The Agent Skills Directory

[SAFE] (SAFE): The skill architecture is well-structured and follows standard Claude Code extension patterns for agent-based workflows. No obfuscation or malicious intent was found.\n- [COMMAND_EXECUTION] (SAFE): Employs local shell and Python scripts for vault detection and file validation. Validation logic in scripts/validate-base.sh uses yaml.safe_load() to prevent potential remote code execution from malicious YAML input.\n- [DATA_EXFILTRATION] (SAFE): User preferences and feature usage patterns are tracked locally in the .claude/obsidian-memory.json file. Data collection is limited to feature telemetry (e.g., counting callouts or wikilinks), and no external network exfiltration was detected.\n- [INDIRECT_PROMPT_INJECTION] (SAFE): The skill interacts with user-controlled files (.md, .base, .canvas), creating a surface for indirect prompt injection. However, it processes these files using restricted pattern matching and standard agent safety boundaries.\n
Ingestion points: Agents use Read(*) permissions and post-tool-use.sh parses tool results.\n
Boundary markers: Absent in agent instructions but mitigated by model-level guardrails.\n
Capability inventory: Agents have access to Write, Edit, and restricted Bash permissions.\n
Sanitization: Content is analyzed via grep for pattern tracking without execution or direct interpolation into system prompts.

obsidian