v0.2.1 — skills that compound
Skills now close the full Act → Measure → Analyze → Adjust loop:
knitbrain_skill_outcome(26th tool) — report a concrete worked/failed signal after using a skill; 'task complete' is not an outcome- Failure notes become pitfalls — folded into the playbook automatically, dated
- Health verdicts — skills that keep failing are flagged
needs-revision;knitbrain_runwarns instead of silently re-serving them - Constraints — non-negotiable guardrails on skills ('never run migrations directly') that merge on update and propagate into every agent's mission brief
knitbrain help— bare invocation starts the MCP server silently; new users now get a command map- knit-mcp is deprecated on npm in favor of knitbrain
Verified: 199 tests · all 26 tools live over stdio · evals PASS on 4,567 real blocks · production audit 50/50 · cold-registry check green.