caveman-optimize
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes shell commands via the
cavemanCLI (e.g.,caveman opportunities list) to retrieve optimization data. This is a core function of the skill and is restricted to a specific toolset. - [INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external data from two sources: output from the
cavemanCLI and local files located in.caveman/proposals/*.md. - Ingestion points: The agent reads server-provided observation arrays from CLI output and historical proposal files from the repository.
- Boundary markers: The instructions explicitly label repository files as 'untrusted historic context' and warn the agent not to let this data override current session logic or operator choices.
- Capability inventory: The skill possesses the ability to execute shell commands (via CLI) and modify the repository source code (applying approved candidates).
- Sanitization: The risk is mitigated by a mandatory human-in-the-loop requirement where the operator must explicitly choose an observation and approve both the evaluation design and the final code change.
Audit Metadata