模型端是否真可用？
    └─ 是 → Dify 服务器到模型端网络通吗？
              └─ 是 → Dify 模型供应商凭据配对了吗？
                        └─ 是 → 模型 schema 兼容吗？（不兼容/选不到）
                                  └─ 是 → 默认 LLM 指向 qwen3.6-27b 了吗？
                                            └─ 是 → plugin_daemon 加载 patch 了吗？
                                                      └─ 是 → 问答时 system 消息格式正确吗？
                                                                └─ 是 → 通。

#!/usr/bin/env bash
set -euo pipefail

curl -s http://116.62.10.183:8000/v1/models \
  -H "Authorization: Bearer 123456" | head

{
  "object": "list",
  "data": [
    {"id": "qwen3.6-27b", "object": "model"}
  ]
}

#!/usr/bin/env bash
set -euo pipefail

curl -s -X POST http://116.62.10.183:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer 123456" \
  -d '{
    "model": "qwen3.6-27b",
    "messages": [{"role": "system", "content": "你是一个助手"},{"role": "user", "content": "你好"}],
    "max_tokens": 128
  }'

#!/usr/bin/env bash
set -euo pipefail

curl -s -X POST http://116.62.10.183:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer 123456" \
  -d '{
    "model": "qwen3.6-27b",
    "messages": [{"role": "user", "content": "上海今天天气怎么样"}],
    "tools": [{
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "查询天气",
        "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}
      }
    }],
    "tool_choice": "auto"
  }'

#!/usr/bin/env bash
set -euo pipefail

curl -s -m 10 http://116.62.10.183:8000/v1/models \
  -H "Authorization: Bearer 123456" | head

/dify/dify-main/docker/volumes/plugin_daemon/cwd/langgenius/openai-1.0.3@<hash>/models/llm/llm.py

def get_customizable_model_schema(
        self, model: str, credentials: dict
    ) -> AIModelEntity:
        schema = {item.model: item for item in self.predefined_models()}.get(
            _base_model(model)
        )
        if schema is None:
            # Fallback for custom / non-OpenAI base models served via an
            # OpenAI-compatible endpoint (e.g. Qwen, DeepSeek, local vLLM):
            # clone a chat-capable predefined schema so the model is listable
            # and can be set as the default model.
            template = next(
                (m for m in self.predefined_models()
                 if m.model_properties.get("mode") == "chat"),
                None,
            )
            if template is None:
                template = self.predefined_models()[0]
            props = dict(template.model_properties)
            mp = (credentials or {}).get("model_properties") or {}
            try:
                if "context_size" in mp:
                    props["context_size"] = int(mp["context_size"])
                if "max_tokens" in mp:
                    props["max_tokens"] = int(mp["max_tokens"])
            except (ValueError, TypeError):
                pass
            schema = template.model_copy(deep=True, update={"model_properties": props})
        return schema.model_copy(
            deep=True,
            update={
                "model": model,
                "label": I18nObject(zh_hans=model, en_us=model),
                "fetch_from": FetchFrom.CUSTOMIZABLE_MODEL,
            },
        )

#!/usr/bin/env bash
set -euo pipefail

cd /dify/dify-main/docker && docker compose restart plugin_daemon

#!/usr/bin/env bash
set -euo pipefail

docker exec docker-db_postgres-1 psql -U postgres -d dify -c \
  "SELECT model_type, provider_name, model_name FROM tenant_default_models;"

   model_type   |          provider_name           | model_name
----------------+----------------------------------+-------------
 llm            | langgenius/openai/openai         | qwen3.6-27b
 text-embedding | langgenius/xinference/xinference | bge-m3

#!/usr/bin/env bash
set -euo pipefail

# 获取当前登录用户对应的 tenant_id
TENANT_ID=$(docker exec docker-db_postgres-1 psql -U postgres -d dify -tc \
  "SELECT id FROM tenants WHERE email='38205036@qq.com'" | xargs)

# 把 llm 默认模型改成 qwen3.6-27b
docker exec docker-db_postgres-1 psql -U postgres -d dify -c \
  "UPDATE tenant_default_models
   SET provider_name='langgenius/openai/openai', model_name='qwen3.6-27b'
   WHERE tenant_id='$TENANT_ID' AND model_type='llm';"

#!/usr/bin/env bash
set -euo pipefail

cd /dify/dify-main/docker && docker compose restart api worker plugin_daemon

#!/usr/bin/env bash
set -euo pipefail

docker ps --filter name=plugin_daemon --format '{{.Names}} {{.Status}}'
# 应显示 docker-plugin_daemon-1 Up ...

