fix: correct GEPA arg, constraint input, and rate limit handling in evolve_skill

[MwC-Trexx]

Summary

Three bugs in evolution/skills/evolve_skill.py that together cause every optimization run to (a) never actually use GEPA, (b) always reject the evolved skill as invalid, and (c) accumulate 0.0-scored trials from rate limit errors.

Bug 1 — GEPA called with wrong keyword argument

dspy.GEPA(metric=..., max_steps=N) raises TypeError: GEPA.__init__() got an unexpected keyword argument 'max_steps' on every run in DSPy 3.1.3. The optimizer silently falls back to MIPROv2. The correct parameter is max_metric_calls.

# Before
optimizer = dspy.GEPA(metric=skill_fitness_metric, max_steps=iterations)
# After
optimizer = dspy.GEPA(metric=skill_fitness_metric, max_metric_calls=iterations)

Bug 2 — Constraint validator receives body-only text

_check_skill_structure checks that text starts with --- and contains name:/description:. Both call sites passed skill["body"] (the markdown after frontmatter is stripped), which never has frontmatter. The constraint therefore always fails and rejects every evolved skill — even valid ones — writing to evolved_FAILED.md instead of deploying.

Fix: pass skill["raw"] for the baseline check and evolved_full (the reassembled skill with frontmatter, already computed at line 185) for the evolved check.

# Before
validator.validate_all(skill["body"], "skill")
validator.validate_all(evolved_body, "skill", baseline_text=skill["body"])
# After
validator.validate_all(skill["raw"], "skill")
validator.validate_all(evolved_full, "skill", baseline_text=skill["raw"])

Bug 3 — Concurrent calls saturate OpenAI free-tier rate limits

DSPy's parallelizer fires all evaluation calls simultaneously. On a free-tier OpenAI account (3 RPM / 60 K TPM for gpt-4.1-mini), the majority of calls in each trial fail with RateLimitError. DSPy records those trials as 0.0, poisoning MIPROv2's Bayesian optimizer (trials 4, 5, and 6 all scored 0.0 in the observed run despite the underlying prompt being reasonable).

Fix: add num_retries=8 to the LM so litellm retries transient limit hits, and num_threads=1 to MIPROv2 to serialize evaluation and eliminate concurrent burst pressure.

# Before
lm = dspy.LM(eval_model)
optimizer = dspy.MIPROv2(metric=skill_fitness_metric, auto="light")
# After
lm = dspy.LM(eval_model, num_retries=8)
optimizer = dspy.MIPROv2(metric=skill_fitness_metric, auto="light", num_threads=1)

Test plan

• python -m evolution.skills.evolve_skill --skill github-code-review --dry-run — setup validates without errors
• python -m evolution.skills.evolve_skill --skill github-code-review --iterations 3 --eval-source synthetic — baseline constraint check shows all ✓ including skill_structure; evolved skill saves to output/<skill>/<timestamp>/ instead of evolved_FAILED.md; no 0.0-scored MIPROv2 trials from rate limit errors
• With a valid DSPy GEPA install, the GEPA fallback message no longer appears

🤖 Generated with Claude Code