feat(pipeline): add keep-first-think-only toggle for multi-round CoT

src/langbot/libs/wecom_ai_bot_api/ws_client.py

@@ -433,7 +433,9 @@ async def send_template_card(self, chat_id: str, card_payload: dict[str, Any]) -
         body['chatid'] = chat_id
         return await self._send_reply(req_id, body, cmd=CMD_SEND_MSG)
 
-    async def push_stream_chunk(self, msg_id: str, content: str, is_final: bool = False) -> bool:
+    async def push_stream_chunk(
+        self, msg_id: str, content: str, is_final: bool = False, keep_stream: bool = False
+    ) -> bool:
         """Push a streaming chunk for a given message ID.
 
         Compatible interface with WecomBotClient.push_stream_chunk.
@@ -442,6 +444,9 @@ async def push_stream_chunk(self, msg_id: str, content: str, is_final: bool = Fa
             msg_id: The original message ID.
             content: The cumulative content from the pipeline.
             is_final: Whether this is the final chunk.
+            keep_stream: When True, keep the stream session alive even on
+                is_final so a subsequent round (e.g. tool-call loop) can
+                reuse the same stream_id.
 
         Returns:
             True if the stream session exists and chunk was sent.
@@ -490,7 +495,7 @@ async def push_stream_chunk(self, msg_id: str, content: str, is_final: bool = Fa
             # every frame must contain the complete snapshot, not only a delta.
             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:
+            if is_final and not keep_stream:
                 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/pipeline/respback/respback.py

@@ -46,12 +46,22 @@ async def process(self, query: pipeline_query.Query, stage_inst_name: str) -> en
         try:
             if await query.adapter.is_stream_output_supported() and has_chunks:
                 is_final = [msg.is_final for msg in query.resp_messages][-1]
+                bot_msg = query.resp_messages[-1]
+                # Read the keep_stream hint that the runner may have set on
+                # the MessageChunk's provider_specific_fields.  This tells
+                # adapters (e.g. WeComBot WS) to keep the stream session
+                # alive across multi-round tool-call loops.
+                keep_stream = False
+                psf = getattr(bot_msg, 'provider_specific_fields', None)
+                if psf:
+                    keep_stream = bool(psf.get('keep_stream', False))
                 await query.adapter.reply_message_chunk(
                     message_source=query.message_event,
-                    bot_message=query.resp_messages[-1],
+                    bot_message=bot_msg,
                     message=message_chain,
                     quote_origin=quote_origin,
                     is_final=is_final,
+                    keep_stream=keep_stream,
                 )
             else:
                 await query.adapter.reply_message(

src/langbot/pkg/platform/sources/dingtalk.py

@@ -571,6 +571,7 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ):
         message_id = bot_message.resp_message_id
         msg_seq = bot_message.msg_sequence

src/langbot/pkg/platform/sources/discord.py

@@ -1261,6 +1261,7 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ):
         msg_id = (
             bot_message.get('resp_message_id')

src/langbot/pkg/platform/sources/http_bot.py

@@ -468,6 +468,7 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ) -> dict:
         message_is_final = is_final and getattr(bot_message, 'tool_calls', None) is None
         return await self._emit_reply(message_source, message, is_final=message_is_final, stream=True)

src/langbot/pkg/platform/sources/lark.py

@@ -2072,6 +2072,7 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ):
         """
         回复消息变成更新卡片消息

src/langbot/pkg/platform/sources/qqofficial.py

@@ -532,6 +532,7 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ):
         # Periodically clean up stale stream contexts
         await self._cleanup_stale_streams()

src/langbot/pkg/platform/sources/telegram.py

@@ -563,6 +563,7 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ):
         message_id = bot_message.resp_message_id
         msg_seq = bot_message.msg_sequence

src/langbot/pkg/platform/sources/web_page_bot_adapter.py

@@ -63,10 +63,11 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ) -> dict:
         if self._ws_adapter is not None:
             return await self._ws_adapter.reply_message_chunk(
-                message_source, bot_message, message, quote_origin, is_final
+                message_source, bot_message, message, quote_origin, is_final, keep_stream=keep_stream
             )
         return {}
 

src/langbot/pkg/platform/sources/websocket_adapter.py

@@ -252,6 +252,7 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ) -> dict:
         """回复消息块 - 流式"""
         # 获取会话和pipeline信息

