Qwen3.8-27B + SGLang (RadixArk/MiaAI DGX Spark): System message must be at the beginning — working plugin coalesce #45055

Summary

OpenCode 1.18.23 still emits multiple role: "system" fragments to OpenAI-compatible backends. Serving Qwen3.8-27B with SGLang on NVIDIA DGX Spark using the stock HuggingFace chat template (strict: only messages[0] may be system) makes every agent turn fail with:

Bad Request: System message must be at the beginning.

The same OpenCode config / same short model id worked under vLLM. After switching the inference stack to the MiaAI / RadixArk SGLang-on-DGX-Spark recipe, OpenCode broke — the model weights are Qwen3.8, the engine + raw HF template are what changed.

Related: #15059, #16560, #20785, #20813 · PRs #15018, #16981, #23656.

Serving stack (keywords)

Deployed from / based on:

• Repo: https://github.com/MiaAI-Lab/Qwen3.8-27B-SGLang-DGX-Spark
• Cookbook: https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-27B
• Weights (default QUANT=nvfp4): RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead
  (also: RadixArk/Qwen3.8-27B-NVFP4, Qwen/Qwen3.8-27B, Qwen/Qwen3.8-27B-FP8)
• Docker image: lmsysorg/sglang:qwen38-27b
• Engine: SGLang (not vLLM) — / returns "SGLang is running", /v1/models → owned_by: "sglang"
• Hardware: NVIDIA DGX Spark / GB10 (aarch64)
• Spec modes in that repo: DSpark / EAGLE-MTP / DFlash2 (we hit the same template guard regardless)
• Reasoning / tools (server): --reasoning-parser qwen3, tool parser qwen3_coder
• Context: native 262144
• Default served name in MiaAI scripts: SERVED_MODEL_NAME=qwen3.8-27b-sglang
• Our OpenCode alias (intentionally shortened to match older config): API model id qwen3.8-27b → OpenCode dgx/qwen3.8-27b

Live identity probes also reported e.g. model_type: qwen3_5, architecture Qwen3_5ForConditionalGeneration, SGLang version string like 0.0.0.dev0+qwen38.27b.….

Repro

Server (template):

# OK — single system at [0]
curl …/v1/chat/completions -d '{
  "model":"qwen3.8-27b",
  "messages":[{"role":"system","content":"Be brief."},{"role":"user","content":"Say OK"}]
}'

# 400 — system not first
# "System message must be at the beginning."

# 400 — agent-like second system after tools
# [system, user, assistant+tools, tool, system, user]

OpenCode (before workaround):

opencode run -m dgx/qwen3.8-27b 'Reply with exactly: PONG'
# → Error: Bad Request: System message must be at the beginning.

Workaround that works (plugin, no OpenCode fork)

Last plugin in plugin[] coalesces via experimental hooks (in-place mutate; reassign is a no-op):

1. experimental.chat.system.transform — join all output.system[] with \n\n → one fragment
2. experimental.chat.messages.transform — merge all system/developer messages → one messages[0]

Optional model config (documents intent; may be ignored until core lands):

"systemMessage": "single"

After plugin + restart:

opencode run -m dgx/qwen3.8-27b 'Reply with exactly: PONG'
# → PONG

Plugin sketch

export default async function QwenSingleSystem() {
  const mergeSystemFragments = (system) => {
    if (!Array.isArray(system) || system.length <= 1) return
    const merged = system
      .map((s) => (typeof s === "string" ? s : String(s ?? "")))
      .map((s) => s.trim())
      .filter(Boolean)
      .join("\n\n")
    system.splice(0, system.length, merged || "")
  }

  const collapseMessageSystems = (messages) => {
    if (!Array.isArray(messages) || messages.length === 0) return
    const systemIdx = []
    for (let i = 0; i < messages.length; i++) {
      const role = messages[i]?.info?.role
      if (role === "system" || role === "developer") systemIdx.push(i)
    }
    if (systemIdx.length <= 1) {
      if (systemIdx.length === 1 && systemIdx[0] !== 0) {
        const [sys] = messages.splice(systemIdx[0], 1)
        messages.unshift(sys)
      }
      return
    }
    const texts = []
    for (const i of systemIdx) {
      for (const p of messages[i]?.parts || []) {
        if (p?.type === "text" && p.text) texts.push(String(p.text).trim())
      }
    }
    const mergedText = texts.filter(Boolean).join("\n\n")
    for (let k = systemIdx.length - 1; k >= 0; k--) messages.splice(systemIdx[k], 1)
    messages.unshift({
      info: { role: "system" },
      parts: mergedText ? [{ type: "text", text: mergedText }] : [],
    })
  }

  return {
    "experimental.chat.system.transform": async (_input, output) => {
      mergeSystemFragments(output.system)
    },
    "experimental.chat.messages.transform": async (_input, output) => {
      collapseMessageSystems(output.messages)
    },
  }
}

Ask

Please land a core equivalent (systemMessage: "single" / merge before provider send for openai-compatible / non-Anthropic), so SGLang + raw HF Qwen3.8 (RadixArk NVFP4 on DGX Spark, lmsysorg/sglang:qwen38-27b) works without a local plugin.

Happy to test nightlies against this stack.

Note

We will drop the plugin when upstream merge ships and stays stable. If this coalesce causes context loss / instability, we will revert and report.