Cedar kool klub

Before you commit: agent-configuration changes

Cedar's configuration agent runs on the knowledge in apps/server/.claude/skills/cedar-configuration/. Those files assert concrete facts about this codebase — tool ids, enum members, constants, values that are retired. When the code changes and those files don't, the agent keeps acting on a claim that is no longer true, and it fails silently: it writes a config that compiles, reads correctly to the user, gets confirmed as working, and does nothing. That is how a retired task channel ended up in a live customer's playbook.

If your change touches the agent-configuration surface, run /refresh-agent-knowledge before committing. It re-verifies every claim against the code, repairs what it can, and only asks you about things that need a decision.

That means any change to:

  • triggers, subagent dispatch, or orchestrator ordering
  • tool allowlists or what an agent is permitted to call
  • the operation registry, or any button, task, field, or stage vocabulary
  • playbook or subagent write paths, and their verification
  • org/user scope resolution, or who may act on whom
  • limits an agent is expected to design around — runtime caps, query caps, rate limits

Retiring or renaming a value counts, and is the case most worth catching. A dead value that the agent still believes in is indistinguishable from a working one until a customer hits it.