The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/validation_control.py dynamically modifies the system path and imports several modules (model_routing, run_state, provider_routing, browser_preflight, review_evidence) from a directory specified by the --shared-dir command-line argument. This mechanism allows for the loading and execution of code from arbitrary local paths if the agent is provided with a malicious directory.
[COMMAND_EXECUTION]: The skill instructions and its unit tests (tests/test_validation_control.py) utilize subprocess.run to execute the controller script and interact with repository tools like git. This established pattern of executing shell commands provides a functional capability that could be leveraged for unintended command execution if inputs are not strictly controlled.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data such as acceptance contracts, validation plans, and evidence packets (Markdown and JSON). These files represent an attack surface where adversarial instructions could be embedded to influence the agent's validation decisions. While the skill employs SHA-256 hashing to maintain integrity between steps, it remains vulnerable to malicious content in the initial input files.
[EXTERNAL_DOWNLOADS]: The file references/model-research.md contains numerous links to external model catalogs, benchmarks, and documentation from well-known services and organizations, including OpenRouter, HuggingFace, Vercel Labs, and Microsoft. These references are used to inform model selection and research policies during the validation workflow.

validate-e2e