pi-wandb

Weights & Biases run tracking and metric history tools for pi

Package details

extension

Install pi-wandb from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:pi-wandb

Pi manifest JSON

{
  "extensions": [
    "./index.ts"
  ]
}

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

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

pi install npm:pi-wandb

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

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)

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)

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)