The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes validation briefs and experiment results which are potentially untrusted data sources. This surface is mitigated by design:
Ingestion: Brief files (ideas/*.md) and experiment data recorded via CLI.
Boundary Markers: The skill uses a code() function to wrap all untrusted evidence and statements in markdown backtick fences when generating reports.
Sanitization: The visible() function in ideas.py identifies and escapes hidden Unicode characters (e.g., control characters, zero-width spaces, RTL overrides) into their \uXXXX representations to prevent visual deception or steganographic prompt injection.
Least Privilege: The core ideas.py tool is strictly limited to the Python standard library and contains internal checks and test suite monitors to ensure it never performs network operations or spawns subprocesses.
[REMOTE_CODE_EXECUTION]: An automated scanner flagged a curl | sh pattern. Investigation confirms this string is part of a static test fixture (HOSTILE variable in assets/test_ideas_flow.py) used to verify that the tool correctly fences and neutralizes malicious payloads during reporting. It does not represent an execution path within the functional code.
[OBFUSCATION]: A scanner alert for Unicode steganography in assets/test_ideas_flow.py relates to the same HOSTILE test fixture. The fixture contains Right-to-Left Override (RLO) and zero-width characters to ensure the tool's sanitization logic (visible() function) properly detects and escapes them.
[COMMAND_EXECUTION]: The assets/contract_check.py utility uses subprocess.run to interact with git and verify Python versions. These calls are restricted to specific, non-shell-mediated commands necessary for the tool's validation logic and do not process unsanitized user input in a way that allows command injection.
[DATA_EXFILTRATION]: The skill enforces a .gitignore containing * in its state directory (.skill-contract/ideas/) to prevent accidental commits of experiment results, which may contain sensitive human-provided data such as interview notes.

idea-validation