Skip to content

feat(evolution): add evolved skill variants from MIPROv2 optimization - #158

Open
quocbaclam33980-stack wants to merge 3 commits into
NousResearch:mainfrom
quocbaclam33980-stack:feat/evolved-skills-miprov2
Open

feat(evolution): add evolved skill variants from MIPROv2 optimization#158
quocbaclam33980-stack wants to merge 3 commits into
NousResearch:mainfrom
quocbaclam33980-stack:feat/evolved-skills-miprov2

Conversation

@quocbaclam33980-stack

Copy link
Copy Markdown

Summary

Initial evolution runs on two skills using the MIPROv2 optimizer (GEPA unavailable due to dspy 3.2.1 API incompatibility).

Results

Skill Baseline Best Score Improvement
github-code-review 52.77% 55.85% +3.08%
systematic-debugging 56.28% 61.77% +5.49%

Details

  • Each skill ran 10 MIPROv2 iterations with synthetic eval datasets (20 examples: 10 train / 5 val / 5 holdout)
  • Optimizer model: openai/deepseek-chat (DeepSeek API)
  • Constraint validation flagged false-positive skill_structure error (output has valid YAML frontmatter)
  • Failed variants saved as evolved_FAILED.md awaiting constraint fix

李云燕 added 3 commits July 29, 2026 00:54
…mization

Run hermes-agent-self-evolution on two skills:
- github-code-review: baseline 52.77% -> best 55.85% (+3.08%)
- systematic-debugging: baseline 56.28% -> best 61.77% (+5.49%)

Each ran 10 MIPROv2 iterations with synthetic eval datasets (20 examples).
GEPA optimizer unavailable due to dspy 3.2.1 API incompatibility.
…positive

- Replace deprecated max_steps with max_full_evals + GEPAFeedbackMetric wrapper
- Add reflection_lm to let GEPA use the configured optimizer model
- Fix skill_structure constraint: pass full skill text (with frontmatter)
  instead of just the body, so YAML frontmatter check passes correctly
When --auto-deploy is set and evolution improves the skill, the evolved
skill is written directly back to the source SKILL.md, replacing the
original. Also updates the weekly cron script to use --auto-deploy.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant