Route2
Reasoning models arrived with a dial: low, medium, or high effort. The dial was polite. It didn't disturb your cached context. But it moved money, because reasoning tokens are the expensive part. And it came without instructions, so turn by turn you had to decide how much thinking each task deserved, mostly by gut feel.
The first fix was to do it by hand. People wrote blog posts with rules of thumb: low effort for file reads and small edits, high for planning and debugging. Sensible advice. But it gave you a new job. You became the router, deciding dozens of times per session whether this turn deserved expensive thinking. Nobody wants that job, and nobody is good at it.
Then came Jev, a decision model built for exactly this kind of small decision. Look at the task, pick from a fixed set of options. No text generation, no parsing. It seemed like the right shape for routing. But Jev was a cloud API, which meant every routing decision sent your code and prompts to someone else's server. For the component that decides how your agent thinks, that felt wrong.
So I built route2. Same idea, but the classifier lives on the machine. A 0.6B model looks at each turn and picks low, medium, or high reasoning effort for Codex. Because it runs locally, the routing decision costs almost nothing, and your code never leaves your laptop. And the decision isn't final. On failure, stalled progress, or new instructions mid-task, it re-evaluates. A misroute corrects instead of compounding.
It's early, and the numbers could change. But the direction seems right. Picking reasoning effort by hand was always going to be a transitional phase. The judgment of how much to think is itself a small, cheap decision, and it belongs on your machine, not in your head.