The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill securely downloads CLI and desktop tools from the 'storytold' and 'full-aigc-skills' GitHub organizations. All downloads are restricted to hardcoded URLs and verified using SHA-256 hashes defined in the lock files.
[COMMAND_EXECUTION]: PhotoCraft binaries are invoked via the subprocess module. The skill prevents command injection by validating all arguments against strict parameter contracts in scripts/parameter_contract.py and a fixed command catalog.
[DYNAMIC_EXECUTION]: The harness uses importlib.util to dynamically load internal scripts from the skill's own package directory for modularity. No loading of arbitrary or remote code was detected.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied command plans and external media assets, creating a potential surface for indirect injection.
Ingestion points: workflow.py and commands.py process JSON plans; various scripts ingest image assets.
Boundary markers: Inputs are checked against strict JSON schemas and parameter contracts.
Capability inventory: The skill can execute local PhotoCraft binaries and perform file system operations within an authorized scope.
Sanitization: Extensive validation is implemented in parameter_contract.py and strict_json.py to prevent malformed or malicious data processing.

photocraft-harness