Huggingface

Hugging Face

Workflow

1. Read STATE.md; reset it from references/template/STATE.template.md for a new run. Classify the request as discovery, local inference, authentication, publishing, training preparation, or troubleshooting.
2. Read only the matching domain reference:
   • references/discovery.md for model, dataset, or Space selection.
   • references/inference.md for local inference.
   • references/auth.md for private or gated access.
   • references/publish.md for uploads.
   • references/troubleshooting.md for errors.
3. Inspect the actual runtime, package versions, hardware, authentication state, and requested asset metadata before proposing or running a command.
4. Execute the smallest reproducible workflow that satisfies the request. Use $download-models for managed long-running model downloads when available.
5. Validate model or dataset identity, task compatibility, revision, output shape or files, and authorization scope as applicable.
6. Record commands, assumptions, artifacts, validation, and unresolved issues in STATE.md.

Rules

• Never print or persist raw access tokens.
• Pin revisions when reproducibility matters.
• Do not install missing dependencies unless the user explicitly requests environment setup; follow the target project's dependency policy.
• Get explicit confirmation before publishing or overwriting remote Hub assets.

Output

Return concise results with exact commands used, verified artifact identifiers, environment assumptions, and the next diagnostic step when unresolved.