from dify_api import Dify, pp

d = Dify("38205036@qq.com", "Difyai123456")
d.login()

body = {
    "model": "qwen3.6-27b",
    "model_type": "llm",
    "credentials": {
        "openai_api_key": "123456",
        "openai_api_base": "http://116.62.10.183:8000/v1",
        "openai_organization": "",
        "api_protocol": "chat",
    },
    "name": "qwen3.6-27b",
}
r = d.post(
    "/workspaces/current/model-providers/langgenius/openai/openai/models/credentials",
    body,
)
pp("add qwen cred", r)   # 期望 200/201

r = d.get("/workspaces/current/models/model-types/llm")
pp("llm list", r)
d.close()

from dify_api import Dify, pp
d = Dify("38205036@qq.com", "Difyai123456")
d.login()
r = d.post(
    "/workspaces/current/default-model",
    body={
        "model_settings": [{
            "model_type": "llm",
            "provider": "langgenius/openai/openai",
            "model": "qwen3.6-27b",
        }]
    },
)
pp("set default llm", r)
d.close()

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

docker exec -it docker-db_postgres-1 psql -U postgres -d dify \
  -c "UPDATE tenant_default_models SET model='qwen3.6-27b' WHERE model_type='llm';"

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

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

/app/storage/cwd/langgenius/openai-<版本>@<hash>/models/llm/llm.py
# 宿主机常见映射：
/dify/dify-main/docker/volumes/plugin_daemon/cwd/langgenius/openai-*@*/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 {}
            # Align context_size to the actual vLLM --max-model-len (65536) so
            # Dify never lets prompts exceed what the served model can accept.
            # The credential-update API does NOT persist model_properties for
            # this provider, so we hardcode the safe ceiling as the default and
            # still honor an explicit override when one is supplied.
            VLLM_MAX_CTX = 64000
            try:
                props["context_size"] = int(mp["context_size"]) if "context_size" in mp else VLLM_MAX_CTX
            except (ValueError, TypeError):
                props["context_size"] = VLLM_MAX_CTX
            # NOTE: do NOT put max_tokens into model_properties. Dify API-side
            # ModelPropertyKey only allows {mode, context_size, ...} and does
            # NOT include MAX_TOKENS; adding it breaks Pydantic validation when
            # the API parses the schema from plugin_daemon. max_tokens is
            # already exposed as a parameter_rule by the cloned template.
            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

curl -s -N -X POST 'http://192.168.110.207/v1/chat-messages' \
  -H 'Authorization: Bearer app-v9yYVTQTrUoSFHNqAhufSGmh' \
  -H 'Content-Type: application/json' \
  -d '{"inputs":{},"query":"你好，介绍一下你自己","response_mode":"streaming","user":"debug"}'

.../openai-1.0.3@<hash>/models/llm/chat.py

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 prompt_messages]),
        stream=stream,
        **params
    )

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

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

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

#!/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

docker exec -it docker-db_postgres-1 psql -U postgres -d dify \
  -c "UPDATE tenant_default_models SET model='qwen3.6-27b' WHERE 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 logs docker-api-1 2>&1 | grep -iE 'gpt-5.6|does not exist|error' | tail -20

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题一 · 步骤 2：控制台 API 添加自定义 Qwen3.6-27B 凭据。"""
from __future__ import annotations

import sys

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


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        body = {
            "model": C.QWEN_MODEL,
            "model_type": "llm",
            "credentials": {
                "openai_api_key": C.VLLM_KEY,
                "openai_api_base": C.VLLM_BASE,
                "openai_organization": "",
                "api_protocol": "chat",
            },
            "name": C.QWEN_MODEL,
        }
        path = "/workspaces/current/model-providers/%s/models/credentials" % C.OPENAI_PROVIDER
        r = d.get(
            "/workspaces/current/model-providers/%s/models/credentials?model_type=llm&model=%s"
            % (C.OPENAI_PROVIDER, C.QWEN_MODEL)
        )
        cid = (r.get("json") or {}).get("current_credential_id")
        if cid:
            body["credential_id"] = cid
            r = d.req("PUT", path, body=body)
            pp("update qwen cred", r)
        else:
            r = d.post(path, body=body)
            pp("add qwen cred", r)
        r2 = d.get("/workspaces/current/models/model-types/llm")
        pp("llm list", r2)
        return 0 if r.get("code") in ("200", "201") else 1
    finally:
        d.close()


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

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题一 · 步骤 3：设默认 LLM（API + tenant_default_models）。"""
from __future__ import annotations

import sys

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


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        r = d.post(
            "/workspaces/current/default-model",
            body={
                "model_settings": [{
                    "model_type": "llm",
                    "provider": C.OPENAI_PROVIDER,
                    "model": C.QWEN_MODEL,
                }]
            },
        )
        pp("set default llm (API)", r)

        sql = (
            "UPDATE tenant_default_models SET model='%s' WHERE model_type='llm';"
            % C.QWEN_MODEL
        )
        out, err = d.sh(
            "docker exec %s psql -U postgres -d dify -c \"%s\""
            % (C.PG_CONTAINER, sql)
        )
        print("tenant_default_models:\n", out or err)

        out2, err2 = d.sh(
            "cd %s && docker compose restart api plugin_daemon" % C.DIFY_DOCKER_DIR,
            t=120,
        )
        print(out2 or err2)
        return 0
    finally:
        d.close()


