Find and (on request) fix complexity/performance hotspots with behavior preserved and proven by tests. Bias toward small, verified wins over broad rewrites.
Core rule
Never change observable behavior. Every optimization preserves outputs, ordering, error semantics, and public APIs — and is proven by a test that passed before and after.
Loop: Triage → Prove → Optimize → Verify
- Triage — get candidate leads fast, then reason about them:
- First pass (cheap, optional):
python3 scripts/scan_hotspots.py <root> --format jsonfor a ranked multi-language lead list. Treat output as leads, not proof. - Confirm with
Grep/Glob/Read: nested loops over the same data, membership tests in a loop (list vs set), sort-in-loop, pairwise comparisons, repeated scans, render recomputation, N+1 queries. Rank by impact (hot path × input size), not raw count.
- First pass (cheap, optional):
- Prove — before touching hot code, run/author a focused test that pins current behavior and, if feasible, a quick timing or complexity witness.
- Optimize — apply the smallest localized
Editthat lowers complexity (seereferences/optimization-playbook.md). No drive-by refactors or formatting churn. - Verify — re-run the focused test, then the broad gates (type-check, lint, build, suite). If any regresses, revert that change.
Subagent fan-out
For large repos, shard files/dirs across Task subagents (read-only). Each returns a ranked findings
list: file:line, pattern, estimated current → target complexity, risk. The main agent dedupes and
merges into one ranked plan, then optimizes serially (it owns all edits). Cap concurrency at ~4-6
subagents; prefer fewer, broader shards over many tiny ones.
Optimization safety checklist
- Same outputs for the same inputs (including ordering and ties).
- No swallowed/altered exceptions; same error types.
- No new unbounded memory (a set/dict cache must be bounded by the same data).
- Equivalent for empty / single / duplicate / large inputs.
- Public API and call sites unchanged.
Analysis-only by default
Default to a report (see references/report-template.md): ranked findings with location, current and
target complexity, risk, and a proposed fix — but do not edit. Apply fixes only when the user says
implement / fix / apply / refactor (or pre-approves low-risk wins like list→set membership).
Resources
scripts/scan_hotspots.py— heuristic multi-language hotspot scanner (leads only).references/optimization-playbook.md— transformation catalog + correctness checks + what-not-to-do.references/report-template.md— analysis-only report skeleton.