chore: bump version to 4.22.3

* feat(provider/vllm_rerank): add configurable rerank_api_suffix option

Add rerank_api_suffix config option to the VLLM Rerank provider so
users can control the API URL path suffix instead of having /v1/rerank
hardcoded.

- Default value is /v1/rerank (preserves existing behavior)
- Users can set it to empty string to disable auto-append
- Handles suffix without leading slash by auto-adding one
- Schema, default config, and i18n metadata all updated

Issue: Fixes #7238

* fix(provider/vllm_rerank): handle null suffix and improve hint descriptions

- Add explicit None check for rerank_api_suffix (fixes HIGH from Gemini)
- Update rerank_api_base hint to describe actual behavior without
  mentioning specific provider options (fixes 3x MEDIUM from Gemini)
- Add ru-RU i18n for rerank_api_suffix (fixes P2 from Codex)

Co-authored-by: gemini-code-assist[bot]
Co-authored-by: chatgpt-codex-connector[bot]

---------

Co-authored-by: LehaoLin <linlehao@cuhk.edu.cn>

astrbot/cli/__init__.py

@@ -1 +1 @@
-__version__ = "4.22.2"
+__version__ = "4.22.3"

astrbot/core/agent/mcp_client.py

@@ -51,18 +51,29 @@ def _prepare_stdio_env(config: dict) -> dict:
     """Preserve Windows executable resolution for stdio subprocesses."""
     if sys.platform != "win32":
         return config
-
-    pathext = os.environ.get("PATHEXT")
-    if not pathext:
-        return config
-
     prepared = config.copy()
     env = dict(prepared.get("env") or {})
-    env.setdefault("PATHEXT", pathext)
+    env = _merge_environment_variables(env)
     prepared["env"] = env
     return prepared
 
 
+def _merge_environment_variables(env: dict) -> dict:
+    """合并环境变量，处理Windows不区分大小写的情况"""
+    merged = env.copy()
+
+    # 将用户环境变量转换为统一的大小写形式便于比较
+    user_keys_lower = {k.lower(): k for k in merged.keys()}
+
+    for sys_key, sys_value in os.environ.items():
+        sys_key_lower = sys_key.lower()
+        if sys_key_lower not in user_keys_lower:
+            # 使用系统环境变量中的原始大小写
+            merged[sys_key] = sys_value
+
+    return merged
+
+
 async def _quick_test_mcp_connection(config: dict) -> tuple[bool, str]:
     """Quick test MCP server connectivity"""
     import aiohttp

astrbot/core/config/default.py

@@ -5,7 +5,7 @@ from typing import Any, TypedDict
 
 from astrbot.core.utils.astrbot_path import get_astrbot_data_path
 
-VERSION = "4.22.2"
+VERSION = "4.22.3"
 DB_PATH = os.path.join(get_astrbot_data_path(), "data_v4.db")
 PERSONAL_WECHAT_CONFIG_METADATA = {
     "weixin_oc_base_url": {
@@ -1761,6 +1761,7 @@ CONFIG_METADATA_2 = {
                         "enable": True,
                         "rerank_api_key": "",
                         "rerank_api_base": "http://127.0.0.1:8000",
+                        "rerank_api_suffix": "/v1/rerank",
                         "rerank_model": "BAAI/bge-reranker-base",
                         "timeout": 20,
                     },
@@ -1826,7 +1827,12 @@ CONFIG_METADATA_2 = {
                     "rerank_api_base": {
                         "description": "重排序模型 API Base URL",
                         "type": "string",
-                        "hint": "AstrBot 会在请求时在末尾加上 /v1/rerank。",
+                        "hint": "最终请求路径由 Base URL 和路径后缀拼接而成（默认为 /v1/rerank）。",
+                    },
+                    "rerank_api_suffix": {
+                        "description": "API URL 路径后缀",
+                        "type": "string",
+                        "hint": "追加到 base_url 后的路径，如 /v1/rerank。留空则不追加。",
                     },
                     "rerank_api_key": {
                         "description": "API Key",

