The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands to interact with the macOS service manager (launchctl), manage filesystem links (ln), and process JSON data (jq). These operations are used to manage the ambient watcher daemon and install git hooks.
[PROMPT_INJECTION]: The skill facilitates an indirect prompt injection surface by design. It aggregates data from external, potentially untrusted sources (such as git commit messages, CI/CD logs, or external webhooks) and injects this content into the agent's context during session startup.
Ingestion points: Event files stored in ~/.claude/events/*.event populated by external scripts and hooks.
Boundary markers: No specific boundary markers or 'ignore' instructions are defined in the provided documentation for the injected event content.
Capability inventory: The skill can manage system services via launchctl, modify local git hooks, and read/write to specific application directories in the user's home folder.
Sanitization: Sanitization of the injected event messages is not documented and depends on the implementation of the external drain-events.sh and claude-event scripts.

ambient-context