The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands using Python's subprocess module to interact with development tools.
Scripts such as pr_review_github_transport.py, pr_review_gitlab_transport.py, and pr_review_subprocess.py execute gh, glab, and git binaries.
Command execution is performed via subprocess.Popen using argument lists to avoid shell interpolation, which is a defensive practice.
Includes utility logic in pr_review_subprocess.py to terminate process trees, specifically using the Windows taskkill utility when running on that platform.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and act upon data from external sources (reviewer comments), which creates an indirect injection surface.
Ingestion points: Untrusted data enters the agent context through pr_review_gh_transport.py and pr_review_gitlab_transport.py when fetching PR threads, review bodies, and merge request notes.
Boundary markers: While the skill uses structured JSON snapshots and Pydantic models for data management, the agent is instructed in SKILL.md to assess the "complete evidence set" of human and bot inputs without explicit textual delimiters or "ignore embedded instructions" warnings for the comment bodies themselves.
Capability inventory: The skill possesses significant capabilities, including repository mutation (replies, resolutions) via CLI tools and the ability to drive source code edits based on its assessment of reviews.
Sanitization: The skill performs rigorous schema validation using Pydantic, but does not implement semantic filtering or sanitization of the natural language content within the review comments before processing.

receiving-pr-reviews