Preface

Running agent tasks in DeepSeek Harness (DSH), the toolset and prompts exposed to the model in the first request directly affect subsequent reasoning trajectories. Community evaluations (xiaobright/modeltest) show: the Standard preset exposes 25 tools in the first round, scoring around 91/92 on Project2 V4.1b; the Minimal preset has only a line of prompt plus 2 tools in the first round, scoring around 99/96, but the toolset is too narrow to support real-world engineering. If you want to retain the Minimal first-round trajectory advantage while still being able to call the complete Standard tool catalog later, here’s the community plugin dsh-anchored-standard.

What is this

dsh-anchored-standard is a DSH workflow plugin released by maintainer Jungod1121 (current version v1.2.1, MIT license). It implements a two-stage agent preset: the first round uses a streamlined prompt aligned with Minimal and a few tools to complete “anchoring,” then after the first persistent tool/call or response in the session record, it automatically switches to the full Standard tool catalog (25 tools).

The design concept comes from the community project xiaobright/dsh-anchored-standard. In the same reference evaluation, this approach scored around 98/99 on Project2 V4.1b (n=2, single frozen task, not representative of all scenarios).

How it works

The core of the plugin is two-stage promotion:

sequenceDiagram
    participant U as User
    participant A as Agent loop
    participant M as DeepSeek V4 Pro

    U->>A: first message (new session)
    A->>M: request #1<br/>task-matched anchor persona<br/>tools: bash + read (± edit/write)
    Note over M: reasoning chain anchors on the task-matched scaffold
    M->>A: first tool/call
    Note over A: promotion signal (durable session event)
    A->>M: request #2<br/>same persona · all 25 Standard tools
    Note over M: full-capability work continues on the anchored trajectory
  1. The first user message in the session determines the anchoring type:
Task Type Anchor Persona First-round Tools
spec — Fix / Maintain / Debug Minimal original prompt bash + read + edit
react — Build from scratch / Create Hands-on prompt bash + read + write
weak — Ambiguous intent Model self-classifies (Pro uses classify instruction) bash + read

glob / grep do not appear in any bootstrap phase tool catalog; edit / write are considered anchor-safe tools. In the first request, the anchor persona is the only prompt segment, and runtime context is cleared.

  1. After the session produces the first persistent tool/call, subsequent requests see the full Standard catalog; the anchor persona remains unchanged, and other segments like plan-mode are restored. The mode and promotion state are written to persistent session events and remain valid after refreshing or restoring the session.

Installation and activation

The plugin is distributed via the DSH installer bundle. Recommended command:

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

After execution, restart DeepSeek Harness. The bundle copies the preset to $DSH_HOME/.agent-presets/anchored-standard/ (idempotent operation, won’t overwrite local modifications). Create a new blank session and select Anchored Standard (experimental) to use it.

To set it as the default preset for new sessions, edit $DSH_HOME/settings.yaml:

# $DSH_HOME/settings.yaml
agent-presets:
  default: anchored-standard

You can also manually copy the preset directory:

dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
cp -R preset "$dsh_home/.agent-presets/anchored-standard"

Uninstallation:

dsh plugin --profile web remove dsh-anchored-standard
rm -rf "${DSH_HOME:-$HOME/.dsh}/.agent-presets/anchored-standard"

Verify two-stage switching

Export session JSONL and check request/header events. The first round header contains only bash / read (and possibly edit / write), while after promotion, the header includes all 25 tools:

seq 18  -> 2 tools  ['bash', 'read']                       # request #1
seq 137 -> 25 tools ['ask_user_question', 'bash', ...]     # request #2 (promoted)

Trade-offs with Minimal and Standard

Compared to Gains Costs
vs Minimal 23 more tools (glob / grep / edit / write / web_search / subagents / workflows / skills / goals / todo etc.) One more round of reasoning in bootstrap phase
vs Standard Task-matched trajectory anchoring (reference evaluation 98/99 vs 91/92) + cleaner first request First-round tool catalog determined by keyword classification; ambiguous tasks fall back to weak anchoring

Use cases and notes

Who it’s for

  • Using DeepSeek V4 Pro (reasoningEffort: max) in DSH to handle engineering tasks, wanting the first-round reasoning trajectory to be close to Minimal while still being able to call full Standard capabilities later.
  • Verified on Harness 0.1.0-rc.6; recommend rechecking compatibility before upgrading upstream.

Who it’s not for

  • V4 Flash models: README notes that Flash is insensitive to harness style changes and doesn’t need this preset.

Pre-installation notes

  • The plugin runs with current DSH process permissions; check the source code and MIT license before installation.
  • This is a community project, not an official DeepSeek preset, with no affiliation to DeepSeek / High-Flyer.
  • The 98/99 evaluation scores come from xiaobright/modeltest Project2 V4.1b, n=2, single task, which is reproducible reference evidence, not a general guarantee.

Conclusion

dsh-anchored-standard uses a two-stage strategy of “first-round anchoring + subsequent full tools” to balance Minimal’s reasoning trajectory with Standard’s tool breadth in DSH. If you’re debugging the first-round behavior of Pro models, you can start by selecting this preset in a new session.