Fix context summarization failing with mid-conversation system messages

Only treat messages[0] as the initial system prompt when determining the
summarization range. Previously, the code scanned the entire context for
the first system-role message, which caused failures when the only system
message was a mid-conversation injection (e.g. "The user has been quiet").
In that case summary_start exceeded summary_end, producing an empty range
and "No messages to summarize" errors.

Fixes #4286

src/pipecat/processors/aggregators/llm_context_summarizer.py

@@ -429,18 +429,17 @@ class LLMContextSummarizer(BaseObject):
         config = self._auto_config.summary_config
         messages = self._context.messages
 
-        # 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,
-        )
+        # Preserve the first message if it is a system message (initial system prompt).
+        # Only messages[0] is treated as the system preamble — system messages at
+        # other positions are mid-conversation injections and are not preserved
+        # separately (they will be part of the summary or the recent messages).
+        first_system_msg = None
+        if (
+            messages
+            and not isinstance(messages[0], LLMSpecificMessage)
+            and messages[0].get("role") == "system"
+        ):
+            first_system_msg = messages[0]
 
         # Get recent messages to keep
         recent_messages = messages[last_summarized_index + 1 :]

src/pipecat/utils/context/llm_context_summarization.py

@@ -522,25 +522,19 @@ class LLMContextSummarizationUtil:
         if len(messages) <= min_messages_to_keep:
             return LLMMessagesToSummarize(messages=[], last_summarized_index=-1)
 
-        # 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 not isinstance(msg, LLMSpecificMessage) and msg.get("role") == "system"
-            ),
-            -1,
+        # Check if the first message is a system message (initial system prompt).
+        # Only messages[0] is treated as the system message to preserve — system
+        # messages at other positions are mid-conversation injections and should be
+        # included in the summarization range.
+        first_msg = messages[0] if messages else None
+        first_is_system = (
+            first_msg is not None
+            and not isinstance(first_msg, LLMSpecificMessage)
+            and first_msg.get("role") == "system"
         )
 
-        # Messages to summarize are between first system and recent messages
-        # We exclude the first system message itself
-        if first_system_index >= 0:
-            summary_start = first_system_index + 1
-        else:
-            summary_start = 0
+        # Start summarization after the initial system message if present
+        summary_start = 1 if first_is_system else 0
 
         # Get messages to keep (last N messages)
         summary_end = len(messages) - min_messages_to_keep

tests/test_context_summarization.py

@@ -136,6 +136,58 @@ class TestContextSummarizationMixin(unittest.TestCase):
         self.assertEqual(len(result.messages), 0)
         self.assertEqual(result.last_summarized_index, -1)
 
+    def test_get_messages_to_summarize_mid_conversation_system_message(self):
+        """Test that a system message mid-conversation is included in summarization.
+
+        Regression test for #4286: when the only system message is a mid-conversation
+        injection (not at index 0), get_messages_to_summarize should start from index 0,
+        not after the mid-conversation system message.
+        """
+        context = LLMContext()
+
+        # No system message at index 0 (using system_instruction instead)
+        context.add_message({"role": "user", "content": "Hello"})
+        context.add_message({"role": "assistant", "content": "Hi there"})
+        context.add_message({"role": "user", "content": "Tell me a joke"})
+        context.add_message({"role": "assistant", "content": "Why did the chicken..."})
+        # Mid-conversation system injection (e.g. "The user has been quiet")
+        context.add_message({"role": "system", "content": "The user has been quiet for a while"})
+        context.add_message({"role": "user", "content": "Latest message"})
+        context.add_message({"role": "assistant", "content": "Latest response"})
+
+        # Keep last 2 messages
+        result = LLMContextSummarizationUtil.get_messages_to_summarize(context, 2)
+
+        # Should summarize from index 0 (no system preamble to skip)
+        # through index 4, keeping last 2 messages
+        self.assertEqual(len(result.messages), 5)
+        self.assertEqual(result.messages[0]["content"], "Hello")
+        # The mid-conversation system message should be included in summarization
+        self.assertEqual(result.messages[4]["content"], "The user has been quiet for a while")
+        self.assertEqual(result.last_summarized_index, 4)
+
+    def test_get_messages_to_summarize_system_at_index_0_with_mid_system(self):
+        """Test that only messages[0] system message is preserved, not later ones."""
+        context = LLMContext()
+
+        context.add_message({"role": "system", "content": "You are helpful"})
+        context.add_message({"role": "user", "content": "Hello"})
+        context.add_message({"role": "assistant", "content": "Hi"})
+        # Mid-conversation system injection
+        context.add_message({"role": "system", "content": "The user seems frustrated"})
+        context.add_message({"role": "user", "content": "Help me"})
+        context.add_message({"role": "assistant", "content": "Sure"})
+
+        # Keep last 2 messages
+        result = LLMContextSummarizationUtil.get_messages_to_summarize(context, 2)
+
+        # Should skip index 0 (system preamble), summarize indices 1-3
+        self.assertEqual(len(result.messages), 3)
+        self.assertEqual(result.messages[0]["content"], "Hello")
+        # Mid-conversation system message is summarized, not treated as preamble
+        self.assertEqual(result.messages[2]["content"], "The user seems frustrated")
+        self.assertEqual(result.last_summarized_index, 3)
+
     def test_format_messages_for_summary(self):
         """Test message formatting for summary."""