hou13851865503-design/domestic-global-ai-workbench

Domestic & Global AI Collaboration Workbench is a local-first Windows desktop tool for coordinating AI agents. Route planning, prompt optimization, execution, and independent review to different backends, then follow real call and handoff status in one dashboard.
国内外 AI 协同工作台是一个本地优先的 Windows 桌面工具。它把任务拆分为评估、提示词优化、执行和独立验收，可按角色分配不同 AI 后端，并在工作台实时查看 Agent 的调用、接力等待与最近结果。
> This repository contains the app and source code. It does not include personal API keys, task history, local logs, or runtime state.
## Highlights
Multi-agent workflow: separate assess, optimize, execute, and review roles; configure a backend for each role.
Backend choices: Codex CLI and OpenAI-compatible APIs for text roles; Codex CLI, WorkBuddy, and TRAE for file-producing or handoff work.
Live participation monitor: see which role is actually making a request, its task, model, elapsed time, recent outcome, and call count.
Honest handoff status: TRAE and WorkBuddy file handoffs show as waiting; the dashboard does not claim to observe activity inside those external apps.
Prompt improvement loop: freeze acceptance criteria, compare review scores across a bounded number of rounds, and retain the best prompt.
Independent review and traceability: review outputs separately and inspect task transitions and public-safe process events.
Local-first operation: the dashboard binds to 127.0.0.1; API keys and workflow state stay on the local machine.
No web framework dependencies: the Python service uses the standard library.
## Start on Windows
1. Install Python 3.10 or newer and make it available on PATH. The desktop launcher also checks AIRELAYPYTHON and common per-user installation paths.
2. Double-click AI多Agent流水线.exe in the repository root. It starts the local services and opens the dashboard at .
3. Open 多 Agent 分工 to choose role backends and configure OpenAI-compatible endpoint details. API keys are saved locally under ai-relay/state/agents.json.
4. Configure and sign in to Codex CLI separately when using Codex. WorkBuddy and TRAE use task-file handoffs and may require you to complete the work in those applications.
The live monitor is available from AI 状态. It refreshes independently of the task list. “尚未参与” means no tracked call has occurred yet; it is not a backend connectivity test. Use the explicit backend test control for backends that support a connectivity check.
### Run from source
powershell
cd ai-relay
python startall.py --interval 45
This starts the local platform, the Codex-compatible proxy helper, and task polling. To start only the platform, run:
powershell
python platform.py --port 8767
Alternatively, double-click ai-relay/start.bat to start all services. The dashboard and APIs are local-machine services, not a hosted web application.
## Build the Windows launcher
The checked-in executable is ready to launch the workbench. To rebuild it, install Python and PyInstaller, then run build.ps1 in PowerShell:
powershell
python -m pip install pyinstaller
./build.ps1
Set AIRELAYBUILDPYTHON to the desired python.exe if Python is not on PATH.
## Workflow
text
Task request
-> Assess and select a model (optional)
-> Optimize a structured prompt