---

name: performance-review description: Review code changes only for performance and efficiency risks (N+1 queries, repeated work, bad asymptotics, missing batching/caching, blocking I/O, memory growth). Use when the user explicitly asks for a performance review of a diff/PR.

Performance Review

In rag-reviewer:review-pr use the PR-session tools above. Standalone (no PR session): use the session-less tools per the reviewer-grounding block when reviewer is connected and the index is fresh; otherwise fall back to grep/Read.

Goal

Look only for performance and efficiency risks in the selected changes. Ignore style, architecture, tests, and general correctness unless they materially affect performance.

Prioritize findings such as:

• N+1 queries and repeated remote calls;
• unnecessary loops or repeated work;
• bad asymptotic behavior on hot paths;
• redundant rendering, serialization, parsing, allocations, or avoidable copies;
• missing batching, caching, pagination, or streaming where the diff makes that risk likely;
• blocking I/O or CPU-heavy work on latency-sensitive paths;
• memory growth or large payload handling.

Method

1. Read the diff first.
2. Open only the nearby code needed to understand whether the changed path is performance-sensitive. In rag-reviewer:review-pr use the reviewer MCP tools: read_file, search_code, find_callers.
3. Prefer concrete findings over vague perf speculation.
4. If a concern depends on an assumption, state that assumption explicitly.
5. If a path is probably not performance-sensitive, do not invent issues.

Severity

• critical / high: likely severe latency, throughput, or resource regression on an important path.
• medium: meaningful inefficiency or scaling risk that should probably be fixed.
• low: worthwhile optimization or preventive note, not a blocker.

Output

Return only actionable findings.

Return ONLY the findings JSON used by the review pipeline, with "category": "performance":

• Calibrate confidence against a measurable, reproducible effect: a hot path you can point to (loop bound, query inside a loop) → 0.8+; a plausible but data-dependent cost → 0.5–0.7; no measurable/reproducible effect → ≤ 0.4 (drop). Set "category" to "performance"; "side" is always "RIGHT".