Labali Image Desensitize

labali-image-desensitize

Treat this skill as a layered system.

Layer Contract

1. SKILL.md is the policy layer.
2. references/architecture.md is the strategy layer.
3. scripts/sanitize-image.sh is the execution layer.

NEVER

• Never overwrite the input file — always write to a separate output path.
• Never report PASS on the sensitive-info scan without completing Step 4.
• Never skip Step 3 diagnostics — before/after comparison is required output regardless of result.
• Never describe this skill as OCR, visible-content review, screenshot redaction, or pixel-level masking.

Pipeline Steps

Execute in fixed order:

If the intent or design rationale of any pipeline step is unclear, load references/plan.md before proceeding.

Step 1 — Validate prerequisites

• Confirm magick and exiftool are available in PATH.
• Fail fast with explicit error if either is missing.
• Confirm input image path is readable.

Step 2 — ImageMagick re-encode

• Run scripts/sanitize-image.sh with input and output paths.
• Apply -resize 99% -resize 101% -strip -quality 90.
• Require only two runtime inputs: input image path and output image path.

Step 3 — Diagnostics

• Print before/after comparison:
  • file size,
  • dimensions,
  • EXIF line count,
  • SHA256 hash,
  • EXIF diff.

Step 4 — Reviewer: sensitive metadata scan

• Run post-sanitize hidden metadata scan on output image via scripts/check-sensitive-info.sh.
• Check EXIF/XMP/IPTC metadata for sensitive key/value patterns.
• Treat the scan as metadata-only:
  • do not infer anything from visible UI text, buttons, usernames, or other pixels,
  • do not present this skill as a substitute for blur, mosaic, crop, or mask workflows.
• Load references/model-review.md for review rubric.
• Print Sensitive info review: PASS|REVIEW_REQUIRED.
• Print MODEL_REVIEW_SUMMARY + MODEL_REVIEW_JSON payload.
• Apply in-agent model judgment (no API key required):
  • read payload from script output,
  • decide PASS|REVIEW_REQUIRED|BLOCK,
  • report verdict with reason and evidence.

Step 5 — Return result

• In normal mode: report verdict and summary.
• In --strict mode: exit non-zero when sensitive metadata candidates are detected.

Success Criteria

A run is successful only when all conditions hold:

1. Input image is readable.
2. Output image is generated at target path.
3. Before/after diagnostics are printed.
4. EXIF diff section is printed (empty diff is acceptable).
5. Post-sanitize sensitive metadata scan section is printed.
6. Model review verdict section is printed.
7. In strict mode, suspicious hidden metadata produces non-zero exit.

Runtime Inputs

Use skill.yaml as input schema source of truth.

Operational Mode

• Deterministic local shell execution only.
• No browser, no network, no API.
• Keep output reproducible and script-driven.
• Metadata-only scope: no OCR, no visible-content analysis, no pixel redaction.

Resources

• Architecture: references/architecture.md
• Workflow plan: references/plan.md
• Model review rubric: references/model-review.md
• Deterministic script: scripts/sanitize-image.sh
• Post-check script: scripts/check-sensitive-info.sh
• Optional CLI wrapper: scripts/run.ts
• Regression checks: tests/test_regression.sh