audit-vision-config

Audit Vision Config

Use this skill before changing a threshold, model, device, or calibration number. Report facts and mismatches; do not silently rewrite the user's config.

Workflow

1. Identify the exact YAML file and the executable that consumes it. Run from the repository root because model and demo paths are relative.
2. Compare the YAML against the actual reads in code, not just the README. Search tools::read, yaml[, and tools::load in the relevant module.
3. Check file existence and compatibility for classify_model and yolov5_model_path; ensure detector_name selects the intended detector and device is available.
4. Validate camera fields (camera_name, vid_pid, exposure, gain/gamma, frame rate) and gimbal fields (com_port, optional CRC setting) without opening hardware. Hand hardware failures to $diagnose-camera-stack or $diagnose-gimbal-link.
5. Validate calibration structure and units: 3x3 camera_matrix, distortion vector, 3x3 rotations, 3-vector translation, and consistent coordinate/frame conventions. Check rotation orthonormality and determinant near +1; flag implausible values rather than auto-correcting them.
6. Check that OpenVINO lookup paths are consistent in root CMakeLists.txt, tasks/auto_aim/CMakeLists.txt, and tasks/omniperception/CMakeLists.txt.
7. Produce a report with missing, unused, conditionally required, and suspicious value categories. A syntactically valid YAML file is not a valid deployment configuration.

Boundaries

• Never overwrite configs/standard3.yaml or calibration values during an audit.
• Use a copied temporary config for CPU fallback or an experiment, and identify every changed key.
• Do not call a calibration result correct solely because reprojection runs; preserve the measured error and coordinate-frame assumptions.

Read config-map.md for project-specific keys and checks.