IOP issue with LM Studio + Qwen3.8 27b Splash model #53524

Activity

1. github-actions

   assigned

   neriousy

   on Oct 6, 2026
2. github-actions commented on Oct 6, 2026

   

   github-actionsBot

   on Oct 6, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • #46415: similar assistant-prefill continuation sent to a provider that rejects it after a completed response.
   • #19981: tracks avoiding assistant-prefill payloads for models and providers that do not support them.
3. opencodejr commented on Oct 27, 2026

   

   opencodejrBot

   on Oct 27, 2026

   Contributor

   More actions

   Thanks for the detailed report and the LM Studio logs. I could reproduce this on 2.0.24 when the agent has a step limit (steps, or the older maxSteps, in the agent config). It still happens on the latest dev (f7d0ce133).

   When an agent reaches its step limit, OpenCode adds a final "maximum steps reached" instruction as an assistant message at the end of the request. LM Studio's Splash engine treats a final assistant message as prefill and rejects it.

   The script below uses a mock server that acts like LM Studio in this respect. With steps: 3, the last request goes from 4 to 7 messages and ends with that assistant message, and OpenCode shows the same error you saw. That matches your 59 → 62 jump. Without a step limit, every request ended with a user or tool message and the session finished normally.

   A few notes:

   • Could you check whether your agent config (or the agent you use) sets steps or maxSteps? If it does, removing it or raising it should avoid the error for now. If it doesn't, please tell us, because that would mean another code path is involved.
   • v2 has no capabilities.assistantPrefill option, and the npm provider field doesn't change how v2 builds these requests, so those settings can't help here. chat_template_kwargs.enable_thinking is unrelated.

   I've passed this to the maintainers.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Reproduces #53524: when an agent hits its `steps` limit, OpenCode v2 appends the
   # MAX_STEPS_PROMPT as a trailing *assistant* message (assistant prefill). LM Studio's
   # Splash engine rejects such requests. A mock OpenAI-compatible server stands in for
   # LM Studio: it rejects requests whose last message is from the assistant and logs every request.
   # Needs: bun, and `opencode` on PATH (bun add -g @opencode/[email protected]).
   set -euo pipefail
   W=$(mktemp -d); cd "$W"
   cat > mock.ts <<'EOF'
   const sse = (c: unknown[]) => new Response(c.map((x) => `data: ${JSON.stringify(x)}\n\n`).join("") + "data: [DONE]\n\n", { headers: { "content-type": "text/event-stream" } })
   Bun.serve({ port: 12345, async fetch(req) {
     if (new URL(req.url).pathname.endsWith("/models")) return Response.json({ data: [{ id: "qwen3.8-27b-splash" }] })
     const b: any = await req.json(); const m = b.messages; const last = m.at(-1)
     console.log(`messages=${m.length} roles=${m.map((x: any) => x.role[0]).join("")} last=${last.role} tool_choice=${JSON.stringify(b.tool_choice)}`)
     if (last.role === "assistant") { console.log("  REJECTED trailing assistant:", JSON.stringify(last.content).slice(0, 60))
       return Response.json({ error: "Engine protocol promptTemplating capability gap for 'splash-server': Splash does not support assistant response prefill/continuation." }, { status: 400 }) }
     const base = { id: "x", object: "chat.completion.chunk", created: 0, model: b.model }
     if (b.tools?.some((t: any) => t.function.name === "glob") && b.tool_choice !== "none")
       return sse([{ ...base, choices: [{ index: 0, delta: { role: "assistant", tool_calls: [{ index: 0, id: `c${Math.random()}`, type: "function", function: { name: "glob", arguments: '{"pattern":"*"}' } }] }, finish_reason: null }] },
                   { ...base, choices: [{ index: 0, delta: {}, finish_reason: "tool_calls" }] }])
     return sse([{ ...base, choices: [{ index: 0, delta: { role: "assistant", content: "Done." }, finish_reason: "stop" }] }])
   } })
   EOF
   mkdir proj home; echo hi > proj/hello.txt
   cat > proj/opencode.json <<'EOF'
   {
     "provider": { "lmstudio": { "npm": "@ai-sdk/openai-compatible", "options": { "baseURL": "http://127.0.0.1:12345/v1", "apiKey": "x" },
       "models": { "qwen3.8-27b-splash": { "name": "Qwen3.8-27B-Splash" } } } },
     "agent": { "capped": { "mode": "primary", "steps": 3 } }
   }
   EOF
   bun mock.ts > mock.log 2>&1 & MOCK=$!; trap 'kill $MOCK' EXIT; sleep 1
   cd proj
   HOME=$W/home XDG_CONFIG_HOME=$W/home/.config XDG_DATA_HOME=$W/home/.data \
     timeout 120 opencode run --standalone --auto --agent capped -m lmstudio/qwen3.8-27b-splash "list files" || true
   echo "--- requests seen by the mock LM Studio ---"; cat ../mock.log

   Run it: bash repro.sh
4. opencodejr

   added

   reproduced

   on Oct 27, 2026
5. noahnao commented on Nov 1, 2026

   

   noahnao

   on Nov 1, 2026

   More actions

   Please fix