[Pepps233]

Summary

Evaluate structurally valid evolved-skill drafts with separate retrieval and behavioral gates before they can enter the accepted candidate state, without adding a skill writer or activation path.

Reasoning

Evidence qualification and deterministic draft validation establish recurrence, provenance, structure, scope, and security, but they do not show that a skill is retrieved for the right tasks or improves downstream outcomes. Exact retrieval accuracy is also insufficient because a selected skill can still regress execution or be misapplied to neighboring and unrelated tasks.

Evaluation therefore uses exact source replay for reproducibility, independently held-out intended cases for improvement, and hard-negative and unrelated cases for non-regression. Baseline and candidate-available arms remain pinned to the same harness and deterministic verifier configuration.

Changes

• Add immutable skill-evaluation cases, outcomes, metrics, gates, runtime configuration, and report contracts.
• Require complete source-replay coverage plus held-out intended, hard-negative, and unrelated cases before execution.
• Evaluate offline selection separately from paired baseline and full-pool behavioral execution.
• Require held-out intended improvement, source replay reproducibility, and no negative-case regression or candidate misapplication.
• Persist the complete evaluation artifact with the exact harness, candidate, neighboring skills, model, runtime, retriever, executor, and verifier configuration.
• Advance candidates through resumable evaluation-pending checkpoints to accepted or quarantined, and invalidate stale evaluations when proposal revisions change.

Stack

Position: 8 of 12
Base: #147
Next: #149

1. feat(coding-agent): activate versioned evolved skills #134
2. feat(agent): preserve target-neutral reflection evidence #136
3. feat(agent): curate and checkpoint skill candidates #137
4. feat(agent): persist and process skill candidate registry #138
5. feat(agent): synthesize outcome-aware skill drafts #139
6. feat(agent): validate evolved skill drafts #140
7. feat(agent): measure procedural information loss #147
8. feat(agent): evaluate evolved skill candidates #148 (this PR)
9. feat(agent): activate evolved skill candidates #149
10. feat(agent): observe evolved skill lifecycle outcomes #150
11. feat(agent): define evolved skill memory handling #151
12. feat(coding-agent): add evolved skill operator controls #152

Testing

• cd packages/evolution && node ../../node_modules/vitest/dist/cli.js --run test/skill-evaluator.test.ts test/skill-candidate-registry.test.ts test/skill-candidate-worker.test.ts test/filesystem-store.test.ts test/skill-draft-validator.test.ts test/skill-synthesizer.test.ts
• npm run check (blocked by pre-existing model-catalog TypeScript failures reproduced at feat(agent): measure procedural information loss #147 head c09ee8035; no packages/evolution errors)

This pull request is AI-generated by Diffex.