dsh-computer-use_codex-style

Computer Use for DeepSeek Harness

Drive Windows desktop apps and Chromium tabs from DeepSeek Harness.

Based on the Codex window2 tool surface — the Codex-style Computer Use experience on DeepSeek Harness.

English | [中文](README.zh-CN.md)

As of this plugin's release, the official harness has shipped Computer Use in an experimental release, with a noticeably different design. This plugin is coexistence-safe: both can be installed side by side, but running both drivers at once is not recommended.

---

Contents

• [What you get](#what-you-get)
• [Quick start](#quick-start)
• [How it works](#how-it-works)
• [Tool surface](#tool-surface)
• [The act-and-refresh loop](#the-act-and-refresh-loop)
• [Safety model](#safety-model)
• [Experience layer](#experience-layer)
• [Configuration](#configuration)
• [Browser automation](#browser-automation)
• [Verification](#verification)
• [Build from source](#build-from-source)
• [Repository layout](#repository-layout)
• [Troubleshooting](#troubleshooting)
• [Security and privacy](#security-and-privacy)
• [Compatibility and known gaps](#compatibility-and-known-gaps)
• [License and attribution](#license-and-attribution)

---

What you get

| | |

|---|---|

| The full Codex surface | All 13 official Codex window2 methods — list_windows, get_window, list_apps, launch_app, get_window_state, click, press_key, type_text, scroll, set_value, drag, perform_secondary_action, activate_window — with the same names, parameters, defaults, return shapes and error strings as Codex. |

| Real input, real screenshots | SendInput for input, UI Automation for the accessibility tree, Windows.Graphics.Capture for screenshots that work even when the window is occluded. Screenshots arrive as vision image parts, never as base64 inside JSON. |

| A visible, cancellable overlay | A status pill plus an animated synthetic cursor. The pill is hidden for the duration of each screenshot, so the model never reads its own status pill back while you keep seeing it. Esc interrupts the turn at any moment. |

| Verified against the real thing | Behaviour is gated against the official plugin and its helper: AX tree grammar, overlay pixels, cursor motion, capture freshness, transport budgets, approval strings, turn lifecycle. |

| Harness-native, not MCP-shaped | DSH tools with a bundled skill, not a JavaScript REPL. The plugin is a DSH bundle: package.json + cordis.patch.yml + entry modules at the package root. |

| Two planes, one sidecar | The desktop is process-wide, so the pointer, overlay, capture and approvals live on the HOST plane; the model-facing tools live in an agent preset, so an ordinary coding session never gets the mouse. |

| An experience layer | Local, machine-specific notes written after a task and replayed as advisory context before the next one — see [Experience layer](#experience-layer). |

---

Quick start

Requirements

| | |

|---|---|

| OS | Windows 10 1809+ or Windows 11, x64 |

| Harness | DeepSeek Harness (a source checkout or an installed dsh CLI) |

| Node | 20 or newer |

| Python | 3.10+ — only for the optional browser catalog; the 13 desktop tools are Rust + Node |

| Rust | only if you rebuild the helper (a release binary already ships in helper-rs/bin/win32-x64/) |

1. Install the bundle

# from GitHub (no build step, no install-time script)
dsh plugin --profile web add github:wushi2333/dsh-computer-use_codex-style

# from a local checkout
dsh plugin --profile web add /path/to/dsh-computer-use_codex-style

# from a packed tarball
dsh plugin --profile web add ./dsh-computer-use-0.1.0.tgz

The bundle declares dsh.bundle.patch, so dsh plugin adds the HOST-plane row to the profile and links the package. The plugin is plain ESM JavaScript plus a prebuilt helper, so it ships no prepare or postinstall script — pnpm never asks you to allow a build.

2. Add the tool row to an agent preset

The desktop tools deliberately live in an agent preset, not in the host composition. Pick the preset you want to give the mouse to and add one row to its agent.cordis.yml (presets live in $DSH_HOME/.agent-presets//):

# Computer Use tools consume the HOST dshComputerUse sidecar. They publish
# nothing, so they sit loose like tool-bash. Install the dsh-computer-use
# bundle in the profile first; otherwise this row waits for the service.
- id: tool-computer-use
  name: dsh-computer-use/tool
  config:
    surfaces:
      - computer
    browserSkill: computer-use-browser

Keep a separate preset for Computer Use if you prefer: sessions started with the standard preset then keep their normal tool set and never touch your mouse.

3. Deliver the skills

The two skills ship inside the package and are copied into the skill roots the harness reads:

# from the profile directory
node node_modules/dsh-computer-use/scripts/sync-skills.mjs --write --defaults