#!/usr/bin/env bash
set -euo pipefail

docker logs docker-plugin_daemon-1 2>&1 | grep -iE 'error|syntax|traceback|openai' | tail -20

#!/usr/bin/env bash
set -euo pipefail

cd /dify/dify-main/docker && docker compose restart plugin_daemon
# 等 5-10 秒
sleep 8
docker ps --filter name=plugin_daemon --format '{{.Names}} {{.Status}}'

def _ensure_system_first(messages: list[PromptMessage]) -> list[PromptMessage]:
    """Qwen3/vLLM chat templates accept ONLY one system message, at the start.

    Dify (esp. Agent apps) stacks several system/developer prompts: KB assistant
    prompt + agent tool prompt + an empty system + config manifest... Merging them
    into a single system message (joined by blank lines, empties dropped) keeps it
    at the front; every other message keeps its original relative order.
    """
    system_parts: list[str] = []
    rest: list[PromptMessage] = []
    for m in messages:
        if isinstance(m, (SystemPromptMessage, DeveloperPromptMessage)):
            text = _text_content(m)
            if text.strip():
                system_parts.append(text)
        else:
            rest.append(m)
    if system_parts:
        rest.insert(0, SystemPromptMessage(content="\n\n".join(system_parts)))
    return rest

    response = client.chat.completions.create(
        model=model,
        messages=cast(Any, [message(item) for item in _ensure_system_first(prompt_messages)]),
        stream=stream,
        ...
    )

#!/usr/bin/env bash
set -euo pipefail

cd /dify/dify-main/docker && docker compose restart plugin_daemon
# 若改过 default model 也要重启：
docker compose restart api worker plugin_daemon

#!/usr/bin/env bash
set -euo pipefail

docker logs docker-plugin_daemon-1 2>&1 | grep -iE 'error|syntax|traceback|openai' | tail -20
# 应无 SyntaxError / Traceback，能看到 openai 插件正常 register

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题五 · 步骤 1：验证 vLLM 三层（列模型 / 单轮对话 / 工具调用）。"""
from __future__ import annotations

import json
import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def curl_json(d: Dify, cmd: str, label: str) -> None:
    out, err = d.sh(cmd, t=120)
    print("\n=== %s ===" % label)
    print((out or err)[:1500])


