The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from a user's Obsidian vault, including note properties and configuration files, which could potentially contain embedded instructions.
Ingestion points: The skill reads .base files and note properties from the user's vault in SKILL.md during the inspection step.
Boundary markers: The skill requires an explicit review process where a transaction bundle is inspected and then applied only after the user provides a matching APPROVAL_SHA256 hash.
Capability inventory: The agent can modify files in the vault using the claude-obsidian.py transaction tool.
Sanitization: The transaction-based workflow ensures that the agent does not write to the file system directly, allowing for user oversight of the proposed changes.
[COMMAND_EXECUTION]: The skill utilizes a local Python script for managing vault transactions and validation.
Evidence: The workflow involves executing python3 "$CORE" transaction inspect and python3 "$CORE" transaction apply.
Context: Execution is restricted to a script (claude-obsidian.py) located within the product's own installation directory (PRODUCT_ROOT), and all mutations are gated by a mandatory user-approved SHA256 plan hash.

obsidian-bases