Enforce System Message Ordering for Qwen3

[seanbruno]

It looks like recent Qwen3 models are enforcing strict ordering on their system messages. This probably means a new module needs to be written that specifically deals with this quirk.

I did a bit of hackery in 1fa74da to try and test out possible solutions but its not really practical to add this exception to the openai module and means that a new Qwen module is likely the best path forward.

[olimorris]

When you say "strict ordering", are we talking about there only being one system message for the entire request?

DeepSeek does enforce that, so to that end, their adapter has been built differently to most. Where it can, it extends the OpenAI adapter but has to conform to the API requirements for building out the messages payload.

[seanbruno]

The error JSON coming back from Qwen looks like:

Error 09:05:04 AM notify.error CodeCompanion [ERROR] 2026-09-12 09:05:04 Error: {"error":{"message":"litellm.BadRequestError: OpenAIException - System message must be at the beginning.. Received Model Group=nrp/qwen3\nAvailable Model Group Fallbacks=None","type":null,"param":null,"code":"400"}} Error 09:05:04 AM notify.error CodeCompanion [ERROR] 2026-09-12 09:05:04 [chat::_submit_http] Error: { body = '{"error":{"message":"litellm.BadRequestError: OpenAIException - System message must be at the beginning.. Received Model Group=nrp/qwen3\\nAvailable Model Group Fallbacks=None","type":null,"param":null,"code":"400"}}',exit = 0, headers = { "date: Sat, 12 Sep 2026 15:05:04 GMT", "content-type: application/json", "content-length: 215", "x-litellm-call-id: 7315108b-d122-4d99-a26e-febd98d46131", "x-litellm-response-cost: 0", "x-litellm-key-spend: 0.0", "x-litellm-timeout: 1200", "x-frame-options: DENY", "content-security-policy: frame-ancestors 'none'", "x-content-type-options: nosniff", "strict-transport-security: max-age=31536000; includeSubDomains" },status = 400}

When I rewrote the handler to put all system messages at the start of the transaction, I was able to get chat completion to work, but then started tripping over tool messages.

I'm going to sign up for the free tier of Qwen and reproduce the failure case.

[olimorris]

Did you extend from the DeepSeek adapter instead of the OpenAI one?