# Next Steps

## Current status

- Phase 1 validates OAuth → Job → Bucket.
- Phase 2 validates private Space creation/upload → live API validation.
- Phase 3 validates Pi inside an HF Job editing `app.py` before upload.
- Phase 4 validates a closer article-style workflow: Pi receives the HF Spaces Agent Quickstart gist, uses `hf` CLI, creates/uploads a private Space, and the wrapper independently validates the live API.

## Next pass: Phase 5

Add a narrow model-card workflow:

1. User provides a model ID.
2. Worker fetches `model_info`, README, pipeline tag, library name, and sibling files.
3. Worker selects one of a few safe templates.
4. Pi applies the same gist-style loop to create a private Space.
5. Wrapper validates the live API.

Recommended first model families:

- simple Transformers text generation / text classification
- ASR with `pipeline("automatic-speech-recognition")`
- Diffusers text-to-image only after the CPU/simple path is stable

Keep these out of scope for Phase 5:

- Docker Spaces
- multi-GPU
- fine-tuning
- ComfyUI
- video generation
- public publication automation


## Phase 6

Adds a no-build runtime recommender Job that analyzes model metadata and writes `runtime_recommendation.json` to the Bucket.


## Phase 7 — LongCat article reproduction

Phase 7 attempts an article-style LongCat Space build: private target Space, Pi-guided app adaptation, ZeroGPU first, fixed GPU fallback when explicitly enabled, and live `/health` API validation. Full video generation remains a manual-review step until model-specific runtime validation is complete.


## V8 follow-up

- Run the LongCat Phase 8 once with `Qwen/Qwen3-Coder-Next`; if the provider/model is unavailable, switch back to `moonshotai/Kimi-K2.5`.
- If hardware request returns 401, set hardware manually in the Space settings and use the report/logs to continue debugging.
- Inspect `runs/<run_id>/logs/space_logs_runtime.txt`, `space_logs_build.txt`, `space_runtime.json`, and `tests/http_health.json`.
- Decide whether the next iteration should build a real full inference Space or formalize a safe scaffold mode.