gepa-anywhere

Build a trustworthy, anchored evaluator for an agent whose output you want to optimize, when no reliable metric/judge/gold exists yet. It learns exactly what the agent produces, gives the evaluator the SAME inputs the agent saw (building + testing those surfaces), drafts the evaluator as an EXTERNAL markdown rubric, has a DIVERSE expert panel (incl. an adversary) harden it, and calibrates it against a real anchor while hill-climbing self-consistency — then registers it so gepa can optimize the agent against it. Use whenever the user wants to optimize / improve / "make better" an agent's prompt or output and the way to MEASURE quality is missing or weak — e.g. "set up a judge for my extraction agent", "how do I score whether my summarizer is good", "optimize this prompt but I have no gold labels", "build an evaluator for <agent>", or BEFORE any `gepa run` whose metric is absent or untrustworthy. NOT for laying out the repo (that's gepa-init) or driving the optimization loop (that's gepa-run); this is only the build-the-metric step. The evaluator is the bottleneck on every optimization — do not skip building one well.

Drive a GEPA optimization run to automatic completion, stopping when a stop-policy condition is met (budget exhausted, corroborated saturation, or too many invalid patches). Use when the user says "autorun gepa", "run until done", "keep optimizing automatically", or "stop when saturated", AND a `.gepa/config.yaml` exists. Wraps gepa-run with a decide_stop check between every iteration.

Inspect a finished gepa-anywhere run's Pareto frontier and deliberately promote the winning candidate back onto the artifact. Use after `gepa run` reaches exit 0, when the user says "show the frontier", "which candidate won", "promote the winner", "apply the best prompt", or wants to compare candidates / check the holdout report before committing.

Lay out a repository to hill-climb (optimize) ARBITRARILY MANY agents with gepa, each with its own grounded evaluator — by hand-creating the `.gepa/agents/<name>/` directories + a registry convention. Use once per repo, before building evaluators or running optimizations, when the user says "I want to optimize several agents/prompts in this repo", "set up a multi-agent gepa project", "initialize gepa for many agents", or "add gepa to this project" AND more than one agent will be optimized. NOT for building the evaluator itself (that's evaluator-discovery) or running the loop (that's gepa-run). For a single prompt + a code metric, prefer `gepa scaffold` (the flat `.gepa/config.yaml` quickstart) — this skill is the multi-agent superset and produces a DIFFERENT, incompatible layout, so don't mix them in one repo.

Drive a GEPA reflective-optimization loop on any text artifact in any repo, using THIS Claude Code session as the free (Max-billed) reflection LM. The Python CLI `gepa run` handles GEPA's optimization math + checkpointing; when it needs the reflection LM it writes a pending envelope and exits 42 — you read it, propose improved artifact text, write the response, and re-invoke. Use when the user says "run gepa", "optimize this prompt/instructions", "evolve the artifact", "improve the extraction prompt", AND a `.gepa/config.yaml` exists (or they want one — then use gepa-scaffold first).