The Agent Skills Directory

[SAFE]: The skill implements robust security boundaries by validating all resolved file paths and rejecting absolute paths or symlink escapes, effectively preventing directory traversal attacks during indexing and retrieval.
[SAFE]: The extension prioritizes privacy by requiring explicit opt-in flags and configuration for any off-machine data transmission (e.g., --allow-egress for Anthropic API or --allow-remote-ollama), preventing silent data exfiltration.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-controlled content from the vault to build a retrieval index, creating an attack surface for indirect prompt injection if the indexed content contains malicious instructions.
Ingestion points: Reads markdown and text content from the $VAULT/wiki/ directory.
Boundary markers: Pages are split on paragraph boundaries and stored with page-level prefixes; however, no specific LLM-facing delimiters are described for the final synthesis phase.
Capability inventory: The skill executes local Python scripts, performs file system writes to .vault-meta/, and supports optional authenticated network requests to external LLM providers.
Sanitization: Enforces strict relative path resolution within designated subdirectories and requires hash matching (chunk-body, page, and index) before cached records are served.

wiki-retrieve