The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill fetches and applies code changes from external GitHub Pull Requests using gh pr diff and git apply, subsequently executing the resulting code via yarn vitest. This workflow allows for the execution of arbitrary, potentially malicious code submitted by untrusted contributors.
[DYNAMIC_EXECUTION]: The skill programmatically generates a temporary Vitest specification file (_compat-review-{fn}.spec.ts) at runtime and executes it. This dynamic script generation and execution pattern is used to compare library behaviors.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted data from GitHub PR bodies to extract test cases and verification claims without sanitization or boundary markers.
Ingestion points: GitHub PR metadata and body via gh pr view (SKILL.md, Workflow Step 1).
Boundary markers: None identified; the agent is instructed to directly extract logic and test cases from the raw PR description.
Capability inventory: Bash (used for GitHub CLI, Git, Yarn, and file deletion), Write (used to create the test script), Grep, and Glob.
Sanitization: None; the skill does not filter or validate the content extracted from the PR body before using it to guide its actions.
[COMMAND_EXECUTION]: The skill uses Bash to execute multiple powerful CLI tools, including git, gh, and yarn, to modify the local environment and run tests based on external inputs.

compat-review