[MomoSolaris8]

DashScope retired qwen3-max and qwen-vl-plus, so chat, tools, claim review and image analysis were failing with model_not_found. This PR replaces the retired models and adds DeepSeek as its own provider.

What changed

• Removed qwen3-max and added qwen3.8-max, the latest Qwen max model
• New deepseek provider:
  • It calls DeepSeek's official API through the existing OpenAI-compatible client.
  • It uses its own key: DEEPSEEK_API_KEY, plus DEEPSEEK_BASE_URL with a default of https://api.deepseek.com.
  • It adds the model deepseek-v4-pro, with text and tool capabilities.
• Example configs, docker-compose.prod.yml, CI and the README use qwen3.8-max, deepseek-v4-pro and qwen3-vl-plus instead of the retired models

Testing

• 4 new registry tests:
  • DeepSeek resolves to its own provider, key and base URL.
  • DeepSeek is left out of the model list when its key is missing.
  • qwen3.8-max resolves.
  • qwen3-max in CHAT_MODELS now fails at startup.
• Before writing the registry entries, I called every candidate model for real, once for chat and once for a tool call.
• I called deepseek-v4-pro through complete_chat with a real key. It worked and reported reasoning tokens.

Deployment

Production was already fixed by changing config only, before this PR (revision 0000017):

• CHAT_MODEL, TOOL_MODEL and CLAIM_REVIEW_MODEL are qwen-plus.
• VISION_MODEL is qwen3-vl-plus.
• CHAT_MODELS is qwen-plus,qwen-max,kimi-k2.6.

After this PR is merged and deployed:

1. Add the secret deepseek-api-key to the Container App, and set DEEPSEEK_API_KEY to read from it.
2. Add qwen3.8-max,deepseek-v4-pro to CHAT_MODELS.
3. Smoke test complete_chat with requested_model="deepseek-v4-pro" inside the container.

Deploy the code before changing the config. The running code doesn't know deepseek-v4-pro, and an unknown id in CHAT_MODELS stops the backend from starting.

Follow-ups (not in this PR)

• Log a warning when a model is skipped because its provider has no key. Today it is dropped silently, so a typo in the variable name would go unnoticed.
• Rename the qwen provider to dashscope, since it is the platform and not the model family.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
support-agent-one-coyl	Ready	Preview	Oct 30, 2026 2:49pm UTC