從本地模型開始，
建立自己的 OpenHands。

wsl --install -d Ubuntu

curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env
uv --version

uv tool install openhands --python 3.12
openhands --version

docker version
docker run --rm hello-world

mkdir -p ~/projects/my-project
cd ~/projects/my-project
printf '# My Project\n' > README.md
pwd
ls -la

cd ~/projects/my-project
openhands serve --mount-cwd

http://localhost:3000

請先執行 pwd 與 ls -la。
確認目前的工作目錄。
接著讀取 README.md，說明這個專案目前包含哪些內容。
不要修改任何檔案。

---
name: coordinator
model: openai/qwen3.8:27b
tools:
  - delegate
---

# Coordinator
Coordinate the task.
Delegate research, implementation, and evaluation.
Do not perform implementation work directly.

---
name: researcher
model: openai/qwen3.8:27b
tools:
  - terminal
  - browser_navigate
  - browser_get_content
  - invoke_skill
---

# Researcher
Collect evidence from the workspace and authoritative external sources.
Separate observed evidence, inference, and hypothesis.
Do not modify project code.

---
name: engineer
model: openai/qwen3.8:27b
tools:
  - terminal
  - file_editor
  - task_tracker
  - invoke_skill
---

# Engineer
Implement controlled changes and verify them.
Do not silently redefine the research objective.

---
name: evaluator
model: openai/qwen3.8:27b
tools:
  - terminal
---

# Evaluator
Independently inspect artifacts and metrics.
Do not modify the implementation being evaluated.

Use invoke_skill only when an installed skill materially improves the current task.

Prefer the minimum necessary set.
Do not invoke skills merely because they are available.

C=$(docker ps --filter "name=oh-agent-server" --filter "status=running" --format '{{.Names}}' | head -1)
docker logs -f "$C"

Run one minimal multi-agent verification cycle.

Researcher:
- inspect the workspace
- identify one existing file
- use one authoritative external source
- do not modify files

Coordinator:
- summarize the evidence
- delegate one harmless implementation task

Engineer:
- create a temporary file containing exactly ENGINEER_OK
- read it back

Evaluator:
- independently verify the file and exact content
- do not modify anything

Coordinator:
- report whether every role respected its boundary

Do not run training.
Do not modify datasets.
Do not modify .agents.