Resolving Tool-Use Compatibility for Qwen3-Coder on OpenClaw

Published on

Abstract

This document details the successful integration of the qwen3-coder:30b model (hosted via Ollama) with the OpenClaw agent platform. The primary challenge involved incompatible tool-calling formats: Qwen3-Coder natively outputs XML-wrapped tool calls, whereas OpenClaw expects standard OpenAI JSON formats. The solution involved a two-layer strategy: customizing the Ollama Modelfile to enforce JSON output and configuring OpenClaw to treat the Ollama endpoint as a native OpenAI provider.

1. Problem Statement

When deploying qwen3-coder:30b for OpenClaw’s agent-browser skill, the following issues were observed:

  1. Format Mismatch: The model outputted tool calls in XML format (e.g., <tools>...</tools>), which OpenClaw parsed as plain text content rather than executable instructions.
  2. Execution Failure: OpenClaw displayed the raw JSON/XML to the user but failed to trigger the actual tool execution logic.
  3. Context Pollution: Failed attempts left the conversation context in a state where the model refused to attempt further tool calls.

2. Solution Architecture

The final working configuration relies on “deceiving” OpenClaw into treating the Ollama endpoint as a standard OpenAI API, while simultaneously forcing the underlying model to adhere to OpenAI’s strict JSON schema for tool calls.

2.1 Ollama Modelfile Configuration

We created a custom model variant, qwen3-coder-json, which overrides the default System Prompt and Template to strictly enforce JSON output without XML wrappers.

Final Modelfile Content:

FROM qwen3-coder:30b
PARAMETER num_ctx 65536
PARAMETER temperature 0.1
PARAMETER top_p 0.9

# Critical: Remove XML-related stop sequences to prevent premature truncation
PARAMETER stop <|im_end|>

# Custom Template to ensure clean message formatting
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
"""

# System Prompt to enforce OpenAI-compatible JSON structure
SYSTEM """You are a function calling AI.
When the user asks you to do something that requires a tool:
1. DO NOT explain what you are doing.
2. DO NOT output XML tags like <tools>.
3. OUTPUT ONLY a raw JSON object matching this schema:
{"tool_calls": [{"id": "call_unique_id", "type": "function", "function": {"name": "tool_name", "arguments": "{\"arg_name\": \"value\"}"}}]}

Example:
User: Search for apple
Assistant: {"tool_calls": [{"id": "call_123", "type": "function", "function": {"name": "web_search", "arguments": "{\"query\": \"apple\"}"}}]}
"""

Deployment Command:

ollama create qwen3-coder-json -f ~/Modelfile

2.2 OpenClaw Configuration

The openclaw.json configuration was adjusted to use the openai provider type instead of ollama. This bypasses OpenClaw’s Ollama-specific logic (which might not support tool calls fully) and uses the standard OpenAI SDK path.

File Path: ~/.openclaw/openclaw.json

{
  "provider": "openai",
  "base_url": "https://ollama.gindody.com/v1",
  "api_key": "ollama",
  "model": {
    "id": "qwen3-coder-json",
    "name": "Qwen3 Coder JSON"
  }
}

3. Implementation Steps

  1. Model Creation: Execute the ollama create command with the Modelfile above to generate the qwen3-coder-json model.

  2. Configuration Update: Modify ~/.openclaw/openclaw.json to point to this new model using the openai provider.

  3. Context Reset (Crucial): Before testing, clear any polluted conversation history that might contain failed XML outputs.

    • Command: /new in the OpenClaw chat interface.
    • Alternative: rm ~/.openclaw/agents/main/sessions/*.jsonl

  4. Verification:

    • Prompt: “Use Agent Browser to search for Shanghai weather.”
    • Expected Behavior: The agent should not output JSON code blocks. Instead, it should silently execute the tool and return the natural language result (e.g., “Shanghai is currently clear with a temperature of 4°C”).

4. Key Learnings