Skillkeeper · Local developer tool
Improve an agent skill without silently changing it.
An evaluation and release workflow that separates proposed repairs from live behavior.
Product evidence
See the working surface.
Product capture
The problem
A change to an agent skill can improve one example while breaking another. Developers need an inspectable path from a proposed change to a deliberate release.
Choices in the implementation
- Keep a repair candidate separate from the live skill.
- Evaluate replay and held-out cases before explicit human promotion.
- Preserve revisions, audit evidence and rollback as part of the workflow.
What exists
- A local CLI for inventory, evaluation, staged repair, promotion and rollback.
- A React/FastAPI team cockpit and immutable context revisions.
The next question
Measure regression detection and decision quality with developers maintaining real agent skills.