What problem does it solve? Keeping Pi's wandb provider configuration in models.json aligned with the models actually served by W&B Serverless Inference is manual and error-prone: model IDs change, context windows and modalities are documented separately, and stale entries silently persist. This Skill automates fetching live endpoint IDs and documented catalog metadata, reconciling the configuration, and validating the result. ## Core Features & Use Cases - Live catalog fetch: A Python helper queries the authenticated W&B models endpoint and parses the W&B documentation tables into a snapshot JSON with context windows, modalities, and display names. - Deterministic reconciliation: Rules for adding, removing, and updating models.json entries, with strict guards against inventing metadata or leaking API keys. - Model recommendations: Compares W&B candidates against preferred models on coding specialization, reasoning, modality, and context window, updating SUITABLE_MODELS.md. - Validation workflow: jq-based checks confirm exact endpoint coverage, explicit context windows, and Pi model loading. - Use Case: When W&B adds a new inference model, run the Skill to fetch the catalog, update providers.wandb in models.json with documented context windows, and verify with pi --list-models wandb. ## Quick Start Ask the agent to sync the wandb provider in models.json with the current W&B Serverless Inference catalog and validate the result.
