AI Agent Hub
Back to plugins
🤖

dsh-v4-anchor

Model Inference Updated 2026.08.25

Run the following command in DeepSeek Harness:

dsh plugin install Starfie1d1272/dsh-v4-anchor

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

In DeepSeek Harness, install the plugin with the command dsh plugin install Starfie1d1272/dsh-v4-anchor; source code: https://github.com/Starfie1d1272/dsh-v4-anchor.

About this plugin

dsh-v4-anchor is a deliberately minimal DeepSeek Harness plugin designed to address potential inaction in DeepSeek V4 models during standard sessions due to overly complex system prompts. Existing experiments indicate that exposing a minimal RL-shaped surface in the first request—such as stripping runtime contexts and hiding AGENTS.md and Skill Catalog—can significantly boost tool call rates and practical task execution. This plugin replicates that mechanism: for qualifying DeepSeek V4 sessions, it automatically applies bootstrap filtering on the first request and restores full Standard capabilities after the first real tool call, while injecting a one-time reminder to ensure the model re-engages with the newly restored tools. It is not intended as a router or complex steering replacement but as a timely behavioral patch, helping researchers and developers easily test and validate the impact of RL-shaped bootstrap on model trajectories.

Use Cases

  • When starting a new task in a DeepSeek V4 standard session.
  • When aiming to improve tool call rates and practical action-taking.
  • During RL-shaped bootstrap behavioral experiments.

Best For

  • Researchers using DeepSeek V4.
  • Developers looking to optimize model behavior.
  • Engineers interested in model inference mechanisms.