v0.9.0 Model Lab MVP cutline: HF provider polish, search, passports, profiles

[Hmbown]

Problem

Hugging Face should be first-class in CodeWhale, but the v0.9.0 Model Lab/HF backlog needs an ordered MVP. Without a cutline, agents may jump to Spaces/Jobs before provider correctness, search, and model metadata are usable.

MVP order

1. Provider polish and drift check — v0.9.0 Hugging Face provider polish: route validation, aliases, docs drift #2706
   • route validation
   • aliases/env vars
   • docs/config consistency
   • no secrets in status/errors
2. Direct HF search and picker — v0.9.0 Direct Hugging Face search commands and picker flow #2712
   • /hf search or equivalent
   • model/dataset/space filters
   • graceful offline/cache behavior
3. Model passport metadata — v0.9.0 Hugging Face Hub browser and model passport metadata #2707
   • model card basics
   • license/gated/private status
   • tags/tasks/context/provider availability
   • last updated/download/star where available
4. HF HarnessProfiles — v0.9.0 Hugging Face HarnessProfiles: model-family profiles across HF routes #2711
   • model-family profile matching across HF routes
   • open-weight/local model posture defaults
5. HF MCP/workset integration — v0.9.0 Hugging Face MCP and Hub workset integration #2709
6. Spaces/Jobs workflows — v0.9.0 Hugging Face Spaces and Jobs workflows for demos/evals/experiments #2710, after the basics are stable

Acceptance criteria

• CodeWhale can search/browse HF models without switching the active provider or downloading anything implicitly.
• A selected HF model can show a concise passport in the TUI/CLI.
• HF provider route remains separate from model id and HarnessProfile.
• Gated/private models are clearly marked and never attempted without user intent.
• Docs explain when to use Hugging Face router vs OpenRouter/local vLLM/SGLang/Ollama.
• The issue links implementation PRs back to v0.9.0 EPIC: Make Hugging Face a first-class CodeWhale surface #2705.

Non-goals

• No automatic model downloads in the MVP.
• No silent uploads to HF.
• No Spaces/Jobs orchestration until search/passports/provider polish are working.

[Hmbown]

Status update: #2853 (caa1d4a99) records the v0.9 cutline: ship HF provider/MCP/concepts docs foundation only, defer native Model Lab/HF search/passports/Spaces/Jobs/eval/export runtime. It also corrected docs/MODEL_LAB.md so it no longer says the HF provider route is unimplemented; the deferred part is native Hub browser/search/passports/workflows/UI.

Credits preserved from harvested/audited work: @hongqitai (#2780/#2799 env aliases) and @idling11 (#2782/#2802 HF MCP/concepts).