The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs extensive shell operations using tools like npm, jq, and the claude CLI to diagnose installation health and manage configuration.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from local session transcripts and repository memory files (CLAUDE.md). It implements robust mitigations, including mandatory user confirmation gates and strict input sanitization to prevent the execution of instructions embedded in the processed data.
[CREDENTIALS_UNSAFE]: The skill reads local settings and MCP configuration files known to contain secrets like API keys. It adheres to a "key-scoped read" policy, ensuring only necessary non-sensitive keys are accessed and that secrets are never inlined into shell commands or proposed reports.
[PRIVILEGE_ESCALATION]: The skill identifies and proposes changes to the agent's security posture, such as enabling "auto mode" or adding command "allow" rules. These high-privilege operations are strictly isolated and require explicit, multi-step user approval.
[EXTERNAL_DOWNLOADS]: The skill fetches versioning information from well-known sources, including the npm registry, Anthropic's release domain, and Homebrew's API to verify if the installation is current.

doctor