The Agent Skills Directory

[SAFE]: The skill represents a well-structured, local-only workflow for verifying knowledge citations. No malicious patterns, obfuscation, or unauthorized access attempts were detected.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill explicitly adheres to a 'zero-network' architecture for its verification gate. It operates on local project directories and wiki files, with no network operations (no curl, wget, or external API calls) detected in the instructions or orchestrated scripts.
[COMMAND_EXECUTION]: Shell commands are used strictly for local file management (cp, rm, test) and executing internal Python scripts. The logic includes safeguards against path expansion issues and uses atomic file replacement for manifest updates.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an ingestion surface for untrusted data (draft content and wiki claims), which is a characteristic of its primary function.
Ingestion points: Reads <project>/output/draft-vN.md, <project>/.metadata/citation-manifest.json, and source files in wiki/sources/*.md.
Boundary markers: None explicitly defined in the orchestrator; however, data is handled through structured JSON manifests and sharded verifier tasks.
Capability inventory: Uses Task to dispatch internal agents (wiki-verifier, revisor), Bash for local script execution, and standard file Read/Write tools.
Sanitization: Employs strict JSON schema validation (versions 0.1.0 and 0.1.1) and verification counts to ensure integrity during data processing. The risk is assessed as safe due to the offline, structured nature of the pipeline.

knowledge-verify