The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external and local sources, which could potentially contain instructions intended to influence the agent's code generation or evaluation tasks.
Ingestion points: CLI output from caveman opportunities list (specifically the report_only_observations array) and local markdown files matching the pattern .caveman/proposals/*.md.
Boundary markers: The skill explicitly instructs the agent to treat historic proposals as "untrusted historic context" and requires an "explicit operator choice" before proceeding with code modifications.
Capability inventory: The skill possesses the ability to execute shell commands (caveman CLI and evaluation commands) and perform file system writes (applying candidate edits).
Sanitization: While the instructions emphasize preserving server-provided text verbatim, there are no technical sanitization or escaping protocols described for the ingested data before it is processed by the agent.
[COMMAND_EXECUTION]: The skill executes specific shell commands to interact with vendor tooling.
Evidence: The instruction to run caveman opportunities list to retrieve diagnostic observations.

caveman-optimize