Local optimization laboratory
Profile a flow. Prove the change.
CodeVetter gives a coding agent a bounded loop for finding a local bottleneck, proposing one small experiment, and retaining it only when correctness and repeated before/after evidence agree. The compact current path targets repository-owned JavaScript, Node.js, and Go tests and benchmarks; it does not treat a plausible edit as an optimization.
One bounded loop
The machine receipt is authoritative. Every stage records source identity, tested boundary, decision, and limitations so another agent can continue without guessing.
- 01DiscoverSelect one executable local test or benchmark flow.
- 02MeasureCapture repeatable timing, CPU, allocation, and memory evidence supported by that flow.
- 03DiagnoseRank source-backed bottlenecks while separating observation from inference.
- 04ExperimentGive the coding agent one bounded candidate and reject disproportionate patches.
- 05VerifyRun exact correctness and repeated before/after measurements before retaining anything.
JavaScript and Node.js
Exact Node, Vitest, and Jest selections with repeated console metrics, CPU source attribution, sampled allocation evidence, and process-tree memory where observable.
Go
Repository-owned tests and benchmarks with ns/op, bytes/op, allocs/op, bounded profile sources, paired measurements, and explicit evidence gaps.
Shipping guardrails
A faster patch still has to earn its size.
- Change cost: default maximum of 3 files, 160 added lines, 200 gross changed lines, and no new production dependency.
- Correctness first: exact repository-owned checks run before repeated paired performance verification.
- No confidence is a result: noisy, incomplete, stale, or externally dominated evidence cannot authorize shipping.
- Local means local: no production or customer-impact claim is inferred from a development fixture or synthetic benchmark.