The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the caveman CLI tool (e.g., caveman cloud costs, caveman cloud score, caveman cloud traces search) to retrieve evidence from Caveman Cloud. These commands are limited to data retrieval and do not include administrative or destructive actions.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a surface for indirect prompt injection by ingesting data from external cloud reports and execution traces.
Ingestion points: External data enters the context through MCP tools like caveman_report and CLI commands like caveman cloud traces search in SKILL.md.
Boundary markers: The skill provides a structured reporting template in Step 5, which acts as a boundary for the agent's final response.
Capability inventory: The agent can execute the caveman CLI and various caveman_* MCP tools. However, the instructions explicitly restrict the agent to a "read-only operator" role.
Sanitization: While no explicit sanitization is mentioned, the skill mandates that completion payloads (which are higher-risk sources of injection) should not be fetched unless explicitly requested by the user.

caveman-evidence-review