Skip to content

The agent engineering platform.

fix(langchain): keep tool call / `AIMessage` pairings when summarizing (#34609)

Mason Daugherty2b6911d9af58
diff --git a/libs/langchain_v1/langchain/agents/middleware/summarization.py b/libs/langchain_v1/langchain/agents/middleware/summarization.py
index 37c782842b..8f6fb2789c 100644
--- a/libs/langchain_v1/langchain/agents/middleware/summarization.py
+++ b/libs/langchain_v1/langchain/agents/middleware/summarization.py
@@ -7,6 +7,7 @@ from functools import partial
from typing import Any, Literal, cast
from langchain_core.messages import (
+ AIMessage,
AnyMessage,
MessageLikeRepresentation,
RemoveMessage,
@@ -478,13 +479,37 @@ class SummarizationMiddleware(AgentMiddleware):
def _find_safe_cutoff_point(self, messages: list[AnyMessage], cutoff_index: int) -> int:
"""Find a safe cutoff point that doesn't split AI/Tool message pairs.
- If the message at cutoff_index is a ToolMessage, advance until we find
- a non-ToolMessage. This ensures we never cut in the middle of parallel
- tool call responses.
+ If the message at `cutoff_index` is a `ToolMessage`, search backward for the
+ `AIMessage` containing the corresponding `tool_calls` and adjust the cutoff to
+ include it. This ensures tool call requests and responses stay together.
+
+ Falls back to advancing forward past `ToolMessage` objects only if no matching
+ `AIMessage` is found (edge case).
"""
- while cutoff_index < len(messages) and isinstance(messages[cutoff_index], ToolMessage):
- cutoff_index += 1
- return cutoff_index
+ if cutoff_index >= len(messages) or not isinstance(messages[cutoff_index], ToolMessage):
+ return cutoff_index
+
+ # Collect tool_call_ids from consecutive ToolMessages at/after cutoff
+ tool_call_ids: set[str] = set()
+ idx = cutoff_index
+ while idx < len(messages) and isinstance(messages[idx], ToolMessage):
+ tool_msg = cast("ToolMessage", messages[idx])
+ if tool_msg.tool_call_id:
+ tool_call_ids.add(tool_msg.tool_call_id)
+ idx += 1
+
+ # Search backward for AIMessage with matching tool_calls
+ for i in range(cutoff_index - 1, -1, -1):
+ msg = messages[i]
+ if isinstance(msg, AIMessage) and msg.tool_calls:
+ ai_tool_call_ids = {tc.get("id") for tc in msg.tool_calls if tc.get("id")}
+ if tool_call_ids & ai_tool_call_ids:
+ # Found the AIMessage - move cutoff to include it
+ return i
+
+ # Fallback: no matching AIMessage found, advance past ToolMessages to avoid
+ # orphaned tool responses
+ return idx
def _create_summary(self, messages_to_summarize: list[AnyMessage]) -> str:
"""Generate summary for the given messages."""
diff --git a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_summarization.py b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_summarization.py
index 7c2995061a..c0ac6c82f5 100644
--- a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_summarization.py
+++ b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_summarization.py
@@ -281,8 +281,8 @@ def test_summarization_middleware_profile_inference_triggers_summary() -> None:
]
-def test_summarization_middleware_token_retention_advances_past_tool_messages() -> None:
- """Ensure token retention advances past tool messages for aggressive summarization."""
+def test_summarization_middleware_token_retention_preserves_ai_tool_pairs() -> None:
+ """Ensure token retention preserves AI/Tool message pairs together."""
def token_counter(messages: list[AnyMessage]) -> int:
return sum(len(getattr(message, "content", "")) for message in messages)
@@ -297,7 +297,7 @@ def test_summarization_middleware_token_retention_advances_past_tool_messages()
# Total tokens: 300 + 200 + 50 + 180 + 160 = 890
# Target keep: 500 tokens (50% of 1000)
# Binary search finds cutoff around index 2 (ToolMessage)
- # We advance past it to index 3 (HumanMessage)
+ # We move back to index 1 to preserve the AIMessage with its ToolMessage
messages: list[AnyMessage] = [
HumanMessage(content="H" * 300),
AIMessage(
@@ -314,14 +314,15 @@ def test_summarization_middleware_token_retention_advances_past_tool_messages()
assert result is not None
preserved_messages = result["messages"][2:]
- # With aggressive summarization, we advance past the ToolMessage
- # So we preserve messages from index 3 onward (the two HumanMessages)
- assert preserved_messages == messages[3:]
+ # We move the cutoff back to include the AIMessage with its ToolMessage
+ # So we preserve messages from index 1 onward (AI + Tool + Human + Human)
+ assert preserved_messages == messages[1:]
- # Verify preserved tokens are within budget
- target_token_count = int(1000 * 0.5)
- preserved_tokens = middleware.token_counter(preserved_messages)
- assert preserved_tokens <= target_token_count
+ # Verify the AI/Tool pair is preserved together
+ assert isinstance(preserved_messages[0], AIMessage)
+ assert preserved_messages[0].tool_calls
+ assert isinstance(preserved_messages[1], ToolMessage)
+ assert preserved_messages[1].tool_call_id == preserved_messages[0].tool_calls[0]["id"]
def test_summarization_middleware_missing_profile() -> None:
@@ -666,7 +667,7 @@ def test_summarization_middleware_binary_search_edge_cases() -> None:
def test_summarization_middleware_find_safe_cutoff_point() -> None:
- """Test _find_safe_cutoff_point finds safe cutoff past ToolMessages."""
