From 258917a879fc728675e51403203e8b1e4f8bc808 Mon Sep 17 00:00:00 2001
From: Nishitha M <32355027+imnishitha@users.noreply.github.com>
Date: Wed, 5 Aug 2026 18:59:16 +0000
Subject: [PATCH] fix(langchain): stop `SummarizationMiddleware` pulling closed
 turns back into retention

`_snap_cutoff_past_open_thinking_turn` scanned backward from the
candidate cutoff even when that cutoff already landed on a
`HumanMessage`/`SystemMessage` - a boundary right after a completed
assistant turn, with nothing open to preserve. For a history like
`Human -> AI(thinking+tool) -> Tool -> Human` with the cutoff on the
final human message, the scan walked through the finished turn and,
because it contained a thinking block, incorrectly pulled the cutoff
back to the earlier `AIMessage`, retaining an old turn beyond the
configured retention budget (and undoing the orphan-tool forward
fallback).

The helper now returns the cutoff unchanged whenever
`messages[cutoff_index]` is itself a human/system boundary, before any
backward scanning happens.

Co-authored-by: open-swe[bot] <open-swe@users.noreply.github.com>
---
 .../agents/middleware/summarization.py        |  7 ++++
 .../implementations/test_summarization.py     | 34 +++++++++++++++++++
 2 files changed, 41 insertions(+)

diff --git a/libs/langchain_v1/langchain/agents/middleware/summarization.py b/libs/langchain_v1/langchain/agents/middleware/summarization.py
index 912e269a2c..b773d18057 100644
--- a/libs/langchain_v1/langchain/agents/middleware/summarization.py
+++ b/libs/langchain_v1/langchain/agents/middleware/summarization.py
@@ -844,6 +844,13 @@ class SummarizationMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, R
         if cutoff_index <= 0 or cutoff_index >= len(messages):
             return cutoff_index
 
+        # A cutoff that already lands on a HumanMessage/SystemMessage is itself a
+        # turn boundary right after the preceding (completed) assistant turn -
+        # there's nothing open to preserve, so leave it alone rather than scanning
+        # backward into that already-finished turn.
+        if isinstance(messages[cutoff_index], (HumanMessage, SystemMessage)):
+            return cutoff_index
+
         turn_start = cutoff_index
         while turn_start > 0:
             previous_message = messages[turn_start - 1]
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 17b629d39a..bfa7c0f8dc 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
@@ -1017,6 +1017,40 @@ def test_summarization_middleware_find_safe_cutoff_point_no_snap_without_thinkin
     assert middleware._find_safe_cutoff_point(messages, 3) == 3
 
 
+def test_summarization_middleware_find_safe_cutoff_point_closed_turn_not_reopened() -> None:
+    """A cutoff already on a Human/System boundary must not be pulled backward.
+
+    `Human -> AI(thinking+tool) -> Tool -> Human` with the cutoff on the final
+    `HumanMessage` is already a safe boundary right after a *completed* assistant
+    turn. Scanning backward from there (as opposed to refusing to scan at all)
+    would walk through that finished turn and, because it contains a thinking
+    block, incorrectly pull the cutoff back to the earlier `AIMessage` - retaining
+    an old turn beyond the configured retention budget.
+    """
+    model = FakeToolCallingModel()
+    middleware = SummarizationMiddleware(
+        model=model, trigger=("messages", 10), keep=("messages", 2)
+    )
+
+    def thinking_block(text: str) -> dict[str, str]:
+        return {"type": "thinking", "thinking": text, "signature": "sig"}
+
+    messages: list[AnyMessage] = [
+        HumanMessage(content="first request", id="h0"),  # index 0
+        AIMessage(
+            content=[thinking_block("plan")],
+            tool_calls=[{"name": "get_data", "args": {}, "id": "call_get"}],
+            id="a1",
+        ),  # index 1
+        ToolMessage(content="raw data", tool_call_id="call_get", id="t1"),  # index 2
+        HumanMessage(content="second request", id="h1"),  # index 3 - naive cutoff, closed turn
+    ]
+
+    # The candidate cutoff (index 3) already sits right after the completed turn
+    # (index 1-2). It must be returned unchanged, not pulled back to index 1.
+    assert middleware._find_safe_cutoff_point(messages, 3) == 3
+
+
 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