Mlflow

OpenCode plugin that sends LLM tracing data to MLflow with full observability: token counts, costs, model info, generation parameters, and GenAI semantic conventions.

What it captures

Category	Fields
Tokens	input, output, reasoning, cache read/write, total
Model	model ID, provider, agent name
Cost	USD cost per turn
Generation	temperature, topP, topK, maxTokens
Finish	stop, tool-calls, etc.
Request/Response	Full prompt and completion text
MLflow metrics	message_count, tool_count, duration_ms

Setup

1. Place send_trace.py and dist/ in your plugin directory
2. Add to opencode.json:

{
  "plugin": ["C:/path/to/opencode-mlflow-plugin/dist/index.js"]
}

1. Ensure MLflow is running on http://localhost:5000 (or configure trackingUri)

Options

{
  "plugin": [["C:/path/to/dist/index.js", {
    "trackingUri": "http://localhost:5000",
    "experimentName": "opencode-sessions",
    "logSpans": true,
    "logTokens": true,
    "logToolDetails": false
  }]]
}

Option	Default	Description
trackingUri	http://localhost:5000	MLflow tracking server URL
experimentName	opencode-sessions	MLflow experiment name
logSpans	true	Create MLflow traces with spans
logTokens	true	Capture token counts
logToolDetails	false	Log individual tool execution durations as metrics

How it works

• Listens to OpenCode events (chat.message, message.updated, chat.params, tool events)
• Accumulates token counts across parallel assistant messages in a turn
• Uses a 2-second debounce to handle multiple assistant messages per turn
• Sends traces to MLflow via a Python subprocess using the MLflow SDK (synchronous mode)
• Creates LLM-type spans with GenAI semantic convention attributes

Requirements

• Python 3.10+ with mlflow installed (pip install mlflow)
• MLflow tracking server running
• OpenCode with plugin support

License

MIT