Find a concrete source of unnecessary work, change it at its owner, and verify equivalent completed investigations. A useful outcome can also be a measured decision to leave an implementation alone.

Follow the checkout's performance priorities, tool design, and testing guide. This skill supplies an optimization workflow, not additional permission to execute targets, publish, or merge. Use the current contracts, provider prerequisites, and user authorization.

Read the Kuna case study when selecting a reuse strategy or interpreting a benchmark. Its pinned observations illustrate tradeoffs; they are not current-provider guarantees or universal speedups.

Describe the analyst's completed task, including the facts and relationships needed to answer it, requested detail, source access, and meaningful unknowns. Choose a representative real artifact and a small or empty case. Record the baseline revision, target digest, engine identity, options, host, and entrypoint. Use a separate checkout for changes without disturbing other work.

Establish an oracle independent of the changed implementation: source-built fixture expectations, independently mapped file bytes, or a native producer paired with checks of its interpretation. Comparing two wrappers around the same faulty decoder does not establish correctness. Preserve address widths, same-address alternatives, unresolved endpoints, coverage, and source linkage. An empty result proves absence only within the actually examined scope.

Trace one public CLI or MCP task through shared application code to its producer. Measure relevant phases: admission and identity checks, native analysis, decoding, lookup, projection, serialization, Evidence retention, and cleanup. Include catalog discovery and extra calls when they affect the completed task.

Use the evidence to choose the smallest structural correction:

Do not introduce an index, generic cache, batch endpoint, or orchestration layer without identifying the work it removes and its maintenance/resource cost. Projection after materializing a full report may reduce output while leaving analysis time and peak memory unchanged. Fix missing or incorrect facts before benchmarking an alternative composition against a better-covered workflow.

Let the producer boundary validate and normalize facts once. Consumers query that representation through the shared CLI/MCP implementation; adapters should not rebuild relationships or independently reinterpret raw reports. Keep lookup identity separate from display formatting. Reuse contracts when their intent matches, honoring requested detail and complete Evidence access.

For every cache or shared producer, state its key, owner, invalidation, retained size, and lifetime. Account for result-affecting target bytes, engine contents, analysis options, architecture/load mapping, and parser/schema versions where they affect reuse. Validate live state before calling a cached observation current. Cache a refusal only when it is deterministic for that validated key; malformed, partial, transient, or cancelled attempts need explicit retry policy.

One cancelled waiter must not invalidate a successful peer. Preserve producer failure context and each caller's cancellation reason. Retain ownership and capacity charges until shared work and cleanup settle; preserve completed observations when cleanup fails. Exercise invalidation, concurrent consumers, last-waiter cancellation, close, and cleanup retry when those paths change.

Bound the process family and retained state, not just a parent heap or worker count. Account for decoding and serialization expansion before allocation. Choose CPU, memory, concurrency, and disk limits from the host and workload; stop before OOM and verify owned cleanup. A guard-stopped task is incomplete. Do not replace a transport buffer cap with an unbounded retained inventory or hide partial results behind a success status.

Run the affected real public workflow before and after with equivalent inputs, results, coverage, and Evidence semantics. Include small inputs so setup costs and regressions are visible. Distinguish fresh sessions, warm runtime caches, and OS file caches. Alternate or repeat runs when variance affects the decision. Choose meaningful boundary regressions for cases the actual engine cannot reliably reproduce; a fake engine does not qualify a real provider.

Report task outcome, relevant phase timings, total elapsed and CPU time, peak process-family memory and its accounting method, resource stops, and cleanup. Check the payload exit status as well as the profiler's. Keep failed attempts and identify unmeasured costs; a passing retry does not explain a failure.

For token claims, count schemas/discovery, arguments, replies, retries, and continuations needed to finish the task. Label tokenizer estimates separately from actual model usage and record the client/catalog behavior. Smaller JSON alone establishes neither fewer model tokens nor better investigation quality. Do not claim a complex-software advantage without completed representative investigations and comparable correctness/coverage.

Finish with the bottleneck, owning change, preserved semantics, executed checks, before/after measurements, and remaining limits. Keep any integration record current with its tested revision and distinguish merged production work from experiment-only code. Stop when the demonstrated bottleneck is resolved or further benefit remains speculative.