fix: stabilize native streaming and improve bot controls

src/langbot/libs/wecom_ai_bot_api/api.py

@@ -2123,7 +2123,7 @@ class WecomBotClient:
 
         Args:
             msg_id: 原始企业微信消息 ID。
-            content: 模型产生的片段内容。
+            content: 流水线传入的累计正文快照。
             is_final: 是否为最终片段。
 
         Returns:
@@ -2138,12 +2138,8 @@ class WecomBotClient:
             return False
 
         previous_content = self._stream_last_content.get(msg_id, '')
-        if previous_content and content.startswith(previous_content):
-            next_content = content
-        elif previous_content and not content:
-            next_content = previous_content
-        else:
-            next_content = previous_content + content if previous_content else content
+        # Pipeline delivery provides a snapshot, not a provider delta.
+        next_content = content or previous_content
         if len(next_content) > _MAX_STREAM_CONTENT_CHARS:
             next_content = next_content[-_MAX_STREAM_CONTENT_CHARS:]
 

src/langbot/libs/wecom_ai_bot_api/ws_client.py

@@ -11,6 +11,7 @@ Official Node.js SDK: https://github.com/WecomTeam/aibot-node-sdk
 from __future__ import annotations
 
 import asyncio
+import hashlib
 import json
 import secrets
 import time
@@ -134,6 +135,9 @@ class WecomBotWsClient:
         self._stream_ids: dict[str, str] = {}  # msg_id -> req_id|stream_id
         # Dedup: skip sending when content hasn't changed
         self._stream_last_content: dict[str, str] = {}  # msg_id -> last content sent
+        # Preserve completion identity after removing active stream mappings.
+        # A runner may emit both a final delta and a completed message snapshot.
+        self._completed_stream_content: dict[str, str] = {}
         # Stream session info for feedback tracking
         self._stream_sessions: dict[str, dict] = {}  # msg_id -> session info
         # Feedback tracking: feedback_id -> session info
@@ -661,16 +665,15 @@ class WecomBotWsClient:
         """
         key = self._stream_ids.get(msg_id)
         if not key:
+            if is_final and self._completed_stream_content.get(msg_id) == hashlib.sha256(content.encode()).hexdigest():
+                return True
             return False
         req_id, stream_id = key.split('|', 1)
         try:
             previous_content = self._stream_last_content.get(msg_id, '')
-            if previous_content and content.startswith(previous_content):
-                next_content = content
-            elif previous_content and not content:
-                next_content = previous_content
-            else:
-                next_content = previous_content + content if previous_content else content
+            # Pipeline delivery provides complete snapshots, including edits
+            # that need not start with the previous text (e.g. reasoning tags).
+            next_content = content or previous_content
             if len(next_content) > _MAX_STREAM_CONTENT_CHARS:
                 next_content = next_content[-_MAX_STREAM_CONTENT_CHARS:]
 
@@ -720,6 +723,8 @@ class WecomBotWsClient:
             await self.reply_stream(req_id, stream_id, next_content, finish=is_final, feedback_id=feedback_id)
             self._stream_last_content[msg_id] = next_content
             if is_final:
+                self._completed_stream_content[msg_id] = hashlib.sha256(next_content.encode()).hexdigest()
+                self._cap_mapping(self._completed_stream_content, _STREAM_CACHE_MAX)
                 self._stream_ids.pop(msg_id, None)
                 self._stream_last_content.pop(msg_id, None)
                 self._stream_sessions.pop(msg_id, None)

src/langbot/pkg/box/service.py

@@ -811,6 +811,10 @@ class BoxService:
         to the sandbox (and vice-versa). It is ``None`` / not a local dir for
         E2B and remote runtimes, where we must fall back to the exec channel.
         """
+        # secure_fs relies on POSIX openat/O_NOFOLLOW. On Windows use the
+        # sandbox execution channel rather than misclassifying EACCES as a link.
+        if os.open not in os.supports_dir_fd or not hasattr(os, 'O_NOFOLLOW'):
+            return None
         root = self._tenant_workspace(self._query_execution_context(query))
         if not root or not os.path.isdir(root) or os.path.islink(root):
             return None

src/langbot/pkg/pipeline/process/handlers/chat.py

@@ -6,6 +6,7 @@ import traceback
 
 
 from .. import handler
+from ..stream_results import coalesce_stream_results
 from ... import entities
 from ... import plugin_diagnostics
 
@@ -132,7 +133,10 @@ class ChatMessageHandler(handler.MessageHandler):
 
