optimize

• SKILL.mdcore · 30,337 bytes
• SKILL.mdcore · 30,337 bytes
• SKILL.mdcore · 30,337 bytes
• agents/openai.yamlasset · 196 bytes
• references/experiment-mode.mddoc · 6,480 bytes
• references/lenses.mddoc · 10,570 bytes
• references/state-and-resume.mddoc · 2,707 bytes
• references/tooling.mddoc · 7,611 bytes

# Optimize

Five mode names share one authority: reversible local writes to the resolved target, never push, tag, publish, deploy, or mutate a remote. Full and `extremely-optimize` intentionally select the same measured-floor campaign route, rather than separate procedures; the retired name remains an explicit compatibility mode value. That route runs an isolated five-lens search with an integrated 1.05x gate and an atomic commit. Quick mode runs a single measure-identify-fix-verify loop with a noise-aware keep/revert and a CI guard for the obvious-bottleneck case. `fastopt` mode reports suspected-hot-path hypotheses without mutating code. `fastopt-extreme` mode reports estimated hot paths and complexity-neutral cold-path simplifications while refusing complexity theater.

## Contract

| Field | Bound contract |
|---|---|
| Trigger | The user asks to optimize code, make a path faster, make this as fast as possible, reduce allocations, fix a performance regression, profile and optimize a path, symbol, or diff, or a performance requirement, slowness report, Core Web Vitals miss, or profiling evidence identifies a bottleneck; asks for a measured-floor campaign, hypothesis-only suspected-hot-path analysis, or estimated hot/cold analysis that refuses complexity theater; or says "extremely optimize" or "grill every inefficiency". |
| Authority | Full and `extremely-optimize` modes write only the exact VCS-tracked source-target set shown before mutation plus `.outline/optimize/<target>/` (destructive; rollback is version control); quick mode writes only the named target and an authorized CI performance-budget or field-monitor guard; `fastopt` and `fastopt-extreme` modes are read-only with no file, VCS, credential, paid, published, deployed, or remote mutation. Full mode may commit one winning optimization, but no mode tags, publishes, deploys, or mutates a remote. |
| Side effect | Full and `extremely-optimize` modes append measurements and decisions to `.outline/optimize/<target>/log.jsonl`, benchmark isolated candidates, gate all boundary surfaces before touch, revert non-wins, apply the winner, and create one atomic commit. Quick mode applies one targeted fix and may add a CI performance budget or field monitor. `fastopt` mode emits labeled hypotheses and benchmark designs. `fastopt-extreme` mode emits hot-path hypotheses, benchmark designs, cold-path simplifications, and complexity-theater refusals. |
| Done | Full and `extremely-optimize` modes rebuild hot units with a measured 1.05x-or-better win, grade cold units fixed/at-floor/left, remove residue, pass the verifier, and land the target atomically. Quick mode: the identified bottleneck is measurably improved past noise, tests remain green, no new regressions exist, and a CI budget or field monitor guards the metric. `fastopt` mode: every named path has a labeled hypothesis and benchmark design, with no mutation or speed claim. `fastopt-extreme` mode: every hot estimate has a falsifiable benchmark, every cold candidate is complexity-neutral or rejected, and every complexity-theater proposal is refused. |
| Inputs | The required mode-specific inputs are listed below; the `mode` input selects exactly one route and must not combine a mutating route with a read-only route. |