The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill performs extensive scanning of the local codebase in Phase 1 using Glob, Grep, and Read tools. This data is used to identify the technology stack and recommend relevant skills. While this represents a potential attack surface where data in the codebase could influence the agent's logic, it is a core functional requirement of the stack-detection feature.
[COMMAND_EXECUTION]: The skill executes a bundled maintenance script (scripts/write-operator-permissions.mjs) to update Claude Code's security settings. This process is protected by a strict security protocol requiring a mandatory "dry-run" preview and explicit user confirmation via the AskUserQuestion tool before any writes occur.
[PERSISTENCE]: The skill modifies configuration files (.claude/settings.json, .mcp.json) to persist project settings, MCP server configurations, and security permission rules. These modifications are standard for application configuration and are performed with user oversight through the setup wizard.
[DATA_EXFILTRATION]: The skill includes an optional telemetry phase (Phase 9) for streaming session events to a user-provided webhook. This implementation adheres to privacy best practices by requiring explicit, informed opt-in, mandating HTTPS for endpoints, and using environment variables rather than configuration files for sensitive authentication tokens.
[SAFE]: Static analysis identified a potential fork-bomb pattern in references/operator-permissions.json. Upon review, this was confirmed to be a false positive; the pattern exists only as a string within a metadata changelog field explaining why a specific security rule was deprecated, and it is not in an executable context.

setup