The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill includes an automated installer in scripts/bootstrap.py that fetches the lightcraft-cli binary from official releases on GitHub. It enforces strict security controls, including mandatory SHA256 integrity verification of both the downloaded archive and the extracted binary against a hardcoded lockfile (scripts/runtime.lock.json). The download is restricted to specific, regex-verified URLs from the project's own repositories.
[COMMAND_EXECUTION]: The skill wraps the Lightcraft CLI, executing commands through a supervision layer in scripts/native_process.py. All subprocess calls are made using argument lists (avoiding shell interpolation) and are subject to timeouts and logging. This structure prevents arbitrary command injection and ensures that only allowed CLI subcommands are executed.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data in the form of JSON plans and image files. It mitigates potential indirect injection attacks by implementing a strict JSON parser that rejects ambiguous formats and validating all plan parameters against internal schemas in scripts/command_gateway.py. It also checks that photo modification values fall within allowed ranges defined in scripts/photo_workflows.py.
[DYNAMIC_EXECUTION]: To maintain modularity across its internal components, the skill uses importlib.util to load its own local Python scripts (e.g., command_gateway.py, mode_contract.py). This dynamic loading is restricted to files within the skill's own directory and does not execute untrusted or remote code.
[DATA_EXPOSURE]: The skill manages file operations within a defined output_root. It includes pre-flight logic in scripts/command_gateway.py to ensure that modification commands do not conflict with or overwrite original photo files, maintaining data integrity during the editing process.

lightcraft-cli