The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/hook-runtime.py script executes the tools.codegraph_voyage sidecar using asyncio.create_subprocess_exec. The execution is highly controlled, using strictly hardcoded arguments such as status or index. The runtime environment is scoped to the project root, and the command only executes if specific project marker files (.codegraph/codegraph.db) are present.
[INDIRECT_PROMPT_INJECTION]: The skill's lifecycle hooks ingest tool metadata (tool names and project paths) from the agent client's stdin to determine when the index needs a refresh. The script includes logic to parse varied JSON dialects from different clients and uses this data solely to mark the project state as 'dirty' or trigger a status check. It does not provide a path for external data to influence the agent's instructions.
[PERSISTENCE]: The skill includes an optional installer (scripts/hooks-fanout.py) that modifies project-local configuration files (e.g., .claude/settings.json) to register lifecycle hooks. This behavior is the primary intended purpose of the sidecar functionality and is implemented with security-conscious practices, including symlink rejection and atomic writes with automatic backups.
[CREDENTIALS_SAFE]: The skill documentation explicitly instructs users to manage the VOYAGE_API_KEY using environment variables or secret managers like 1Password. It provides clear warnings against hardcoding keys in source files, logs, or command-line arguments. The code itself does not contain any hardcoded secrets or unsafe credential handling patterns.

codegraph-voyage