The Agent Skills Directory

[COMMAND_EXECUTION]: The provided Python and Node.js templates use os.system and execSync to perform file operations and build steps. The code specifically uses shell interpolation (e.g., execSync(cp -r ${inputDir}/* ${buildDir})) which could be vulnerable to command injection if the metadata paths or CLI arguments contain malicious characters.
[DATA_EXFILTRATION]: The skill documents a privileged: true flag that grants the deployment container internet access and injects the project API key. While a legitimate feature of the platform, it increases the risk of credential exposure if the build environment is compromised.
[PROMPT_INJECTION]: The skill uses instructions such as "Always read... Never hard-code..." to guide agent behavior. These are benign instructional markers for scaffolding but are monitored for safety.
[INDIRECT_PROMPT_INJECTION]: The skill scaffolds a workflow where an agent processes untrusted external data (project models, SDKs, and metadata) from Edge Impulse Studio.
Ingestion points: Reads deployment-metadata.json and files from metadata.folders.input (SKILL.md).
Boundary markers: None; the scripts directly parse and interpolate metadata values into shell commands.
Capability inventory: Uses os.system, execSync, shutil, and fs for file system manipulation and command execution across the Python and Node.js templates (SKILL.md).
Sanitization: The templates do not implement sanitization or validation for the values read from the metadata JSON before using them in shell commands.

build-custom-deployment-blocks