parity-code-review

name	parity-code-review
description	Lisa-native code review of the current git diff. Walks every changed hunk and reports correctness bugs, security issues, and obvious defects as severity-ranked findings with file:line references. Vendor-neutral — the cross-agent equivalent of the upstream code-review command, runnable on Codex, agy, Copilot, Cursor, and Claude.
allowed-tools	["Read","Bash","Grep","Glob"]

Parity Code Review

Review the code that is about to ship — the current uncommitted/branch diff — for defects a reviewer would block on. This is a focused defect hunt: correctness, security, and obvious mistakes. It is not a style audit and not a refactor pass (use parity-code-simplifier for quality-only cleanup).

Not drift-trackable. This skill intentionally carries no synced-from pin. The upstream code-review@claude-plugins-official plugin publishes no semver (its cache version resolves to unknown), so a pin would be unparseable and meaningless to scripts/plugin-parity-drift.mjs. Drift is tracked manually — re-review the upstream command by hand when the curated plugin set is refreshed. This is a Lisa-native reimplementation, not a port of upstream code.

Step 1: Establish the diff

Determine exactly what changed. Prefer the broadest accurate view of the work-in-progress:

# Branch changes vs the merge base (preferred for a PR-style review)
git merge-base HEAD origin/main 2>/dev/null && \
  git diff "$(git merge-base HEAD origin/main)"...HEAD

# Plus anything still uncommitted in the working tree
git diff HEAD
git status --short

If there is no diff at all, say so plainly and stop — do not invent findings. If the diff is enormous, review in full but prioritize the files with the most logic changes; never silently skip files (note any you deprioritized).

Step 2: Read for real context

Do not review hunks in isolation. For each changed file, open enough surrounding code to understand:

• What the function/module is supposed to do and who calls it.
• Invariants and preconditions the change might violate.
• Error/edge paths touched by the change.

Use Read, Grep, and Glob to follow call sites and trace data flow. A finding you can't ground in the actual code is a guess — drop it.

Step 3: Hunt for defects

For every changed hunk, evaluate against these lenses:

1. Correctness — Off-by-one errors, inverted conditions, wrong operator, missing await, unhandled null/undefined, incorrect default, broken control flow, type coercion bugs, mutation of shared state, race conditions.
2. Security — Unsanitized input at trust boundaries; injection (SQL/shell/template); secrets, tokens, or keys committed or logged; missing authn/authz on new endpoints; unsafe deserialization; path traversal; overly broad permissions; SSRF.
3. Edge cases & failure modes — Empty collections, zero, negative numbers, very large input, concurrent calls, partial failures, timeouts, retries that aren't idempotent.
4. Obvious defects — Dead code paths, unreachable branches, swallowed errors, resource leaks (unclosed handles/connections), TODO/FIXME left in shipping code, debug logging left on, broken or missing tests for the new behavior.
5. Contract & API — Breaking changes to public signatures, changed return shapes, altered error semantics callers depend on.

Step 4: Output — severity-ranked findings

Group findings by severity. Within each group, list the most impactful first. Every finding must carry a file:line reference.

Critical (must fix before merge)

Bugs that break correctness, leak/expose data, or introduce a security hole.

Warning (should fix)

Likely to cause problems later, or a real defect with limited blast radius.

Suggestion (nice to have)

Minor correctness nits or defensive improvements.

Finding format

For each finding:

• What — precise description of the defect.
• Where — path/to/file.ts:42 (and a span if it covers multiple lines).
• Why — the concrete failure it causes, with an example input or sequence that triggers it.
• Fix — a specific, actionable suggestion (or a short code sketch).

Example:

Critical — Unhandled null dereference Where: src/auth/session.ts:88 Why: findUser() returns null when the id is unknown, but line 88 reads user.roles directly. An unknown session id (expired token replay) throws and 500s instead of returning 401. Fix: Guard if (!user) return unauthorized() before reading user.roles.

Rules

• Ground every finding in the diff. No speculative findings, no generic best-practice lectures unrelated to the change.
• Be honest about coverage. If you deprioritized files or couldn't fully trace a path, say so.
• If the diff is clean, say so clearly — "No blocking issues found across N changed files" — do not manufacture problems.
• This is review-only: report findings, do not edit files. Apply fixes via the normal implementation flow or parity-code-simplifier (quality) after triage.