astrbot/core/provider/sources/gemini_source.py

@@ -241,15 +241,10 @@ class ProviderGoogleGenAI(Provider):
                 thinking_config = types.ThinkingConfig(
                     thinking_budget=thinking_budget,
                 )
-        elif model_name in [
-            "gemini-3-pro",
-            "gemini-3-pro-preview",
-            "gemini-3-flash",
-            "gemini-3-flash-preview",
-            "gemini-3-flash-lite",
-            "gemini-3-flash-lite-preview",
-        ]:
-            # The thinkingLevel parameter, recommended for Gemini 3 models and onwards
+        elif any(model_name.startswith(p) for p in ("gemini-3-", "gemini-3.")):
+            # The thinkingLevel parameter, recommended for Gemini 3 models and onwards.
+            # Use prefix match so new variants (3.1, 3-flash-lite-preview, etc.) are
+            # covered without needing to keep an exhaustive list up to date.
             # Gemini 2.5 series models don't support thinkingLevel; use thinkingBudget instead.
             thinking_level = self.provider_config.get("gm_thinking_config", {}).get(
                 "level", "HIGH"
@@ -517,7 +512,11 @@ class ProviderGoogleGenAI(Provider):
         ):
             chain.append(Comp.Plain("这是图片"))
         for part in result_parts:
-            if part.text:
+            # Skip thinking parts — their text is already captured via
+            # _extract_reasoning_content above.  Including them here would
+            # leak the model's internal reasoning into the user-facing message,
+            # which also causes duplicate/triple replies on some platforms.
+            if part.text and not part.thought:
                 chain.append(Comp.Plain(part.text))
 
             if (

astrbot/core/provider/sources/openai_source.py

@@ -532,6 +532,7 @@ class ProviderOpenAIOfficial(Provider):
             **payloads,
             stream=True,
             extra_body=extra_body,
+            stream_options={"include_usage": True},
         )
 
         llm_response = LLMResponse("assistant", is_chunk=True)
@@ -539,12 +540,10 @@ class ProviderOpenAIOfficial(Provider):
         state = ChatCompletionStreamState()
 
         async for chunk in stream:
-            if not chunk.choices:
-                continue
-            choice = chunk.choices[0]
-            delta = choice.delta
+            choice = chunk.choices[0] if chunk.choices else None
+            delta = choice.delta if choice else None
 
-            if dtcs := delta.tool_calls:
+            if delta and (dtcs := delta.tool_calls):
                 for idx, tc in enumerate(dtcs):
                     # siliconflow workaround
                     if tc.function and tc.function.arguments:
@@ -574,7 +573,7 @@ class ProviderOpenAIOfficial(Provider):
                 _y = True
             if chunk.usage:
                 llm_response.usage = self._extract_usage(chunk.usage)
-            elif choice_usage := getattr(choice, "usage", None):
+            elif choice and (choice_usage := getattr(choice, "usage", None)):
                 # Workaround for some providers that only return usage in choices[].usage, e.g. MoonshotAI
                 # See https://github.com/AstrBotDevs/AstrBot/issues/6614
                 llm_response.usage = self._extract_usage(choice_usage)
@@ -863,6 +862,9 @@ class ProviderOpenAIOfficial(Provider):
 
     def _finally_convert_payload(self, payloads: dict) -> None:
         """Finally convert the payload. Such as think part conversion, tool inject."""
+        model = payloads.get("model", "").lower()
+        is_gemini = "gemini" in model
+
         for message in payloads.get("messages", []):
             if message.get("role") == "assistant" and isinstance(
                 message.get("content"), list
@@ -874,11 +876,24 @@ class ProviderOpenAIOfficial(Provider):
                         reasoning_content += str(part.get("think"))
                     else:
                         new_content.append(part)
-                message["content"] = new_content
-                # reasoning key is "reasoning_content"
+                # Some providers (Grok, etc.) reject empty content lists.
+                # When all parts were think blocks, fall back to None.
+                message["content"] = new_content or None
                 if reasoning_content:
                     message["reasoning_content"] = reasoning_content
 
+            # Gemini 的 function_response 要求 google.protobuf.Struct（即 JSON 对象），
+            # 纯文本会触发 400 Invalid argument，需要包一层 JSON。
+            if is_gemini and message.get("role") == "tool":
+                content = message.get("content", "")
+                if isinstance(content, str):
+                    try:
+                        json.loads(content)
+                    except (json.JSONDecodeError, ValueError):
+                        message["content"] = json.dumps(
+                            {"result": content}, ensure_ascii=False
+                        )
+
     async def _handle_api_error(
         self,
         e: Exception,

astrbot/core/provider/sources/vllm_rerank_source.py

@@ -20,6 +20,11 @@ class VLLMRerankProvider(RerankProvider):
         self.auth_key = provider_config.get("rerank_api_key", "")
         self.base_url = provider_config.get("rerank_api_base", "http://127.0.0.1:8000")
         self.base_url = self.base_url.rstrip("/")
+        self.api_suffix = provider_config.get("rerank_api_suffix", "/v1/rerank")
+        if self.api_suffix is None:
+            self.api_suffix = "/v1/rerank"
+        if self.api_suffix and not self.api_suffix.startswith("/"):
+            self.api_suffix = "/" + self.api_suffix
         self.timeout = provider_config.get("timeout", 20)
         self.model = provider_config.get("rerank_model", "BAAI/bge-reranker-base")
 