                 # Use AgentRunOrchestrator to run the agent
                 # This replaces direct runner lookup and PluginRunnerWrapper
-                async for result in self.ap.agent_run_orchestrator.run_from_query(query):
+                results = self.ap.agent_run_orchestrator.run_from_query(query)
+                if is_stream:
+                    results = coalesce_stream_results(results)
+                async for result in results:
                     has_result = True
                     self._check_response_size(result)
 
@@ -149,6 +153,11 @@ class ChatMessageHandler(handler.MessageHandler):
                         )
                         result.attachments = attachments
 
+                    if is_stream and result.tool_calls:
+                        # A model round ending in tool calls does not end the
+                        # reply. Keep the native stream open for the answer.
+                        result = result.model_copy(update={'is_final': False})
+
                     result.resp_message_id = str(resp_message_id)
 
                     # For streaming mode, pop previous response before adding new chunk
@@ -192,6 +201,11 @@ class ChatMessageHandler(handler.MessageHandler):
 
                     if is_stream:
                         yield entities.StageProcessResult(result_type=entities.ResultType.CONTINUE, new_query=query)
+                        if result.is_final and not result.tool_calls:
+                            # Native adapters release their card on completion.
+                            # A distinct later message needs a fresh card identity.
+                            is_create_card = False
+                            resp_message_id = uuid.uuid4()
 
                 # Log final summary after streaming completes
                 if is_stream:

src/langbot/pkg/pipeline/process/stream_results.py

@@ -0,0 +1,44 @@
+"""Coalesce a runner's terminal delta and completed snapshot before delivery."""
+
+from collections.abc import AsyncIterable, AsyncIterator
+
+from langbot_plugin.api.entities.builtin.provider.message import Message, MessageChunk
+
+
+async def coalesce_stream_results(
+    results: AsyncIterable[Message | MessageChunk],
+) -> AsyncIterator[Message | MessageChunk]:
+    pending: MessageChunk | None = None
+    snapshot = ''
+    async for result in results:
+        if pending is not None:
+            # Completion snapshots can add attachments and usage to the final
+            # delta. Deliver that richer result once, before closing the card.
+            content = pending.all_content if pending.all_content is not None else pending.content
+            same_message = (
+                isinstance(result, Message)
+                and result.role == pending.role
+                and result.tool_calls == pending.tool_calls
+                and (not content or result.content == content)
+            )
+            if not same_message:
+                yield pending
+            pending = None
+            snapshot = ''
+
+        if isinstance(result, MessageChunk) and isinstance(result.content, (str, type(None))):
+            # Runner output is already cumulative (unlike raw provider deltas).
+            # Explicit all_content also supports runners forwarding raw deltas.
+            # Never guess delta semantics from a missing all_content field.
+            snapshot = result.all_content if result.all_content is not None else (result.content or snapshot)
+            result = result.model_copy(update={'all_content': snapshot})
+        elif isinstance(result, Message):
+            snapshot = ''
+
+        if isinstance(result, MessageChunk) and result.is_final:
+            pending = result
+        else:
+            yield result
+
+    if pending is not None:
+        yield pending

src/langbot/pkg/platform/adapters/wecom/manifest.yaml

@@ -8,9 +8,9 @@ metadata:
     zh_Hans: 企业微信
     zh_Hant: 企業微信
   description:
-    en_US: WeCom application message adapter with event-driven orchestration support
-    zh_Hans: 企业微信内部应用消息适配器，支持事件驱动编排
-    zh_Hant: 企業微信內部應用訊息適配器，支援事件驅動編排
+    en_US: Serve employees through a custom WeCom app
+    zh_Hans: 通过企业自建应用，为内部员工提供服务
+    zh_Hant: 透過企業自建應用，為內部員工提供服務
   icon: wecom.png
 
 spec:

src/langbot/pkg/platform/adapters/wecombot/adapter.py

@@ -148,6 +148,16 @@ class WecomBotAdapter(WecomBotAPIMixin, abstract_platform_adapter.AbstractPlatfo
         raw = await self.bot.send_message(str(target_id), content)
         return platform_events.MessageResult(raw={'result': raw})
 
