The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied data, such as code diffs and project structures, to generate PR body content. This creates a surface for indirect prompt injection where malicious content in a diff could attempt to influence the agent's output. However, the risk is negligible because the skill possesses no dangerous capabilities (e.g., file writing, network access, or command execution) and its output is intended for human review in a PR context.
Ingestion points: Code diffs, file trees, and component structures provided by the user are interpolated into the template via the SKILL.md instructions.
Boundary markers: The skill uses Markdown code blocks and specific headers (## Summary, ## Evidence) to structure and delimit the generated content.
Capability inventory: The skill is restricted to text generation; no tools for network access, shell execution, or file system modification are utilized.
Sanitization: The instructions do not specify explicit sanitization or escaping of input data, relying on the agent's standard text processing.
[SAFE]: A thorough review of the skill's instructions and metadata found no evidence of malicious patterns, obfuscation, credential harvesting, or unauthorized persistence mechanisms.

pr