The Agent Skills Directory

[SAFE]: The skill is a specialized PR verification tool that manages native agent sessions in isolated environments.
[CREDENTIALS_UNSAFE]: The skill interacts with native harness credentials (such as Codex auth.json) but protects them by using symbolic links within isolated run homes. It includes validation logic to detect if candidate code attempts to redirect or overwrite these credentials.
[DATA_EXFILTRATION]: The skill performs authorized network operations to GitHub using the gh CLI for reporting. It mitigates data exfiltration risks by sanitizing all public evidence through a publishable predicate that enforces length and character set limits, ensuring only non-sensitive data is published.
[INDIRECT_PROMPT_INJECTION]: Although the skill processes untrusted PR data, it mitigates injection risks by using machine-checked "recipes" and assertions that rely on session tool-call records and file system state rather than the agent's textual claims.
[OBFUSCATION]: A deterministic detector flagged a base64 decoding pattern in scripts/catalog.test.ts. This is a false positive, as the code is a legitimate test case verifying browser rendering of nonces and does not contain malicious hidden payloads.
[DYNAMIC_EXECUTION]: The skill generates and executes simple helper scripts from static templates for testing purposes. These are used within a controlled environment to simulate CLI behavior.

verify-open-pstack