The Higress team at Alibaba Cloud has open-sourced HiRoute, a routing engine that runs locally between a coding agent and the model services it calls. The project was previewed at AGNTCon + MCPCon China and released at the Apsara Conference; it lists Codex, Claude Code, Qoder, DeepSeek Harness and Pi as front-end agents.
HiRoute separates two decisions. Model routing chooses which model handles each stretch of work; task routing lets the main agent hand independent research, implementation or testing to another agent. Either layer can run on its own.
Routing, configuration and logs stay local, while inference stays at whichever service is connected. A subscription can be folded into the same plan as a pay-per-call API key through CLIProxyAPI. A decision model — from Alibaba Cloud's Bailian or TypeSafe's Jev, reached through OpenRouter — classifies each task's category and difficulty and scores the preceding stage against thresholds the user sets; simple work goes to an economy group, hard work to a primary group.
One constraint is named explicitly: switching models mid-task usually breaks the KV cache. ContextHold keeps the chosen model stable while the context prefix continues, and re-decides only when compaction or history rebuilding breaks the old prefix. A Sessions page records which model actually ran, token usage and how earlier stages were scored.
Our takeThe binding constraint here is caching, not routing: every switch re-prefills the context, so the router only moves at points where a rebuild is already happening. That ties routing granularity to context structure rather than to task difficulty — a real limit, and the honest place to put it. Whether one set of thresholds survives across projects is the open question.