cyrusnuevodia/gepa-research/optimize

Run the GEPA-backed optimization loop. The plugin calls gepa.optimize_anything under the hood; each candidate it proposes is applied in a fresh git worktree, the benchmark is run, gates are checked, and the result is backported into .gepa-research/<run>/graph.json so the dashboard continues to render the lineage DAG.

Host conventions

• Slash commands shown in user-facing copy (e.g. /gepa-research:optimize) — translate to your host's mention syntax when speaking to the user (e.g. $gepa-research optimize on Codex — plugin namespace then skill name, separated by a space).

Configuration

All arguments are optional. Invoked as /optimize [max-metric-calls=N] [stall=N] [reflection-lm=MODEL].

• max-metric-calls — GEPA evaluator-call budget for this run (default: 50).
• stall — consecutive iterations without improvement before auto-stopping (default: 5).
• reflection-lm — model string passed to ReflectionConfig.reflection_lm (default: gepa's default, currently openai/gpt-5.1). Use e.g. anthropic/claude-opus-4-7 for Claude.

The legacy subagents, budget, and per-subagent knobs are no longer accepted — GEPA owns the search strategy.

Prerequisites

• Workspace must be initialized (gepa-research status should succeed).
• A baseline experiment must be committed (run /discover first). GEPA's

seed candidate is read from the current best committed node's target file.

• The gepa library on the Python path (auto-installed as a transitive dependency when the CLI is installed from GitHub: uv tool install "git+https://github.com/CyrusNuevoDia/gepa-research#subdirectory=plugins/gepa-research").
• A reflection LM API key in the environment (OpenAI/Anthropic/etc., depending

on the reflection-lm value). Without this the first GEPA iteration will fail.

Architecture

Orchestrator (this skill):
  1. Reads current best committed node from .gepa-research/<run>/graph.json
  2. Extracts seed_candidate: {target_relpath: file_contents}
  3. Calls gepa.optimize_anything(seed, evaluator=adapter.evaluate,
                                  objective=config["optimization_objective"],
                                  config=GEPAConfig(stop_callbacks=...))
  4. Reports the final best candidate and updates the graph
 
GepaResearchAdapter.evaluate (called by gepa per candidate):
  a. allocate_experiment(parent_id=best_committed)
     -> creates .gepa-research/<run>/worktrees/exp_NNNN and a fresh branch
  b. write candidate dict contents into the worktree
  c. run config["benchmark"] as subprocess; parse_score from stdout
  d. run inherited gates (collect_gates_from_path)
     -> on failure, return (0.0, {"gate_failures": [...], "traces": ...})
  e. on score improvement + all gates pass: maybe_commit_worktree + mark "committed"
  f. return (score, side_info) so gepa can reflect on stdout/stderr/traces

side_info returned to gepa includes:

• experiment_id — so diagnostics reference .gepa-research/<run>/experiments/<id>/
• stdout / stderr — trailing 4 KB of each
• benchmark_result — parsed JSON (score + per-task breakdown, when available)
• task_traces — contents of task_*.json from the SDK, when instrumented
• gate_failures — list of gate names that rejected the candidate

GEPA uses this side_info in its reflection prompt to propose the next candidate. For that to work well, your benchmark must write diagnostic output (stack traces, task-level failure reasons, etc.) rather than just a terminal score.

The loop

Run once per /optimize invocation:

1. Verify prerequisites

gepa-research status                         # confirms workspace exists + shows current best
gepa-research-version-check                  # confirms CLI matches plugin manifest

If status shows no committed node, stop and tell the user to run /discover first.

2. Resolve the reflection LM

If the user passed reflection-lm=…, use that value verbatim. Otherwise, read config.json for a reflection_lm field. If neither is set, leave it unset — GEPA will use its own default (currently openai/gpt-5.1).

Before handing off to GEPA, verify the corresponding API key is in the environment (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.). If missing, stop and tell the user which env var to export; do not try to proceed — GEPA will burn the metric budget failing to call the LM.

3. Invoke GEPA via the plugin command

The optimize entry point wraps the full call:

gepa-research optimize \
  --max-metric-calls 50 \
  --stall 5 \
  [--reflection-lm anthropic/claude-opus-4-7]

This resolves parent_id = best_committed_node(graph, metric) internally, builds the seed candidate from that node's target file, and calls run_gepa_optimize from gepa_research.gepa_adapter. Each candidate GEPA proposes allocates a new worktree under .gepa-research/<run>/worktrees/.

4. Stream progress

While the loop is running, the dashboard (if started) auto-refreshes as new experiment nodes appear in graph.json, and the BUDGET / STALL hero cards update from .gepa-research/<run>/progress.json. Surface the URL to the user if they don't already have it.

Do not run multiple gepa-research optimize invocations concurrently against the same workspace — they would race on graph/meta file locks and corrupt the lineage.

5. Final summary

When gepa-research optimize exits, print:

• Best score and the experiment id that produced it
• Total candidates evaluated (from GEPA's total_metric_calls)
• The winning diff: gepa-research diff <best_exp_id>
• Whether the stop was due to budget, stall, or user interrupt

Suggest follow-up actions: raise the budget, switch reflection LM, or introduce additional gates if the winning candidate regressed something the objective didn't capture.