A.I. Smart Router

Expertise-aware model router with semantic domain scoring, context-overflow protection, and security redaction. Automatically selects the optimal AI model using weighted expertise scoring (Feb 2026 benchmarks). Supports Claude, GPT, Gemini, Grok with automatic fallback chains, HITL gates, and cost optimization.

This skill appears mostly consistent with a multi-provider model router, but there are several red flags you should address before installing: 1) The metadata claims only ANTHROPIC_API_KEY is required, but the code expects additional optional provider keys and uses router/log state env vars that aren't declared — don't export credentials you don't intend to use. 2) The SKILL.md contains prompt-injection patterns (e.g., 'ignore-previous-instructions', 'system-prompt-override'); review the SKILL.md and all code for any instructions that attempt to change the agent's system prompt or execute unbounded commands. 3) The HITL gate references Telegram notifications but no TELEGRAM_TOKEN is declared — find where notifications are sent and verify you control that endpoint. 4) Because the skill can read/write ~/.openclaw state and invoke sessions_spawn (spawning sub-agents), run it in an isolated environment or sandbox (or inspect and strip the code) if you decide to try it. 5) If you trust the author and want to proceed, at minimum: (a) audit the code for external network calls/endpoints, (b) remove or neutralize any prompt-override directives, (c) only set provider keys you intend to allow, and (d) run it with least privilege and logging enabled so you can observe behavior. If you want, provide the omitted files or the rest of SKILL.md and I can re-scan for external endpoints, Telegram logic, or other undisclosed behaviors.

Type: OpenClaw Skill Name: smart-router Version: 0.1.2 The OpenClaw Smart Router skill bundle is classified as benign. It implements a sophisticated AI model routing system with strong emphasis on security and transparency. Key indicators for this classification include a robust `InputSanitizer` in `router_gateway.py` that actively blocks API keys, private keys, and PII, and performs unicode normalization and control character stripping. The `references/security.md` explicitly disclaims arbitrary code execution (e.g., `eval`, `exec`, shell commands from input) and this is consistently upheld by the Python code, which relies on regex, dictionary lookups, and OpenClaw's `sessions_spawn` API for delegation. File operations are confined to standard application state and log directories (`~/.openclaw/router-state`, `~/.openclaw/logs`), and there are no unauthorized network calls or persistence mechanisms. The `SKILL.md` and `README.md` are descriptive and do not contain prompt injection instructions against the agent.