skill-evolve-looplisted

# skill-evolve-loop — sediment a lesson into a living skill, then prove you didn't break it Editing a skill that already ships is not the same as writing one: you can break behavior that used to work, and another session may be mid-edit on the same file. This loop covers the *edit path*: fold one lesson in, run a five-class regression, backfill only what failed. ## When this triggers - The human says: "沉淀到 skill 里边" / "把这轮教训写进 skill" / "skill 更新了跑一轮验证" / "fold this lesson into the skill". - Or: a deliverable just finalized and a lesson is worth keeping — propose the sedimentation round yourself; don't wait to be asked. ## The five-class behavior test After a skill edit, spin up N parallel subagents — one class each — produce a pass table, then backfill only the failed classes. The **five class names come from a real user's verification design**; the one-line criterion under each is this skill's operationalization: | Class | Operational check | Example | |---|---|---| | **Routing** | New trigger words hit; old scenarios don't stop routing here | A "sediment this lesson" request routes to the edited skill; an unrelated request doesn't | | **Regression** | Old behavior unchanged — same prompt against the pre-edit version gives the same result | A prompt that used to produce a 3-step flow still produces 3 steps | | **Discipline** | The skill's hard constraints (gates, taboos) are still obeyed after the edit | A send-gate the skill mandates still blocks an unstamped send | |