mashanglinyu/luanyujia-workbench

What it solves

Most AI assistants live in a single chat window, which gets messy when you use them for many different kinds of work. This project offers a way to turn that one window into a multi-scene desktop workbench, where each scene is a separate workspace with its own always-on AI session. Each scene's AI knows which scene it belongs to — it reads the right memory, uses the right skills, and remembers where it left off. It also solves the problem of keeping multiple AI agents working together without stepping on each other: it provides rules for collaboration, task assignment, and memory handling so things don't fall apart after a few months of heavy use.

How it works

The workbench is built around a layered structure — roughly a host, a command console, a constitution, a shared layer, scenes, projects, runtime/archive, and audit. Each scene is described by a JSON config that defines its grid layout, panel types, and the AI session attached to it. A key piece is "session anchoring": the AI binds to a specific session using three fields (pinTarget, sessionTitle, sessionId), so it knows its identity and follows the right window. Memory is stored in files rather than inside the conversation, split into three layers, and written back according to clear rules. The project also includes a way to connect free web-based AIs as local tools via MCP, so you can save on API tokens. It provides templates, source code, and a minimal runnable demo — plus a full public workbench with four scenes and complete shell source that you can actually run after installing a DSH engine. The repo emphasizes that this is a method and a shell template, not a plug-and-play software: every path and config needs to be adapted to your own environment, and many things — especially MCP integration — require on-site debugging.

Who it's for

This is for people who want to build their own multi-scene AI desktop workspace, especially those who already use AI for various tasks and want to organize them into separate scenes with dedicated memory and skills. It suits both curious newcomers who just want to see the concept (they can open the zero-dependency single-file demo) and more advanced users who want to copy the templates and adapt them to their own setup. The author also warns that beginners should read the "Newbie Must Read" file before starting, because many paths are placeholders and won't work if copied verbatim.

Highlights

• A complete method plus shell templates for building a scene-based AI workbench — not just a tutorial, but actual source code and configurations.
• 24 real panel types and 11 real scene configurations (all sanitized) shown as examples.
• Three core mechanisms deeply documented: collaboration, memory, and scene system — these determine whether your setup stays organized after months.
• A session anchoring system (three fields) that gives each AI a clear identity and prevents cross-talk between scenes.
• Way to hook free web-based AIs as local tools using MCP, saving API tokens.
• Includes a double-click-run demo (single file, zero dependencies) and a full public workbench with four scenes and complete shell code that runs after installing DSH engine.
• Detailed sinkhole notes: each section has a "symptom lookup" table from real incidents.
• Strict privacy: all personal data, paths, keys, and business data have been removed — validated via automated scans.
• MIT licensed — you can freely use and modify it.