Benchmark Optimize

Accept a function, process, optional language hint, or no target. With no target, choose a bottleneck supported by profiles, slow tests, or realistic resource costs. Resolve material workload ambiguity before optimization.

Define the metric, representative inputs, environment, and acceptance criterion. Capture a baseline before changing the implementation. Measure only resources relevant to the target, such as latency, throughput, memory, disk, or network use.

Use the repository's benchmark framework. Otherwise use a language-appropriate tool: Criterion for Rust, pytest-benchmark or pyperf for Python, and a repeatable Node runner for JavaScript. Use testing.B for Go and JMH for Java. Record the command and sampling settings.

Keep correctness and public behavior stable unless the task authorizes a change. Compare candidates with the same workload and environment. Run correctness tests and applicable quality gates on the selected result.

Continue until a measured improvement satisfies the target or evidence rules out the attempted approaches. Discard unsuccessful optimization edits while preserving unrelated work. Report baseline and final metrics, uncertainty, resource tradeoffs, validation, and the decision to retain or reject the change.

In Plan Mode, inspect existing evidence and define the benchmark and approach without changing tracked files. No extra target-approval checkpoint is required.