[kediwu0331]

sim-bimanual run:

agibot_mixed_028_bulk_seed2026.mp4

sim-transfer runs:

blur.mp4

seg.mp4

edge.mp4

depth.mp4

sim-auto run:

vision_view00_camera_front_wide_120fov.mp4

vision_view06_camera_front_tele_30fov.mp4

vision_view05_camera_cross_left_120fov.mp4

vision_view04_camera_rear_left_70fov.mp4

vision_view03_camera_rear_tele_30fov.mp4

vision_view02_camera_rear_right_70fov.mp4

vision_view01_camera_cross_right_120fov.mp4

transfer-auto wsm:

vision_view00_camera_front_wide_120fov.mp4

vision_view06_camera_front_tele_30fov.mp4

vision_view05_camera_cross_left_120fov.mp4

vision_view04_camera_rear_left_70fov.mp4

vision_view03_camera_rear_tele_30fov.mp4

vision_view02_camera_rear_right_70fov.mp4

vision_view01_camera_cross_right_120fov.mp4

transfer-auto lidar:

transfer_auto_lidar.mp4

lidar_info: {'shape': [67, 128, 1800], 'dtypes': {'range_m': 'F32', 'intensity': 'F16', 'validity': 'U8'}, 'metadata': {'fps': '10.0', 'layout': '[sweeps, beams, azimuth]', 'min_range_m': '5.0', 'max_range_m': '100.0', 'invalid_range_m': '0.0'}}

[pengcuo]

Reviewed the helpers, tests and the four notebooks at cell level. The helper changes are solid and well covered (the MockTransport tests in particular), but I cannot call this mergeable yet:

1. No description. Please add: why these notebooks were removed before and what changed (I assume sgl-project/sglang#39475, merged 2026-10-18), which lmsysorg/sglang:dev image digest or commit the four notebooks were run against, and end-to-end evidence per notebook (Auto WSM, Auto joint LiDAR including the rangemap validation cell, Sim-Auto, Bimanual, Sim-Transfer). The second commit documents multi-GPU Ulysses flags; were those run?
2. Moving image tag. All four notebooks tell readers to docker pull lmsysorg/sglang:dev "before each start" because "the tag moves". A cookbook that cannot be reproduced next month is a support burden. Please pin the first release tag (or a digest) that contains #39475 as soon as one exists, and record the date and commit in the version table until then. The vLLM-Omni notebooks at least pin a dedicated vllm/vllm-omni:cosmos3 tag.
3. CI coverage. validate-cookbooks.yml only lints and tests under simulation/auto. The new bimanual/bimanual_helpers.py (354 lines of request and validation logic) has no lint or unit test in CI; the notebook lint step covers notebooks only. Add it to a ruff step at minimum, ideally with a small unit test for prepare_bimanual_input / encode_extra_params like the Auto helpers have.
4. CI for this fork PR has not run yet; a maintainer needs to approve the workflow before merge.

Non-blocking: the Sim-Auto SGLang notebook re-implements submit_video_job / download_video / probe_video in its "client helpers" cell while this PR adds _submit_video_job / _download to auto_helpers.py; importing them would remove one copy. The vLLM-Omni notebooks carry the same cell, so this is existing convention. Smaller things inline.

[kediwu0331]

Reviewed the helpers, tests and the four notebooks at cell level. The helper changes are solid and well covered (the MockTransport tests in particular), but I cannot call this mergeable yet:

1. No description. Please add: why these notebooks were removed before and what changed (I assume sgl-project/sglang#39475, merged 2026-10-18), which lmsysorg/sglang:dev image digest or commit the four notebooks were run against, and end-to-end evidence per notebook (Auto WSM, Auto joint LiDAR including the rangemap validation cell, Sim-Auto, Bimanual, Sim-Transfer). The second commit documents multi-GPU Ulysses flags; were those run?
2. Moving image tag. All four notebooks tell readers to docker pull lmsysorg/sglang:dev "before each start" because "the tag moves". A cookbook that cannot be reproduced next month is a support burden. Please pin the first release tag (or a digest) that contains #39475 as soon as one exists, and record the date and commit in the version table until then. The vLLM-Omni notebooks at least pin a dedicated vllm/vllm-omni:cosmos3 tag.
3. CI coverage. validate-cookbooks.yml only lints and tests under simulation/auto. The new bimanual/bimanual_helpers.py (354 lines of request and validation logic) has no lint or unit test in CI; the notebook lint step covers notebooks only. Add it to a ruff step at minimum, ideally with a small unit test for prepare_bimanual_input / encode_extra_params like the Auto helpers have.
4. CI for this fork PR has not run yet; a maintainer needs to approve the workflow before merge.

Non-blocking: the Sim-Auto SGLang notebook re-implements submit_video_job / download_video / probe_video in its "client helpers" cell while this PR adds _submit_video_job / _download to auto_helpers.py; importing them would remove one copy. The vLLM-Omni notebooks carry the same cell, so this is existing convention. Smaller things inline.

1-3 are addressed. ulysses 4 is fully tested for auto notebooks