The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on executing various third-party AI command-line interfaces (including Claude Code, OpenAI Codex, and Grok) via subprocess calls in scripts/dispatch_checker.py. It uses shlex.quote to sanitize arguments when wrapping commands in login shells, following security best practices for process execution.
[CREDENTIALS_UNSAFE]: The skill manages API keys for multiple AI providers. It includes a robust scrub_text function using regex and literal value matching to redact credentials from all generated output files, ensuring that secrets are not leaked into the local work directory or the agent's conversation history.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted text from user articles and essays to extract claims for verification.
Ingestion points: Articles are read from the file system and processed by the orchestrator into claims.json (as described in Phase 1 of SKILL.md).
Boundary markers: Verification prompts in references/checker-prompt.md use markdown JSON code blocks to isolate the untrusted claim data from instructions.
Capability inventory: The skill executes local binaries and orchestrates network-enabled research agents across multiple scripts.
Sanitization: While claims are structured as JSON, the raw article text is not sanitized for injection patterns, representing a standard but present attack surface for research-oriented agents.
[PERSISTENCE]: The skill temporarily modifies the local ~/.zcode/cli/config.json file to manage model selection for the zcode harness. This is implemented via a Python context manager in scripts/dispatch_checker.py that ensures the original configuration is restored even if an error occurs during execution.

fact-check