Skip to content

Jungod1121/dsh-anchored-standard

30Last commit Aug 15, 2026

dsh-anchored-standard DSH plugin

Anchored Standard optimizes DeepSeek V4 Pro's reasoning by controlling the first request trajectory. It starts with a minimal persona and only 2-3 tools (bash, read, optionally edit/write), then promotes to all 25 Standard tools after the first tool call. This preserves the clean trajectory of the Minimal preset while giving the agent full capability, achieving 98/99 task scores in reference evaluations.

How to install the dsh-anchored-standard DSH plugin

dsh plugin --profile web add github:Jungod1121/dsh-anchored-standard

Copying does not run this command. Review the repository and version before installing the dsh-anchored-standard DSH plugin.

dsh-anchored-standard DSH plugin data source

dsh-anchored-standard DSH plugin snapshot date: Aug 16, 2026

discovered

What the dsh-anchored-standard DSH plugin can do

  • Two-phase bootstrapping: first request uses task-matched anchor persona and minimal tools, second request unlocks full Standard catalog
  • Task-adaptive anchoring: spec (fix/debug), react (build/create), or weak (ambiguous) personas based on first user message classification
  • Automatic promotion after the first durable tool call, with state preserved across session refresh/resume
  • Achieves 98/99 ability scores in xiaobright/modeltest Project2 V4.1b evaluation
  • No post-anchor injection: measured to avoid harming Pro performance

Where the dsh-anchored-standard DSH plugin fits

  • Fix, maintain, or debug existing codebases with a focused first turn (spec anchor)
  • Build or create new projects from scratch with a hands-on doer persona (react anchor)
  • Handle ambiguous tasks where the model self-selects the best anchoring strategy (weak anchor)
  • Leverage full tool catalog (glob, grep, edit, write, web_search, subagents, workflows, etc.) after the first call
  • Integrate into existing DeepSeek Harness workflows as the default preset for new sessions

Who the dsh-anchored-standard DSH plugin is for

  • DeepSeek V4 Pro users seeking maximum reasoning performance on harness-based tasks
  • Developers and researchers using DeepSeek Harness who want to optimize agent trajectory

dsh-anchored-standard DSH plugin limitations

  • Targeted specifically at DeepSeek V4 Pro with reasoningEffort: max; V4 Flash does not benefit
  • First-request tool catalog is decided by keyword classification – ambiguous tasks fall back to the weak anchor, relying on model self-pick
  • The 98/99 ability scores are from a single frozen task with n=2; not a universal guarantee across models or workloads
  • Requires DeepSeek Harness 0.1.0-rc.6; upstream breaking changes may affect compatibility

dsh-anchored-standard DSH plugin: from the repository README

Quoted from the Jungod1121/dsh-anchored-standard README, the upstream source of the dsh-anchored-standard DSH plugin. Copyright remains with the original authors.

**Anchor the first request. Unlock the full stack.** A two-phase agent preset for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness): bootstrap DeepSeek V4 Pro with the Minimal-aligned prompt and two tools, then expose the complete Standard catalog after the first durable tool call or reply. ![harness](https://img.shields.io/badge/harness-0.1.0--rc.6-4e8cff) [![license](https://img.shields.io/badge/license-MIT-2ea44f)](./LICENSE) [![stars](https://img.shields.io/github/stars/Jungod1121/dsh-anchored-standard?style=flat)](https://github.com/Jungod1121/dsh-anchored-standard/stargazers) [![dsh-plugin](https://img.shields.io/badge/topic-dsh--plugin-8a5cf5)](https://github.com/topics/dsh-plugin) [中文说明](./README.zh-CN.md) · [Landing page](https://jungod1121.github.io/dsh-anchored-standard/) · [Design origin](https://github.com/xiaobright/dsh-anchored-standard) </div> --- ## The one-line idea V4 Pro's ability ceiling is high — but which reasoning strategy it uses is decided by what the **first API request** shows it. The community evaluation at [`xiaobright/modeltest`](https://github.com/xiaobright/modeltest) measured: | Preset | First request sees | Full task sees

Read the full READMERepository license: MIT

dsh-anchored-standard DSH plugin questions

How do I install dsh-anchored-standard?

Run `dsh plugin --profile web add github:Jungod1121/dsh-anchored-standard` and restart DeepSeek Harness. The installer copies the preset into `$DSH_HOME/.agent-presets/anchored-standard/`. Then create a new session and select 'Anchored Standard (experimental)' from the preset list.

How do I make it the default preset for all new sessions?

Edit your `$DSH_HOME/settings.yaml` file and add `agent-presets: default: anchored-standard`. This will automatically select Anchored Standard for every new session without manual selection.

How can I verify that the preset is working correctly?

Export a session JSONL and inspect the `request/header` events. The first header should contain only 2 tools (bash, read), while every later header should list all 25 Standard tools (e.g., ask_user_question, bash, ...). This confirms the two-phase promotion is active.

Is this preset compatible with V4 Flash?

No, it is designed specifically for DeepSeek V4 Pro with `reasoningEffort: max`. V4 Flash generalizes well across harnesses and does not need this optimization. Using it with Flash may not provide any benefit.

What are the differences between Anchored Standard, Standard, and Minimal presets?

Standard starts with full 25 tools and a long prompt. Minimal starts with only 2 tools (bash, read) and a 1-line prompt but never expands. Anchored Standard starts like Minimal (2 tools, clean prompt) but automatically promotes to all 25 tools after the first tool call, achieving higher task scores (98/99 vs 91/92 for Standard).