lisa-parity-coderabbit

name	lisa-parity-coderabbit
description	Thorough PR-style review of the…
allowed-tools	["Read","Bash","Grep","Glob"]
synced-from	coderabbit@claude-plugins-official@1.1.1

Parity CodeRabbit

A comprehensive, PR-grade code review of the entire diff — the kind of pass a senior reviewer (or an automated reviewer like CodeRabbit) gives before approving. Where parity-code-review is a tight defect hunt, this skill is broad and thorough: it covers bugs, security, performance, and maintainability, suggests concrete fixes, and ends with a reviewer-style summary and verdict.

Independent reimplementation. This skill is not a wrapper around CodeRabbit's hosted service and does not port or invoke CodeRabbit's code. It is a Lisa-native review that produces a comparable artifact using only the model and local tooling. No network calls to any review SaaS are made.

Drift tracking. Pinned to coderabbit@claude-plugins-official@1.1.1. scripts/plugin-parity-drift.mjs compares this pin against the upstream version in the plugin cache and flags staleness. Reimplemented from scratch — do not port or copy upstream plugin code.

Step 1: Assemble the full review surface

Gather the complete change set the way a PR review would see it:

# Full branch diff vs merge base
BASE="$(git merge-base HEAD origin/main 2>/dev/null)"
git diff "$BASE"...HEAD
git diff "$BASE"...HEAD --stat        # file-by-file overview
git log "$BASE"..HEAD --oneline       # commit narrative
git diff HEAD                         # uncommitted work

Read the commit messages and (if available) the PR/ticket description to understand intent — a review judges the change against what it was meant to do, not just what it does.

Step 2: Build context per file

For each changed file, Read the surrounding code and use Grep/Glob to follow callers, dependents, and tests. Note the change's blast radius: public API, shared modules, migrations, config, and anything downstream consumers rely on.

Step 3: Review across all dimensions

Apply the convergent-review rule before deciding severity or whether a finding blocks merge. Bias toward merge: block only concrete correctness, security, data-loss, or contract-violation failures with evidence. Lint-owned style, formatting, taste, and speculative maintainability improvements are non-blocking unless the work item or repository rules explicitly make them release criteria.

Walk every meaningful hunk and evaluate each dimension. A finding in any dimension is fair game.

1. Correctness / bugs — Logic errors, off-by-one, inverted conditions, missing await, null/undefined handling, type coercion, broken control flow, incorrect defaults, mutation of shared state, race conditions, broken or missing tests for new behavior.
2. Security — Injection (SQL/shell/template), unsanitized boundary input, secrets/keys/tokens in code or logs, missing or broken authn/authz, unsafe deserialization, path traversal, SSRF, overly broad permissions, dependency vulnerabilities introduced by the change.
3. Performance — N+1 queries, work inside hot loops, unnecessary allocations or re-renders, missing indexes, blocking I/O on hot paths, unbounded growth, accidental O(n²), redundant network/database round-trips, large bundle additions.
4. Maintainability — Duplication, dead code, unclear naming, overly long/complex functions, missing or misleading docs, leaky abstractions, inconsistent patterns, hidden coupling, magic numbers, and violations of the project's conventions (immutability/functional style, statement order, barrel-export integrity, "never edit generated plugin artifacts").
5. Test coverage — Are the new branches and edge cases tested? Do tests assert behavior rather than implementation? Are failure paths covered?

Step 4: Suggest concrete fixes

For each finding, give a specific, actionable fix — ideally a short code sketch or a diff-style suggestion, not vague advice. The reader should be able to act on it without re-deriving the problem.

Step 5: Output — structured review

Produce a review document with these sections:

Summary

2–4 sentences: what the change does, overall quality, and the headline risks. State a verdict: Approve, Approve with nits, or Request changes.

Findings by severity

Group as Critical → Major → Minor → Nit. Every finding includes:

• What — the issue, and which dimension it falls under (bug / security / perf / maintainability / tests).
• Where — path/to/file.ts:line.
• Why — the concrete consequence, with a triggering example where relevant.
• Blocking — yes only for concrete correctness/security/data-loss/contract failures; otherwise no.
• Failure scenario — the concrete user, operator, security, or factory outcome if the change ships as-is.
• Evidence — the cited file, observed behavior, command output, or contract that proves the scenario is reachable.
• Fix — a concrete suggestion or code snippet.

Walkthrough (optional but encouraged)

A brief per-file note on what changed and any file-specific observations — the orientation a human reviewer leaves so the next reader understands the diff quickly.

Strengths

Call out what's done well. A credible review is balanced, not only critical.

Rules

• Cover the whole diff. If you deprioritize anything for size, say which files and why — never imply full coverage you didn't give.
• Ground every finding in the code. No generic checklists detached from the actual change; no speculative findings you can't point to.
• Malformed blockers do not block. If you cannot name a concrete failure scenario and evidence, report it as non-blocking context or omit it.
• Concrete fixes only. "Consider improving error handling" is not a finding; "wrap the fetch in try/catch and return a 502 on network error at api/proxy.ts:31" is.
• No external review service. Use only local git/tooling and the model — this is an independent review, not a CodeRabbit proxy.
• Review-only. Report findings; do not edit files. Route fixes through the implementation flow, parity-code-simplifier (quality), or a follow-up.
• If the diff is clean, approve it plainly and say why — do not invent problems to look thorough.