The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves reading source code and diffs to send them to external LLMs for review and auditing. This creates an inherent surface for indirect prompt injection, as malicious instructions within the code being analyzed could theoretically attempt to influence the reviewer LLM. Ingestion points: scripts/specialist-review.sh and scripts/diff-test-audit.sh collect content from the local project files. Boundary markers: The prompt templates use standard markdown code fences but lack robust instructions to the external LLM to disregard potentially adversarial content inside the code blocks. Capability inventory: The skill utilizes shell scripts to invoke CLI tools and git commands for committing changes. Sanitization: External code content is interpolated directly into prompt templates without transformation.
[DYNAMIC_EXECUTION]: The shell scripts scripts/specialist-review.sh and scripts/diff-test-audit.sh generate and execute static Python code on-the-fly to perform template substitution for the review prompts. Evidence: The scripts use `python3
<<'PYEOF'` to run inline Python logic for string manipulation. Context: This dynamic execution is used for utility purposes and the executed code is fixed within the skill's scripts.

agent-loops