System message must be at the beginning - LiteLLM/vLLM/Qwen3

[JKratto]

Short description

vLLM/Qwen3 throws "System message must be at the beginning" error - plugin prepends context before system message (similar to cortexkit/aft#150)

What happened?

When using magic-context with OpenCode, I'm encountering a BadRequestError from vLLM when using Qwen3.6 27B model:

litellm.BadRequestError: Hosted_vllmException - {"error":{"message":"System message must be at the beginning.","type":"BadRequestError","param":null,"code":400}}

This error persists even with model fallbacks (Qwen3.6-27B → Qwen3.5-9B). The issue appears to be that magic-context is prepending context messages before the system message in the message list, which violates vLLM's/chat tempplate's requirement that the system message must be the first message.

Related Issue: This appears similar to cortexkit/aft#150, which also involved system message positioning.

Environment:

• OpenCode with magic-context plugin
• Model: Qwen3.6-27B served by vLLM through LiteLLM

Diagnostics

Plugin version

v0.36.1

OpenCode version

1.18.18

Platform

macOS arm64

Client

OpenCode TUI (CLI)

Log output (optional)

[coleleavitt]

Verified fixed on current master (b27a6d76). Commit 7d2d8444 changed Magic Context guidance injection to concatenate into output.system[0] instead of creating a second system entry; system-prompt-hash.test.ts covers the real OpenAI-compatible converter. Current targeted result: 44 pass, 0 fail. This issue can be closed.