HOI Object Reconstruction Doctor

Inspect evidence first and ask later. Work from reconstruction/, preserve the job directory, and diagnose the first bad stage instead of rebuilding everything.

Discover context automatically

1. Read the exact command and error from the conversation or available log.
2. If no job directory is named, inspect the newest directory under data/outputs/hoi_recon/ and other output paths referenced by recent commands.
3. Infer the mode from the command and artifacts: merged_recon/ means BundleSDF; sam3d/ means SAM3D.
4. Infer the input from the recorded command or job metadata.
5. Ask one focused question only if neither logs nor a job directory can be discovered.

Run the fast diagnostic set

git rev-parse --short HEAD
python --version
docker version
nvidia-smi
df -h .
docker ps -a --no-trunc
python -c 'from v2d_hoi_object_reconstruction.docker.run_reconstruction import main'

When the input is known:

python ../.claude/skills/hoi-object-reconstruction-setup/scripts/preflight_input.py \
  <absolute-mapping-data-dir>

Inspect the job tree, nonempty files, timestamps, and the last relevant log. Separate observed evidence, root cause, repair, and resume command.

Classify the first failure

Evidence	Classification	Narrow repair
Docker unavailable or GPU invisible in a container	setup/runtime	Repair Docker or NVIDIA Container Toolkit; re-run the GPU probe
missing image or weight directory	setup	Use hoi-object-reconstruction-setup for the selected mode only
preflight errors	input contract	Fix the named metadata, calibration, synchronization, or file path issue
CuSFM scan-quality gate rejects BundleSDF input	capture/input	Do not bypass it; inspect whether the stationary-rotate-stationary scan pattern exists
no kernel image is available	image/GPU compatibility	Rebuild for that architecture or use a validated GPU; do not blame the dataset
libEGL.so.1 or offscreen renderer failure	SAM3D image	Rebuild v2d_sam3d, run its EGL probe, then resume rendering/selection
Stage-1 BundleSDF mesh is already bad	masks/depth/CuSFM/Stage-1	Inspect masks, trajectory, depth range, and Stage-1 extraction first
Stage-1 is good but merged mesh is bad	FoundationPose/alignment	Inspect tracking overlays and world poses before rerunning merged reconstruction
SAM3D candidate is bad	frame/mask/inference	Inspect source frame and mask, then SRT result and render overlay
wrapper exited but final review files are absent	incomplete pipeline	Resume from the first missing artifact; do not mark complete

Detect a stall correctly

Check the wrapper PID, active containers, nvidia-smi, disk growth, and artifact timestamps. A quiet log alone is not a stall. Report which signal stopped changing and for how long.

Repair and re-verify

Prefer the smallest reversible repair. Do not delete partial outputs, weaken quality gates, reset unrelated state, or force all stages to rerun. After a repair:

1. Re-run the failed diagnostic probe.
2. Resume with the original input, prompt, mode, GPU, and job directory.
3. Verify the first formerly failing stage produces its complete artifact set.
4. Return to the hoi-object-reconstruction-run skill for completion and visual QA.

If the evidence shows a product defect rather than an environment or data problem, preserve the reproducer and report the source revision, exact command, first bad artifact, logs, hardware, and expected versus observed behavior.