diagnose-camera-stack

Diagnose Camera Stack

Use this skill for camera-only failures. Keep the diagnosis read-only until the user explicitly asks to change a device setting.

Workflow

1. Establish the exact symptom and whether it reproduces with hardware disconnected. Record the config path, camera name, VID:PID, expected resolution/FPS, and the first error line.
2. Run scripts/camera_preflight.sh [config] from the repository root. Treat missing devices, permissions, SDK libraries, and model binaries as separate findings.
3. Confirm the build before touching hardware:
   • cmake --build build-ninja --target camera_test -j$(nproc)
   • cmake --build build-ninja --target camera_detect_test -j$(nproc)
   • ./build-ninja/camera_test -h
4. Exercise the smallest relevant executable. camera_test proves acquisition; camera_detect_test adds the detector. Do not jump to infantry until both are understood.
5. Compare observed frame timestamps/FPS and image dimensions with exposure_ms, gain or gamma, frame_rate, camera_name, and vid_pid. For HikRobot also check /opt/MVS/lib/64 and LD_LIBRARY_PATH; MindVision uses the repository SDK library.
6. If acquisition is healthy, hand off algorithm symptoms to $replay-auto-aim. If the failure is a config/model mismatch, hand off to $audit-vision-config.

Evidence and stop conditions

• Capture command, exit status, stderr, device listing, and the first failing layer. Redact machine-specific serials and private paths when reporting.
• A successful process start is not proof of streaming. Require non-empty frames and a measured rate.
• Stop before changing exposure, gain, camera firmware, udev rules, or SDK installation. Ask the user before any privileged or hardware-mutating action.
• GUI tests need a display. Report DISPLAY/headless limitations instead of substituting an unverified result.

Read camera-matrix.md for repository-specific facts and expected commands.