src/langbot/pkg/platform/sources/wecombot.py

@@ -410,6 +410,7 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ):
         content = await self.message_converter.yiri2target(message)
         _ws_mode = not self.config.get('enable-webhook', False)
@@ -463,7 +464,7 @@ async def reply_message_chunk(
             return {'stream': False, 'form': True, 'fallback': True}
 
         if _ws_mode:
-            success = await self.bot.push_stream_chunk(msg_id, content, is_final=is_final)
+            success = await self.bot.push_stream_chunk(msg_id, content, is_final=is_final, keep_stream=keep_stream)
             if not success and is_final:
                 event = message_source.source_platform_object
                 req_id = event.get('req_id', '')

src/langbot/pkg/provider/runners/localagent.py

@@ -481,7 +481,12 @@ async def run(
         except AttributeError:
             is_stream = False
 
-        remove_think = ((query.pipeline_config.get('output') or {}).get('misc') or {}).get('remove-think', False)
+        _misc = (query.pipeline_config.get('output') or {}).get('misc') or {}
+        remove_think = _misc.get('remove-think', False)
+        keep_first_think_only = _misc.get('keep-first-think-only', False)
+        # When keep-first-think-only is on, the first (pre-loop) round keeps
+        # its CoT; subsequent tool-call rounds strip it as usual.
+        _strip_think_first_round = remove_think and not keep_first_think_only
 
         # Build ordered candidate list (primary + fallbacks)
         candidates = await self._get_model_candidates(query)
@@ -500,23 +505,30 @@ async def run(
                 candidates,
                 req_messages,
                 query.use_funcs,
-                remove_think,
+                _strip_think_first_round,
             )
             final_msg = msg
         else:
             # Streaming: invoke with fallback
-            stream_accumulator = _StreamAccumulator(msg_sequence=1, remove_think=remove_think)
+            stream_accumulator = _StreamAccumulator(msg_sequence=1, remove_think=_strip_think_first_round)
 
             stream_src, use_llm_model = await self._invoke_stream_with_fallback(
                 query,
                 candidates,
                 req_messages,
                 query.use_funcs,
-                remove_think,
+                _strip_think_first_round,
             )
             async for msg in stream_src:
                 chunk = stream_accumulator.add(msg)
                 if chunk:
+                    # When keep-first-think-only is active and this round
+                    # produced tool calls, signal the adapter to keep the
+                    # stream session alive so the next round can reuse it.
+                    if keep_first_think_only and chunk.is_final and chunk.tool_calls:
+                        psf = dict(chunk.provider_specific_fields or {})
+                        psf['keep_stream'] = True
+                        chunk = chunk.model_copy(update={'provider_specific_fields': psf})
                     yield chunk
                     initial_response_emitted = True
 

src/langbot/templates/default-pipeline-config.json

@@ -107,7 +107,8 @@
             "at-sender": true,
             "quote-origin": true,
             "track-function-calls": false,
-            "remove-think": false
+            "remove-think": false,
+            "keep-first-think-only": false
         }
     }
 }

src/langbot/templates/metadata/pipeline/output.yaml

@@ -145,4 +145,14 @@ stages:
         type: boolean
         required: true
         default: false
+      - name: keep-first-think-only
+        label:
+          en_US: Keep Only First CoT
+          zh_Hans: 仅保留第一条思维链
+        description:
+          en_US: 'Only effective when "Remove CoT" is enabled. Keeps the chain-of-thought on the first LLM round but strips it from all subsequent rounds in a multi-round tool-call loop.'
+          zh_Hans: '仅在启用"删除思维链"时生效。多轮工具调用中，保留第一轮的思维链，后续轮次的思维链一律删除。'
+        type: boolean
+        required: true
+        default: false
 

tests/factories/platform.py

@@ -230,6 +230,7 @@ async def reply_message_chunk(
         message: platform_message.MessageChain,
         quote_origin: bool = False,
         is_final: bool = False,
+        keep_stream: bool = False,
     ):
         """Simulate streaming reply (captures for assertions)."""
         if self._raise_error:
@@ -244,6 +245,7 @@ async def reply_message_chunk(
                 'message': message,
                 'quote_origin': quote_origin,
                 'is_final': is_final,
+                'keep_stream': keep_stream,
             }
         )