AI Agent Hub
Back to plugins
🤖

oh-my-dsh

Model Inference Updated 2026.08.25

Run the following command in DeepSeek Harness:

dsh plugin install llmpolska/oh-my-dsh

Paste the following prompt into your AI chat to install this plugin:

Run `dsh plugin install llmpolska/oh-my-dsh#v0.1.0` in DSH Desktop or the dsh CLI, then fully restart DeepSeek Harness (Cmd+Q → relaunch); new sessions will have the /omd command and settings tab ready. Source: https://github.com/llmpolska/oh-my-dsh

About this plugin

Running every step of a coding session on a single frontier model is expensive and slow. oh-my-dsh splits each session into a Think tier (planning, architecture, hard debugging) and a Build tier (day-to-day implementation), routing automatically between them — no manual session switching, no YAML editing.

Beyond tiered routing, it ships a vision-delegation layer: paste a screenshot and say "build this frontend," and a vision model describes the image in the background while the working tier writes code from that description. The image stays in the chat, later turns never fail on unsupported content, and the session model never switches. It also includes an image-generation tool, a high-impact guard that blocks rm -rf, sudo, and force-pushes while the Build tier runs, automatic failure escalation to the Think tier, and on-demand Advisor / Reviewer consultation — all configured from a settings tab in DSH Desktop.

Built for developers who run coding agents on DSH Desktop, dsh web, or the terminal TUI and want "cheapest that works, strongest when it counts." One codebase covers all three surfaces; after install and a full restart, every new session picks up the /omd command and the full omd_* toolset with zero extra dependencies.

Use Cases

  • Auto-route planning steps to a strong model and implementation to a cheaper one, cutting per-session token cost
  • Paste a screenshot and say "build this frontend" — the vision model describes in the background, the working tier writes code
  • Block rm -rf, sudo, and force-pushes while the Build tier runs; escalate to Think tier on repeated step failures

Best For

  • Developers running coding agents on DSH Desktop, dsh web, or the terminal TUI every day
  • Small teams that want to match model strength to task difficulty and keep inference budgets in check
  • AI coding workflows that need screenshot-to-code or text-to-image multimodal steps without manual session switching