DeepSeek Harness — Claude TUI

Claude Code muscle memory. DeepSeek Harness underneath.

English · 简体中文

An unofficial, high-fidelity Claude Code-style terminal interface for DeepSeek Harness,
reconstructed from real PTY captures and verified cell by cell.

[!NOTE] This is an independent community project. It is not affiliated with, endorsed by, or sponsored by Anthropic or DeepSeek. “Claude Code” identifies the version-pinned compatibility target; no Anthropic source code is included. See the trademark and compatibility notice.

Why this exists

DeepSeek Harness has a composable Agent, Session, tool, approval, question, and subagent runtime. Claude Code has a terminal workflow many developers already know by muscle memory.

This plugin joins those two ideas without forking Harness core:

• Feels familiar: Claude-shaped shell, prompt, menus, transcript, approvals, questions, and agent states.
• Runs on Harness: live Harness models, durable Sessions, commands, permissions, tools, and subagents.
• Proves fidelity: 23 frames captured from a real Claude Code 2.1.227 PTY; 21 are automated comparison anchors.
• Tests the terminal, not a mockup: buffer choice, cell geometry, RGB styles, hardware cursor, and interaction transitions.

It is an external Harness bundle—not a web skin and not a hardcoded terminal recording.

Try it

Prerequisites: a working DeepSeek Harness CLI, Node.js 24, and pnpm 11.

git clone https://github.com/cogine-ai/dsh-claude-tui.git
cd dsh-claude-tui

corepack pnpm install --frozen-lockfile
corepack pnpm check

dsh plugin --profile claude-tui add "$PWD"
DSH_TOOLS_MODE=code dsh --profile claude-tui

Sending a real model request requires the credentials for the Harness model provider you select.

What already works

Surface	Implemented behavior
Main shell	normal-buffer scrollback, Claude orange logo, responsive header, editor, status footer
Prompt	multiline editing, submit/steer, cancellation, reverse history search
Completion	slash commands and bounded @ workspace file mentions
Transcript	user, assistant, reasoning, tool call/result, usage, request and turn outcomes
Protocols	real Harness approval and structured-question providers
Agents	foreground/background subagent states, expandable output, active-agent roster
Sessions	create, exact-id resume, interactive picker, graceful flush and terminal restoration

Useful controls:

Key	Action
Enter	submit while idle or steer a running Agent
Shift+Enter	insert a newline
Esc / Ctrl+C	interrupt a running turn
Ctrl+R	search prompt history
Ctrl+O	expand or compact tool details
Left Arrow	hide or show the active-agent roster
Ctrl+D	press twice on an empty prompt to exit cleanly

Fidelity

Verified against Claude Code 2.1.227 in a true-color xterm-compatible PTY:

• 23 reference frames and 21 automated visual/semantic anchors.
• 41/41 terminal tests at 80x24 and 100x30.
• Real Harness runs for approvals, questions, and foreground/background subagents.
• One intentional difference: a blank top row prevents logo clipping.

Full qualification report

Architecture

@deepseek-ai/dsh-base
  → dsh-claude-tui bundle
    → startup argument service
    → worker-thread code runtime
    → one terminal-owned root Agent
      → durable Session event projection
      → pi-tui normal-buffer renderer
      → approval + question protocol adapters
      → subagent lifecycle + active-run roster

The plugin waits for Loader settlement, binds every event to the exact root Agent, delegates unknown slash commands to Harness, and reverses registrations before restoring raw terminal state.

Compatibility

Targets the observed Claude Code 2.1.227 TUI only. Harness remains the source of truth for runtime data and capabilities; unsupported Claude-only states are not simulated. New versions require requalification.

Development

corepack pnpm install --frozen-lockfile
corepack pnpm check

The check gate runs TypeScript no-emit validation, all Vitest terminal tests, and the production build.

To refresh the reference corpus with the locally installed Claude Code 2.1.227 executable:

corepack pnpm capture:claude-reference -- --cwd /path/to/a/trusted/workspace

Dynamic capture scenarios use an isolated Claude config, a dummy key, and a loopback-only Anthropic mock. They make no external model request and do not require a user API key.

Roadmap

• richer attachment and completion surfaces;
• broader session management and rename flows;
• additional plan, todo, and background-job states;
• more terminal emulators and operating-system qualification;
• release packaging beyond installation from a checkout.

Issues and focused pull requests are welcome. Visual-parity changes should include an independently captured reference or an explicit, documented Harness-semantic boundary.

License

Original project code is available under the MIT License. Product names and marks remain the property of their respective owners; the MIT License does not grant rights to third-party trademarks.