The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to automate UI rendering and Git operations. Specifically, it uses git fetch, git show, git hash-object, git mktree, and git commit-tree to create ephemeral commits. It also performs a git push --force to specific reference namespaces (refs/pr-screenshots/ and refs/evidence/) to host images for pull request reviews.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection as it ingests untrusted data from the repository to perform visual comparisons.
Ingestion points: SKILL.md (Step 2) reads ticket acceptance criteria and design specifications directly from project files and passes them to a secondary agent context.
Boundary markers: None explicitly defined in the instructions for separating the spec text from agent instructions.
Capability inventory: The skill has access to file system reads, shell command execution (Git and headless Chromium), and network communication via git push to the remote repository.
Sanitization: There is no mention of sanitization or filtering for the external spec text before it is interpolated into the judge agent's context.
[EXTERNAL_DOWNLOADS]: The skill fetches content from the remote Git repository (origin) and generates URLs for raw.githubusercontent.com to embed screenshots in pull requests. These operations target well-known services and are consistent with the skill's primary purpose.

pixel-review