The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes system binaries (such as git, node, jq, rg, ast-grep) to perform environment checks. It uses spawn with shell: false and strict argument validation to prevent command injection.
[PERSISTENCE]: The skill can modify .claude/settings.local.json to install hooks that run during session lifecycle events. This is the intended purpose of the tool's hook-adapter functionality, enabling automated diagnostics.
[DYNAMIC_EXECUTION]: The hook adapter identifies and executes scripts from other skills in the suite (e.g., relay.mjs) to manage session state. Execution is restricted to pre-defined paths within the skill's own directory structure.
[INDIRECT_PROMPT_INJECTION]: The skill processes user prompts to suggest relevant tools ('nudges'). While this creates an ingestion surface for untrusted data in scripts/_lib/hook-adapter.mjs, the resulting output injected into the agent context is limited to static suggestions based on regex matching.
Ingestion points: User prompts are processed in scripts/_lib/hook-adapter.mjs via the UserPromptSubmit hook.
Boundary markers: Explicit delimiters are not used when injecting nudge suggestions into the context.
Capability inventory: The skill has capabilities for subprocess execution (execute.mjs), file writing (files.mjs), and configuration modification (doctor.mjs).
Sanitization: Prompt data is used for regex matching in router.mjs but is not directly interpolated into executable commands or raw system prompts.

doctor