+ """Test `_find_safe_cutoff_point` preserves AI/Tool message pairs."""
model = FakeToolCallingModel()
middleware = SummarizationMiddleware(
model=model, trigger=("messages", 10), keep=("messages", 2)
@@ -676,7 +677,7 @@ def test_summarization_middleware_find_safe_cutoff_point() -> None:
HumanMessage(content="msg1"),
AIMessage(content="ai", tool_calls=[{"name": "tool", "args": {}, "id": "call1"}]),
ToolMessage(content="result1", tool_call_id="call1"),
- ToolMessage(content="result2", tool_call_id="call2"),
+ ToolMessage(content="result2", tool_call_id="call2"), # orphan - no matching AI
HumanMessage(content="msg2"),
]
@@ -684,8 +685,14 @@ def test_summarization_middleware_find_safe_cutoff_point() -> None:
assert middleware._find_safe_cutoff_point(messages, 0) == 0
assert middleware._find_safe_cutoff_point(messages, 1) == 1
- # Starting at a ToolMessage advances to the next non-ToolMessage
- assert middleware._find_safe_cutoff_point(messages, 2) == 4
+ # Starting at ToolMessage with matching AIMessage moves back to include it
+ # ToolMessage at index 2 has tool_call_id="call1" which matches AIMessage at index 1
+ assert middleware._find_safe_cutoff_point(messages, 2) == 1
+
+ # Starting at orphan ToolMessage (no matching AIMessage) falls back to advancing
+ # ToolMessage at index 3 has tool_call_id="call2" with no matching AIMessage
+ # Since we only collect from cutoff_index onwards, only {call2} is collected
+ # No match found, so we fall back to advancing past ToolMessages
assert middleware._find_safe_cutoff_point(messages, 3) == 4
# Starting at the HumanMessage after tools returns that index
@@ -699,6 +706,65 @@ def test_summarization_middleware_find_safe_cutoff_point() -> None:
assert middleware._find_safe_cutoff_point(messages, len(messages) + 5) == len(messages) + 5
+def test_summarization_middleware_find_safe_cutoff_point_orphan_tool() -> None:
+ """Test `_find_safe_cutoff_point` with truly orphan `ToolMessage` (no matching `AIMessage`)."""
+ model = FakeToolCallingModel()
+ middleware = SummarizationMiddleware(
+ model=model, trigger=("messages", 10), keep=("messages", 2)
+ )
+
+ # Messages where ToolMessage has no matching AIMessage at all
+ messages: list[AnyMessage] = [
+ HumanMessage(content="msg1"),
+ AIMessage(content="ai_no_tools"), # No tool_calls
+ ToolMessage(content="orphan_result", tool_call_id="orphan_call"),
+ HumanMessage(content="msg2"),
+ ]
+
+ # Starting at orphan ToolMessage falls back to advancing forward
+ assert middleware._find_safe_cutoff_point(messages, 2) == 3
+
+
+def test_summarization_cutoff_moves_backward_to_include_ai_message() -> None:
+ """Test that cutoff moves backward to include `AIMessage` with its `ToolMessage`s.
+
+ Previously, when the cutoff landed on a `ToolMessage`, the code would advance
+ FORWARD past all `ToolMessage`s. This could result in orphaned `ToolMessage`s (kept
+ without their `AIMessage`) or aggressive summarization that removed AI/Tool pairs.
+
+ The fix searches backward from a `ToolMessage` to find the `AIMessage` with matching
+ `tool_calls`, ensuring the pair stays together in the preserved messages.
