The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts visual-parity-diff.mjs and visual-parity-diff.eval.mjs utilize createRequire and require.resolve to load the pngjs and pixelmatch libraries. The paths for these libraries are computed at runtime based on the REPO_ROOT environment variable or the current working directory, which allows for dynamic loading of code from potentially untrusted directory structures.
[COMMAND_EXECUTION]: The test harness visual-parity-diff.eval.mjs uses node:child_process to spawn the main script. While this is standard for automated testing, it involves executing shell commands where the script path is derived from the local environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes external PNG images to provide a 'less wrong' verdict, creating a vulnerability surface for adversarial visual content.
Ingestion points: External PNG files read from directories defined by REFERENCE_DIR and CANDIDATE_DIR environment variables.
Boundary markers: The subagent instructions in references/subagent-visual-review.md use neutral labeling (Image A vs Image B) but do not include specific warnings or delimiters to prevent the agent from being influenced by instructions embedded within the images themselves.
Capability inventory: The skill can perform file system read/write operations to generate artifacts and can execute local Node.js scripts.
Sanitization: Content is parsed using the pngjs library, but there is no mechanism to sanitize or validate the visual content against prompt injection before the agent evaluates it.

compare-screenshots