cuopt-skill-evolution

Skill evolution workflows capture generalizable learnings from non-trivial interactions like user corrections or failures, generating structured proposals to safely update target skills. Proposals follow a four-field format: Target, Trigger, Scored, and Diff.

The structured proposal format for skill updates uses four fields: Target (the skill to update), Trigger (the interaction condition), Scored (the evaluated learning), and Diff (the specific change). This ensures review-ready edits that comply with safety governance.

Trigger skill evolution after meaningful interactions such as user corrections, workflow failures, undocumented API behavior, or workflow thrash. These non-trivial events surface reusable patterns that can be extracted as generalizable learnings for proposal generation.

No, skill evolution generates review-ready proposals but does not automatically apply changes. You must review the four-field proposal before applying any updates, ensuring all safety constraints are preserved and the evolution skill itself remains unmodified.

Yes, skill evolution applies across cuOpt skills and workflows to surface reusable patterns. It extracts generalizable learnings from non-trivial interactions and generates structured proposals for any target skill within the cuOpt ecosystem.

Skill evolution cannot modify its own evolution skill and must preserve all safety constraints. Proposals are limited to the four-field format and require manual review before applying changes, preventing autonomous or unverified skill modifications.