Skip to content

playbook 3.2.0

Choose a tag to compare

@andresfrei andresfrei released this 19 Nov 19:45
· 98 commits to main since this release
4040347

Learned rules from repeated review findings

When review keeps finding the same kind of mistake, the agent now learns from it (work repeated-corrections, #34).

  • Evidence. When review has findings, it starts the new agent playbook-agents:classify (sonnet), which records each finding under a kind of mistake in docs/wiki/evidence.md. A SubagentStop check stops a review that had findings and started no classifier.
  • Rules at the start. A SubagentStart hook gives write and review the rules of level 2 and up, from the project's page and from the playbook's general rules (plugins/playbook-agents/evidence-rules.md).
  • Consolidation by level. evidence.sh counts distinct reviews per kind and level. At the close of a work, a kind that reached 3 distinct reviews is proposed as one concrete rule one level up; the rows it covers are removed, and git keeps them. The owner approves each one, or delegates the new kind consolidate.
  • Delegation. When the owner asks for a work in their own words, the agent opens it without asking (this repository only).

Nothing to migrate: the wiki format stays 5, and a project with no evidence page sees no change. Subagent nesting must stay on (CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH not set to 1).