Why it mattersA 24-day-old community plugin now controls which reasoning mode a coding harness uses per task, and it ships a script that measures verification decay on your own session logs.
dsh-routing-suite, a three-part add-on for DeepSeek Harness, has reached 7,116 GitHub stars since developer yjh051108 created the repository on 14 August. The counts come from the GitHub REST API, retrieved 7 September: 7,116 stars, 151 forks and 38 open issues, 24 days after creation. DeepSeek Harness itself, the coding harness the suite plugs into, is in developer preview, and its maintainers have said breaking changes will happen.
The repository bundles three components under an MIT licence. A runtime injector adds, hot-reloads and removes dsh plugins without restarting the harness, and can promote a staged plugin into a permanent install. Two routing presets, router-standard and router-spec, classify each incoming task and pick a reasoning mode for it, so light questions and heavy planning work each get a matching mode. A graded task protocol walks long jobs through fixed stages: a brainstorm quiz to align on the goal, a specified plan, check-ins, a group close-out and a final check, with a red-team gate and an audit endpoint along the way.
A measurement of models getting lazy, with the script to rerun it
The project's research pages carry the author's own numbers: routing accuracy of 96 percent for the presets and prompt-cache hit rates of 92 to 94 percent for the guidance they inject. The more interesting claim is what the author calls model attention slack. In long sessions run without the protocol, the model's verification effort decayed toward the end of the task, with late segments showing one tool call and no image checks, while sessions run under the protocol kept verifying throughout. The comparison rests on three sessions of the same 3D task, so it is an observation, and the repository says so. What makes it useful is the measurement script that ships with it: anyone can run it against their own session log and get the same metrics for their own transcripts.
The current v0.3.0 release mostly repairs the suite's own assembly chain. Before it, the first message of every session went unclassified, and the injected guidance created a second API call for each user message, which doubled cost. Both fixes are tracked in linked issues. The suite targets dsh versions from 0.1.0-rc.6 up to, but not including, 0.2.0. The installer is a PowerShell script, and most of the documentation is in Chinese, with an English README maintained alongside.
For a team using or evaluating DeepSeek Harness, the growth number is the signal: a routing layer at 7,116 stars in 24 days means the harness's users want per-task control over reasoning modes, and they are getting it from community scaffolding rather than the harness itself. The pin to a preview version range is the cost, since any dsh release can break the suite. The laziness measurement travels beyond this harness: a script that scores late-task verification decay on your own session logs turns a hunch about model behaviour into a number you can track.
Source
Primary source: yjh051108/dsh-routing-suite on GitHub. Star, fork and issue counts from the GitHub REST API, retrieved 2026-09-07.
Source: GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.