AI Agent Hub
Back to plugins
🤖

dsh-shift-router

Model Inference Updated 2026.08.25

Run the following command in DeepSeek Harness:

dsh plugin install green-dalii/dsh-shift-router

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

Run dsh plugin install green-dalii/dsh-shift-router in DeepSeek Harness to install the plugin; the source is available at https://github.com/green-dalii/dsh-shift-router .

About this plugin

Routine turns should not burn flagship tokens, yet consequential turns should not be left to a cheap model. dsh-shift-router invokes a lightweight LLM Judge before every top-level agent step to classify the message as fast or smart, then routes the entire turn-thinking, tool calls, code edits-through the appropriate tier. Upgrades are instant; downgrades require a sliding-window majority, preventing thrash.

Built-in runtime resilience handles 429, 5xx, and quota errors with exponential-backoff cooldowns (1 min to 6 h cap), re-resolving to the next healthy model within the same tier without cross-tier drift or context loss. Complex tasks engage a CTO orchestration mode where the Smart model plans and delegates to Fast subagents, with plugin-enforced hard caps on rounds and escalations that deny the subagent tool outright when exceeded. Per-tier token and throughput telemetry, plus an optional USD pricing table, make session cost transparent.

Built for developers and teams running multi-model inference on DeepSeek Harness who want both quality and cost efficiency. Zero-config startup means routing works immediately; every parameter is adjustable live via the GUI settings card or /router config commands with no restart required.

Use Cases

  • Route routine turns to a low-cost model and escalate complex ones to a flagship tier automatically
  • Survive 429 or 5xx outages with exponential-backoff cooldown and same-tier failover
  • Delegate implementation to Fast subagents under a Smart-tier CTO plan with enforced round caps

Best For

  • Developers running multi-model inference on DeepSeek Harness
  • Teams seeking to reduce token spend by tiering model calls by task complexity
  • Platform engineers designing fault-tolerant model fallback and degradation pipelines