The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions to intercept and override deployment-related user commands (e.g., 'push to production', 'deploy') to force a security audit. While this is the intended functionality for a gatekeeping tool, it explicitly directs the agent to ignore the user's initial intent and prioritize the audit protocol.
[COMMAND_EXECUTION]: The skill includes a Python script scripts/ship_gate_scanner.py used to scan the local filesystem for security patterns. The script uses standard file operations and regular expressions to identify potential risks. It does not perform network operations or execute external code, but it does execute complex regex scans across the project directory.
[INDIRECT_PROMPT_INJECTION]: The skill scans untrusted codebase files and presents extracted code snippets to the AI agent in its final report. This creates a surface where an attacker could place malicious instructions inside code comments or strings to influence the agent's behavior when it processes the report.
Ingestion points: Local codebase files (e.g., .js, .py, .sql) are read and processed by scripts/ship_gate_scanner.py using grep_files.
Boundary markers: Results are structured in a markdown report block defined in SKILL.md. No explicit delimiters or 'ignore embedded instructions' warnings are applied to the code snippets displayed in the report.
Capability inventory: The skill primarily reads files and reports findings. It includes capabilities for filesystem traversal (os.walk) and file reading (open), but does not perform automated fixes or network requests.
Sanitization: Code snippets are extracted verbatim (truncated to 120 characters) and displayed in the report without escaping or validation.

ship-gate