The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents the execution of a local Python script (../../shared/scripts/sanitize_edit.py) to validate edit requests. This script is an internal functional component used to process annotation data and feedback before generating an edit contract.
[INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted user data, such as image annotations and text-based feedback, which is an inherent injection surface. The skill is specifically designed to mitigate this risk through a defined evidence chain:
Ingestion points: The skill processes user-provided geometry (bounding boxes, polygons, masks) and semantic text feedback via edit-request.json.
Boundary markers: It employs a structured EditContract and a multi-state verdict system (ready, clarify, reject, veto) to explicitly control the flow to downstream tools.
Capability inventory: The skill's capabilities are restricted to local coordinate resolution, semantic mapping, and script-based validation.
Sanitization: It enforces strict rules such as material stability outside target regions, mutation budgets, and mandatory vetoes for global restyling requests or ambiguous targets.

edit-sanitizer