# Model routing

> One routing and policy system decides which model does the work: your constraints first, then what your accounts can run, then the work itself.

- Canonical page: https://www.winskel.com/docs/model-routing
- Applies to: Winskel public download 0.7.0-alpha.1 (alpha channel), Apple Silicon, macOS 12 or later
- Content updated: 2026-10-06
- Agent reference: https://www.winskel.com/agents.md

## One routing and policy system

Every routing decision comes from one policy, applied in a fixed order:

1. A model or agent system you named. If it cannot run, nothing is substituted.
2. Availability: only models your installed Claude Code and Codex can run, skipping a provider that has reported it is out of quota.
3. The work: one strong model by default. Today that is Sonnet 5.5 when your Mac can run it, otherwise Opus 5.5 or GPT-5.6 Sol.
4. Your default model from setup, when none of those can run.

## When more models join

Winskel decides when the work warrants more than one model. Today that happens when:

- The work involves security, authentication, migrations, infrastructure, or architecture: Winskel plans it, may split it or run parts at the same time, and adds an independent review when the risk is high.
- The objective is too large for one step: Winskel plans it into steps.
- A step's checks fail: the same model gets one repair turn, then the step moves once to a stronger model. If that also fails, the objective stops as Failed and shows the reason.
- You ask for separate or parallel parts, two agent systems, or a review: Winskel plans it and keeps your split.

## What each decision leaves behind

- The model chosen and a one-line reason, shown in the chat when the work starts
- Whether the work finished or failed, and why it stopped
- The checks Winskel observed
- Whether a step moved to a stronger model, and any retry
- Your Approve or Request changes on the result

This record explains routing. Using it to improve future routing is still in development.

## Learning from outcomes

Winskel records which model ran, whether the work completed, what checks passed, what review found, and whether a repair was needed. Learning from those outcomes is still in development: routing follows a fixed, versioned policy today, and outcomes from other developers do not influence your routing. Your explicit constraints always win.

## Your choice wins

A model or agent system you name in the objective overrides routing, retries, and escalation. See [choosing models](https://www.winskel.com/docs/choosing-models).

## What is not exposed

The product reports which model owned each step and a short reason. It does not publish private planner prompts or hidden evaluation instructions.