The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill instructs the main agent to delegate tasks to subagents (cavecrew-investigator, cavecrew-builder, cavecrew-reviewer) and inject their tool-results back into the main conversation context verbatim. This establishes a path where malicious instructions embedded in the code or files being analyzed by a subagent could be passed to the main agent, potentially influencing its subsequent actions.
Ingestion points: Main agent context via subagent tool-results (described in README.md and SKILL.md).
Boundary markers: Absent; the instructions state that results are injected "verbatim" to save tokens.
Capability inventory: The subagents are described as being able to read and edit files, and the main thread is directed to use their output for further decision-making.
Sanitization: No sanitization or validation of subagent output is specified in the instructions.
[NO_CODE]: The analyzed files (README.md and SKILL.md) consist entirely of markdown documentation and instructions for agent behavior. No executable scripts, binaries, or active code components are included in this skill.

cavecrew