Choose a model workflow

This is the umbrella entry point, not a launcher. Start from the repository root. Identify the requested outcome before running anything: cached inputs, working inference, or training are different outcomes with different evidence.

Select the supported path

Common preflight

1. Confirm the operator's goal, model license, repository ID, immutable full revision, complete required-file list, cache path and real reader UID/group. Do not accept a directory or config.json alone as complete model coverage.
2. Confirm the target and authorization for each action. Cache preparation, downloads, package installation, GPU allocation, service startup and inference requests are separate actions. A skill or validator grants none of them.
3. Reuse the existing cache interface: share only HF_HUB_CACHE, use a designated writer and pass the local snapshot path to consumers. Mount the whole cache read-only; keep authentication and framework caches private. Never put training outputs or checkpoints in the shared input cache.
4. Follow the selected skill's failure branches. Stop on absent authorization, incomplete files, space pressure, unknown endpoint ownership or required site authentication. Never remove controls, change infrastructure or delete other users' files to turn a failure into a pass.

Report only the evidence obtained

Record which validator ran, its exit code, schema version, explicit inputs and checks. The cache guide and serving guide define their JSON contracts; the fine-tuning readiness guide separates preparation from training. This umbrella does not invent a combined success field.

A readable cache is preparation only. A completion is inference only. Neither proves that a particular revision was loaded, a GPU was used, or training succeeded. A checkpoint filename by itself is not training validation. Keep startup logs private; do not include credentials or sensitive prompts in reports. No multi-node hooks or implicit fallback to a different workflow are provided.