Merge pull request #3794 from omChauhanDev/fix/context-summarization-llm-specific-message

skipping provider-specific messages during summarization

src/pipecat/processors/aggregators/llm_context_summarizer.py

@@ -18,7 +18,7 @@ from pipecat.frames.frames import (
     LLMContextSummaryResultFrame,
     LLMFullResponseStartFrame,
 )
-from pipecat.processors.aggregators.llm_context import LLMContext
+from pipecat.processors.aggregators.llm_context import LLMContext, LLMSpecificMessage
 from pipecat.utils.asyncio.task_manager import BaseTaskManager
 from pipecat.utils.base_object import BaseObject
 from pipecat.utils.context.llm_context_summarization import (
@@ -290,8 +290,18 @@ class LLMContextSummarizer(BaseObject):
         """
         messages = self._context.messages
 
-        # Find the first system message to preserve
-        first_system_msg = next((msg for msg in messages if msg.get("role") == "system"), None)
+        # Find the first system message to preserve. LLMSpecificMessage instances are excluded
+        # because they are not dict-like and never represent a system message; they hold
+        # service-specific metadata (e.g. thinking blocks) that is always paired with a
+        # standard message.
+        first_system_msg = next(
+            (
+                msg
+                for msg in messages
+                if not isinstance(msg, LLMSpecificMessage) and msg.get("role") == "system"
+            ),
+            None,
+        )
 
         # Get recent messages to keep
         recent_messages = messages[last_summarized_index + 1 :]

src/pipecat/utils/context/llm_context_summarization.py

@@ -15,7 +15,7 @@ from typing import List, Optional
 
 from loguru import logger
 
-from pipecat.processors.aggregators.llm_context import LLMContext
+from pipecat.processors.aggregators.llm_context import LLMContext, LLMSpecificMessage
 
 # Token estimation constants
 CHARS_PER_TOKEN = 4  # Industry-standard heuristic: 1 token ≈ 4 characters
@@ -188,6 +188,11 @@ class LLMContextSummarizationUtil:
         total = 0
 
         for message in context.messages:
+            # LLMSpecificMessage holds service-specific data (e.g. thinking blocks,
+            # thought signatures). Skipping them here for now.
+            if isinstance(message, LLMSpecificMessage):
+                continue
+
             # Role and structure overhead
             total += TOKEN_OVERHEAD_PER_MESSAGE
 
@@ -248,6 +253,12 @@ class LLMContextSummarizationUtil:
 
         for i in range(start_idx, len(messages)):
             msg = messages[i]
+            # LLMSpecificMessage instances (e.g. thinking blocks) never carry tool_call or
+            # tool_call_id fields, so they cannot affect the pending-call tracking. Skipping
+            # them avoids an AttributeError.
+            if isinstance(msg, LLMSpecificMessage):
+                continue
+
             role = msg.get("role")
 
             # Check for tool calls in assistant messages
@@ -296,9 +307,17 @@ class LLMContextSummarizationUtil:
         if len(messages) <= min_messages_to_keep:
             return LLMMessagesToSummarize(messages=[], last_summarized_index=-1)
 
-        # Find first system message index
+        # Find first system message index. LLMSpecificMessage instances are excluded because
+        # they are not dict-like and never represent a system message; they hold
+        # service-specific metadata (e.g. thinking blocks) that is always paired with a
+        # standard message.
         first_system_index = next(
-            (i for i, msg in enumerate(messages) if msg.get("role") == "system"), -1
+            (
+                i
+                for i, msg in enumerate(messages)
+                if not isinstance(msg, LLMSpecificMessage) and msg.get("role") == "system"
+            ),
+            -1,
         )
 
         # Messages to summarize are between first system and recent messages
@@ -356,6 +375,14 @@ class LLMContextSummarizationUtil:
         transcript_parts = []
 
         for msg in messages:
+            # LLMSpecificMessage holds service-specific internal data (e.g. Anthropic thinking
+            # blocks, Gemini thought signatures). This data is not meaningful as plain text for
+            # a summarization transcript, and the summarizer LLM would not know how to interpret
+            # it. The conversational content of those turns is already captured by the
+            # accompanying standard assistant message.
+            if isinstance(msg, LLMSpecificMessage):
+                continue
+
             role = msg.get("role", "unknown")
             content = msg.get("content", "")