The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PR review comments from CodeRabbit to drive code remediation, creating a surface where malicious instructions in comments could influence the agent's actions.
Ingestion points: Review comments are fetched via GitHub REST and GraphQL APIs in scripts/pr-review.ts and stored in local markdown files.
Boundary markers: The generated markdown files for issues do not include explicit delimiters or 'ignore' instructions for the agent to isolate the untrusted comment body.
Capability inventory: The skill has the capability to write files, run arbitrary build/test commands (pnpm run), execute shell commands, and commit to the git repository.
Sanitization: No sanitization or escaping is performed on the comment text before it is interpolated into the markdown task files.
[COMMAND_EXECUTION]: The skill relies on several shell command executions to perform git operations and interact with the GitHub CLI.
Evidence: SKILL.md contains instructions for the agent to run git commit, pnpm exec, and bash scripts.
Evidence: scripts/pr-review.ts uses child_process.execSync to retrieve repository configuration via git config.
[DYNAMIC_EXECUTION]: The skill uses dynamic script generation and execution for file processing.
Evidence: scripts/resolve_pr_issues.sh executes Python code blocks provided via shell heredocs to perform regex-based updates on the issue tracking files.

fix-coderabbit-review