The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes documentation files which are considered untrusted data and interprets content within them as executable claims or commands.
Ingestion points: The agent reads documentation pages (Markdown files) to extract file paths, commands, config keys, and versions in the 'Check each page against its system' step.
Boundary markers: There are no specific delimiters or instruction-ignore markers used when the agent reads and processes the content of the documentation files.
Capability inventory: The skill is authorized to run shell commands, including python3 scripts, git commands, and arbitrary commands extracted from the docs themselves (e.g., run the command with --help).
Sanitization: The skill includes instructional constraints such as 'run no command that changes state' and 'run the command with --help', but these rely on the agent's interpretation rather than technical sanitization or sandboxing.
[DYNAMIC_EXECUTION]: The skill features an 'Overlay' mechanism that dynamically loads additional instructions from a path defined at runtime.
Evidence: The skill checks the $POWERSUIT_OVERLAY environment variable and reads files from $POWERSUIT_OVERLAY/docs/ and docs/review-sweep/ within that path.
Logic Override: The instructions explicitly state, 'Where the overlay and this file disagree, follow the overlay,' allowing instructions from a computed path to supersede the original safety and operational rules.
[COMMAND_EXECUTION]: The skill routinely executes local scripts and shell commands to perform its tasks.
Evidence: It executes python3 scripts/build_docs.py --stale, python3 scripts/build_docs.py --check, and python3 .github/actions/gate/run.py --dry-run. It also uses git log to identify potential page owners.

review-sweep