CHANGELOG.md

V10

• Added Phase 10 universal model-card builder.
• Generalized the LongCat full-inference gate to arbitrary Hugging Face models.
• Added INFERENCE_CONTRACT.json/TECHNICAL_BLOCKERS.json contract prompts and private Space generation flow.

Changelog

V9

• Added a stricter LongCat full-inference gate.
• full-inference-gated is now the default LongCat implementation mode.
• Pi must either attempt the closest real inference path or produce TECHNICAL_BLOCKERS.json with evidence.
• The prompt now requires investigation of PyTorch SDPA, xformers, and HF Kernels flash-attn alternatives before declaring flash-attn blocking.
• Health-only validation is no longer reported as full inference success. The worker writes inference_gate.json and uses statuses such as technical_blocker, health_only, or full_inference_candidate_health_passed.

V8

• Added robust LongCat Phase 8 behavior.
• Added HTTP /health validation before Gradio Client fallback.
• Added best-effort Space runtime/log diagnostics collection.
• Made hardware requests non-blocking and stop retrying on clear OAuth/auth 401 failures.
• Added recursive workspace upload so Pi-generated packages are not lost.
• Added IMPLEMENTATION_MODE for full-inference-attempt vs safe-scaffold.
• Switched default Pi model for LongCat runs to Qwen/Qwen3-Coder-Next.

V7.3

• Fix Phase 7 worker syntax error in README metadata sanitizer by escaping embedded newlines in the generated worker script.
• Add a compile test for the extracted LongCat worker payload.

v7.1

• Fixed Phase 7 UI launcher import: launch_longcat_article_job is now imported in app.py.
• Updated app description title to V7.

Changelog

v5.2

• Fix Phase 5 generated Space runtime dependency conflict by pinning huggingface_hub>=0.34.0,<1.0.0 in target Space requirements.txt.
• Add a Pi instruction not to remove the huggingface_hub compatibility pin.

V5

• Added Phase 5: model_id → model metadata analysis → Pi-adapted Gradio template → private Space → live API validation.
• Added model_analysis.json in the Bucket run folder.
• Increased default Job timeout to 30 minutes for model Space builds.

Changelog

V3

• Added Phase 3 Pi smoke test.
• Added launch_pi_space_smoke_job in src/jobs.py.
• Added PI_SPACE_WORKER_SCRIPT in src/worker_payload.py.
• Added a new Gradio tab to run Pi in a controlled HF Job.
• Pi modifies a minimal Gradio app before the worker creates a private Space.
• The live Gradio API validation now requires the Pi marker phrase.
• Pi output and traces are copied to the Bucket with best-effort redaction.

V2.3

• Fixed Gradio API endpoint discovery.
• Avoided hardcoded /predict by using client.view_api() and fallbacks.

V2.2

• Confirmed Gradio Client auth compatibility helper.
• Improved stdout logging for worker events.

v4 — Pi gist recipe

• Added Phase 4 tab: Pi reads the HF Spaces Agent Quickstart gist and is asked to use the hf CLI to create/upload a private target Space.
• Added a new pi_gist_recipe worker payload.
• The wrapper still performs independent final validation through the live Gradio API before declaring success.
• Saved artifacts include generated/GOAL.md, optional generated/PI_SUMMARY.md, Pi logs, traces, API schema, and API test result.

v5.1

• Fixed Phase 5 Pi invocation for Pi 0.73.x: use pi -p instead of removed --prompt.
• No architecture changes; Phase 5 remains wrapper-owned for Hub operations and live API validation.

V6

• Added Phase 6 Runtime Recommender.
• Adds a no-build HF Job that analyzes model metadata, estimated file sizes, task/library, risks, and recommends CPU Basic / CPU Upgrade / ZeroGPU candidate / manual review.
• Writes model_analysis.json, runtime_recommendation.json, state.json, events.jsonl, and report.md to the Bucket.

V7 — LongCat article reproduction pass

• Added Phase 7: LongCat article-style reproduction workflow.
• Adds a dedicated HF Job worker that asks Pi to adapt a LongCat Space scaffold using the HF Spaces gist.
• Creates the target Space privately.
• Requests zero-a10g first, with optional fixed GPU fallback (l40sx1 by default).
• Validates a cheap /health endpoint live via gradio_client before marking success.
• Stores hardware attempts, model analysis, generated files, Pi logs, traces, and report in the bucket.

V7.2

• Fixed Phase 7 README metadata upload failures by sanitizing short_description to <= 60 characters before upload.
• Moved Space hardware requests after file upload to avoid newly-created private Space propagation issues.
• Made hardware requests more tolerant by passing hardware strings directly and retrying with backoff.
• Added read-billing to OAuth scopes metadata to help with paid hardware/fallback checks.