ccch1mneyyy/dsh-TUI

dsh‑TUI – a Claude‑style terminal UI for DeepSeek Harness

What it is

• A plug‑in that adds a rich, graphical‑looking Text‑User‑Interface (TUI) on top of the official dsh CLI (DeepSeek Harness). It does not modify the core engine; it simply mounts a UI layer that talks to the existing DSH services.
• The UI mimics the look of Claude Code: a pixel‑art whale banner, glowing title, live status bar, progress bars, TPS (tokens‑per‑second) gauge, and support for streaming Markdown, tool cards, and image attachments.

Why it matters

• Makes long‑running AI‑agent sessions usable from a terminal – you can see model tokens, cache hit rates, and the current “agent” state in real time.
• Provides a full command set (/new, /resume, /model, /audit, …) that mirrors the capabilities of DeepSeek Harness, including tool execution, sub‑agents, and approval flows, all from within the same window.
• Designed for low‑resource environments: virtual‑scrolling keeps rendering cost O(visible‑window) even for very long chats.

Key features (as described in the README)

Feature	What you get
Native terminal interaction	Streaming Markdown, structured tool cards, command/file completion (including @‑file references), image attachment handling, history search, inline/alternate‑screen rendering, language toggle (/lang).
Observable agent status	Real‑time work‑status line, context‑segment progress bar, TPS gauge, cache‑hit rate, inference level, token counts, Git/session info.
Complete session workflow	Commands for new/resume/compact/export sessions, side‑questions (/btw), model switching, double‑Esc time‑rewind/fork.
DeepSeek Harness integration	Access to Agent presets, Skills, MCP (remote tool servers), Goals, Todos, sub‑agents, and the ask_user_question questionnaire flow.
Long‑session design	Event‑driven projection, diff‑based terminal output, message virtualization, bounded cache, preventing memory blow‑up.
Theme system	Auto‑detect terminal background, built‑in light/dark/ansi themes, and custom JSON themes placed in ~/.dsh‑tui/themes/.
Keyboard & mouse shortcuts	Full list of hotkeys (Enter, Ctrl‑C, Esc, Ctrl‑O, Ctrl‑R, etc.) plus mouse drag‑to‑copy, scroll, and click‑to‑select.
Plugin ecosystem	Docs for building extensions, a community repo (dsh‑tui‑ecosystem), and a template starter.

Installation & quick start

# install the core DSH CLI and the TUI plugin globally
npm install -g @deepseek-ai/dsh @deepseek-harness-tui/dsh-tui

# launch – the first run creates a profile automatically
dsh-tui

• Requires a TTY terminal, pnpm 10+, and a DEEPSEEK_API_KEY for model calls.
• On Windows you can run the bundled dsh-tui.cmd.
• The command dsh-tui --resume restores the last session; /update updates the plugin in‑place.

How it works under the hood

1. dsh loads a profile that chains the base DSH implementation → the dsh‑TUI Cordis patch → the selected Agent preset.
2. Session events flow from the DSH backend to the TUI via an in‑process event stream.
3. The UI renders with a ported Ink/Yoga renderer, using incremental diff updates and layout virtualization to keep performance constant.
4. All persistence (chat logs, tool results, etc.) remains the responsibility of the DSH service; the TUI only displays state.

Typical workflow

• Start dsh-tui → see the whale banner and status line.
• Type a prompt (Enter to send, Shift‑Enter for newline). The model streams back Markdown; tool cards appear inline.
• Use /model to pick a different model, /preset to change the agent mode, or /new to begin a fresh conversation.
• Press double‑Esc on an empty line to rewind the session (fork a new branch from a previous point).
• Export the whole conversation with /export (Markdown) or search history with /.

Community & support

• Officially featured in the DeepSeek Harness public‑beta WeChat channel.
• A dedicated QQ group (572549239) and a WeChat group for quick help.
• Documentation covers installation, configuration, theme customization, interaction details, architecture, VS Code integration, plugin development, and contribution guidelines.

License

• MIT – free to use, modify, and redistribute.

---

All information above is taken directly from the repository’s README; no additional features have been inferred.