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

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题一 · 步骤 5：验证 vLLM 可达且 Dify LLM 列表中 qwen active。"""
from __future__ import annotations

import json
import sys

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


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        print("=== 1/2 从 207 curl vLLM /v1/models ===")
        out, err = d.sh(
            "curl -s -m 10 -H 'Authorization: Bearer %s' %s/models"
            % (C.VLLM_KEY, C.VLLM_BASE)
        )
        print(out or err)
        try:
            data = json.loads(out)
            ids = [m.get("id") for m in data.get("data", [])]
            print("model ids:", ids)
            if C.QWEN_MODEL not in ids:
                print("[WARN] vLLM 列表无", C.QWEN_MODEL)
        except Exception:
            pass

        print("\n=== 2/2 Dify model-types/llm ===")
        d.login()
        r = d.get("/workspaces/current/models/model-types/llm")
        pp("llm list", r)
        found = False
        for p in (r.get("json") or {}).get("data") or []:
            if p.get("provider") != C.OPENAI_PROVIDER:
                continue
            for m in p.get("models") or []:
                if m.get("model") == C.QWEN_MODEL:
                    found = True
                    print("qwen status:", m.get("status"), "props:", m.get("model_properties"))
        if not found:
            print("[WARN] Dify LLM 列表未找到", C.QWEN_MODEL)
            return 1
        return 0
    finally:
        d.close()


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

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题一 · 步骤 4 · 修改前：确认 llm.py 尚未打 Qwen schema 补丁。"""
from __future__ import annotations

import sys

import _bootstrap  # noqa: F401
import _config as C
import _plugin_util as PU
from dify_api import Dify, pp


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        path = PU.find_plugin_file(d, "models/llm/llm.py")
        print("llm.py:", path)
        content = PU.container_read(d, C.PLUGIN_CONTAINER, path)
        if PU.LLM_PATCH_MARKER in content:
            print("[INFO] 已存在 Qwen schema fallback 补丁，无需再改")
        else:
            print("[WARN] 尚未打补丁 — 运行 专题一_04_改llm_schema_修改后.py")
            snippet = PU.grep_in_container(d, C.PLUGIN_CONTAINER, path, "get_customizable_model_schema")
            print(snippet[:1200])

        r = d.get("/workspaces/current/models/model-types/llm")
        pp("llm list（不兼容时 status 非 active）", r)
        return 0
    finally:
        d.close()


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

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题一 · 步骤 4 · 修改后：patch llm.py 克隆 chat schema，重启 plugin_daemon。"""
from __future__ import annotations

import sys

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


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        path = PU.find_plugin_file(d, "models/llm/llm.py")
        print("target:", path)
        content = PU.container_read(d, C.PLUGIN_CONTAINER, path)
        if PU.LLM_PATCH_MARKER in content:
            print("already patched")
            return 0
        d.sh(
            "docker exec %s cp %s %s.bak_qwenfix"
            % (C.PLUGIN_CONTAINER, path, path)
        )
        new_content = PU.patch_llm_schema(content)
        PU.container_write(d, C.PLUGIN_CONTAINER, path, new_content)
        out, err = d.sh(
            "docker exec %s python -m py_compile %s" % (C.PLUGIN_CONTAINER, path)
        )
        print("py_compile:", out or err or "ok")
        print("restarting plugin_daemon ...")
        PU.restart_plugin_daemon(d)
        print("done — 回控制台确认 qwen 不再「不兼容」")
        return 0
    finally:
        d.close()


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

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题一 · 步骤 6 · 修改前：确认 chat.py 尚未合并多条 system。"""
from __future__ import annotations

import sys

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


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        path = PU.find_plugin_file(d, "models/llm/chat.py")
        print("chat.py:", path)
        if PU.CHAT_PATCH_MARKER in PU.container_read(d, C.PLUGIN_CONTAINER, path):
            print("[INFO] 已有 _ensure_system_first 补丁")
        else:
            print("[WARN] 尚未打补丁 — Agent 问答可能 400 System message must be at the beginning")
            print(PU.grep_in_container(d, C.PLUGIN_CONTAINER, path, "chat.completions.create")[:800])
        return 0
    finally:
        d.close()


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

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""专题一 · 步骤 6 · 修改后：patch chat.py 合并 system，重启 plugin_daemon。"""
from __future__ import annotations

import sys

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


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        d.login()
        path = PU.find_plugin_file(d, "models/llm/chat.py")
        print("target:", path)
        content = PU.container_read(d, C.PLUGIN_CONTAINER, path)
        if PU.CHAT_PATCH_MARKER in content:
            print("already patched")
            return 0
        d.sh(
            "docker exec %s cp %s %s.bak_systemfix"
            % (C.PLUGIN_CONTAINER, path, path)
        )
        new_content = PU.patch_chat_system(content)
        PU.container_write(d, C.PLUGIN_CONTAINER, path, new_content)
        out, err = d.sh(
            "docker exec %s python -m py_compile %s" % (C.PLUGIN_CONTAINER, path)
        )
        print("py_compile:", out or err or "ok")
        print("restarting plugin_daemon ...")
        PU.restart_plugin_daemon(d)
        print("done — 新开 Agent 对话验证不再 400")
        return 0
    finally:
        d.close()


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