@@ -45,8 +50,9 @@ class VLLMRerankProvider(RerankProvider):
         if top_n is not None:
             payload["top_n"] = top_n
         assert self.client is not None
+        rerank_url = f"{self.base_url}{self.api_suffix}"
         async with self.client.post(
-            f"{self.base_url}/v1/rerank",
+            rerank_url,
             json=payload,
         ) as response:
             response_data = await response.json()

changelogs/v4.22.3.md

@@ -0,0 +1,68 @@
+- [更新日志(简体中文)](#chinese)
+- [Changelog(English)](#english)
+
+<a id="chinese"></a>
+
+## What's Changed
+
+### 新增
+
+- 新增 `vllm_rerank` 提供商支持可配置的 `rerank_api_suffix`，可灵活适配不同 API 路径。（[#7278](https://github.com/AstrBotDevs/AstrBot/pull/7278)）
+- 适配 llama.cpp 的 token 用量提取支持，增强 token 统计与计费分析能力。（[#7358](https://github.com/AstrBotDevs/AstrBot/pull/7358)）
+- WebUI 捆绑 Monaco Editor 的 worker 文件，防止从 CDN 下载失败。（[#7249](https://github.com/AstrBotDevs/AstrBot/pull/7249)）
+- 全新的模型/平台总览统计面板。（[#7152](https://github.com/AstrBotDevs/AstrBot/pull/7152)）
+
+### 优化
+
+- 优化 OpenAI 消息处理，若 `content` 列表为空时将其置为 `None`，并降低空内容下的异常率。（[#6551](https://github.com/AstrBotDevs/AstrBot/pull/6551)）
+- 优化 provider page 布局与样式，提升整体设置页面排版一致性。（[#7248](https://github.com/AstrBotDevs/AstrBot/pull/7248)）
+- 用户已登录时将直接跳转欢迎页，减少再次进入登录页的步骤。（[commit](https://github.com/AstrBotDevs/AstrBot/commit/9896b48c)）
+
+### 修复
+
+- 修复 Windows 下 MCP runner 未继承完整系统环境变量导致启动异常的问题。（[#7054](https://github.com/AstrBotDevs/AstrBot/pull/7054)）
+- 修复 Gemini 思考片段泄露到用户可见消息链路的问题。（[#7196](https://github.com/AstrBotDevs/AstrBot/pull/7196)）
+- 修复 Gemini(OpenAI 兼容 API) 下，`function_response` 格式要求导致的 `400 Invalid argument` 错误。（[#7216](https://github.com/AstrBotDevs/AstrBot/pull/7216)
+- 修复知识库初始化失败时的错误提示与恢复问题。（[#7243](https://github.com/AstrBotDevs/AstrBot/pull/7243)）
+- 修复 Bailian Rerank 对新旧响应格式兼容性，避免解析失败。（[#7217](https://github.com/AstrBotDevs/AstrBot/pull/7217)）
+- Docker 镜像补充 AMR 编码器依赖，修复部分部署场景下的音频转码缺失问题。（[#7247](https://github.com/AstrBotDevs/AstrBot/pull/7247)）
+- 修复 pip 安装缺少 `pysocks` 依赖导致的 SOCKS5 代理不可用问题。（[#7221](https://github.com/AstrBotDevs/AstrBot/pull/7221)）
+- 修复 `cached_tokens` 在 `_extract_usage` 中的处理逻辑，避免 token 统计错乱。（[#6719](https://github.com/AstrBotDevs/AstrBot/pull/6719)）
+- 增强 SSE 断开保护，压缩传输下主动发送心跳，减少 WebChat 误断连。（[#7003](https://github.com/AstrBotDevs/AstrBot/pull/7003)）
+- 修复严格 API 下空 assistant 消息导致的 400 错误。（[#7202](https://github.com/AstrBotDevs/AstrBot/pull/7202)）
+- 新增二维码功能缺失的 `qrcode` 依赖修复，恢复二维码生成功能。（[commit](https://github.com/AstrBotDevs/AstrBot/commit/5886c437)）
+- 修复配置器数字输入组件的交互体验问题，避免异常输入场景下的处理偏差。（[#7153](https://github.com/AstrBotDevs/AstrBot/pull/7153)）
+- 修复 `Record.path` 在 Pydantic v2 下被错误视为必填字段的问题。（[#7048](https://github.com/AstrBotDevs/AstrBot/pull/7048)）
+
+<a id="english"></a>
+## What's Changed (EN)
+
+### New Features
+
+- Added configurable `rerank_api_suffix` for the `vllm_rerank` provider to support custom API paths.（[#7278](https://github.com/AstrBotDevs/AstrBot/pull/7278)）
+- Added token usage extraction support for llama.cpp to improve token accounting.（[#7358](https://github.com/AstrBotDevs/AstrBot/pull/7358)）
+- Bundled Monaco Editor workers in WebUI so code editing does not depend on external CDN downloads.（[#7249](https://github.com/AstrBotDevs/AstrBot/pull/7249)）
