Dsh Oc Tui

Deepseek Harness opencode-like TUI Plugin

An opencode-inspired **terminal UI (TUI)** for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness), shipped as a dsh profile app plugin. It boots a chat client inside the dsh process: it creates/resumes agents through `ctx.agents`, renders the durable `session/event` stream (user messages, streaming assistant tokens, tool cards, todo lists), routes human input back via `agent.followup()`, and answers `approval/request` prompts inline.

The UI follows a DeepSeek blue-white dark design language (brand blue `#4D6BFE` accents on a blue-tinted canvas, with a blue→white gradient on the DeepSeek Harness title): a responsive session rail, markdown transcript and tool activity, a bordered multiline composer with slash-command suggestions, and a telemetry footer for session tokens, average time to first token (TTFT), decode throughput, KV-cache hit rate, and a **context meter** showing the current context length over the context-length limit. Thinking and running tools (read/write/...) show flowing spinner animations, the status row carries a flowing wave, and the composer border becomes a flowing gold marching-ants frame (dashes chasing clockwise around the box) while the agent is working.


中文用户手册见 [docs/用户手册.md](docs/用户手册.md)。

Requirements

• Node.js >= 22, an installed @deepseek-ai/dsh CLI, and an interactive terminal (Windows Terminal / ConPTY, iTerm2, GNOME Terminal, ...).
• pnpm on PATH for the one-time plugin install (dsh plugin forwards to pnpm).
• A configured model route: the profile reuses $DSH_HOME/settings.yaml (llm-pi-ai providers or llm-deepseek) and $DSH_HOME/.credentials.yaml — the same setup the Web GUI uses.

Quick start

# 1. One-time setup: create the tui profile and install this plugin into it.
dsh plugin --profile tui add dsh-oc-tui

# 2. Launch.
dsh --profile tui

The first command initializes `$DSH_HOME/profiles/tu