The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/_runtime.py implements a runtime discovery mechanism that searches for Python interpreters in local .venv directories within the current or parent folders. It uses os.execve to replace the current process with the discovered interpreter. While intended for dependency management, this allows for execution of binaries outside of the skill package.
[COMMAND_EXECUTION]: In scripts/_runtime.py, the skill uses subprocess.run to execute commands on discovered interpreters. This is used to verify module availability before switching the runtime.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests and processes untrusted data from user-provided images and architectural names.
Ingestion points: User inputs such as location names, object descriptions, and photo metadata are extracted during the photo_preflight.py and research-and-annotation.md phases.
Boundary markers: The skill instructions in references/image-generation-workflow.md define a structured prompt format, but user data is directly interpolated into instructions for the image model.
Capability inventory: The agent has the capability to call image generation APIs (gpt-image-2) and write files to the output directory.
Sanitization: There is no evidence of explicit sanitization or escaping of user-provided text before it is included in the final generation prompts.

yingzao