pi harness gateway path collapses multi-family providers to anthropic-only, breaking non-Anthropic models (404 with no body)

[angiarao]

Summary

When a provider config declares both an anthropic: and openai: family, _apply_provider_to_pi (omnigent/runtime/workflow.py) always selects anthropic: as pi's single auth/base-URL source, regardless of which model the agent spec actually requests. Any non-Anthropic model then gets routed to the anthropic-messages endpoint and fails with a bare 404 status code (no body).

Root cause

The current gateway can serve multiple models, both Anthropic and OpenAI-family, from the same base URL. PiExecutor's gateway path generates a fresh models.json into a temp PI_CODING_AGENT_DIR, built entirely from HARNESS_PI_GATEWAY_BASE_URLS — which _apply_provider_to_pi populates with at most one family's base URL, preferring anthropic: whenever both are present on the provider. This throws away pi's native ability to route different models to different wire protocols from a single models.json.

Repro

1. Configure one provider with both families pointed at the same gateway base URL:

   providers:
     my-gateway:
       anthropic: {api_key_ref: ..., base_url: https://gateway.example.com/}
       openai:    {api_key_ref: ..., base_url: https://gateway.example.com/, models: {default: some-glm-model}, wire_api: chat}
       default: true
       kind: gateway

2. Run an agent with executor: {model: some-glm-model, auth: {type: provider, name: my-gateway}, config: {harness: pi}}
3. Observe inner executor error: 404 status code (no body)

Current workaround

Split into two providers, one per family, and pin the right one per agent spec via executor.auth. Works, but forces N providers for N wire families sharing one gateway.

Suggested fix

_apply_provider_to_pi (or PiExecutor._build_models_json) should emit a models.json with one entry per configured family, each carrying its own base URL/auth/wire API, instead of collapsing to a single family upfront.

Environment

• Harness: pi (SDK/headless path via PiExecutor)
• Relevant files: omnigent/runtime/workflow.py::_apply_provider_to_pi, omnigent/inner/pi_executor.py::_build_models_json

[github-actions]

Thanks for reporting this. #4343 may be related, and has already been fixed — so the fix may have shipped after the build you're on.

Could you check whether you're on a version that includes it? If you are and this still happens, say so here — that makes it a regression rather than a duplicate, and we'll keep this open.

[github-actions]

🤖 Automated triage

• Bot assessment: Medium impact
• Priority: P2-medium
• Why: Pi users requesting non-Anthropic models through a provider containing multiple model families cannot complete the submit-request-and-receive-results CUJ because requests fail with a 404, but they can work around it by splitting families into separate providers and selecting the appropriate provider.

This automated assessment uses the issue content and repository signals. Maintainers can override the priority label.