adapter_workflow.md

Adapter Workflow

This is the shortest practical guide for adding or improving an adapter in Codette.

What To Define

For a new adapter, document:

• the adapter name
• the reasoning role it fills
• what kinds of prompts should route to it
• how it should differ from base behavior

Where It Hooks In

Key files:

• inference/codette_orchestrator.py
• inference/adapter_router.py
• configs/adapter_registry.yaml
• training scripts under training/

Typical Steps

1. Add or update the adapter entry in the registry/config.
2. Add or adjust routing logic so the adapter is reachable.
3. Update prompts or mode instructions if needed.
4. Add tests for:
   • routing selection
   • response shape
   • any failure modes or constraints
5. Add benchmark or transcript evidence if the adapter changes major behavior.

Behavioral LoRA Notes

If you train or update a behavioral adapter:

• record the dataset source
• record the training objective
• document what permanent or semi-permanent behavior it is meant to stabilize
• add at least one reproducible validation artifact

Verification

At minimum:

python3 -m unittest tests.test_adapters
python3 -m unittest tests.test_phase7_executive_controller

Then add a benchmark or demo artifact if the adapter meaningfully changes public behavior.