dsh-web-search-ddg
aooyoo
Zero-token DuckDuckGo search provider for the DeepSeek Harness (DSH) web seam — local headless browser, no API key, no m…
PROJECT TOPICS
PROJECT README
English | 中文
Native Windows Subprocess Runtime & Optimized Agent Presets for DeepSeek Harness (DSH)
dsh-subprocess-win32 is a specialized Cordis runtime plugin and preset management bundle designed for DeepSeek Harness on Windows. It solves platform-specific process management limitations and provides streamlined, token-efficient Agent presets out of the box.
subprocess-win32)Minimal (Windows): Lightweight, minimal token overhead with persistent Git Bash and str_replace_editor.Anchored Standard (Windows): Two-phase progressive loading preset. Starts with zero tool bloat and unlocks heavy tools (PowerShell, web, subagents) on demand.doctor), automated setup (setup), atomic updates (update), rollback, and clean uninstallation (uninstall)—no hidden background changes or silent PATH tampering.dsh-subprocess-win32?DeepSeek Harness stock terminal tools rely on Unix-oriented process handling, which can cause process leaks, improper signal handling, or terminal stalls on Windows.
dsh-subprocess-win32 introduces a dedicated Windows process adapter registered directly in DSH's Cordis runtime plugin inventory. It pairs this robust backend with optimized Agent presets so Windows users get an out-of-the-box, enterprise-grade development experience.
This repository serves as the core runtime and distribution bundle for Windows DSH:
| Repository | Role in DSH | What It Provides | When to Use |
|---|---|---|---|
dsh-subprocess-win32 (This repo) |
Cordis Runtime Plugin | subprocess-win32 engine, lifecycle manager script, and managed copies of both presets |
Essential for Windows. Install this first to enable Windows runtime support and both presets. |
dsh-anchored-standard-windows |
Agent Preset Package | Standalone agent.cordis.yml, phase-gating modules, and isolated preset test suites |
Optional. Used only if you want to inspect, customize, test, or distribute the Anchored preset independently. |
💡 Quick Takeaway: Most Windows users only need to install
dsh-subprocess-win32. Its built-in lifecycle manager will automatically configure and render theAnchored Standard (Windows)andMinimal (Windows)presets for you.
Prerequisites:
Clone this repository, then add the local package to the profile. On DSH builds that expose the plugin command, the usual form is:
dsh plugin --profile web add C:\src\dsh-subprocess-win32
If the installed CLI does not provide add, add the package as a local
dependency in that profile's package.json and add dsh-subprocess-win32 to
dsh.profile.bundles, then run the profile's package-manager install. Do not put
the package in the DSH core release directory.
The bundle patch disables the stock subprocess row only on Windows and inserts
subprocess-win32. On Linux and macOS the stock row remains unchanged.
For the managed layout created by the Windows installer, the explicit lifecycle entry point is:
$setup = "$env:LOCALAPPDATA\DeepSeekHarness\setup\dsh-subprocess-win32.ps1"
& $setup -Action doctor
& $setup -Action setup -PackageSource C:\src\dsh-subprocess-win32
After setup, start a new empty session and select Minimal (Windows) or
Anchored Standard (Windows). Existing sessions retain their original Agent
composition.
Core DSH updates and this bundle's updates are deliberately separate. Keep the current package directory and install a new source tree beside it; do not replace the active directory in place. Before applying a change, run:
& $setup -Action doctor
& $setup -Action update -DryRun -PackageSource C:\src\dsh-subprocess-win32
Only apply the update after reviewing the manifest and test results:
& $setup -Action update -PackageSource C:\src\dsh-subprocess-win32
& $setup -Action rollback
The manager backs up the DSH profile, keeps the previous package, and restores the active pointer and profile files on failure. It does not automatically update third-party DSH bundles or Codex plugins.
Preview first, then confirm the destructive operation in an elevated PowerShell only if Windows asks for it:
& $setup -Action uninstall -DryRun
& $setup -Action uninstall
The managed uninstall reads its install manifest, stops only registered DSH
processes/tasks, removes the two DSH presets and the dedicated
%LOCALAPPDATA%\DeepSeekHarness root, and performs a residue audit. It does not
delete project source, .agents\skills, .codex\skills, Codex plugins, Codex
MCP configuration, or other Edge applications. Use the manager's export option
when you need a backup before removal; do not delete the root manually while a
host process is running.
0xC0000142 when DSH's restricted
workspace-write token is applied to this Windows process tree. The full-access
validation passed; the package does not widen permissions, bypass approval,
or silently select danger-full-access.landlock; treat command output as
untrusted and keep the DSH approval policy enabled..codex-plugin manifests into this package. Codex plugins,
DSH Cordis bundles, Agent presets, Skills, and MCP definitions use different
loaders. Only explicitly compatible SKILL.md and MCP Tools configurations may
be shared.Use the Node runtime from your DSH installation (or Node 22.19+):
npm test
The tests cover preset phase gating, context suppression, tool discovery, compaction recovery, and the Windows subprocess adapter. They do not claim that the host's model provider, API key, or sandbox policy is configured.
This project is an integration and Windows adaptation. It is not affiliated with DeepSeek, OpenAI, or the upstream maintainers.
sjh9714/dsh-win32, fixed source commit
f8a68a9836b84fdfec5c1f36ab60cea9923c689f, for the Windows subprocess direction.xiaobright/dsh-anchored-standard,
fixed source commit f57a1bde2dbaba3039bdae8631f78a0cb3ae3ebe, for the
two-phase Agent preset idea.Thanks to the DeepSeek Harness and community contributors for publishing the
runtime seams, preset experiments, tests, and issue reports that made this
Windows package possible. See NOTICE for license and provenance
details.
CLASSIFICATION EVIDENCE
系统优先读取 GitHub Topics,再与站内分类词典和词根规则比对。