DeepSeek Harness 麒麟部署与改造

tar -xJf node-v22.23.3-linux-arm64.tar.xz -C /opt
mv <NODE_ARCHIVE_DIR> <NODE_PREFIX>

ln -sf <NODE_PREFIX>/bin/node     /usr/local/bin/node
ln -sf <NODE_PREFIX>/bin/npm      /usr/local/bin/npm
ln -sf <NODE_PREFIX>/bin/npx      /usr/local/bin/npx
ln -sf <NODE_PREFIX>/bin/corepack /usr/local/bin/corepack

cat > /etc/profile.d/node.sh <'EOF'
export PATH=<NODE_PREFIX>/bin:$PATH
EOF

node -v   # v22.23.3

npm i -g @deepseek-ai/dsh

dsh web --host <HOST_IP> --port 80
# 启动被 host 校验拒绝

cd <NODE_GLOBAL_MODULES>/@deepseek-ai/dsh/node_modules/@deepseek-ai/dsh-host-webserver/lib
# 打开 index.js,定位到被校验的 host 枚举(第 142 行附近)

cd <NODE_GLOBAL_MODULES>/@deepseek-ai/dsh/node_modules/@deepseek-ai
cp -n dsh-client-connection/lib/client.js          dsh-client-connection/lib/client.js.bak
cp -n dsh-client-ui-settings/lib/client.js         dsh-client-ui-settings/lib/client.js.bak
cp -n dsh-client-ui-settings-general/lib/client.js dsh-client-ui-settings-general/lib/client.js.bak

curl -s -o /dev/null -w "%{http_code}\n" --noproxy '*' http://<MODEL_SERVER_IP>:8080/v1/models
# 200

{"ok":true,"value":[{"id":"qwen-local","name":"qwen-local"}]}

[Unit]
Description=DeepSeek Harness Web UI
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
Environment=PATH=<NODE_PREFIX>/bin:/usr/local/bin:/usr/bin:/bin
Environment=DSH_HOME=<DSH_HOME>
# 出网代理(仅在需要代理出网时设置;本文不展开具体方案)
Environment=HTTP_PROXY=http://127.0.0.1:PROXY_PORT
Environment=HTTPS_PROXY=http://127.0.0.1:PROXY_PORT
# 关键:必须用显式 IP,不能用 CIDR
Environment=NO_PROXY=localhost,127.0.0.1,<HOST_IP>,<MODEL_SERVER_IP>
# 本地模型凭据(见 §8.4)
Environment=LOCAL_QWEN_KEY=<REDACTED_MODEL_API_KEY>
ExecStart=<NODE_PREFIX>/bin/dsh web --host <HOST_IP> --port 80 --no-open --trusted-host <HOST_IP>
Restart=always
RestartSec=3

[Install]
WantedBy=multi-user.target

systemctl daemon-reload
systemctl enable --now dsh-web
systemctl is-active dsh-web        # active

firewall-cmd --permanent --add-port=80/tcp
firewall-cmd --permanent --add-port=3080/tcp
firewall-cmd --reload

{"id":"qwen-local","object":"model","owned_by":"llamacpp",
 "meta":{"n_ctx":262144,"n_ctx_train":262144,"n_params":27320697856,
         "size":20207181824,"ftype":"Q5_K - Medium"}}

dsh --profile web --dump-config | grep -A2 -i "pi-ai"
# - id: llm-pi-ai
#   name: '@deepseek-ai/dsh-llm-pi-ai'

- id: ui-settings-general
  name: "@deepseek-ai/dsh-client-ui-settings-general"
  config:
    welcomeNoticeVersion: 2026-09-28.1
- id: llm-pi-ai
  name: "@deepseek-ai/dsh-llm-pi-ai"
  config:
    providers:
      local-qwen:                       # 手写路由名:小写连字符
        displayName: "Qwen Local (llama.cpp)"
        api: openai-completions         # 线协议
        baseURL: "http://<MODEL_SERVER_IP>:8080/v1"
        apiKeyEnv: LOCAL_QWEN_KEY       # 凭据引用名
        models:                         # 手写路由必须显式给模型列表
          - id: qwen-local
            name: "Qwen3.8-27B Uncensored"
            contextWindow: 262144
            reasoningEfforts: false     # 声明为非思考模型(无思考等级)

dsh --profile web --dump-config | grep -A15 "id: llm-pi-ai"

umask 077
printf 'LOCAL_QWEN_KEY=<REDACTED_MODEL_API_KEY>\n' > <DSH_HOME>/.env
chmod 600 <DSH_HOME>/.env
# 同时在 dsh-web.service 里设置 Environment=LOCAL_QWEN_KEY=...

systemctl daemon-reload
systemctl restart dsh-web

{
  "type": "client-request",
  "rpcId": "<任意字符串>",
  "method": "<namespace>/<method>",   // 如 llm/listProviders
  "payload": { "args": { ... } }      // 注意 payload 必须且只能有 args
}

curl -s -b cookie.txt -X POST "http://<HOST_IP>/api/llm/listProviders" \
  -H "content-type: application/json" \
  -d '{"type":"client-request","rpcId":"1","method":"llm/listProviders","payload":{"args":{}}}'
# → deepseek-official, deepseek-account, local-qwen

curl -s -b cookie.txt -X POST "http://<HOST_IP>/api/session/modelCatalog" \
  -H "content-type: application/json" \
  -d '{"type":"client-request","rpcId":"2","method":"session/modelCatalog","payload":{"args":{}}}'

{"default":{"provider":"deepseek-official","model":"deepseek-flash"},
 "routableProviders":["deepseek-official","local-qwen"],
 "groups":[{"id":"local-qwen","name":"Qwen Local (llama.cpp)",
            "models":[{"id":"qwen-local","name":"Qwen3.8-27B Uncensored"}]}]}

# 1) 建会话
curl ... /api/session/create \
  -d '{"type":"client-request","rpcId":"c1","method":"session/create","payload":{"args":{"request":{}}}}'
# ⇒ sessionId

# 2) 切到本地模型
curl ... /api/session/selectModel \
  -d '{"type":"client-request","rpcId":"sm","method":"session/selectModel",
       "payload":{"args":{"request":{"sessionId":"<sid>","provider":"local-qwen","model":"qwen-local"}}}}'
# ⇒ {"selected":{"provider":"local-qwen","model":"qwen-local"}}

# 3) 发提示
curl ... /api/session/prompt \
  -d '{"type":"client-request","rpcId":"p2","method":"session/prompt",
       "payload":{"args":{"request":{"requestId":"req-2","sessionId":"<sid>","mode":"queue",
                    "content":[{"type":"text","text":"Reply with exactly: OK"}],"clientTimeZone":"UTC"}}}}'

remote.session.prompt({
  requestId, sessionId, mode,      // mode 常用 "queue"
  content,                         // [{ type:"text", text }]
  clientTimeZone
}, signal)

{"modelSelection":{"lastUsed":{"provider":"local-qwen","model":"qwen-local"},
                   "next":{"provider":"local-qwen","model":"qwen-local"}},
 "sessionStats":{"turns":2,"steps":2,"llmMs":11963,
                 "ttftMs":10598,"decodeMs":1365,"decodeTokens":29},
 "turnOutline":[{"turn":2,"prompt":"Reply with exactly: OK","response":"OK"}]}