The Agent Skills Directory

[COMMAND_EXECUTION]: The templates execute the gjc CLI using subprocess calls. This is handled safely by passing arguments as an array rather than a shell string in templates/direct-sdk.py and templates/direct-sdk.ts, which prevents command injection vulnerabilities.
[DATA_EXFILTRATION]: The scripts include a redaction mechanism in templates/direct-sdk.py and templates/direct-sdk.ts that scans for common sensitive field names such as 'secret', 'token', and 'password', redacting their values before displaying output to the user.
[INDIRECT_PROMPT_INJECTION]: The skill processes output from the gjc CLI and mitigates potential injection risks by parsing the output as JSON and redacting sensitive information before presenting it to the agent.
Ingestion points: CLI stdout is read into the script context in templates/direct-sdk.py and templates/direct-sdk.ts.
Boundary markers: The output is parsed as structured JSON in both templates, providing a clear boundary between external data and code logic.
Capability inventory: The scripts perform subprocess execution via subprocess.run in templates/direct-sdk.py and Bun.spawn in templates/direct-sdk.ts, restricted to the specified repository directory.
Sanitization: A redact function is implemented in both templates to filter out sensitive keys based on a regex allowlist before outputting results.

gjc-sdk-author