Deploying MiniCPM-Desk-Pet (开发者从源码)

重要：这个 Skill 只针对开发者。普通最终用户走的是 下载 dmg → Onboarding 流程，不应使用本 Skill。如果对方只是想"用一下这个应用"，请引导他们去 Releases 下载 dmg。

v0.8 起推理后端从 PyTorch / transformers 切换到 llama.cpp。本 Skill 已按 v0.9 平铺布局更新；sidecar 结构与构建方式以 minicpm-sidecar/README.md 为准。

This skill walks the agent through deploying the MiniCPM-Desk-Pet project on a colleague's machine from source. The project has 3 moving parts that need to come up in the right order: llama-server (compiled from vendored llama.cpp), the FastAPI gateway, and the Electron clawd-on-desk app.

Quick assessment first

Before doing anything, check what state the colleague's machine is in:

# Are we in the repo root?
ls README.md go.sh clawd-on-desk minicpm-sidecar 2>&1 | head -5

# What's already installed?
command -v node && node -v          # need 18+
command -v uv && uv --version       # go.sh installs if missing
command -v cmake && cmake --version # go.sh installs if missing (needed for llama-server)

If the user isn't in the repo root, cd to wherever they cloned it first.

Critical prerequisite: GGUF model

The model weight (~600 MB – 1 GB GGUF) is NOT in the repo. Without it, the gateway boots in "waiting for model" mode and Onboarding prompts the user to download one.

Check first:

ls -la models/*.gguf 2>&1

If missing, three ways to obtain it:

1. Already have a .gguf locally (most likely if the colleague worked on MiniCPM before):

   mkdir -p models
   ln -s /absolute/path/to/your/minicpm5-0.9b.Q4_K_M.gguf models/
   

2. Let Onboarding download — just run ./go.sh and follow the wizard's "Download model" step (~2–5 minutes from Hugging Face).

3. From a teammate's machine (scp):

   scp teammate@host:/path/to/minicpm5-0.9b.Q4_K_M.gguf ./models/
   

The gateway accepts both a .gguf file path and a directory containing .gguf files.

Install path (uv + cmake, what go.sh uses)

./go.sh

go.sh is idempotent. It will:

1. Check Node 18+ (auto-install via brew or fnm if missing)
2. Auto-install uv if missing (curl install.sh)
3. Auto-install cmake if missing (brew / apt)
4. First-time: clone vendored llama.cpp + cmake build llama-server (~5–10 min)
5. uv sync in minicpm-sidecar/ (a few dozen MB: fastapi / uvicorn / httpx / huggingface_hub — no torch)
6. npm install in clawd-on-desk/ (~1 minute, first time only)
7. Export MINICPM_SIDECAR_DIR + MINICPM_PYTHON + MINICPM_MODEL_DIR
8. npm start to launch Electron

If only the dependency-install part is needed (no launch yet):

./go.sh setup

Environment-check only (no install, no launch):

./go.sh doctor

Force a rebuild of llama-server (after pulling a new vendor pin):

./go.sh build-llama

Verifying the deploy

Once Electron starts, you should see:

1. Menu bar icon — paw, top-right
2. Floating pet — sitting on the desktop, can be dragged
3. Sidecar log in terminal — [sidecar] lines, eventually llama-server ready + gateway listening on :18765

Quick sanity check (in another terminal):

curl -s http://127.0.0.1:18765/api/health | python3 -m json.tool

Expected ok: true plus reported backend (metal on Apple Silicon, cuda, or cpu).

Then click the pet (or ⌘⇧M), type "你好" — pet should think and reply within a few seconds.

Common deploy failures & fixes

Cannot find module 'electron' / npm start fails immediately

cd clawd-on-desk && npm install

ModuleNotFoundError: fastapi (or uvicorn / httpx)

cd minicpm-sidecar && uv sync

llama-server: command not found / bin/<triple>/llama-server missing

The vendored llama.cpp didn't build. Re-run:

./go.sh build-llama

If cmake errors out, check cmake --version (need 3.20+) and re-install via brew / apt.

Sidecar log says model not found / No GGUF files in MODEL_DIR

The gateway couldn't locate a .gguf. Drop one into models/ or override:

MINICPM_MODEL_DIR=/abs/path/to/dir ./go.sh start
# or point directly at one file via Onboarding's "选本地模型"

Pet starts but "气泡聊天没反应 / 卡住"

Probably the sidecar didn't come up. In the clawd-on-desk terminal, look for [sidecar] lines.

• If gateway can't find Python: ensure minicpm-sidecar/.venv/bin/python exists (cd minicpm-sidecar && uv sync)
• If llama-server crashed during model load: check memory (Q4_K_M needs ~1 GB RAM resident; F16 needs ~2 GB)

port 18765 in use (or 23333)

Old sidecar / clawd HTTP server didn't die cleanly:

lsof -ti:18765 | xargs -r kill -9   # sidecar
lsof -ti:23333 | xargs -r kill -9   # clawd HTTP server

Cursor / Claude / Codex hook didn't auto-register

This is supposed to happen on first launch. Re-trigger via Settings → 🐾 MiniCPM, or restart the pet. Manual fallback (Claude as example):

cd clawd-on-desk
node -e "require('./hooks/install.js').installAll({silent:false})"

Updating someone else's already-deployed install

If a teammate already has an older checkout and wants the latest:

git pull
./go.sh setup     # re-syncs deps + rebuilds llama-server if vendor pin changed
./go.sh start

uv sync, npm install, and cmake are idempotent — no-op if nothing changed.

What "success" looks like

The deploy is done when:

• Pet visible on desktop
• curl -s http://127.0.0.1:18765/api/health returns ok:true
• Click pet → bubble pops up → type a message → reply within a few seconds
• In Cursor: ask Cursor anything → after Cursor finishes → pet says a one-line reaction (proves cursor-hook is wired)
• Settings → 🐾 MiniCPM shows current model + adapter status

If any of these fail, walk back through the corresponding "Common failure" section above.