| name | review-performance |
| description | Use when the user explicitly asks for a performance review, benchmark review, profiling review, perf audit, pprof analysis, allocation review, latency regression check, hot-path review, throughput review, "is this fast enough", or "review perf". This skill is opt-in only — never run automatically as part of review-all unless the user explicitly requests it. |
Performance Review
Structured review of hot-path performance, allocation patterns, and benchmark coverage. Opt-in only — performance is rarely the bottleneck for new code, and a perf review on every PR is noise. Run when the user explicitly asks, or when a feature is performance-sensitive by nature (hot path, fanout, large-data).
Out of scope (defer to siblings):
- Concurrency safety (race, deadlock, leak) →
review-reliability
- Cardinality / metric cost →
review-observability
- Database query plans / index design →
review-database
review-performance looks at code-level allocation and CPU hotspots, benchmark sufficiency, and obvious algorithmic complexity issues.
Workflow
1. Scope and explore
- Confirm scope with the user: full codebase, specific packages/directories, changed files only (PR or branch diff), or a specific hot path.
- Resolve scope. Same conventions as siblings:
- Changed files (PR or branch):
git diff --name-only --diff-filter=d <base>...HEAD; filter to source files.
- Explicit paths/packages: include all relevant source files.
- Full codebase: walk source files.
- If invoked from review-all (only when the user opted in): receive
file_list, package_paths, has_changes, base_ref, REVIEW_DIR, and pr_url.
This skill is primarily Go-focused because the existing review-* corpus targets Go; it has cross-language sections in reference.md but the deepest checklist is for Go.
2. System overview
Produce a brief summary covering:
- Languages and runtimes in scope
- Identifiable hot paths (RPC handlers, queue consumers, batch jobs)
- Existing benchmark coverage (Go:
Benchmark* functions; presence of testing.B)
- Observed SLOs / latency targets if documented in
REVIEW.md, README.md, or runbooks
3. Launch investigation subagent
Launch a single investigation subagent (subagent_type="generalPurpose", model: sonnet per subagent-model-routing) with the system overview and in-scope file list.
Prompt it to:
- Read all in-scope source files.
- Apply the checklists in reference.md:
- Allocation hotspots (Go-specific patterns)
- Algorithmic complexity (N+1, nested loops over unbounded inputs, O(n²) on hot paths)
- Concurrency efficiency (goroutine-per-request anti-patterns, oversharing of locks, mutex contention)
- Benchmark sufficiency
- I/O batching and connection reuse
- Identify which findings are real concerns vs. premature optimization. Bias toward "show me the profile" over "I think this allocates". Findings without evidence are P2/nit.
- For each finding, search nearby code for existing benchmarks, comments, or TODOs.
- Return findings using the performance findings template.
4. Run static analyzers and benchmarks
command -v staticcheck >/dev/null && staticcheck -checks "U1000,SA1019,SA6005" <package_paths>
command -v gocyclo >/dev/null && gocyclo -over 15 <source_dirs>
grep -rl 'func Benchmark' --include='*_test.go' <package_paths>
Static analyzers feed the subagent. Do not run benchmarks as part of the review unless the user explicitly asks — benchmarks need isolated, repeatable hardware to produce useful numbers.
5. Present results
Resolve the review output directory (skip if REVIEW_DIR was provided by the review-all orchestrator):
REVIEW_DATE=$(date +%Y-%m-%d)
REVIEW_DIR=".reviews/${REVIEW_DATE}"
if [ -d "$REVIEW_DIR" ]; then REVIEW_DIR=".reviews/${REVIEW_DATE}-$(date +%H%M)"; fi
~/.claude/scripts/ensure-gitignore.sh '.reviews/'
mkdir -p "$REVIEW_DIR"
Capture run metadata (see Run metadata header) and prepend to ${REVIEW_DIR}/PERFORMANCE-REVIEW.md.
Output structure:
- Run metadata header
- System overview (from step 2)
- Findings table (grouped by category: allocation / complexity / concurrency / benchmarks / I/O)
- Recommended fix order, with explicit "needs profiling evidence" markers on findings that should not be acted on without a profile
Present the report to the user.
Run metadata header
RUN_DATETIME=$(date -u +"%Y-%m-%d %H:%M UTC")
GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD)
GIT_COMMIT=$(git rev-parse --short HEAD)
GIT_COMMIT_FULL=$(git rev-parse HEAD)
GIT_SUBJECT=$(git log -1 --pretty=%s)
> **Run:** {RUN_DATETIME}
> **Branch:** {GIT_BRANCH} @ {GIT_COMMIT} (`{GIT_COMMIT_FULL}`)
> **Subject:** {GIT_SUBJECT}
> **Base:** {BASE_REF} @ {BASE_COMMIT} <!-- omit when scope is not diff-based -->
> **Scope:** {scope description}
Finding link wrapping (PR mode)
When pr_url is provided (or gh pr view --json url -q .url 2>/dev/null returns one for standalone runs), wrap every path:line reference inside the finding tables as a Markdown link:
~/.claude/scripts/pr-deeplink.sh "$pr_url" <path> <line>
The display text stays path:line. Pass L for findings about removed code. Findings follow terse-comments: concrete fix, optional bug:/risk:/nit:/unsure: prefix.
Output Templates
Performance findings
| Priority | Category | Finding | Impact | Effort | Evidence | Tracked |
|----------|----------|---------|--------|--------|----------|---------|
| P0 | allocation | Description with code references | Allocations on every request | small | profile needed | — |
| P1 | complexity | O(n²) loop in hot path at file:line | Linear-in-N regression for N > 1k | moderate | bench needed | TODO in file:line |
Category column values: allocation, complexity, concurrency, benchmarks, io.
Evidence column values: profile shown, bench shown, profile needed, bench needed, static-only. Use this to flag findings that need empirical confirmation before action.
Re-evaluation table (for follow-up reviews)
| Finding | Status | What Changed |
|---------|--------|--------------|
| ~~1. Description~~ | FIXED | Benchmark shows X% improvement |
| 2. Description | Still applicable | No changes |
Guidelines
- Bias toward measurement. If a finding can't be backed by a profile or benchmark, mark it
static-only and downgrade priority unless the issue is glaring (e.g. O(n²) over user-controlled input).
- Don't run benchmarks on shared hardware unless the user explicitly requests it. Benchmark noise is high; bad numbers are worse than no numbers.
- Premature optimization is a real concern; explicitly say "needs profile" when a finding is intuition-based.
- Include effort estimates.
- When the user asks for a follow-up review, find the most recent review directory and append the re-evaluation table.
- For detailed framework categories, see reference.md.
- Opt-in opt-out: this skill should never run as part of
review-all unless the user explicitly requests it. The orchestrator must check include_performance before fanning out to this skill.
- REVIEW.md integration: If a
REVIEW.md context section was provided by the review-all orchestrator (or exists at the repository root when running standalone), treat its rules as additional review criteria. "Always check" items are HIGH severity; domain-specific items (Performance section) are MEDIUM severity. "Skip" patterns exclude matching files from review scope.
- Findings must cite probed evidence (
path:line, grep output, command result), not pattern-matched suspicion. Per ~/.claude/rules/probe-not-assume.md.