+    async def reply_to_event(
+        self, *, context: dict, target_type: str, target_id: str, message: platform_message.MessageChain
+    ) -> platform_events.MessageResult:
+        """Bind Agent replies to the originating callback and its loading stream."""
+        message_id = ((context.get('delivery') or {}).get('reply_target') or {}).get('message_id')
+        if context.get('event_type') == 'message.received' and message_id:
+            source = await self.get_message(target_type, target_id, message_id)
+            return await self.reply_message(source, message)
+        return await self.send_message(target_type, target_id, message)
+
     async def reply_message(
         self,
         message_source: platform_events.MessageEvent,
@@ -160,9 +170,14 @@ class WecomBotAdapter(WecomBotAPIMixin, abstract_platform_adapter.AbstractPlatfo
         items = await WecomBotMessageConverter.yiri2target(message)
         content = self._join_text_components(items)
         raw = None
+        from .stream_text import format_stream_text
+
+        content = format_stream_text(content)
         if not self.config.get('enable-webhook', False) and event.get('req_id'):
             if content:
-                raw = await self.bot.reply_text(event.get('req_id'), content)
+                finished = await self.bot.push_stream_chunk(event.message_id, content, is_final=True)
+                if not finished:
+                    raw = await self.bot.reply_text(event.get('req_id'), content)
             for item in self._iter_media_components(items):
                 await self._send_media(self.bot, event.get('req_id'), item)
         else:
@@ -181,7 +196,9 @@ class WecomBotAdapter(WecomBotAPIMixin, abstract_platform_adapter.AbstractPlatfo
         if not isinstance(event, WecomBotEvent):
             raise ValueError('WeComBot reply_message_chunk requires a WecomBotEvent source object')
         items = await WecomBotMessageConverter.yiri2target(message)
-        content = self._join_text_components(items)
+        from .stream_text import format_stream_text
+
+        content = format_stream_text(self._join_text_components(items))
         success = await self.bot.push_stream_chunk(event.message_id, content, is_final=is_final)
         if not success and is_final and not self.config.get('enable-webhook', False) and event.get('req_id'):
             await self.bot.reply_text(event.get('req_id'), content)

src/langbot/pkg/platform/adapters/wecombot/manifest.yaml

@@ -8,9 +8,9 @@ metadata:
     zh_Hans: 企业微信智能机器人
     zh_Hant: 企業微信智慧機器人
   description:
-    en_US: WeCom AI Bot adapter with event-driven orchestration support
-    zh_Hans: 企业微信智能机器人适配器，支持事件驱动编排、长连接和 Webhook 两种接入方式
-    zh_Hant: 企業微信智慧機器人適配器，支援事件驅動編排、長連線和 Webhook 兩種接入方式
+    en_US: Answer questions and assist teams in WeCom chats and groups
+    zh_Hans: 在企业微信单聊和群聊中提供智能问答与协作
+    zh_Hant: 在企業微信單聊和群聊中提供智慧問答與協作
   icon: wecombot.png
 
 spec:
@@ -27,10 +27,9 @@ spec:
         zh_Hans: 一键创建机器人
         zh_Hant: 一鍵建立機器人
       description:
-        en_US: 'Scan QR code with WeCom to automatically create a bot and fill in BotId and Secret. Note: Robot Name needs to
-          be filled in manually.'
-        zh_Hans: 使用企业微信扫码自动创建机器人并填写 BotId 和 Secret。注意：机器人名称需手动填写。
-        zh_Hant: 使用企業微信掃碼自動建立機器人並填寫 BotId 和 Secret。注意：機器人名稱需手動填寫。
+        en_US: Scan with WeCom to create a bot and fill in BotId and Secret automatically.
+        zh_Hans: 使用企业微信扫码自动创建机器人并填写 BotId 和 Secret。
+        zh_Hant: 使用企業微信掃碼自動建立機器人並填寫 BotId 和 Secret。
       type: qr-code-login
       login_platform: wecombot
       required: false
@@ -60,8 +59,12 @@ spec:
         en_US: Robot Name
         zh_Hans: 机器人名称
         zh_Hant: 機器人名稱
+      description:
+        en_US: Optional. Only removes matching @name text in group messages; bot mentions are identified by BotId.
+        zh_Hans: 可选。仅用于清理群消息中的 @名称 文本，机器人提及通过 BotId 识别。
+        zh_Hant: 選填。僅用於清理群訊息中的 @名稱 文字，機器人提及透過 BotId 識別。
       type: string
-      required: true
+      required: false
       default: ""
     - name: enable-webhook
       label:
@@ -103,6 +106,10 @@ spec:
       type: string
       required: false
       default: ""
+      show_if:
+        field: enable-webhook
+        operator: neq
+        value: true
     - name: Corpid
       label:
         en_US: Corpid
@@ -115,6 +122,10 @@ spec:
       type: string
       required: false
       default: ""
+      show_if:
+        field: enable-webhook
+        operator: eq
+        value: true
     - name: Token
       label:
         en_US: Token
@@ -127,18 +138,26 @@ spec:
       type: string
       required: false
       default: ""
+      show_if:
+        field: enable-webhook
+        operator: eq
+        value: true
     - name: EncodingAESKey
       label:
         en_US: EncodingAESKey
         zh_Hans: 消息加解密密钥 (EncodingAESKey)
         zh_Hant: 訊息加解密密鑰 (EncodingAESKey)
       description:
-        en_US: Required for webhook mode. Optional for WebSocket mode when encrypted files need to be decrypted.
-        zh_Hans: Webhook 模式必填。WebSocket 模式下如需解密文件则填写。
-        zh_Hant: Webhook 模式必填。WebSocket 模式下如需解密檔案則填寫。
+        en_US: Required for webhook mode. Long connections use the decryption key included in each message.
+        zh_Hans: Webhook 模式必填。长连接模式使用消息自带的解密密钥。
+        zh_Hant: Webhook 模式必填。長連線模式使用訊息自帶的解密金鑰。
       type: string
       required: false
       default: ""
+      show_if:
+        field: enable-webhook
+        operator: eq
+        value: true
     - name: enable-stream-reply
       label:
         en_US: Enable Stream Reply

