The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides APIs to execute shell commands and capture their output for verification.
Evidence: verify_command_line("python --version") in references/python/api.md allows the agent to run arbitrary shell commands during testing.
Evidence: Failure reporters (e.g., GenericDiffReporter) are documented to launch external applications like BeyondCompare, VSCode, and P4Merge to display differences between files.
[DYNAMIC_EXECUTION]: The Python and Java implementations support dynamic execution for testing purposes.
Evidence: verify_command_line in Python spawns a subprocess to execute strings as commands.
Evidence: Custom reporters and scrubbers allow the execution of arbitrary functions passed as parameters at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and verify complex data, which creates an indirect prompt injection surface if the verified data originates from untrusted sources.
Ingestion points: Untrusted data enters the context through verify(), verify_as_json(), verify_all(), and verify_xml/html() across all supported languages.
Boundary markers: None are explicitly enforced within the data content; however, the skill encourages the use of scrubbers to normalize data.
Capability inventory: The skill has the capability to write to the file system (creating .received and .approved files) and launch external GUI diff tools via reporters.
Sanitization: The skill provides a robust "Scrubber" framework to replace dynamic or sensitive content (GUIDs, dates, regex patterns) with placeholders, though this is primarily for test stability rather than security sanitization.

approval-tests