The Agent Skills Directory

[COMMAND_EXECUTION]: The fair_packager.py script allows writing the generated manifest to an arbitrary file path via the --out argument. The implementation does not restrict the output path to the working directory and explicitly allows absolute paths and directory traversal (..). This provides a capability to overwrite sensitive user or system files (e.g., .bashrc, .ssh/config) with JSON content if the agent is misled or compromised.
Evidence: scripts/fair_packager.py lines 147-150: out_path = Path(args.out); _reject_control(str(out_path), 'out'); out_path.write_text(json.dumps(manifest, indent=2, sort_keys=True) + '\n', encoding='utf-8').
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input and file system metadata, which are subsequently returned to the agent in a structured JSON envelope. This creates a surface where malicious data could influence downstream agent logic.
Ingestion points: Command-line arguments (--project-name, --engine, --units, etc.) and file paths passed to the script in scripts/fair_packager.py.
Boundary markers: Data is encapsulated in a JSON object with specific keys (inputs, results.manifest), providing clear structural boundaries.
Capability inventory: The skill has the capability to read file metadata (existence, size, SHA256) and write a single JSON file to an arbitrary path via out_path.write_text in scripts/fair_packager.py.
Sanitization: The script implements basic sanitization by rejecting control characters (_reject_control) and enforcing an allowlist for unit keys and values (SAFE_FIELD).

fair-simulation-packager