ethan-li-k/astra_bench/collect-pens-bicoord

collect_pens_bicoord (long_horizon)

End-to-end VLM-driven pipeline (no expert_replay anywhere), run through the 3-role triangle (LHPlanner → LHExecutor → LHReviewer):

1. LHPlanner — decomposes into ordered pick-and-place atomics (one per pen), pairing each pen with the closer arm.
2. LHExecutor — for each atomic, runs the active executor (zeroshot rollout VLM tool loop, or trained policy once eval flips) on the shared env with a per-atomic Reviewer judging success.
3. LHReviewer — issues the overall verdict and optionally proposes a base-skill fix.
4. posttrain/ — skeleton; once we have N successful long-horizon traces, replay into RL fine-tuning.

Crucially: every success bit (per-atomic and overall) is decided by a fresh Claude subprocess looking at images — no sim-privileged state.