Back to home

mikeyoubeach

dsh-minimal-v3

Windows-friendly full-minimal DeepSeek Harness agent preset (Minimal V3): minimal persona, platform shell (pwsh/bash), common file/search tools, no auto-inject/compaction.

Stars
0
Language
Created
Aug 15, 2026
Updated
Aug 15, 2026

Introduction

dsh-minimal-v3

A Windows-friendly, full-minimal agent preset for DeepSeek V4 Pro. It starts from the official minimal preset and keeps everything that makes minimal clean — a fixed complete persona, no automatic instruction injection, no context compaction — but swaps minimal's pure bash persistent terminal for a platform-adaptive shell (Windows → pwsh, Linux/macOS → bash), and brings in a handful of standard's common file/search tools.

This is a community project. It is not an official DeepSeek preset and is not affiliated with or endorsed by DeepSeek.

Why this matters: DeepSeek V4 Pro conditions strongly on the API-visible tool catalog. Official minimal scores well in evaluations (see modeltest) but only runs on Linux + persistent bash — on Windows it breaks because bash is absent. This preset keeps minimal's "purist" character while working on Windows too, and includes the tools you actually need for engineering (read/write/edit/glob/grep) instead of leaving you with just two.

Compared to official minimal

DimensionOfficial minimaldsh-minimal-v3
personacomplete promptsame (You are a helpful software engineer assistant., complete: true, includeRuntimeContext: false)
shellpersistent bash (unusable on Windows)platform-adaptive: Windows→pwsh, others→bash (non-persistent)
auto-injectionnonenone (does NOT load AGENTS.md/CLAUDE.md/local overlay)
context compactionnonenone
toolsonly bash + str_replace_editorpwsh|bash, read, write, edit, read_image, glob, grep, ask_user_question, todo_write
subagent/workflow/goals/skills/webnonenone (kept minimal)

Final tool catalog

  • Windows: pwsh, read, write, edit, read_image, glob, grep, ask_user_question, todo_write
  • Linux/macOS: bash, read, write, edit, read_image, glob, grep, ask_user_question, todo_write

Constant for the whole session — no two-phase bootstrap/promotion.

Compatibility

Developed and verified against:

  • DeepSeek Harness (DSH) 0.1.0-rc.6
  • Node.js ≥ 22.19 (only needed for the local self-test)

DeepSeek Harness is currently a developer preview and explicitly permits breaking changes. This preset references official packages @deepseek-ai/dsh-persona, dsh-tool-bash, dsh-tool-pwsh, dsh-tool-fs, dsh-tool-fs-search, dsh-tool-ask-user, dsh-tool-todo; if a newer Harness fails to load it, double-check whether any of those package names changed.

Install

Clone this repository, then copy the preset directory into the user preset root under the id minimal-v3.

PowerShell (Windows):

$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\minimal-v3'
if (Test-Path -LiteralPath $target) { throw "Preset already exists: $target" }
New-Item -ItemType Directory -Force -Path (Split-Path -Parent $target) | Out-Null
Copy-Item -Recurse -LiteralPath '.\preset' -Destination $target

Linux/macOS:

dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets/minimal-v3"
cp -R preset/. "$dsh_home/.agent-presets/minimal-v3/"

Fully restart DeepSeek Harness, create a blank session, and select Minimal V3 in the preset picker. Do not switch an active session from a different preset.

Usage notes

  • Best paired with DeepSeek V4 Pro (the edge here is mostly about Pro's sensitivity to the tool catalog; V4 Flash is already stable across harnesses and thinking modes).
  • On the first request, consider having it read a file first to anchor the trajectory before giving it the real task.
  • This preset has no skill tool: if you need open-design-style custom skills, use a full preset or add tool-skill back in your own copy.
  • Because auto-injection is off, project AGENTS.md/CLAUDE.md are not loaded automatically — read them explicitly or write the rules into the conversation. That is intentional ("pure minimal").

Verify

Local zero-dependency self-test:

npm test

To confirm the preset is discoverable (outside the running Harness):

import { discoverPresets } from "@deepseek-ai/dsh-agent-presets/lib/types/discovery.js";
const presets = await discoverPresets([{ path: "~/.dsh/.agent-presets", trust: "user" }]);
console.log(presets.map(p => ({ id: p.id, broken: p.broken ?? false })));

minimal-v3 should appear with broken === false.

For deeper runtime verification, export the session JSONL and check that the request/header tools match the catalog above, and that user/message contains no <system-reminder> AGENTS.md / skills injection.

Important behavior

  • If the first model response makes no tool call, nothing changes (this preset is constant; there is no two-phase promotion).
  • A failed tool execution does not stop the model from calling the others.
  • The preset carries the same trust level as shell access — review preset/agent.cordis.yml before installing.
  • This preset performs no network requests and adds no telemetry.

Official ecosystem guidance

DeepSeek currently asks community plugin authors to publish plugins in their own GitHub projects and add the dsh-plugin topic for discovery. The official repo does not currently accept external PRs and does not mandate a community template. See the official CONTRIBUTING.md.

License

MIT. preset/agent.cordis.yml was written following the shape of the official DeepSeek Harness minimal preset; original copyright and MIT notice are retained in NOTICE.