People should get a runnable model and harness combination without understanding harness compatibility first. Current agent creation can reuse the last-used harness and model, and the backend template defaults to Pi. This can produce a default that does not match the newly selected provider.
Use these defaults when a model provider is selected:
- OpenAI uses the OpenAI harness and requires an OpenAI provider connection with an API key.
- Anthropic uses Claude Code.
- Other providers use Pi.
- Keep the current Bedrock behavior unchanged until its Claude compatibility is confirmed.
- Remove the duplicate Pi and OpenAI harness choice for the OpenAI onboarding path so OpenAI is the only default shown there.
- Recalculate the default when the provider changes, but preserve a later explicit harness choice where it remains compatible.
- Add tests for new-agent creation, onboarding, provider changes, and stored last-used preferences.
Related: #5786 tracks validation of invalid model, harness, and authentication combinations. This issue covers default selection.
People should get a runnable model and harness combination without understanding harness compatibility first. Current agent creation can reuse the last-used harness and model, and the backend template defaults to Pi. This can produce a default that does not match the newly selected provider.
Use these defaults when a model provider is selected:
Related: #5786 tracks validation of invalid model, harness, and authentication combinations. This issue covers default selection.