openclaw - 💡(How to fix) Fix Feature: Document and support intent-based model routing as a first-class pattern [2 comments, 1 participants]

Summary

OpenClaw users running multi-provider setups (e.g. Codex + Anthropic + xAI) would benefit from a documented pattern and optional built-in support for intent-based model routing — classifying each prompt and routing it to the cheapest appropriate model.

Background

Filed originally as part of #61262 (now closed — the auth/billing-disable bug is fixed). The routing concept was noted as a separate product follow-up.

Problem

Today, model selection in OpenClaw is manual (/model <name>) or static (default model in config). Users who want cost-optimized routing across providers have to build external classifiers and wire them in through AGENTS.md scripting. This works but is fragile and undocumented.

Proposal

1. Document the pattern. Add a guide showing how to build an external intent classifier that integrates with OpenClaw's /model and session_status(model=) APIs. Cover: classification → model switch → execution → reset flow.

2. Consider a built-in routing hook. A lightweight config option (e.g. models.router: "auto" or a pre-turn hook) that lets users define routing rules without external scripts. Could be as simple as keyword/regex rules mapping to models, or as flexible as calling a local classifier endpoint.

3. Expose routing telemetry. When model switches happen mid-session (via router or manual), surface per-model token usage in /status so users can measure cost impact.

Prior art

• I built openclaw-intent-router as an external classifier. Early results: ~37% Codex, 33% Opus, 30% xAI — significant cost reduction vs. Opus-for-everything.
• The AGENTS.md integration pattern works but requires per-turn CLI calls and manual state management.

What this is NOT

• Not asking for OpenClaw to pick models automatically with no user control.
• Not asking for provider-specific logic in core — just a hook point and documentation.