From aa6d3b38b38793f948c1aa8ac5fb3b1b8d644e5d Mon Sep 17 00:00:00 2001
From: filipi87 <filipi87@gmail.com>
Date: Fri, 27 Feb 2026 12:53:25 -0300
Subject: [PATCH] Add explanatory comments for LLMSpecificMessage guards in
 context summarization, amd fixed the missing guard in
 LLMContextSummarizer._apply_summary when searching for the first system
 message.

---
 .../aggregators/llm_context_summarizer.py        | 16 +++++++++++++---
 .../utils/context/llm_context_summarization.py   | 15 ++++++++++++++-
 2 files changed, 27 insertions(+), 4 deletions(-)

diff --git a/src/pipecat/processors/aggregators/llm_context_summarizer.py b/src/pipecat/processors/aggregators/llm_context_summarizer.py
index a1a613ccc..7886fcf12 100644
--- a/src/pipecat/processors/aggregators/llm_context_summarizer.py
+++ b/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 :]
diff --git a/src/pipecat/utils/context/llm_context_summarization.py b/src/pipecat/utils/context/llm_context_summarization.py
index 06551e3bb..537cc91ab 100644
--- a/src/pipecat/utils/context/llm_context_summarization.py
+++ b/src/pipecat/utils/context/llm_context_summarization.py
@@ -188,6 +188,8 @@ 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
 
@@ -251,6 +253,9 @@ 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
 
@@ -302,7 +307,10 @@ 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
@@ -367,6 +375,11 @@ 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