ai-native-reviewer.md — Pass 5 (AI-native-coding practices, opt-in)

Each is a thin shim that reads this SKILL.md (and, for Pass 5, also reference/ai-native-rubric.md) and applies its scoped pass; the /coding-skills:review slash command spawns the four default passes in parallel under model: opus, adding Pass 5 only on explicit request.