evolve_skill always fails skill_structure constraint — validator checks bare body instead of reassembled skill

[ycheng87]

Summary

Running python -m evolution.skills.evolve_skill always ends with the skill_structure constraint failing — even when the evolved output is a complete, valid SKILL.md with intact YAML frontmatter:

✗ skill_structure: Skill missing: YAML frontmatter (---), name field, description field
✗ Evolved skill FAILED constraints — not deploying

Root cause

In evolution/skills/evolve_skill.py (~lines 181–186):

evolved_body = optimized_module.skill_text
evolved_full = reassemble_skill(skill["frontmatter"], evolved_body)

# ── 7. Validate evolved skill ───────────────────────────────────────
evolved_constraints = validator.validate_all(evolved_body, "skill", ...)

The validator receives evolved_body — the bare body without the YAML frontmatter — while _check_skill_structure in evolution/core/constraints.py requires the text to start with --- and contain name: / description: in the first 500 chars. A bare body never contains frontmatter, so the check always fails regardless of how good the evolved skill actually is.

Repro

1. python -m evolution.skills.evolve_skill --skill systematic-debugging --iterations 3 --eval-source synthetic
2. The optimizer completes (best score reported), then validation always reports the missing-frontmatter failure.
3. The saved output/systematic-debugging/evolved_FAILED.md is actually a complete file with valid frontmatter — in our run it is byte-identical to the baseline (0 diff), which also confirms the failure is in validation, not in generation.

Suggested fix

Validate the reassembled artifact instead of the bare body:

evolved_full = reassemble_skill(skill["frontmatter"], evolved_body)
evolved_constraints = validator.validate_all(evolved_full, "skill", baseline_text=skill["body"])

(Note: baseline_text comparisons in validate_all may also need the reassembled baseline — reassemble_skill(skill["frontmatter"], skill["body"]) — so size/growth gates compare like with like.)

Environment

• hermes-agent-self-evolution v0.1.0 (main)
• dspy 3.3.0, optuna 4.9.0, Python 3.11.15
• Model: deepseek-v4-flash via an OpenAI-compatible endpoint — the bug reproduces with any model since it lives in validation, not generation.

---

Root cause identified by 宇航员 (Xiaomo), the reporter's Hermes agent; filed by ycheng87.