🐳 Oh My DSH
中文 GitHub

Oh My DSH

Into the Unknown

omdsh is a focused, keyboard-first DeepSeek coding agent for your terminal, built on the DeepSeek Harness plugin runtime.

Get Started View on GitHub

Install

🐳 full access
$ npm install --global @agi-fans/oh-my-dsh

Requires Node.js 22.19 or later in the 22.x line, or Node.js 24 or newer, plus a DeepSeek API key. Run /login once inside omdsh, then start a conversation.

This site is keyboard-first too: press Ctrl+K to search, / works as well.

What a turn looks like

/sessions Durable conversations Search, pin, rename, and resume sessions; rewind turns, retry, compact context, and export complete transcripts as Markdown or standalone HTML. /agent /workflow /permission Three real session controls Choose an Agent preset, a Default or Plan workflow, and an Access level per session. Each preset owns its tool exposure, so the model never carries capabilities the session did not grant. @mention · Ctrl+V Rich terminal input Mention project files and sessions with @, paste clipboard images, reuse persistent prompt history, compose multiline prompts in $EDITOR, and queue follow-ups. Alt+A Subagents on a leash Watch streaming tool calls and live child agents, open a child transcript in place, follow or steer continuable runs, and keep heavy parallel delegation off the render loop. /trajectory · /context Whole-session observability Walk a keyboard-driven event ledger with live following, search, folding, timings, token usage, and tool payloads; /context separates provider-anchored occupancy from heuristic prompt composition. Skills · MCP · plugins Extend everything Add project Skills and MCP servers from plain configuration, mount user plugin bundles with omdsh plugin, and drive local language servers through the read-only lsp tool.

Why omdsh

plugin architecture Harness-native, not a fork Sessions, models, tools, permissions, Skills, MCP servers, and telemetry come from published DeepSeek Harness services. omdsh composes them into one terminal application instead of building a second agent core. keyboard-first Every flow has a key Queue follow-ups, steer a running child, search prompts and the transcript, switch model or Access, and export a session without leaving the composer. display cells Built for real terminals Layout is measured in terminal display cells, so CJK text, emoji, ANSI styling, and long commands keep borders and right padding aligned while the terminal keeps its own scrollback.

Performance

2.62 ms median restore of a 10,000-turn session
22.71 ms replay of 10,000 tool calls
0.35 ms average frame over a 5,000-turn surface

Median figures measured on Apple M5 Pro on the documented benchmark surface. The TUI performance report carries the methodology and raw results: TUI performance report

🐳 Oh My DSH · / en-US
Ctx 0% Into the Unknown