optimize

/optimize — GEPA-style inner loop

Inner-loop skill — used by the asset when an issue is an optimization task. Pareto-frontier search over (perf, cost-time, cost-$) with LLM reflection on execution traces.

Scope

Optimize only the artifact class the dispatch brief names (e.g. a prompt template, a config). Anything outside that class — and anything in the repo's frozen_paths (.claude/sdlc.toml) — is off-limits unless the brief explicitly authorizes it. When in doubt, surface to control rather than widening scope.

Loop

1. Select a candidate from the Pareto frontier over (perf, cost-time, cost-$).
2. Execute a minibatch evaluation, capturing traces (per-example outcomes, error patterns).
3. Reflect: read traces; diagnose where the artifact fails. Be concrete about which examples broke and why — no hand-waving.
4. Mutate: a targeted change, informed by accumulated lessons from earlier failed candidates in this run.
5. Accept: if the new candidate is non-dominated on (perf, cost-time, cost-$), add it to the frontier.

Noise robustness

If the repo's perf metric is noisy (the brief should state the single-run swing, if known), a perf promotion requires a ≥N-replicate mean±band — not a single run. A single-run improvement inside the noise band is not promoted. If the brief doesn't quantify noise, measure it (a few replicates of the baseline) before trusting any delta.

Pareto axes

• perf: the repo's quality metric(s), as defined in the domain brief — replicated if noisy
• cost-time: wall-clock minutes per evaluation
• cost-$: compute / API spend per evaluation

Output

A markdown summary the asset includes in its iteration comment:

Candidate	Mutation	Perf	Cost-time	Cost-$	Dominated by

Plus the current Pareto frontier (frontier candidates only).

The asset decides which (if any) candidate to land. The skill is advisory; the asset commits.