pr-babysitter

Pass

Audited by Gen Agent Trust Hub on Oct 24, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPERSISTENCE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes GitHub PR comments from both human reviewers and automated bots using the scripts/fetch-comments.sh script. This external content directly influences agent actions, such as performing code fixes, replying to comments, and resolving threads. While the skill utilizes a plan file as an audit trail and attempts to filter known bot noise, human comments are treated as major instructions, creating a surface for indirect prompt injection.
  • Ingestion points: PR review comments, issue-level comments, and bot findings fetched via the GitHub API using scripts/fetch-comments.sh.
  • Boundary markers: The skill generates a fix plan at .claude/pr-babysitter/pr-{N}-review-plan.md to document its intended actions, though it proceeds autonomously.
  • Capability inventory: The skill possesses significant capabilities, including modifying source code, performing git push, executing gh pr merge, and creating scheduled tasks.
  • Sanitization: The instructions provide heuristics to identify bot noise, but human comments are processed as directives without additional sanitization or delimiters to prevent instruction overrides.
  • [COMMAND_EXECUTION]: The skill makes extensive use of system commands, relying on the gh and git CLIs for PR management and git operations. It also invokes platform-specific tools such as vercel, flyctl, and the Buildkite bk CLI to fetch logs and manage deployments.
  • [PERSISTENCE]: The skill implements a long-running monitor mode by scheduling recurring tasks via the CronCreate platform tool, enabling it to maintain activity across sessions until a PR is merged or closed.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 24, 2026, 05:17 PM