The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/unseamed_calls.py executes git ls-files using subprocess.run to identify project files for scanning. This operation is restricted to the local directory provided by the user and uses safe argument passing (list-based) rather than a shell string.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted local data by reading the contents of source files to identify "unseamed" calls. This represents a potential surface for indirect prompt injection if an attacker-controlled file contains malicious instructions designed to influence the agent when it reviews the tool's output.
Ingestion points: The scan function in scripts/unseamed_calls.py reads file content using path.read_text().
Boundary markers: The script outputs findings as plain text or JSON, including snippets of the matched code (text field), which may not be explicitly delimited for the agent.
Capability inventory: The script has capabilities for file system reading and executing the local git binary.
Sanitization: The script uses strip_noise to remove comments and docstrings before scanning, and applies specific regex patterns to identify only targeted function calls.

determinism-by-design