The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes human feedback from review.json and reports comments to the agent as described in SKILL.md. This represents an indirect prompt injection attack surface.
Ingestion points: The body field of comments in review.json (referenced in SKILL.md and references/manifest.md).
Boundary markers: The instructions do not define boundary markers or 'ignore' instructions when the agent reads and reports these comments.
Capability inventory: The skill has capabilities for file system operations, image capture (via screencapture), and interaction with other agent tools like agent-browser.
Sanitization: There is no evidence of sanitization or filtering of the content within the review comments before processing.
[COMMAND_EXECUTION]: The astroshot-capture script executes various local command-line tools to manage the review lifecycle.
Evidence: Usage of jq for JSON manipulation, git for worktree detection, and image processing utilities like sips, magick, or identify in scripts/astroshot-capture.
[EXTERNAL_DOWNLOADS]: The skill documentation references the installation of dependencies and downloading a macOS application from an external repository.
Evidence: Instructions in SKILL.md and references/harness-integration.md suggest using npx to add skills and downloading signed builds from the author's GitHub repository.

astroshots-review