New Agent Extensibility: Refactor Agent interface to allow simple(r) Agent integration contributions #281
Description
Problem or use case
We need to restructure the hooks, transcript parsing and lifecycle actions so that it's easy to contribute a new agent.
Currently the boundaries are not fully defined as to what is the framework and what is Agent-specific.
This will get picked up by the core maintainer team - we ask that new agent support please come after this to limit rework.
Acceptance criteria
- Documentation describing
- how the hooks, lifecycle events and actions work together
- what is the responsibility of the framework vs the agent implementation
- a clear interface for mandatory agent concerns
- clarity around optional actions/hooks if necessary (documentation or code-based)
Desired behavior
New agent support can be added without changing 'framework' code.
Proposed solution
- define 'framework' and 'agent' responsibilities in the lifecycle
- Refactor Agent Interface to reflect the above
- Move/create framework code where necessary
- Refactor existing Claude and Gemini agents
Notes
See Pi issue re: log formats when designing the transcript interfaces