The Right Model for Every Message: Automatic Model Routing

Summary
Route each request to the most suitable model automatically — small/fast local model for simple turns, vision model when an image arrives, big model for hard reasoning — instead of pinning every agent to a single fixed model.

Component(s) Affected
done
Core / Backend (app, agents, config, providers, utils, local_models)
done
Console (frontend web UI)
not done
Channels (DingTalk, Feishu, QQ, Discord, iMessage, etc.)
not done
Skills
not done
CLI
done
Documentation (website)
not done
Tests
not done
CI/CD
not done
Scripts / Deploy
Problem / Motivation
Each agent is bound to one model. Users who run several local models (a fast 4B, a strong 9B, a vision model, an OCR model) must manually switch, or create duplicate agents per model. On 8 GB consumer GPUs this is painful: the heavy model burns time and VRAM on trivial messages, and image messages fail outright when the pinned model is text-only (the multimodal probe system already detects this — it just can't act on it).

Proposed Solution
Optional "Auto" model choice per agent with a rules engine:
By content: message contains image → vision-capable model; long document → large-context model.
By difficulty: heuristic or tiny-classifier routing between fast/strong text models.
By availability: prefer local; fall back to a configured cloud model if the local endpoint is down.
User-visible and editable rules (no black box), per-agent overrides, and a "why this model" note in the reply metadata.
Reuses the capability data QwenPaw already stores per model (supports_image, max context, probe results).
Alternatives Considered
Multiple agents, one per model: clutters the agent list and loses shared memory/history.
Manual switching: fine for one chat, unusable across 18 always-on channels where nobody is watching.
Additional Context
The provider system already tracks per-model capabilities and context sizes — routing is a decision layer on top of data that exists today.

Willing to Contribute
not done
I am willing to open a PR for this feature (after discussion).