Model Router

**Judgment:** A practical, lightweight model selection router well-suited for OpenClaw-style, multi-model setups where you want deterministic, cost-aware routing without building your own logic from scratch. **What it does:** - Reads a models configuration (OpenClaw-style or simple JSON) with fields like `capabilities`, `cost_score`, and `tags`. - Heuristically classifies incoming tasks (short/simple vs long/complex) and combines that with model metadata. - Applies explicit overrides from user/caller hints when present. - Outputs the chosen model plus reasoning, via CLI or as a library call. **Strengths:** - Encourages **explicit, auditable routing criteria** instead of ad-hoc `if/else` logic scattered in application code. - **Cost-conscious by design**, defaulting to cheaper models for routine tasks while escalating when complexity or fidelity demands it. - Very small surface area (one main script + example config), which aids comprehension and debugging. **Risks / Limitations:** - **Heuristic complexity detection** (short/simple vs long/complex) may misclassify edge cases, leading to suboptimal model choices unless carefully tuned. - No visible advanced features like dynamic cost tracking, latency feedback, or automatic failover; behavior appears to depend entirely on static metadata and simple rules. - **Safety is configuration-dependent**: if you route sensitive or high-risk content, you must ensure your model list and tags reflect provider/regulatory constraints; the router itself doesn’t enforce content or privacy policies. **Recommended scenarios:** - You maintain **multiple LLM providers/models** and want a central, transparent router instead of hard-coding model selection throughout your codebase. - You’re building **server middleware or batch jobs** that must balance cost vs quality and can accept a simple heuristic for complexity. - You want a **starting point for a custom router** that you can extend with your own scoring logic, rather than designing one from scratch. Less ideal if you need highly sophisticated routing (e.g., live SLO enforcement, automatic failover, or rich safety policy checks) without investing in additional custom logic around this skill.