audit-perf

Imported from ku5ic/dotfiles (claude/skills/audit-perf/SKILL.md). Install upstream with npx skills add ku5ic/dotfiles --skill audit-perf. Copyright stays with the author.

Procedure

0. Resolve external context (rules/workflow.md section 4). Then dispatch the auditor agent (subagent_type: auditor, foreground) with steps 1 onward and the resolved arguments; it writes the report, you relay its summary.

1. Stack is in the repo context your startup produced (agent-context.sh). Get the scratch directory via scratch-dir.sh.

2. Load the patterns skill for the detected stack (react-patterns, django-patterns, etc.) for the anti-pattern reference.

3. Review the target across these categories:

   • Ground each candidate in this project's own precedent before including it.
   • An established pattern is one that repeats consistently elsewhere in the codebase and is not flagged as a problem by CLAUDE.md or existing tests.
   • An established pattern is a deliberate tradeoff, not a finding.
   • Exception: flag it anyway if it is measurably worse at this location than elsewhere.

4. Skip categories with no findings.

Frontend (React and Next.js)

• Unnecessary re-renders: inline object or function props to memoized components, context value churn, parent re-renders that include large subtrees
• Missing memoization where it actually matters (measurably expensive computation, or stable reference needed by memo child)
• Key stability on lists that reorder
• Effect waterfalls: multiple sequential effects that could be one or parallel
• Client bundle bloat: client components importing large libraries, server-only logic leaking to client via shared utilities, unused but imported modules
• Next.js: unintended client component propagation, wrong cache or revalidate config for correctness, missing next/image or next/font
• Core Web Vitals where statically inferable: CLS from missing dimensions, LCP from above-the-fold client-side rendering

Backend (Django, Node APIs)

• N+1 queries: missing prefetch_related, select_related, or ORM equivalent
• Unbounded queries: missing pagination, .all() on large tables
• Synchronous I/O in async contexts
• Missing indexes on filtered or ordered columns (static inference from query shape)
• Serialization cost: deep nested serializers, recursive expansions
• Cache misuse: caching with key collisions, caching before query filtering

General

• Unbounded loops
• Accidental O(n^2) via nested iteration over the same collection
• Reading a whole file when streaming is possible
• Blocking operations in request path

Verification tools

• Command-level perf checks (build scripts, codegen, test runner startup, CLI tools): measure with hyperfine, not eyeballed.
• Render-level perf (React re-renders, bundle cost): measure with React DevTools Profiler or browser perf tools at runtime, not from static analysis.

Output per finding

• Location
• What: the specific pattern
• Why it matters: estimated impact (render count, query count, payload size)
• Verification: how to measure, if not statically obvious (e.g. "profile with React DevTools, expect re-render on every parent state change")
• Fix: concrete code change or pattern switch

Output file

Use the rules/markdown-report.md format. Write to $(scratch-dir.sh)/perf-<target-slug>-<YYYYMMDD-HHMM>.md. Print the path.

Rules

• Do not claim measured improvement. This command does not run anything.
• Flag what to measure, not what to assume.
• Ignore micro-optimizations that change code without measurable benefit.
• A pattern that repeats across the codebase as an established choice is not N separate findings; note it once against the shared source and list the consuming locations.