AI Agent Hub
Back to plugins
🤖

dsh-dual-auto

Model Inference Updated 2026.08.25

Run the following command in DeepSeek Harness:

dsh plugin install lengquan88/dsh-dual-auto

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install lengquan88/dsh-dual-auto in your terminal to install the plugin; source code is available at https://github.com/lengquan88/dsh-dual-auto .

About this plugin

When making frequent model-inference calls through DeepSeek Harness, picking one fixed model either wastes budget on trivial queries or under-delivers on complex ones. dsh-dual-auto introduces a six-criteria routing engine (task length, context size, domain coverage, rule conflict, confidence, and novelty) so that simple requests are served directly by deepseek-v4-flash while complex ones are upgraded to deepseek-v4-pro, with automatic degradation back to flash on failure so the session never breaks.

The distinguishing feature is its escape-learning loop: if a flash-served answer is marked incorrect, the plugin records the request fingerprint, persists it to local JSON logs, and forces an upgrade on the next identical fingerprint. Over time the system learns which task shapes should not be cheaply served, and its storage format stays interoperable with the project's Python model_router.py for easy extension.

If you want your DeepSeek Harness deployments to decide for themselves when to save and when to stay precise without hand-writing routing logic, this lightweight plugin fills exactly that gap.

Use Cases

  • Route a mixed workload of simple and complex queries in an automation pipeline, letting the six-criteria engine pick flash or pro automatically to avoid one-size-fits-all cost
  • Build up a fingerprint memory over time so repeated failing tasks are force-upgraded without manual rule maintenance
  • Switch seamlessly between flash and pro with automatic degradation on failure, keeping sessions alive and auditable
  • Plug into an existing DeepSeek Harness setup that already uses Python model_router.py, since fingerprint formats load directly

Best For

  • Developers running daily inference pipelines on DeepSeek Harness
  • AI application engineers who need to balance response quality against token cost dynamically
  • Small AI teams that want to retire hand-written routing rules in favor of a self-learning router