Performance

Inline two-stage analysis: gather candidate hot paths first (mechanical grep), then prioritize and deepen on the top targets.

Phase 1: Gather candidates

Default scope: src/ (or $ARGUMENTS if provided). Look for likely-slow patterns:

• Sync I/O inside async def (requests.get, time.sleep, open() without aiofiles)
• N+1 query patterns (DB call inside a loop)
• Sequential await calls that could be asyncio.gather
• Compiled regexes or client instantiations inside loops
• Unbounded list / str growth in loops
• Missing connection pooling (new aiomysql connection per request)
• Blocking calls in async handlers

For each candidate: file:line + function name + sample.

Phase 2: Prioritize + deep analysis

1. Rank candidates by likely impact
2. For top candidates, examine call frequency and data volumes
3. If user opts in to runtime profiling: python -m cProfile -o /tmp/profile.out src/main.py then parse with pstats
4. Correlate static findings with runtime hotspots

Phase 3: Report + act

Produce ranked recommendations with concrete fix proposals (not "use async" — show the exact replacement).

For quick wins (1-line fixes like time.sleep -> asyncio.sleep):

• Ask the user before applying
• Apply + run validation (ruff check, pytest)

For structural changes (N+1 fixes, async refactors, pooling):

• Suggest creating a PRP via the generate-prp skill

Rules

• Static analysis is safe and always runs
• Runtime profiling requires explicit user approval (modifies execution)
• Quick wins can be auto-applied with validation
• Structural changes belong in a PRP
• Don't optimize prematurely — focus on measurable bottlenecks