The Agent Skills Directory

[COMMAND_EXECUTION]: The scenario YAML schema supports a teardown block that allows the execution of arbitrary shell commands.
Evidence: The TeardownStep model in references/scenario-schema.md defines a shell type that accepts a command string for execution after scenario completion.
[INDIRECT_PROMPT_INJECTION]: The skill scans external URLs to discover interactive elements, which are then used by the AI to generate test scenarios. This creates a risk where malicious content on a target website could influence the agent's logic or generated code.
Ingestion points: The aat scan --url <URL> command reads external web content into .aat/scan_result.json (SKILL.md).
Boundary markers: The skill does not define explicit delimiters or instructions to the AI to ignore instructions embedded in the scanned content.
Capability inventory: The tool can perform browser automation (clicks, typing), file uploads, and execute shell commands or API calls in teardown steps.
Sanitization: No sanitization of ingested element labels or text is specified before being processed by the AI.
[EXTERNAL_DOWNLOADS]: The skill setup process involves installing external dependencies and browser binaries.
Evidence: SKILL.md instructs the agent to run pip install aat-devqa && playwright install chromium during the auto-setup phase.
[DYNAMIC_EXECUTION]: The skill dynamically generates YAML test scenarios from analyzed web scans and subsequently executes them using a specialized runner.
Evidence: The mandatory 4-step workflow in SKILL.md involves generating scenarios based on scan_result.json and running them via aat run.

awt