Skip to content

The agent engineering platform.

fix(langchain): propagate model middleware control flow (#39199)

ccurme698602eedf12
diff --git a/libs/langchain_v1/langchain/agents/middleware/model_fallback.py b/libs/langchain_v1/langchain/agents/middleware/model_fallback.py
index 7bd1961470..6f5fa29d05 100644
--- a/libs/langchain_v1/langchain/agents/middleware/model_fallback.py
+++ b/libs/langchain_v1/langchain/agents/middleware/model_fallback.py
@@ -19,6 +19,7 @@ import logging
from typing import TYPE_CHECKING, Any
from langchain_core.tools import BaseTool
+from langgraph.errors import GraphBubbleUp
from langchain.agents.middleware.types import (
AgentMiddleware,
@@ -344,6 +345,8 @@ class ModelFallbackMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, R
last_exception: Exception
try:
return handler(request)
+ except GraphBubbleUp:
+ raise
except Exception as e:
last_exception = e
@@ -359,6 +362,8 @@ class ModelFallbackMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, R
)
try:
return handler(fallback_request.override(model=fallback_model))
+ except GraphBubbleUp:
+ raise
except Exception as e:
last_exception = e
continue
@@ -386,6 +391,8 @@ class ModelFallbackMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, R
last_exception: Exception
try:
return await handler(request)
+ except GraphBubbleUp:
+ raise
except Exception as e:
last_exception = e
@@ -401,6 +408,8 @@ class ModelFallbackMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, R
)
try:
return await handler(fallback_request.override(model=fallback_model))
+ except GraphBubbleUp:
+ raise
except Exception as e:
last_exception = e
continue
diff --git a/libs/langchain_v1/langchain/agents/middleware/model_retry.py b/libs/langchain_v1/langchain/agents/middleware/model_retry.py
index 81a54a2ebb..2b4a39e725 100644
--- a/libs/langchain_v1/langchain/agents/middleware/model_retry.py
+++ b/libs/langchain_v1/langchain/agents/middleware/model_retry.py
@@ -7,6 +7,7 @@ import time
from typing import TYPE_CHECKING
from langchain_core.messages import AIMessage
+from langgraph.errors import GraphBubbleUp
from langchain.agents.middleware._retry import (
OnFailure,
@@ -232,6 +233,8 @@ class ModelRetryMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, Resp
for attempt in range(self.max_retries + 1):
try:
return handler(request)
+ except GraphBubbleUp:
+ raise
except Exception as exc:
attempts_made = attempt + 1 # attempt is 0-indexed
@@ -282,6 +285,8 @@ class ModelRetryMiddleware(AgentMiddleware[AgentState[ResponseT], ContextT, Resp
for attempt in range(self.max_retries + 1):
try:
return await handler(request)
+ except GraphBubbleUp:
+ raise
except Exception as exc:
attempts_made = attempt + 1 # attempt is 0-indexed
diff --git a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_model_fallback.py b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_model_fallback.py
index 39eca31ae9..5de942978b 100644
--- a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_model_fallback.py
+++ b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_model_fallback.py
@@ -11,6 +11,7 @@ from langchain_core.language_models.fake_chat_models import GenericFakeChatModel
from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage
from langchain_core.outputs import ChatGeneration, ChatResult
from langchain_core.tools import BaseTool, tool
+from langgraph.errors import GraphInterrupt
from typing_extensions import override
from langchain.agents.factory import create_agent
@@ -216,6 +217,48 @@ def _assert_request_is_sanitized(request: ModelRequest) -> None:
assert dict_tool_extras == {"defer_loading": True}
[email protected]("interrupt_stage", ["primary", "fallback"])
+def test_model_fallback_reraises_graph_bubble_up(interrupt_stage: str) -> None:
+ """Graph control-flow signals propagate from primary and fallback calls."""
+ fallback_model = GenericFakeChatModel(messages=iter([]))
+ middleware = ModelFallbackMiddleware(fallback_model)
+ calls = 0
+
+ def handler(_request: ModelRequest) -> ModelResponse:
+ nonlocal calls
+ calls += 1
+ if interrupt_stage == "fallback" and calls == 1:
+ msg = "try fallback"
+ raise ValueError(msg)
+ raise GraphInterrupt
+
+ with pytest.raises(GraphInterrupt):
+ middleware.wrap_model_call(_make_request(), handler)
+
+ assert calls == (2 if interrupt_stage == "fallback" else 1)
+
+
[email protected]("interrupt_stage", ["primary", "fallback"])
+async def test_model_fallback_async_reraises_graph_bubble_up(interrupt_stage: str) -> None:
+ """Async graph signals propagate from primary and fallback calls."""
+ fallback_model = GenericFakeChatModel(messages=iter([]))
+ middleware = ModelFallbackMiddleware(fallback_model)
+ calls = 0
+
+ async def handler(_request: ModelRequest) -> ModelResponse:
+ nonlocal calls
+ calls += 1
+ if interrupt_stage == "fallback" and calls == 1:
+ msg = "try fallback"
+ raise ValueError(msg)
+ raise GraphInterrupt
+
+ with pytest.raises(GraphInterrupt):
+ await middleware.awrap_model_call(_make_request(), handler)
+
+ assert calls == (2 if interrupt_stage == "fallback" else 1)
+
+
def test_fallback_sanitizes_cache_markers_sync() -> None:
"""Fallback attempts should strip Anthropic cache markers in sync path."""
primary_model = GenericFakeChatModel(messages=iter([AIMessage(content="primary response")]))
diff --git a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_model_retry.py b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_model_retry.py
index 85091b6685..3b6e2fd976 100644
--- a/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_model_retry.py
+++ b/libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_model_retry.py
@@ -9,6 +9,7 @@ from langchain_core.callbacks import CallbackManagerForLLMRun
from langchain_core.messages import AIMessage, BaseMessage, HumanMessage
from langchain_core.outputs import ChatGeneration, ChatResult
from langgraph.checkpoint.memory import InMemorySaver
+from langgraph.errors import GraphInterrupt
from pydantic import Field
from typing_extensions import override
@@ -402,6 +403,41 @@ def test_model_retry_custom_exception_filter() -> None:
assert "2 attempts" in ai_messages[-1].content
+def test_model_retry_reraises_graph_bubble_up() -> None:
+ """Graph control-flow signals propagate without retrying."""
+ retry = ModelRetryMiddleware(max_retries=3, initial_delay=0, jitter=False)
+ calls = 0
+ request = ModelRequest(model=FakeToolCallingModel(), messages=[])
+
+ def handler(_request: ModelRequest) -> ModelResponse:
+ nonlocal calls
+ calls += 1
+ raise GraphInterrupt
+
+ with pytest.raises(GraphInterrupt):
+ retry.wrap_model_call(request, handler)
+
+ assert calls == 1
+
+
+async def test_model_retry_async_reraises_graph_bubble_up() -> None:
+ """Async graph control-flow signals propagate without retrying."""
+ retry = ModelRetryMiddleware(max_retries=3, initial_delay=0, jitter=False)
+ calls = 0
+ request = ModelRequest(model=FakeToolCallingModel(), messages=[])
+
+ async def handler(_request: ModelRequest) -> ModelResponse:
+ nonlocal calls
+ calls += 1
+ raise GraphInterrupt
+
+ with pytest.raises(GraphInterrupt):
+ await retry.awrap_model_call(request, handler)
+
+ assert calls == 1
+
+
def test_model_retry_backoff_timing() -> None:
"""Test ModelRetryMiddleware applies correct backoff delays."""
model = TemporaryFailureModel(fail_count=3)