What happened?

When using the openai-responses API style, the agent always sends the system prompt as role: "developer" whenever model.reasoning is enabled, even when the model has compat.supportsDeveloperRole: false in models.json. Providers that don't accept the developer role (e.g. StepFun) silently drop the system prompt, so the model never sees it.

Steps to reproduce

1. Configure a StepFun model in models.json with "reasoning": true and "compat": { "supportsDeveloperRole": false }.
2. Set api to openai-responses.
3. Start a pi session with that model and a non-empty system prompt.
4. Observe that the model behaves as if no system prompt was provided.

Expected behavior

When supportsDeveloperRole is false, the system prompt should be sent as role: "system" instead of developer.

Version

0.78.1