bug: System prompt sent out of order for vLLM 0.24 #27176

Is there an existing issue for this?

• I have searched the existing issues

Current Behavior

When any system prompt is enabled for a model being served by vLLM inference breaks due to the system prompt being not being sent first.

Relevant Log Output

Jul 10 14:23:05 coder coder[83746]: 2026-07-10 18:23:05.488 [warn]  coderd.aibridged.pool.streaming: openai stream error  model=/opt/vLLM/models/qwen3.6_38b_a3b_fp8/  interception_id=7ade39d0-d5ff-4957-91be-c7b49f23fd45  credential_kind=centralized  error="POST \"https://inference.seer.ci.stuart.fl.us/v1/chat/completions\": 400 Bad Request {\"message\":\"System message must be at the beginning.\",\"type\":\"BadRequestError\",\"param\":null,\"code\":400}"

Expected Behavior

Any system prompt messages should be sent before chat messages/context.

Steps to Reproduce

1. Configure an OpenAI compatible endpoint for the vLLM instance in Coder
2. Configure a model for the the instance
3. Enable Include Coder Agents default system prompt. option for the model.
4. Try to chat with the model inside of Coder Agents

Environment

• Host OS: Alma Linux 10
• Coder version: 2.35.1+8b44bd6
• vLLM version: 0.24

Additional Context

The issue occurs consistently