Skip to content

model-compatibility.md

Latest commit

 

History

History
21 lines (17 loc) · 1.36 KB

model-compatibility.md

File metadata and controls

21 lines (17 loc) · 1.36 KB
Raw

OpenAI-compatible model contract

Custom OpenAI-compatible backends in 0.6 use POST {baseUrl}/chat/completions with bearer-token authentication when an API key is configured. The endpoint must support:

  • JSON requests containing model, messages, stream: true, and OpenAI-style tools;
  • SSE frames prefixed with data:, terminated by data: [DONE];
  • streamed choices[].delta.content, optional reasoning text, and indexed choices[].delta.tool_calls whose function arguments may arrive in fragments;
  • finish reasons stop, tool_calls, or function_call;
  • UTF-8 content and standard HTTP status codes for authentication, rate limits, and availability.

Images are sent as OpenAI content parts only for models that accept them. A backend that rejects image content is retried without image parts and remembered as text-only for the process lifetime. Unsupported proprietary request fields, non-SSE streaming formats, interactive tool approvals, and provider-specific Session semantics are outside this compatibility contract.

OpenGUI also has explicit internal routes for Codex Responses and configured Anthropic Messages models. Those are not evidence that an arbitrary endpoint supports multiple protocols. Validate a custom endpoint with text, Unicode, tool calls, abort, malformed-stream, rate-limit, and context limit cases before offering it to other Host members.