Tool call detection issue with DMR and qwen3-coder-30b-a3b-instruct-gguf:Q3_K_M

[k33g]

Tool call detection issue with DMR and qwen3-coder-30b-a3b-instruct-gguf:Q3_K_M

Context

Configuration used:

• Tool: docker-agent
• Model: huggingface.co/unsloth/qwen3-coder-30b-a3b-instruct-gguf:Q3_K_M
• Backend: Docker Model Runner (DMR) → llama.cpp
• Provider: dmr

Symptoms

When the agent calls a skill or a tool, instead of executing the action, the model returns raw text containing the tool call:

Skill with no parameter (what-time-is-it):

<function=what-time-is-it> </function> </tool_call>

Skill with a parameter (brave-news-search):

<function=brave-news-search>
<parameter=query> small local models </parameter>
</function> </tool_call>

Shell toolset:

<function=shell>
<parameter=cmd> ls -lha </parameter>
</function> </tool_call>

These <function=...> blocks are displayed as-is in the response and are never executed.

Workaround:

• Adjust sampling parameters: lower presence_penalty to 0.0 and slightly raise temperature to improve the coherence of structured outputs (argument JSON):
• Explicit system instruction

agents:
  root:
    model: qwen3-coder
    num_history_items: 50
    skills: true
    description: Bob
    instruction: |
      You are Bob, a useful AI expert.

      IMPORTANT — Tool usage rules:
      - When you need to call a tool, use ONLY the structured tool-calling API.
      - Do NOT output tool invocations as raw text blocks such as <function=...> or <tool_call> tags.
      - The tool calling mechanism is handled automatically by the system. Just invoke the tool directly.
      - If a tool is unavailable, say so in plain text. Never fabricate tool output.

    toolsets:
      - type: shell
      - type: filesystem
      - type: fetch

models:
  qwen3-coder:
    provider: dmr
    model: huggingface.co/unsloth/qwen3-coder-30b-a3b-instruct-gguf:Q3_K_M
    temperature: 0.6
    top_p: 0.8
    presence_penalty: 0.0
    max_tokens: 32768

    provider_opts:
      runtime_flags: ["--context_size=65536"]
      top_k: 20
      repetition_penalty: 1.05