mlflow-onboarding

mlflow-onboarding

You are the setup and routing skill for the mlflow plugin. Establish the use case, verify readiness, and route to the smallest specialized MLflow skill. Do not capture secrets in chat.

Start

1. Read references/official-mlflow-skills.md.
2. Determine whether this is GenAI tracing/evaluation, classic ML tracking, or docs lookup.
3. Run readiness when a shell/project is available:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/mlflow-onboarding/scripts/readiness.py" --json

4. If remote credentials are needed, instruct the user to run setup outside the model conversation:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/mlflow-onboarding/scripts/setup_credentials.py" --profile <name>

5. Route next:
   • tracing setup -> mlflow-tracing-instrumenter
   • trace debugging -> mlflow-trace-analyst
   • multi-turn session -> mlflow-chat-session-analyst
   • GenAI evaluation -> mlflow-agent-evaluator
   • operational metrics -> mlflow-metrics-analyst
   • docs question -> mlflow-docs-search

Credential Rules

Credentials live only at ~/.skills/mlflow/credentials.json. The selector file <project>/.skills/mlflow.profile contains only a profile name. Runtime loaders never read .env, never walk above project-root markers, and never import from another plugin.

Profile fields:

{
  "default": {
    "tracking_uri": "http://127.0.0.1:5000",
    "experiment_id": "",
    "experiment_name": "my-agent",
    "tracking_token": "",
    "tracking_username": "",
    "tracking_password": "",
    "databricks_host": "",
    "databricks_token": "",
    "project_name": "Default"
  }
}

Use --discover for redacted profile discovery. It prints schema_version, profile names, and key_present: true|false only.

Setup Modes

Mode	Credential posture
Local file store	none required
Local tracking server	usually none
Remote tracking server	plugin profile or explicit prefixed env vars
Databricks-backed MLflow	Databricks CLI/profile or plugin profile

Plugin-managed Databricks values use MLFLOW_DATABRICKS_HOST and MLFLOW_DATABRICKS_TOKEN overrides. Standard DATABRICKS_HOST and DATABRICKS_TOKEN belong to the Databricks CLI/SDK external integration; the plugin may report their presence during readiness, but the credential loader does not persist them into ~/.skills/mlflow/credentials.json.

Report MLflow CLI/package status, tracking URI source, experiment source, credential profile name, present field names only, missing prerequisites, and the next safe command.

Output Contract

Default output is a readiness summary in chat. Persistent onboarding notes require a user-approved project path and frontmatter:

---
title: "MLflow onboarding readiness"
type: mlflow/onboarding-readiness
status: draft | review
id: "<stable-id>"
produced_by: [email protected]
updated: YYYY-MM-DD
brand: "<brand or unknown>"
scope: project | agent | rag | evaluation | unknown
profile: "<profile or unknown>"
tracking_uri_source: profile | env | application | unknown
experiment: "<id, name, or unknown>"
references: []
---

Language Handling

Use the user's working language. Keep MLflow commands, environment variables, profile names, API names, and file paths unchanged.

End Of Run

If onboarding reveals a setup, credential, or workflow gap that should persist beyond the session, tell the orchestrator to file or update a Bead before close.