src/langbot/pkg/platform/adapters/wecombot/message_converter.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import datetime
+import re
 
 from langbot.libs.wecom_ai_bot_api.wecombotevent import WecomBotEvent
 import langbot_plugin.api.definition.abstract.platform.adapter as abstract_platform_adapter
@@ -46,12 +47,14 @@ class WecomBotMessageConverter(abstract_platform_adapter.AbstractMessageConverte
             platform_message.Source(id=event.message_id, time=datetime.datetime.now()),
         ]
         if event.type == 'group' and event.ai_bot_id:
+            # Group callbacks identify the mentioned bot by ID, independently
+            # of its display name or the optional text-cleanup configuration.
             components.append(platform_message.At(target=event.ai_bot_id))
 
         if event.content:
             content = event.content
-            if bot_name:
-                content = content.replace(f'@{bot_name}', '').strip()
+            if event.type == 'group' and bot_name:
+                content = re.sub(rf'(?<!\S)@{re.escape(bot_name)}(?=\s|$)', '', content).strip()
             if content:
                 components.append(platform_message.Plain(text=content))
 

src/langbot/pkg/platform/adapters/wecombot/stream_text.py

@@ -0,0 +1,31 @@
+"""Render multi-round reasoning as one WeCom thinking section."""
+
+import re
+
+
+def format_stream_text(content: str) -> str:
+    if content.strip().lower() in {'<', '<t', '<th', '<thi', '<thin', '<think'}:
+        return ''
+    if '<think' not in content.lower() and '</think' not in content.lower():
+        return content
+    # Every invocation receives the full snapshot, so split tags may simply
+    # wait for the next snapshot instead of leaking into the visible answer.
+    content = re.sub(r'<(?:/?t(?:h(?:i(?:n(?:k)?)?)?)?|/?)$', '', content, flags=re.IGNORECASE)
+    tokens = re.split(r'(</?think\s*>)', content, flags=re.IGNORECASE)
+    thoughts: list[str] = []
+    answers: list[str] = []
+    thinking = False
+    for token in tokens:
+        if re.fullmatch(r'<think\s*>', token, flags=re.IGNORECASE):
+            thinking = True
+        elif re.fullmatch(r'</think\s*>', token, flags=re.IGNORECASE):
+            thinking = False
+        elif token.strip():
+            (thoughts if thinking else answers).append(token.strip())
+    if not thoughts:
+        return '\n\n'.join(answers)
+    reasoning = '\n\n'.join(thoughts)
+    answer = '\n\n'.join(answers)
+    if thinking and not answer:
+        return f'<think>\n{reasoning}'
+    return f'<think>\n{reasoning}\n</think>' + (f'\n\n{answer}' if answer else '')

src/langbot/pkg/platform/adapters/wecomcs/manifest.yaml

@@ -8,9 +8,9 @@ metadata:
     zh_Hans: 企业微信客服
     zh_Hant: 企業微信客服
   description:
-    en_US: WeCom Customer Service adapter with event-driven orchestration support
-    zh_Hans: 企业微信客服适配器，支持事件驱动编排，通过统一 Webhook 接收客服会话消息
-    zh_Hant: 企業微信客服適配器，支援事件驅動編排，透過統一 Webhook 接收客服會話訊息
+    en_US: Support customers on WeChat through WeChat Customer Service
+    zh_Hans: 通过微信客服，为微信用户提供咨询与售后服务
+    zh_Hant: 透過微信客服，為微信使用者提供諮詢與售後服務
   icon: wecom.png
 
 spec: