yjh051108/dsh-routing-suite

What it solves

This project addresses the issue of "model laziness" and inconsistency in long-range AI tasks. It provides a framework to manage how an AI agent handles different types of tasks by routing them to specific thinking modes (personas) and enforcing a structured, graded protocol to ensure the agent remains diligent and focused throughout the entire process.

How it works

The suite consists of three primary components:

1. Runtime Injector: A management layer that allows for the injection, hot-reloading, and unloading of tools and routing logic without needing to restart the environment.
2. Thinking Mode Router: A set of presets that route tasks based on the model's persona. It uses specific behaviors (such as spec for planning, react for execution, and mixed for traps) to guide the model's reasoning process and prevent it from drifting off-task.
3. Graded Task Protocol: A session-level protocol that breaks tasks into a strict sequence: brainstorming $\rightarrow$ specification $\rightarrow$ planning $\rightarrow$ injection $\rightarrow$ check-ins $\rightarrow$ closing $\rightarrow$ final verification. It includes "red-team gates" and audit endpoints to monitor and enforce quality.

Who it’s for

Developers and researchers using the DSH (DeepSeek-AI) framework who want to improve the reliability, diligence, and structure of their AI agents during complex, multi-step tasks.

Highlights

• Anti-Laziness Mechanisms: Includes empirical data and measurement tools to track and combat attention decay in long-range tasks.
• Dynamic Routing: Implements task-aware reasoning modes that can be tailored to different models (e.g., Pro vs. Flash).
• Efficient Injection: Uses agent/pre-step to inject guidance alongside user messages, reducing API costs by avoiding extra calls.
• Comprehensive Tooling: Provides a suite of dev_* tools for routing status, mode switching, and sub-agent management.