+- Added a new statistics overview for model and platform usage.（[#7152](https://github.com/AstrBotDevs/AstrBot/pull/7152)）
+- Redirected logged-in users to the welcome page instead of the login page.（[commit](https://github.com/AstrBotDevs/AstrBot/commit/9896b48c)）
+
+### Improvements
+
+- Improved OpenAI message handling by setting message `content` to `None` when content lists are empty, reducing empty-payload edge cases.（[#6551](https://github.com/AstrBotDevs/AstrBot/pull/6551)）
+- Improved provider page layout and styling to keep settings pages consistent.（[#7248](https://github.com/AstrBotDevs/AstrBot/pull/7248)）
+- Refined connection-success handling and messaging when users are already logged in.（[commit](https://github.com/AstrBotDevs/AstrBot/commit/2de23184)）
+
+### Bug Fixes
+
+- Inherited all system environment variables in Windows MCP runner to fix environment-dependent launch issues.（[#7054](https://github.com/AstrBotDevs/AstrBot/pull/7054)）
+- Filtered Gemini thinking parts from user-facing message chains.（[#7196](https://github.com/AstrBotDevs/AstrBot/pull/7196)）
+- Fixed Gemini `function_response` compatibility to prevent `400 Invalid argument` responses.（[#7216](https://github.com/AstrBotDevs/AstrBot/pull/7216)）
+- Improved knowledge base initialization error handling.（[#7243](https://github.com/AstrBotDevs/AstrBot/pull/7243)）
+- Fixed Bailian rerank parsing for both old and new response formats.（[#7217](https://github.com/AstrBotDevs/AstrBot/pull/7217)）
+- Added AMR encoder support in Docker builds to fix missing audio codec support.（[#7247](https://github.com/AstrBotDevs/AstrBot/pull/7247)）
+- Added the `pysocks` dependency so pip installs can use SOCKS5 proxy correctly.（[#7221](https://github.com/AstrBotDevs/AstrBot/pull/7221)）
+- Fixed `_extract_usage` cached token handling.（[#6719](https://github.com/AstrBotDevs/AstrBot/pull/6719)）
+- Sent SSE heartbeat packets to prevent WebChat disconnection under compression.（[#7003](https://github.com/AstrBotDevs/AstrBot/pull/7003)）
+- Fixed `400` errors caused by empty assistant messages on strict APIs.（[#7202](https://github.com/AstrBotDevs/AstrBot/pull/7202)）
+- Added missing `qrcode` dependency to restore QR code generation.（[commit](https://github.com/AstrBotDevs/AstrBot/commit/5886c437)）
+- Fixed number input UX in config renderer.（[#7153](https://github.com/AstrBotDevs/AstrBot/pull/7153)）
+- Fixed `Record.path` being incorrectly treated as required under Pydantic v2.（[#7048](https://github.com/AstrBotDevs/AstrBot/pull/7048)）

dashboard/src/i18n/locales/en-US/features/config-metadata.json

@@ -1075,7 +1075,11 @@
       },
       "rerank_api_base": {
         "description": "Rerank Model API Base URL",
-        "hint": "AstrBot appends /v1/rerank to the request URL."
+        "hint": "The full request URL is formed by combining the Base URL and a path suffix (defaults to /v1/rerank)."
+      },
+      "rerank_api_suffix": {
+        "description": "API URL path suffix",
+        "hint": "Path appended to base_url, e.g. /v1/rerank. Leave empty to disable auto-append."
       },
       "rerank_api_key": {
         "description": "API Key",

