The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes the compile() function in scripts/validate_harness_scripts.py to perform syntax validation on its own Python scripts. Additionally, the test suite in tests/test_scripts.py uses importlib.util and exec_module to dynamically load scripts for unit testing. While these are used for internal validation and testing, they represent dynamic code execution patterns.
[INDIRECT_PROMPT_INJECTION]: The skill's project scanning modules (scripts/scan_project.py and its dependencies) ingest untrusted data from the target repository, including file names, directory structures, and metadata from files like package.json and .git/config. This information is presented to the agent to help build the harness, creating a surface where a malicious repository could attempt to influence agent behavior through specific file naming or metadata.
Ingestion points: Repository file tree traversal and configuration file reading in scripts/scan_project_automation.py and scripts/scan_project_git.py.
Boundary markers: Data is structured into JSON reports, but there are no explicit delimiters or instructions to ignore embedded commands within the ingested file names or metadata.
Capability inventory: The skill possesses file-write capabilities (Markdown, Shell scripts) and the ability to execute validation commands via the generated check.sh.
Sanitization: No sanitization or escaping is applied to file system strings before they are included in the report provided to the agent.
[EXTERNAL_DOWNLOADS]: The skill references and adapts automation recommendations from Anthropic's official plugin resources, as documented in references/automation_recommendation_attribution.md. This reference to a trusted source is documented for attribution purposes.

harness-builder