The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from sources such as {live_results_input} (live-verification-results.json), {waiver_register_input} (gate-waivers.md), and project source code for synthetic journey inference. These files represent an attack surface where a malicious actor could inject instructions to manipulate the quality gate outcome (e.g., forcing a PASS decision) or bias the generated recommendations.
Ingestion points: Evaluation of JSON-formatted live verification results and Markdown-formatted waiver registers.
Boundary markers: The workflow relies on structured data keys and Markdown headings (## {ID}: {title}) but lacks explicit safety delimiters to differentiate between data and instructions during report interpolation.
Capability inventory: Access to read_file, write_file, list_files, search_repo, glob, and uv run for executing local scripts.
Sanitization: The logic performs basic JSON schema and type validation but does not fully sanitize or escape natural language content from requirements or waivers before embedding them in the final report.
[COMMAND_EXECUTION]: The skill frequently executes local scripts using uv run (e.g., resolve_customization.py, resolve_config.py). These scripts are located within the project's _bmad/scripts/ directory. While these are intended for project-specific customization, this capability allows for the execution of arbitrary logic provided in those local files.
[EXTERNAL_DOWNLOADS]: The skill utilizes npx skills add to download and install a sibling skill (bmod-tea) from the vendor's repository (bmad-code-org/bmad-method-test-architecture-enterprise). This involves fetching and potentially executing code from a remote source. This behavior is documented as a standard vendor integration path.
[DYNAMIC_EXECUTION]: The workflow documentation (e.g., in step-02-discover-tests.md and step-04-analyze-gaps.md) contains literal JavaScript-style logic blocks and instructs the agent to "Apply these rules exactly" or "Execute each entry." This requires the agent to interpret and execute complex procedural logic at runtime to process project artifacts and derive statistics.

bmad-testarch-trace