fix: make the GEPA skill-evolution path work end-to-end

[enzo-adami]

Summary

Four commits, each removing a blocker that prevented Phase 1 (skill evolution) from ever producing a validated evolved skill. Exercised end-to-end with dspy 3.3.0 / gepa 0.1.1 against a local OpenAI-compatible endpoint: GEPA now actually runs (no silent fallback) and evolved candidates can pass the constraint gate.

Bug 1 — evolved text never reached the artifact (5915499)

GEPA's candidate space is exactly {name: pred.signature.instructions} over named_predictors() (dspy teleprompt/gepa/gepa.py, seed + candidate build). SkillModule kept the skill body in a plain attribute passed as an InputField, which GEPA never mutates — so optimized_module.skill_text always returned the original text and evolved_skill.md was always identical to the baseline. The optimizer measured improvements it never persisted.

Fix: the skill body becomes the predictor's signature instructions (with_instructions()); skill_text is a read-through property over named_predictors(), so extraction reflects the evolved candidate with no change to the extraction call site.

Bug 2 — constraint gate rejected every candidate (44caaf5)

validate_all() was called with skill['body'] / evolved_body, but _check_skill_structure() looks for YAML frontmatter — which load_skill() strips from body. The structure constraint therefore failed on every run and every evolved candidate ended in 'not deploying'.

Fix: validate skill['raw'] for the baseline and the reassembled evolved_full for the candidate.

Bug 3 — GEPA never ran at all (916661d)

dspy.GEPA has no max_steps parameter: the call raised TypeError on every run and silently fell back to MIPROv2, so the GEPA path advertised in the README could never execute. GEPA also requires exactly one budget parameter and a reflection LM, and its GEPAFeedbackMetric protocol calls the metric with (gold, pred, trace, pred_name, pred_trace).

Fix: pass max_full_evals=iterations and reflection_lm; extend skill_fitness_metric to the GEPA calling convention, returning Prediction(score, feedback) with deterministic missing-terms feedback for predictor-level reflection calls and a plain float everywhere else (holdout and MIPROv2 behavior unchanged).

Enhancement — generation budget for local endpoints (5b7fbd9)

Reasoning models served by local OpenAI-compatible endpoints often sit behind low server-side max_tokens defaults: the thinking phase consumes the whole budget and content comes back empty ('LM returned an empty or null response' on every rollout). New --max-tokens / --temperature options apply to both the eval LM and GEPA's reflection LM; defaults unchanged.

Tests

145 → 154 (9 new: GEPA-visibility of the skill text, full-file vs bare-body constraint regression, metric contract incl. the max_steps regression), all green.

🤖 Generated with Claude Code