edit-sanitizer

Edit Sanitizer

Edit Sanitizer is the fail-close boundary between user feedback/annotations and image-edit execution. It exists because generative image editors may reinterpret more of a source image than the user intended. Do not send a local-edit request directly to prompt-compiler until this Skill has produced a ready EditContract.

Core rule

Convert feedback into the smallest explicit mutation that can satisfy the request.

Do not treat a scribble, arrow, selection, bounding box, or phrase such as fix this as self-explanatory when more than one plausible target exists.

Workflow

1. Identify the approved source checkpoint
2. Resolve annotation coordinates against the source dimensions
3. Map the annotation to one semantic target
4. Split user feedback into atomic requested mutations
5. Reject or veto unrelated global restyling when the request says only, just, this area, or otherwise implies a bounded edit
6. Lock crop, canvas, perspective, identities, exact text, lighting, and unaffected composition unless the user explicitly targets one of them
7. Derive protected regions from the complement of the editable target plus semantic locks
8. Set a conservative mutation budget
9. Require Arabic review for Arabic copy corrections
10. Produce a typed EditContract
11. Only when status is ready, forward the contract to prompt-compiler

Annotation handling

Treat annotation geometry and semantic intent as separate evidence:

• bbox: validate positive area and source bounds
• polygon: require a valid polygon around the intended area
• mask_ref: require an actual mask reference
• semantic: require one clearly identifiable object or text region
• normalized: convert coordinates to native pixels before execution

If two plausible targets have similar confidence, return clarify rather than guessing.

Drift prevention

Every retry starts from source_checkpoint, which must be the last approved source.

Never chain a second corrective edit from a failed or visibly drifted render.

Maximum edit loop: 3 attempts. If the same defect persists or collateral changes grow, stop and return the evidence rather than continuing destructive edits.

Bounded-edit vetoes

Veto execution when:

• a local edit also asks to restyle the whole image
• there are multiple unrelated mutations that should be separate edits
• the target is outside the source or has zero area
• the annotation cannot be mapped confidently
• exact replacement copy is missing for copy correction
• the retry source is not the approved checkpoint

Preservation language

Do not promise literal pixel identity from a generative editor. Require material stability outside the edit target and validate collateral changes after execution.

Output

Return the EditContract plus a short execution verdict:

• ready: may proceed to Prompt Compiler
• clarify: ask one target/scope question, do not execute
• reject: invalid geometry/source/copy contract, do not execute
• veto: conflicting or over-broad mutation request, do not execute

Runtime helper

python3 ../../shared/scripts/sanitize_edit.py < edit-request.json

---

Cross-Skill Neural Connections & References

Peer & Downstream Skills

• Prompt Compiler — Downstream consumer of sanitized EditContracts
• Manipulation Director — Boundary blending physics & light wrap
• Arabic RTL Director — Text ligature verification on Arabic copy edits
• Visual QA — Target accuracy & collateral drift verification
• Designly Director — Lead orchestrator and edit state manager

Schemas & References

• Edit Contract Schema — Bounded edit schema
• Image Editing Guidance — Editing rules
• Routing Graph — Mesh topology & feedback loops