The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads an official DesignCraft CLI binary from the vendor's GitHub releases at runtime. This process is highly secured through the use of a hardcoded runtime.lock.json file that specifies the exact expected SHA256 hashes for both the archive and the executable.
[REMOTE_CODE_EXECUTION]: The downloaded binary is executed to perform document design tasks. The risk of supply chain attacks is mitigated by strict cryptographic verification, a restricted whitelist of allowed subcommands in cli.py, and URL validation that ensures downloads only originate from authorized vendor repositories.
[COMMAND_EXECUTION]: The skill uses subprocess.run to manage the lifecycle of the design tools. The implementation follows defensive coding standards, such as preventing shell execution (no shell=True), using absolute paths where possible, and including resource limits (e.g., file size and entry count limits) to prevent denial-of-service scenarios.
[INDIRECT_PROMPT_INJECTION]: As the skill ingests external outlines and design materials, it is theoretically susceptible to indirect prompt injection. However, it implements a robust 'Receipt and Manifest' system in command_gateway.py. This system hashes all inputs before execution and all outputs after execution, allowing the agent to detect if the design tool was manipulated into making unauthorized changes to the filesystem.
[SAFE]: The skill scripts demonstrate high-quality security hygiene, including ZipSlip protection during extraction, atomic file writing to prevent corrupted state, and explicit warnings against using sudo or modifying system environment variables.

designcraft-use