Tracing Qwen Code

MLflow Tracing provides automatic tracing for Qwen Code, Alibaba's open-source terminal-based coding agent. After setup, MLflow will automatically capture traces of your Qwen Code conversations including:

• User prompts and assistant responses
• Tool usage (file operations, shell commands, etc.)
• Token usage (input, output, and total tokens)
• Session metadata including working directory and user

Traces are created after each conversation turn.

Setup

Requirements

• MLflow tracking server running
• Qwen Code installed (npm install -g @qwen-code/qwen-code)
• Node.js 18+
• The @mlflow/qwen-code npm package

Step 1: Install the Tracing Plugin

bash

npm install -g @mlflow/qwen-code

If you'd rather not install globally, you can invoke the bin through npx (e.g. npx @mlflow/qwen-code setup) — everything below works the same way.

Step 2: Run the Setup Command

Running mlflow-qwen-code setup starts an interactive flow for setting up the tracing hook.

text

~/my-project $ mlflow-qwen-code setup

Configure MLflow tracing for Qwen Code
? Where should MLflow tracing be installed?
  ● Project  ./.qwen/  (default)
  ○ User  ~/.qwen/

  ↑/↓ to move, enter to select

✓ Registered Stop hook in /path/to/project/.qwen/settings.json

  MLflow tracking URI [http://localhost:5000] http://localhost:5000
  MLflow experiment ID [0] 1

✓ Wrote tracing config to /path/to/project/.qwen/mlflow-tracing.json

Next steps
  1. Start the MLflow tracking server in a separate terminal:
       mlflow server --port 5000
  2. Launch qwen — traces appear at http://localhost:5000 after each turn.

Pass --non-interactive / -y to skip the interactive flow and use the default values.

note

Older Qwen Code versions (prior to 0.14) disabled the hook system by default and required a top-level "hooksConfig": { "enabled": true } entry to turn it on. Current versions ignore that setting and run hooks unconditionally — it's safe to leave in if you already have it, but no longer needed.

Step 3: Start the MLflow Tracking Server

The MLflow server must be running to receive traces. Start it in a separate terminal:

bash

mlflow server --port 5000

Step 4: Launch Qwen Code

bash

qwen

Traces become visible at http://localhost:5000 after each turn completes.

note

MLFLOW_TRACKING_URI and MLFLOW_EXPERIMENT_ID environment variables (if set in the shell that launched qwen) override the values written to mlflow-tracing.json. This is convenient for one-off overrides in CI or when switching between local and Databricks endpoints.

How It Works

1. The Stop hook fires after each Qwen Code conversation turn
2. Qwen passes the session ID and transcript path via stdin
3. The plugin reads the chronological ChatRecord JSONL transcript and matches each functionCall to its tool_result by callId
4. An MLflow trace is created with an AGENT root span (qwen_code_conversation) and LLM/TOOL child spans. LLM spans use OpenAI chat format ({model, messages} inputs, {choices} outputs)

Configuration Examples

Write these values via mlflow-qwen-code setup or by editing ~/.qwen/mlflow-tracing.json directly:

Local MLflow server

json

{ "trackingUri": "http://localhost:5000", "experimentId": "0" }

Databricks MLflow

json

{ "trackingUri": "databricks", "experimentId": "123456789" }

Per-run overrides via environment variables (take precedence over the config file):

bash

MLFLOW_TRACKING_URI=databricks MLFLOW_EXPERIMENT_ID=123456789 qwen

Token Usage and Cost

MLflow automatically tracks token usage for each LLM call within Qwen Code conversations. See the Token Usage and Cost Tracking documentation for details.

Troubleshooting

Common Issues

Tracing not working:

• Confirm mlflow-qwen-code is on PATH: run which mlflow-qwen-code (should resolve after npm install -g @mlflow/qwen-code)
• Confirm ~/.qwen/mlflow-tracing.json exists and contains a valid trackingUri. Re-run mlflow-qwen-code setup if the file is missing or stale.
• Verify Node.js 18+ is being used (required for fetch support)
• On Qwen Code versions older than 0.14, hooks are disabled by default — add a top-level "hooksConfig": { "enabled": true } entry to enable them

Missing traces:

• Traces are created per conversation turn, not per session
• Verify the MLflow tracking server is accessible
• Check the Qwen debug log (~/.qwen/debug/latest) for hook execution status

Duplicate traces:

• Ensure the hook is only registered once — check both ~/.qwen/settings.json and .qwen/settings.json for duplicate entries

Disable Tracing

Remove the hooks section from settings.json. Existing traces are preserved.

(On older Qwen Code versions that still honor hooksConfig, setting "hooksConfig": { "enabled": false } will also disable the hook system without removing the entry.)