Set up Cursor Cloud dev environment

[ajhamdi]

Summary

Sets up the development environment for colon-bench-eval and documents non-obvious caveats for future cloud agents.

• Adds AGENTS.md with a ## Cursor Cloud specific instructions section covering: uv-managed install, the essential git lfs pull step, CPU-only torch / skipped EdgeTAM CUDA extension, required secrets (HF_TOKEN, OPENROUTER_API_KEY), how to run the offline plot scripts and the Streamlit viewer, and the absence of first-party lint/test/CI.

No application code is modified.

Environment setup

• Install: uv sync --extra all (CPU torch, EdgeTAM editable build; CUDA extension skipped gracefully — no GPU).
• Data: git lfs pull — several files (edgetam/checkpoints/edgetam.pt, data/colon-bench/*_seg_*.json, benchmark_detection.json, benchmark_segmentation.json) are Git LFS pointers and must be materialized or segmentation plots/eval fail.

Verification

Offline plot scripts (core functionality) reproduce the published leaderboard figures from the shipped result JSONs:

Segmentation benchmark plot

Streamlit viewer runs end-to-end (downloads colonoscopy clips from the gated HF dataset, plays them, and grades VQA answers):

colon_bench_viewer_vqa_demo.mp4

VQA answer feedback (red wrong / green correct)

Testing

• uv sync --extra all + git lfs pull
• uv run python scripts/plot_vqa_accuracy.py / plot_classification_metrics.py / plot_detection_metrics.py / plot_segmentation_metrics.py
• uv run --extra viewer streamlit run viewer/visualize_benchmark.py ... --mode vqa_prompted (manual VQA flow)

_{To show artifacts inline, enable in settings.}