The Agent Skills Directory

[COMMAND_EXECUTION]: The script assets/puzzle-lab/tools/build.py executes shell commands via subprocess.run to orchestrate the build process.
It invokes docker run to compile C source code into a Mega Drive ROM using the SGDK toolchain.
Command-line arguments for platform and build mode are strictly constrained by argparse choices to prevent injection.
Path inputs are validated to ensure the build output directory is outside the skill's source directory, preventing accidental or malicious overwriting of the toolchain itself.
[EXTERNAL_DOWNLOADS]: The build tool is designed to pull and use the ewrogers/sgdk:latest Docker image.
This image is a vendor-owned resource (authored by ewrogers) used as the official, reproducible build environment for the project.
The skill documentation explicitly instructs the agent to pull this image before building.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to translate user descriptions of puzzle mechanics or board layouts into game code and configuration.
Ingestion points: User instructions for room design, tile maps, or rule modifications are processed by the agent within the logic-design workflow.
Boundary markers: The skill provides a dedicated "responsibility boundaries" reference to guide the agent on which parts of the code should be modified versus delegated to other specialized skills.
Capability inventory: The skill possesses file system capabilities via shutil and Path (in prepare.py) and command execution capabilities (in build.py).
Sanitization: The Python preparation script validates that the target output path is not relative to the source root, protecting the integrity of the skill's internal logic.

sgdk-puzzle