Introduction

In the Standard plugin scenario of DeepSeek Harness (DSH), a common requirement is to restrict the first request to a narrower surface, and then restore full Standard capabilities and Skills after the model truly starts calling tools.

Starfie1d1272/dsh-v4-anchor does only this layer: for DeepSeek V4 top-level Standard sessions, it uses a minimal RL-shaped surface before the first persistent tool/call; after the first persistent tool/call, it restores the original Standard prompt, runtime contexts, full tool directory, Skill Catalog, skill loader, and AGENTS.md injection.

What is this

Starfie1d1272/dsh-v4-anchor is a DeepSeek V4 initial RL anchoring plugin for DeepSeek Harness Standard. The maintainer is Starfie1d1272 with an MIT license.

This plugin is not an official DeepSeek project. The project is positioned as a time-sensitive behavior patch, not a long-term Router framework; if DSH official or DeepSeek V4 fixes the corresponding behavior, it can narrow its scope, become a no-op, or stop maintenance and archive.

Scope

The plugin only takes effect for the following sessions:

  • DeepSeek Harness 0.1.0-rc.7
  • Official standard preset
  • Top-level sessions
  • Model is DeepSeek V4, e.g., deepseek-v4-flash, deepseek-v4-pro

Non-DeepSeek V4 models, other presets, and subagent scenarios are no-op.

v0.1.0 requires model selection to be completed before the first request. DSH rc.7 does not expose the Web’s session-local picker state to regular plugins, so it does not promise that dynamic model switching when bootstrap is not yet promoted will strictly maintain the same filtering/no-op semantics.

Core Features

First Request: Minimal RL-shaped Surface

Before the first persistent tool/call, eligible Standard sessions use the following surface:

system:
You are a helpful software engineer assistant.

tools:
bash / pwsh
str_replace_editor

Before the first request, runtime contexts are temporarily removed, and the automatically injected AGENTS.md and Skill Catalog are temporarily hidden.

After promotion: Restore Full Standard

Once the session makes its first persistent tool/call, the plugin restores:

  • Original Standard prompt
  • Runtime contexts
  • Full tool directory
  • Skill Catalog
  • skill loader
  • AGENTS.md injection

str_replace_editor is only used for bootstrap and is removed from the additionally exposed tools after promotion.

Promotion Transition Reminder

After promotion, a one-time promotion transition reminder is injected to remind the model to re-check the restored Skills/Tools.

This part is an experimental fix that is reasonable but has not yet been benchmarked independently.

Installation and Enablement

First, confirm the Node environment meets the requirements:

^22.19.0 || >=24.0.0

The installation command can be written as:

dsh plugin --profile web add dsh-v4-anchor

You can also install a fixed GitHub release:

dsh plugin --profile web add github:Starfie1d1272/dsh-v4-anchor#v0.1.0

After installation, restart the Web profile:

dsh web

This plugin directly reuses the built-in:

@deepseek-ai/dsh-tool-str-replace-editor

It will not install a second DSH runtime package.

Post-Installation Check

This step is used to view the final combination of the profile:

dsh --profile web --dump-config

In a new Standard session that meets the criteria, the first request should see:

system:
You are a helpful software engineer assistant.

tools:
bash / pwsh
str_replace_editor

After the first persistent tool call, the original Standard prompt, runtime contexts, full tool directory, Skill Catalog, skill loader, and AGENTS.md injection should be restored.

Testing and Packaging Check

Testing does not call any LLM/API and does not consume DeepSeek Tokens.

Run check:

npm run check

Check packaging content:

npm run pack:check

Use Cases and Notes

Suitable for scenarios where you are using DeepSeek V4 top-level Standard sessions in the DSH Web profile and want the initial surface to remain narrow while restoring full capabilities later.

Notes before use:

  1. The plugin runs with the current DSH process permissions; you should check the source code, dependencies, and license before installing.
  2. Only applicable to DeepSeek Harness 0.1.0-rc.7, official standard preset, top-level sessions, and DeepSeek V4 models.
  3. v0.1.0 requires model selection to be completed before the first request; dynamic model switching when bootstrap is not yet promoted does not promise to maintain the same semantics.
  4. The promotion transition reminder is an experimental fix that has not yet been benchmarked independently.
  5. The project is a time-sensitive behavior patch, not suitable as a long-term Router framework.

Links

GitHub Repository:

https://github.com/Starfie1d1272/dsh-v4-anchor

The directory page link is not provided with a confirmed URL, so no unverified links are given here.