Message to result is a controlled execution path.
The model reads intent. Policy sets limits. Approval binds arguments. Only then can a Tool act.
- Six-step Trace
- OpenAI-compatible Provider
- Results return in context
- Structured execution records
- 01MESSAGE_RECEIVEDacceptedowner / tui
- 02AGENT_PLANNINGrunningintent → tool
- 03POLICY_CHECKallowedworkspace / argv
- 04APPROVALwaitingexact arguments
- 05TOOL_EXECUTIONrunningisolated action
- 06RESULT_DELIVEREDsucceededsame conversation