The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill instructions include persisting state data to .rcs/state/{scope}/forge-progress.json. This is a standard mechanism for maintaining state across agent iterations within a development environment and does not involve sending data to external or untrusted destinations.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external image data (reference_images and generated_screenshot). There is a theoretical risk of indirect prompt injection if these images contain adversarial text.
Ingestion points: reference_images and generated_screenshot arrays/variables.
Boundary markers: The skill enforces a strict JSON output contract, which acts as a structural constraint against most prompt injection attempts.
Capability inventory: The skill performs visual analysis and local file writes to .rcs/; it does not have network access or arbitrary command execution capabilities defined.
Sanitization: None explicitly mentioned for image content, though the model is constrained by the JSON schema.

visual-verdict