AI Agent Hub
Back to plugins
🤖

dsh-minimal-anchor

Model Inference Updated 2026.08.25

Run the following command in DeepSeek Harness:

dsh plugin install rand0wn/dsh-minimal-anchor

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

Run dsh plugin install rand0wn/dsh-minimal-anchor in DeepSeek Harness to install the plugin; the source repository is at https://github.com/rand0wn/dsh-minimal-anchor

About this plugin

Every fresh DeepSeek Harness session hands the model the full configured toolset on message one — file edits, bash, subagents, jobs — even when the user simply asks to take a look at this repo and see what is going on. That flood of tool-schema descriptions crowds the context window and nudges the model toward premature action instead of quiet exploration. dsh-minimal-anchor addresses exactly this first-turn overload.

It is a fully passive plugin: nothing to invoke, no commands to run. It hooks the Harness system-prompt assembly pipeline and, on the very first assembly for each new scope, does two things. It trims the tool list down to a configurable whitelist (default: read, glob, grep) and prepends a short structural preamble that frames the session as exploration-first. From the second turn onward, every subsequent assembly passes through completely untouched, and the full toolset is automatically restored.

It is well suited for dsh users who typically boot a session with a single natural-language prompt and want the model to spend its first turn reading and orienting rather than reaching for bash or file-edit tools. Install it and every new session gets a calmer, more focused first turn at zero ongoing cost.

Use Cases

  • Boot a session with a single natural-language prompt and want the model to read before it acts
  • Fresh-session tool schemas overload the context window and skew early reasoning
  • Frame turn 1 as codebase orientation before any file edits or shell calls

Best For

  • dsh users who typically start a session with a casual exploratory prompt
  • Power users who want stable first-turn reasoning without tool-schema noise
  • Developers composing dsh profiles in TypeScript who need tunable whitelists and preambles