The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a custom CLI (control-agent-server) to execute arbitrary shell commands for launching the server, managing fixtures, and running test recipes. These commands are executed in a dedicated, isolated run directory with a private environment.
[DYNAMIC_EXECUTION]: To test extensibility, the skill creates Python modules at runtime and instructs the server to import them (e.g., F27.custom-tool-modules). This is a legitimate part of the verification suite's logic for testing custom tool integration.
[EXTERNAL_DOWNLOADS]: The skill clones the public OpenHands/extensions repository from GitHub to verify marketplace and plugin functionality (e.g., F25.public-repo, F26.marketplace-public). These downloads are from a trusted organization repository.
[INDIRECT_PROMPT_INJECTION]: The map run utility processes markdown files and executes shell command blocks contained within them. While this constitutes an injection surface, the markdown files provided are trusted components of the verification suite.
[DATA_EXPOSURE]: The verification harness specifically manages and uses API keys and session tokens to test the server's authentication and encryption features. It includes a built-in redactor to prevent leaking these secrets into logs and evidence files, though users can explicitly bypass this for testing purposes (e.g., api --unsafe-unredacted).

verify-agent-server