dashboard/src/i18n/locales/ru-RU/features/config-metadata.json

@@ -1076,7 +1076,11 @@
             },
             "rerank_api_base": {
                 "description": "Base URL API модели Rerank",
-                "hint": "AstrBot добавляет /v1/rerank к URL запроса."
+                "hint": "Полный URL запроса формируется путём добавления суффикса к Base URL (по умолчанию /v1/rerank)."
+            },
+            "rerank_api_suffix": {
+                "description": "Суффикс пути API",
+                "hint": "Суффикс пути, добавляемый к base_url, например /v1/rerank. Оставьте пустым, чтобы не добавлять."
             },
             "rerank_api_key": {
                 "description": "API Key",

dashboard/src/i18n/locales/zh-CN/features/config-metadata.json

@@ -1077,7 +1077,11 @@
       },
       "rerank_api_base": {
         "description": "重排序模型 API Base URL",
-        "hint": "AstrBot 会在请求时在末尾加上 /v1/rerank。"
+        "hint": "最终请求路径由 Base URL 和路径后缀拼接而成（默认为 /v1/rerank）。"
+      },
+      "rerank_api_suffix": {
+        "description": "API URL 路径后缀",
+        "hint": "追加到 base_url 后的路径后缀，如 /v1/rerank。留空则不追加。"
       },
       "rerank_api_key": {
         "description": "API Key",

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "AstrBot"
-version = "4.22.2"
+version = "4.22.3"
 description = "Easy-to-use multi-platform LLM chatbot and development framework"
 readme = "README.md"
 license = { text = "AGPL-3.0-or-later" }

tests/test_openai_source.py

@@ -2,6 +2,7 @@ from types import SimpleNamespace
 
 import pytest
 from openai.types.chat.chat_completion import ChatCompletion
+from openai.types.chat.chat_completion_chunk import ChatCompletionChunk
 from PIL import Image as PILImage
 
 from astrbot.core.exceptions import EmptyModelOutputError
@@ -1175,6 +1176,93 @@ async def test_parse_openai_completion_raises_empty_model_output_error():
         await provider.terminate()
 
 
+@pytest.mark.asyncio
+async def test_query_stream_extracts_usage_from_empty_choices_chunk(monkeypatch):
+    provider = _make_provider()
+    try:
+        chunks = [
+            ChatCompletionChunk.model_validate(
+                {
+                    "id": "chatcmpl-stream",
+                    "object": "chat.completion.chunk",
+                    "created": 0,
+                    "model": "gpt-4o-mini",
+                    "choices": [
+                        {
+                            "index": 0,
+                            "delta": {
+                                "role": "assistant",
+                                "content": "ok",
+                            },
+                            "finish_reason": None,
+                        }
+                    ],
+                }
+            ),
+            ChatCompletionChunk.model_validate(
+                {
+                    "id": "chatcmpl-stream",
+                    "object": "chat.completion.chunk",
+                    "created": 0,
+                    "model": "gpt-4o-mini",
+                    "choices": [
+                        {
+                            "index": 0,
+                            "delta": {},
+                            "finish_reason": "stop",
+                        }
+                    ],
+                }
+            ),
+            ChatCompletionChunk.model_validate(
+                {
+                    "id": "chatcmpl-stream",
+                    "object": "chat.completion.chunk",
+                    "created": 0,
+                    "model": "gpt-4o-mini",
+                    "choices": [],
+                    "usage": {
+                        "prompt_tokens": 2550,
+                        "completion_tokens": 125,
+                        "total_tokens": 2675,
+                        "prompt_tokens_details": {
+                            "cached_tokens": 2488,
+                        },
+                    },
+                }
+            ),
+        ]
+
+        async def fake_stream():
+            for chunk in chunks:
+                yield chunk
+
+        async def fake_create(**kwargs):
+            return fake_stream()
+
+        monkeypatch.setattr(provider.client.chat.completions, "create", fake_create)
+
+        responses = [
+            response
+            async for response in provider._query_stream(
+                payloads={
+                    "model": "gpt-4o-mini",
+                    "messages": [{"role": "user", "content": "hello"}],
+                },
+                tools=None,
+            )
+        ]
+
+        final_response = responses[-1]
+        assert final_response.completion_text == "ok"
+        assert final_response.usage is not None
+        assert final_response.usage.input_other == 62
+        assert final_response.usage.input_cached == 2488
+        assert final_response.usage.output == 125
+    finally:
+        await provider.terminate()
+
+
 @pytest.mark.asyncio
 async def test_query_filters_empty_assistant_message_without_tool_calls(monkeypatch):
     """Test that empty assistant messages without tool_calls are filtered out."""