+ """
+ model = FakeToolCallingModel()
+ middleware = SummarizationMiddleware(
+ model=model, trigger=("messages", 10), keep=("messages", 2)
+ )
+
+ # Scenario: cutoff lands on ToolMessage that has a matching AIMessage before it
+ messages: list[AnyMessage] = [
+ HumanMessage(content="initial question"), # index 0
+ AIMessage(
+ content="I'll use a tool",
+ tool_calls=[{"name": "search", "args": {"q": "test"}, "id": "call_abc"}],
+ ), # index 1
+ ToolMessage(content="search result", tool_call_id="call_abc"), # index 2
+ HumanMessage(content="followup"), # index 3
+ ]
+
+ # When cutoff is at index 2 (ToolMessage), it should move BACKWARD to index 1
+ # to include the AIMessage that generated the tool call
+ result = middleware._find_safe_cutoff_point(messages, 2)
+
+ assert result == 1, (
+ f"Expected cutoff to move backward to index 1 (AIMessage), got {result}. "
+ "The cutoff should preserve AI/Tool pairs together."
+ )
+
+ assert isinstance(messages[result], AIMessage)
+ assert messages[result].tool_calls # type: ignore[union-attr]
+ assert messages[result].tool_calls[0]["id"] == "call_abc" # type: ignore[union-attr]
+
+
def test_summarization_middleware_zero_and_negative_target_tokens() -> None:
"""Test handling of edge cases with target token calculations."""
# Test with very small fraction that rounds to zero
@@ -814,7 +880,7 @@ def test_summarization_adjust_token_counts() -> None:
def test_summarization_middleware_many_parallel_tool_calls_safety() -> None:
- """Test cutoff safety with many parallel tool calls extending beyond old search range."""
+ """Test cutoff safety preserves AI message with many parallel tool calls."""
middleware = SummarizationMiddleware(
model=MockChatModel(), trigger=("messages", 15), keep=("messages", 5)
)
@@ -826,20 +892,21 @@ def test_summarization_middleware_many_parallel_tool_calls_safety() -> None:
]
messages: list[AnyMessage] = [human_message, ai_message, *tool_messages]
- # Cutoff at index 7 (a ToolMessage) advances to index 12 (end of messages)
- assert middleware._find_safe_cutoff_point(messages, 7) == 12
+ # Cutoff at index 7 (a ToolMessage) moves back to index 1 (AIMessage)
+ # to preserve the AI/Tool pair together
+ assert middleware._find_safe_cutoff_point(messages, 7) == 1
- # Any cutoff pointing at a ToolMessage (indices 2-11) advances to index 12
+ # Any cutoff pointing at a ToolMessage (indices 2-11) moves back to index 1
for i in range(2, 12):
- assert middleware._find_safe_cutoff_point(messages, i) == 12
+ assert middleware._find_safe_cutoff_point(messages, i) == 1
# Cutoff at index 0, 1 (before tool messages) stays the same
assert middleware._find_safe_cutoff_point(messages, 0) == 0
assert middleware._find_safe_cutoff_point(messages, 1) == 1
-def test_summarization_middleware_find_safe_cutoff_advances_past_tools() -> None:
- """Test _find_safe_cutoff advances past ToolMessages to find safe cutoff."""
+def test_summarization_middleware_find_safe_cutoff_preserves_ai_tool_pair() -> None:
+ """Test `_find_safe_cutoff` preserves AI/Tool message pairs together."""
middleware = SummarizationMiddleware(
model=MockChatModel(), trigger=("messages", 10), keep=("messages", 3)
)
@@ -862,15 +929,15 @@ def test_summarization_middleware_find_safe_cutoff_advances_past_tools() -> None
]
# Target cutoff index is len(messages) - messages_to_keep = 6 - 3 = 3
- # Index 3 is a ToolMessage, so we advance past the tool sequence to index 5
+ # Index 3 is a ToolMessage, we move back to index 1 to include AIMessage
cutoff = middleware._find_safe_cutoff(messages, messages_to_keep=3)
- assert cutoff == 5
+ assert cutoff == 1
# With messages_to_keep=2, target cutoff index is 6 - 2 = 4
- # Index 4 is a ToolMessage, so we advance past the tool sequence to index 5
- # This is aggressive - we keep only 1 message instead of 2
+ # Index 4 is a ToolMessage, we move back to index 1 to include AIMessage
+ # This preserves the AI + Tools + Human, more than requested but valid
cutoff = middleware._find_safe_cutoff(messages, messages_to_keep=2)
- assert cutoff == 5
+ assert cutoff == 1
def test_summarization_middleware_cutoff_at_start_of_tool_sequence() -> None: