Reasoning Router

A router that asks a classifier how much reasoning each step of a coding session needs and applies that effort to the outgoing request, so one strong model handles quick edits and hard debugging without you switching effort levels by hand. It is the successor to the same author's archived opencode-jev-router plugin.

Jev is the default classifier, but the README says the choice is configuration rather than a dependency: the project must not depend on Jev or any single provider. Cloudflare Clef, Laya, Kev, OpenAI Decisions, CLM, Microsoft-Decision-1, Liquid d1 and SGLang are also supported.

For each primary request the router sends a bounded summary of the recent conversation to the classifier, which picks a reasoning effort. The choice is added mid-conversation as an OpenAI configuration_update item or an Anthropic effort-only system message, and earlier updates stay in place.

On SWE-bench Verified with GPT-6 Astra, Jev-routed effort solved every task that fixed high effort solved: a short Flask fix used 52% fewer output tokens and about half the time, while harder Django and pytest issues saved only 5-7%. The README calls these selected examples, not an expected saving.

OpenCode V2 through the @reasoning-router/opencode plugin, Pi through @reasoning-router/pi (Claude models only), and Codex CLI or any Responses or Messages client through the standalone proxy at http://127.0.0.1:4320/v1.

Bounded excerpts of recent user and assistant text, up to 8 recent tool results with tool names and error flags, a short failure summary and the model ID. Hosted-tool and computer-use payloads are not sent, and logging keeps metadata only, never prompts or tool output.

The request still runs at the fallback effort, high by default. The README notes the project makes no claim about how well any classifier picks effort for your workload, and lists alpha limits: Node.js 24.x, OpenCode V2 2.0.18, Pi 1.0.4 and Codex CLI 0.159.0 or 0.161.0.