Why it mattersDSH's advantage over closed coding agents is the runtime and the tool contract, and a front end that reads and edits like Claude Code inside the terminal turns that runtime into something a developer would sit in every day.
Ccch1mneyyy's dsh-TUI, published on GitHub on 13 August 2026, is an interactive terminal UI for the DeepSeek Harness runtime, mounted as a Cordis plugin and released under MIT. The repository has picked up 2,881 stars and 164 forks in the twenty-six days since it was created, reached daily rank seven on the TypeScript GitHub Trending board over the same window, and ships on npm as @deepseek-harness-tui/dsh-tui. It touches no DSH core code: installing the plugin turns the interface on, and removing it leaves no patches behind.
What the surface adds
The plugin gives DSH a chat surface that reads like Claude Code inside a terminal. Streaming Markdown renders live. Tool cards fold multi-line command headers to one line plus a count and expand on click or Ctrl+O. Command and file completion work everywhere, @ file references attach text, directory listings and image blocks, and @path#L12-14 line ranges attach only the requested lines and clamp gracefully at end-of-file. History search, message selection, and a /lang switch between simplified Chinese and English are all first-class commands.
Terminal image previews are the piece that reads as new for a coding CLI. On terminals that report Kitty graphics or Sixel through DA1, dsh-TUI renders transcript thumbnails inline, opens a shared preview centered over the transcript on click, and disables graphics cleanly under non-TTY output, tmux, screen or accessibility mode. Sixel uses a bounded 256-colour adaptive palette and encodes only the visible crop while scrolling. A text fallback with matching size preserves layout when graphics are unavailable.
The session and editing model
Double-Esc rewinds through a session fork, so backing up and asking again does not lose the branch that ran. /resume groups history by working directory with search and preview, pinned sessions float in a Pinned group at the top, and pins persist under ~/.dsh-tui. The composer accepts vim editing (/vim), mouse selection, drag-select, Shift+click extension, Ctrl+C copy of the selection, and a full-screen draft editor on Ctrl+Shift+E with line numbers, Enter for newline and Ctrl+Enter to send. Pasting a Finder-copied image stages it as [Image #N] in the composer, and vim x, X, d and u treat the attachment as a single unit.
Why the traction is worth paying attention to
DeepSeek Harness itself is a runtime that speaks the same tool contract as Codex and Claude Code, and its main pitch is that a team can point it at a cheaper model without giving up agent behaviour. Most of the DSH plugins reaching Trending recently sit inside that automation layer, as routers, agent-team orchestrators, or plugin managers. dsh-TUI fills the developer-facing side of the ecosystem: the surface a human keeps open while the automation runs.
The pixel-whale mascot animation, ported from dsh-ui-whale by @lhh010, plays one of three randomised startup intros, wakes and dozes off during the welcome phase, and freezes to a static frame the moment the first agent task begins so there is no ongoing render cost. That last detail matters more than it looks: the README's performance section calls out zero per-frame allocations on the hot paths, message virtualisation, differential output and bounded caches, saving about 200 KB of garbage-collection churn every sixteen-millisecond tick in a 3,200-row session. A TUI that feels good on a long agent run has to earn that on the profiler, and the maintainer has done the work.
The repository is in public beta and the maintainer says so. Read the architecture and limitations page before relying on the permission model in a shared environment. For DSH users, this is the front end that turns the runtime into something worth opening every morning.
Source
Source: GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.