The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external, untrusted project source code and provider implementations. This creates a surface for indirect prompt injection, where malicious instructions could be embedded in code comments or documentation within the target repository.
Ingestion points: According to SKILL.md and extract.md, the agent reads project source code and provider implementations at a pinned revision during the Extract, Compare, and Verify stages.
Capability inventory: The skill allows the agent to read and write files in specified directories and execute local Python scripts for artifact validation (scripts/check_artifacts.py).
Boundary markers: The skill provides a robust procedural boundary by requiring the 'Design' stage to be performed 'from zero' without access to the source code, relying solely on the extracted model.
Sanitization: There is no explicit sanitization of code comments; the skill relies on the agent's reasoning and the structured nature of the artifacts to mitigate risks.
[COMMAND_EXECUTION]: The skill requires the execution of local Python scripts to validate the structure of the generated artifacts.
Evidence: SKILL.md and the individual stage instructions (extract.md, design.md, compare.md) specify the use of scripts/check_artifacts.py to verify that outputs adhere to the required schema.
Safety: Analysis of the provided scripts (scripts/check_artifacts.py and evals/stage_test.py) shows they use standard libraries for regex-based parsing and file system operations, with no network activity or unsafe dynamic execution patterns.

code-architecture