pi-wandb

Weights & Biases integration for pi.

Registers three tools the LLM can call to inspect experiment runs, metrics, and training curves.

Install

pi install npm:pi-wandb

Configuration

Set your W&B API key in your shell:

export WANDB_API_KEY=your-api-key

Optionally set a custom API endpoint:

export WANDB_BASE_URL=https://api.wandb.ai  # default

Tools

wandb_runs

List runs from a project with optional filters and sorting.

| Parameter | Description | |---|---| | entity | W&B entity (username or team) | | project | Project name | | filters | JSON filter string, e.g. '{"state":"finished"}' or '{"tags":{"$in":["baseline"]}}' | | order | Sort order, e.g. +created_at, -summary_metrics.loss | | max_results | Max runs to return (default 20, max 100) |

wandb_run

Get full details of a specific run including config, summary metrics, notes, and system metrics.

| Parameter | Description | |---|---| | entity | W&B entity | | project | Project name | | run_id | Run ID (the short alphanumeric ID, not display name) |

wandb_history

Get sampled metric history (training curves) for a run.

| Parameter | Description | |---|---| | entity | W&B entity | | project | Project name | | run_id | Run ID | | samples | Number of sampled datapoints (default 500, max 10000) |

Examples

• "List my recent runs in project 'diffusion-policy'"
• "Show me the config and final metrics for run abc123"
• "Get the loss curve for that run"
• "Compare finished runs tagged 'baseline' in myteam/robotics"