mlflow-cli

name	mlflow-cli
description	Use when analyzing remote MLflow tracking servers from the CLI to inspect experiments, runs, metrics, params, tags, and artifacts.

MLflow CLI

• Remote-first: assume MLFLOW_TRACKING_URI points to the target server; if it is missing, ask for the server URL before querying.
• Prefer the installed MLflow --help / subcommand help first; use docs only to fill gaps.
• If mlflow is not available, try the project environment (uv run mlflow --help or .venv/bin/mlflow --help) before giving up.
• If MLflow is still unavailable, stop and report this to the user instead of silently falling back to another API/client.
• Read-only only: use search/list/describe/export/download. Do not create, delete, restore, serve, or deploy.
• Prefer --output json where supported and mlflow experiments csv for comparisons.

Workflow

1. Find the experiment with mlflow experiments search, then confirm it with mlflow experiments get -x <id> or -n <name>.
2. Inspect runs with mlflow runs list --experiment-id <id>; use --view active_only|deleted_only|all when needed.
3. Describe candidate runs with mlflow runs describe --run-id <run_id> to capture params, metrics, tags, artifact roots, and metric history context.
4. Inspect artifacts with mlflow artifacts list -r <run_id> [-a <path>], then download interesting paths with mlflow artifacts download.
5. Export the full run table with mlflow experiments csv -x <id> -o <file> for offline sorting/comparison.
6. If the run is/was launched on Slurm, compare MLflow status with the live job state before concluding.

What to extract

• best metrics and the run IDs that achieved them
• final metrics and the gap to best metrics
• metric history shape: steady rise, plateau, peak+decay, or collapse
• runtime/perf: throughput, step time, GPU util, GPU memory, total time
• run completeness: finished, truncated, or still running
• key hyperparameters and data/version tags
• artifact files that explain behavior (MLmodel, requirements.txt, conda.yaml, plots, checkpoints, logs)
• repeated patterns across runs that suggest what changed

Style

• Summarize findings succinctly.
• Call out missing tracking URI or missing experiment/run IDs immediately.
• If the user wants deeper model execution, stop here and ask before using non-read-only MLflow commands.