tk-pr-guidance

Pass

Audited by Gen Agent Trust Hub on Nov 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from GitHub Pull Request bodies and comments, creating a potential surface for indirect instructions.
  • Ingestion points: references/gfm-rendering.md reads raw PR content and comments for validation.
  • Boundary markers: The instructions emphasize treating the input as a "candidate" for rendering and require byte-perfect preservation, though explicit delimiters are not defined in the prompt template.
  • Capability inventory: The skill utilizes the GitHub API (gh api) for remote rendering, file system operations for workspace setup (git worktree), and local file read/write for candidate preparation.
  • Sanitization: The skill implements a robust verification step, requiring the agent to compare the HTML output of the GitHub API against the original Markdown to ensure that code literals or other structures were not improperly escaped or interpreted as instructions.
  • [SAFE]: The skill demonstrates an advanced security posture by including specific prohibitions against common attack vectors.
  • It explicitly forbids the execution of setup hooks found in repositories, noting that reading them is not an authorization to run them.
  • It warns against command injection when using CLI tools by instructing the agent not to interpolate PR text directly into shell strings where characters like backticks or dollar signs could be reinterpreted.
  • It uses pinned commit SHAs when referencing external documentation sources (e.g., github/cmark-gfm, EveryInc/compound-engineering-plugin), ensuring the integrity of the referenced logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 17, 2026, 03:45 PM