Problem
The OSS/MLflow provider currently relies mainly on a static validated model cohort. AI Gateway foundation-model metadata can expose additional chat-completions-only models, reasoning capability, and context information, but the data is partial and must not leave clients uncapped or route native Claude/Gemini/GPT endpoints through MLflow.
Expected behavior
- Discover MLflow chat-completions-only foundation models dynamically.
- Project live capabilities onto matching UC model-service IDs.
- Persist reasoning, context-window, and output-limit metadata for Pi and OpenCode.
- Exclude endpoints that expose native Anthropic, Gemini, or Responses APIs.
- Deduplicate malformed/repeated endpoint entries.
- Keep static GLM/Kimi behavior when metadata is unavailable.
- Complete partial/unknown limits with a conservative bounded pair.
- Keep SSE repair and cache diagnostics in separate changes.
Problem
The OSS/MLflow provider currently relies mainly on a static validated model cohort. AI Gateway foundation-model metadata can expose additional chat-completions-only models, reasoning capability, and context information, but the data is partial and must not leave clients uncapped or route native Claude/Gemini/GPT endpoints through MLflow.
Expected behavior