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        auth = "-H 'Authorization: Bearer %s'" % C.VLLM_KEY
        curl_json(
            d,
            "curl -s %s/v1/models %s | head -c 800" % (C.VLLM_URL, auth),
            "3.1 GET /v1/models",
        )
        chat_body = {
            "model": C.LLM_MODEL,
            "messages": [
                {"role": "system", "content": "你是一个助手"},
                {"role": "user", "content": "你好"},
            ],
            "max_tokens": 128,
        }
        remote = "/tmp/vllm_chat_smoke.json"
        with d.sftp.open(remote, "w") as f:
            f.write(json.dumps(chat_body, ensure_ascii=False))
        curl_json(
            d,
            "curl -s -X POST %s/v1/chat/completions -H 'Content-Type: application/json' %s -d @%s"
            % (C.VLLM_URL, auth, remote),
            "3.2 POST /v1/chat/completions",
        )
        tool_body = {
            "model": C.LLM_MODEL,
            "messages": [{"role": "user", "content": "上海今天天气怎么样"}],
            "tools": [{
                "type": "function",
                "function": {
                    "name": "get_weather",
                    "description": "查询天气",
                    "parameters": {
                        "type": "object",
                        "properties": {"city": {"type": "string"}},
                        "required": ["city"],
                    },
                },
            }],
            "tool_choice": "auto",
        }
        remote2 = "/tmp/vllm_tool_smoke.json"
        with d.sftp.open(remote2, "w") as f:
            f.write(json.dumps(tool_body, ensure_ascii=False))
        curl_json(
            d,
            "curl -s -X POST %s/v1/chat/completions -H 'Content-Type: application/json' %s -d @%s"
            % (C.VLLM_URL, auth, remote2),
            "3.3 POST /v1/chat/completions（带 tools）",
        )
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题五 · 步骤 2：从 Dify 服务器验证到 vLLM 的网络可达性。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        cmd = (
            "curl -s -m 10 %s/v1/models -H 'Authorization: Bearer %s' | head -c 800"
            % (C.VLLM_URL, C.VLLM_KEY)
        )
        out, err = d.sh(cmd, t=30)
        text = out or err
        print(text)
        if C.LLM_MODEL in text:
            print("\n[OK] 从 Dify 服务器可达 vLLM，且列出 %s" % C.LLM_MODEL)
            return 0
        print("\n[FAIL] 未在响应中看到 %s" % C.LLM_MODEL)
        return 1
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题五 · 步骤 3：查询 tenant_default_models 当前默认 LLM。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        sql = (
            "docker exec %s psql -U postgres -d dify -c "
            "\"SELECT model_type, provider_name, model_name FROM tenant_default_models;\"" % C.PG_CONTAINER
        )
        out, err = d.sh(sql)
        print(out or err)
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题五 · 步骤 4 · 修改后：把 tenant 默认 LLM 改为 qwen3.6-27b 并重启 api/worker/plugin。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        tenant_sql = (
            "docker exec %s psql -U postgres -d dify -tc "
            "\"SELECT id FROM tenants WHERE email='%s'\"" % (C.PG_CONTAINER, C.DIFY_EMAIL)
        )
        out, err = d.sh(tenant_sql)
        tenant_id = out.strip()
        if not tenant_id:
            print("未找到 tenant:", err)
            return 1
        print("tenant_id:", tenant_id)
        upd = (
            "docker exec %s psql -U postgres -d dify -c "
            "\"UPDATE tenant_default_models SET provider_name='%s', model_name='%s' "
            "WHERE tenant_id='%s' AND model_type='llm';\"" % (
                C.PG_CONTAINER, C.OPENAI_PROVIDER, C.LLM_MODEL, tenant_id
            )
        )
        out2, err2 = d.sh(upd)
        print(out2 or err2)
        out3, err3 = d.sh(
            "cd %s && docker compose restart api worker plugin_daemon" % C.DIFY_DOCKER_DIR,
            t=180,
        )
        print(out3 or err3)
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题五 · 步骤 5：检查 plugin_daemon 容器状态与 Python 加载错误。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        out1, _ = d.sh(
            "docker ps --filter name=plugin_daemon --format '{{.Names}} {{.Status}}'"
        )
        print("=== plugin_daemon 状态 ===")
        print(out1.strip())
        out2, _ = d.sh(
            "docker logs %s 2>&1 | grep -iE 'error|syntax|traceback|openai' | tail -20"
            % C.PLUGIN_CONTAINER
        )
        print("\n=== 最近 plugin 相关日志 ===")
        print(out2.strip() or "(无匹配行，通常表示 OK)")
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题五 · 步骤 6 · 修改后：patch llm.py get_customizable_model_schema（自定义模型可选）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
import _plugin_patch as P
from dify_api import Dify


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        llm_path = P.find_plugin_file(d, "llm.py")
        print("llm.py:", llm_path)
        content = P.read_remote_file(d, llm_path)
        new_content = P.replace_method(
            content, "get_customizable_model_schema", P.GET_CUSTOMIZABLE_MODEL_SCHEMA
        )
        P.deploy_plugin_file(d, llm_path, new_content)
        print("已应用 llm.py schema 补丁并重启 plugin_daemon")
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题五 · 步骤 7 · 修改后：patch chat.py 合并多条 system（同专题四步骤 4）。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _plugin_patch as P
from dify_api import Dify

import _config as C  # noqa: F401


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        chat_path = P.find_plugin_file(d, "chat.py")
        content = P.read_remote_file(d, chat_path)
        new_content = P.replace_function(content, "_ensure_system_first", P.ENSURE_SYSTEM_FIRST)
        P.deploy_plugin_file(d, chat_path, new_content)
        print("已应用 chat.py 合并 system 补丁")
        return 0
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())