Skip to content

2026-08-20-list-models.md

Latest commit

 

History

History
50 lines (44 loc) · 3.17 KB

2026-08-20-list-models.md

File metadata and controls

50 lines (44 loc) · 3.17 KB
Raw

AutoLLMClient lists the model ids an endpoint serves

  • Date: 2026-08-20
  • Type: feature
  • Scope: auto_client, claude5, gemini3_7, integration, tests
  • PR: #181

中文版

What changed

  • list_models() / listModels() joined AutoLLMClient in both languages, returning the model ids the configured endpoint serves in the order the endpoint returned them. The OpenAI-compatible clients (openai_chat, openai_responses, gpt5_6, minimax_m3, deepseek_v4, glm5_3, kimi_k3, openai_embedding) and the Anthropic Messages clients (claude5, ant_messages) read their SDK's models.list(); gemini3_7 reads the Gemini model list. Paging belongs to the SDKs, so a result spans every page the endpoint serves.
  • A listing is filtered to what the routed client can be used for. A protocol client (openai-chat, openai-responses, ant-messages, openai-embedding) is named explicitly and speaks for whatever the endpoint serves, so its listing comes back whole; a client deduced from a model id keeps only the ids that deduce back to it, so a gateway fronting many vendors narrows to that client's own models.
  • AutoLLMClient._client_class_for_model split the routing conditions out of _create_client_for_model, which now looks up the class and constructs it. The conditions live in one place and answer both questions: which client serves a model id, and whether a listed id belongs to the client in hand.
  • gemini3_7 returns the last path segment of each name, so both models/gemini-3.7-flash and publishers/google/models/gemini-3.7-flash become gemini-3.7-flash — the spelling AutoLLMClient takes as model.
  • UnsupportedOperationError joined errors in both languages, reporting a capability the routed client does not have rather than a rejected parameter value, and claude5 raises it on a bedrock:// base URL because the Bedrock SDK client carries no models resource.
  • The playground lists into its own Model dropdown: a List models control on the Model label row calls POST /api/models and adds each returned id as an option, tagged with the client type the listing ran under. Selecting one keeps that client type on screen in the Client type field, rather than routing by something the panel does not show, and an edit there is written back to the option so it survives the next selection. A rejected listing shows the provider's message under the field instead.
  • Offline tests in both languages route a fake models endpoint through every client and pin the filtering, the Gemini path stripping and the Bedrock rejection: src_py/tests/test_list_models.py, src_ts/tests/list-models.test.ts.

Endpoint by protocol

Protocol Clients Request
OpenAI Chat Completions openai_chat, deepseek_v4, glm5_3, kimi_k3, openai_embedding GET {base}/models
OpenAI Responses gpt5_6, openai_responses, minimax_m3 GET {base}/models
Anthropic Messages claude5, ant_messages GET {base}/v1/models
Gemini generateContent gemini3_7 GET {base}/v1beta/models