Console — provider presets inside the model editor:
- Preset chips for five coding-plan providers (Zhipu GLM / Kimi / DeepSeek /
Volcengine Ark / MiniMax) rendered at the top of the add/edit model dialog
- All five providers expose both an Anthropic-protocol and an
OpenAI-protocol endpoint; a click fills the endpoint matching the current
model type, and switching the type automatically swaps to the same
provider's other endpoint (no stale URL/protocol mismatches)
- Switching to a different provider clears the API key; switching protocols
within the same provider keeps it
- Fills base URL (full-request-URL form for irregular paths), request
protocol, claude-cli custom headers for Anthropic endpoints, token limits,
remark, and the default model id + display name; known preset model ids
merge into the model-name combobox
Server — model discovery for sub-path endpoints:
- model_discovery_url previously replaced the whole path with /v1/models,
which strips the /api/anthropic, /coding, /api/paas/v4 style prefixes every
coding-plan provider uses, so discovery always 404'd for them. It now
appends to the existing path and normalizes endpoint/version suffixes
(/chat/completions, /responses, /messages, /v1, /v3, /v4)
- Verified live against Zhipu (10 models) and Kimi (4 models) coding plans;
existing tests updated to the append semantics plus new URL cases
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
