Add an optional Decision Gate to speed up routine agent decisions

Every time you send a message in aider, the large language model is asked to work out the routine things about that message: what the intent is, whether a tool is needed, and what to do next. For a simple message this still costs one to three seconds and real money, even though the answer was obvious before any deep thinking began.

The idea is to put a small, fast decision model at the front door, before the big model runs. This small model does not write text. In one quick pass it returns typed answers with confidence scores, such as whether the request needs a tool, whether it can be answered from what is already in the conversation, and what the primary intent is. A thin routing layer then sends each message to the cheapest correct path, and whenever the small model is unsure it falls back to the full model exactly as today. Because the decision models are small and open, they can run on ordinary home hardware, a CPU or a small GPU.

The change is designed to be safe to adopt. It is off by default, so a user who does nothing sees the exact current behavior. When it is turned on it runs in shadow mode first, which means it only classifies the turn and logs its recommendation without changing routing, skipping the model call, or altering anything the user sees. It also fails open: on any error, timeout, or unexpected response it simply has no opinion, and the agent continues exactly as if the gate were never there. The agent can therefore only get faster when the decision is obvious, never worse.

We would like to contribute this as an optional feature. I have opened a pull request that adds a small self-contained gate module and a shadow-mode hook at the point where a message is about to be sent to the model. The concept and the reference implementation are by Andrea Bruno and are released under the Creative Commons Attribution 4.0 license. If aider adopts the idea or the code, we kindly ask to keep a short credit to Andrea Bruno and a link to the harness-superfast repository in the documentation. The decision models themselves, Von, OpenJev and Laya, are third-party open models; only the integration architecture and the routing method are ours. The white paper is called harness-superfast.