feat(middleware): scope model fallback to provider/network errors only

This commit is contained in:
CREDO23 2026-05-05 18:04:47 +02:00
parent f695298d30
commit 1745d7dccf
6 changed files with 275 additions and 17 deletions

View file

@ -14,7 +14,6 @@ from deepagents.middleware.subagents import GENERAL_PURPOSE_SUBAGENT
from langchain.agents.middleware import ( from langchain.agents.middleware import (
LLMToolSelectorMiddleware, LLMToolSelectorMiddleware,
ModelCallLimitMiddleware, ModelCallLimitMiddleware,
ModelFallbackMiddleware,
TodoListMiddleware, TodoListMiddleware,
ToolCallLimitMiddleware, ToolCallLimitMiddleware,
) )
@ -56,6 +55,9 @@ from app.agents.new_chat.middleware import (
create_surfsense_compaction_middleware, create_surfsense_compaction_middleware,
default_skills_sources, default_skills_sources,
) )
from app.agents.new_chat.middleware.scoped_model_fallback import (
ScopedModelFallbackMiddleware,
)
from app.agents.new_chat.permissions import Rule, Ruleset from app.agents.new_chat.permissions import Rule, Ruleset
from app.agents.new_chat.plugin_loader import ( from app.agents.new_chat.plugin_loader import (
PluginContext, PluginContext,
@ -217,15 +219,15 @@ def build_main_agent_deepagent_middleware(
if flags.enable_retry_after and not flags.disable_new_agent_stack if flags.enable_retry_after and not flags.disable_new_agent_stack
else None else None
) )
fallback_mw: ModelFallbackMiddleware | None = None fallback_mw: ScopedModelFallbackMiddleware | None = None
if flags.enable_model_fallback and not flags.disable_new_agent_stack: if flags.enable_model_fallback and not flags.disable_new_agent_stack:
try: try:
fallback_mw = ModelFallbackMiddleware( fallback_mw = ScopedModelFallbackMiddleware(
"openai:gpt-4o-mini", "openai:gpt-4o-mini",
"anthropic:claude-3-5-haiku-20241022", "anthropic:claude-3-5-haiku-20241022",
) )
except Exception: except Exception:
logging.warning("ModelFallbackMiddleware init failed; skipping.") logging.warning("ScopedModelFallbackMiddleware init failed; skipping.")
fallback_mw = None fallback_mw = None
registry_subagents: list[SubAgent] = [] registry_subagents: list[SubAgent] = []

View file

@ -31,7 +31,6 @@ from langchain.agents import create_agent
from langchain.agents.middleware import ( from langchain.agents.middleware import (
LLMToolSelectorMiddleware, LLMToolSelectorMiddleware,
ModelCallLimitMiddleware, ModelCallLimitMiddleware,
ModelFallbackMiddleware,
TodoListMiddleware, TodoListMiddleware,
ToolCallLimitMiddleware, ToolCallLimitMiddleware,
) )
@ -77,6 +76,9 @@ from app.agents.new_chat.middleware import (
create_surfsense_compaction_middleware, create_surfsense_compaction_middleware,
default_skills_sources, default_skills_sources,
) )
from app.agents.new_chat.middleware.scoped_model_fallback import (
ScopedModelFallbackMiddleware,
)
from app.agents.new_chat.permissions import Rule, Ruleset from app.agents.new_chat.permissions import Rule, Ruleset
from app.agents.new_chat.plugin_loader import ( from app.agents.new_chat.plugin_loader import (
PluginContext, PluginContext,
@ -792,15 +794,15 @@ def _build_compiled_agent_blocking(
# Fallback chain — primary is the agent's own model; we add cheap # Fallback chain — primary is the agent's own model; we add cheap
# alternatives. Off by default; only the first call site that # alternatives. Off by default; only the first call site that
# configures the chain via env should enable it. # configures the chain via env should enable it.
fallback_mw: ModelFallbackMiddleware | None = None fallback_mw: ScopedModelFallbackMiddleware | None = None
if flags.enable_model_fallback and not flags.disable_new_agent_stack: if flags.enable_model_fallback and not flags.disable_new_agent_stack:
try: try:
fallback_mw = ModelFallbackMiddleware( fallback_mw = ScopedModelFallbackMiddleware(
"openai:gpt-4o-mini", "openai:gpt-4o-mini",
"anthropic:claude-3-5-haiku-20241022", "anthropic:claude-3-5-haiku-20241022",
) )
except Exception: except Exception:
logging.warning("ModelFallbackMiddleware init failed; skipping.") logging.warning("ScopedModelFallbackMiddleware init failed; skipping.")
fallback_mw = None fallback_mw = None
model_call_limit_mw = ( model_call_limit_mw = (
ModelCallLimitMiddleware( ModelCallLimitMiddleware(

View file

@ -0,0 +1,106 @@
"""Fallback only on provider/network errors; let programming bugs raise.
Upstream :class:`langchain.agents.middleware.ModelFallbackMiddleware` catches
every ``Exception``. With a non-provider bug (``KeyError``, ``TypeError``,
``AttributeError`` from middleware/state), every fallback model in the chain
hits the same bug burning latency and tokens before the real cause finally
surfaces. Scoping the catch to provider-style exception types lets bugs fail
fast with clean tracebacks.
Class-name matching (instead of ``isinstance`` against imported provider
types) keeps the dependency surface flat: openai, anthropic, google,
mistral, etc. all ship their own ``RateLimitError`` and we don't want to
import them all.
"""
from __future__ import annotations
from typing import TYPE_CHECKING, Any
from langchain.agents.middleware import ModelFallbackMiddleware
if TYPE_CHECKING:
from collections.abc import Awaitable, Callable
from langchain.agents.middleware.types import ModelRequest, ModelResponse
from langchain_core.messages import AIMessage
_FALLBACK_ELIGIBLE_NAMES: frozenset[str] = frozenset(
{
# Rate / quota
"RateLimitError",
# Server-side
"APIStatusError",
"InternalServerError",
"ServiceUnavailableError",
"BadGatewayError",
"GatewayTimeoutError",
# Network
"APIConnectionError",
"APITimeoutError",
"ConnectError",
"ConnectTimeout",
"ReadTimeout",
"RemoteProtocolError",
"TimeoutError",
"TimeoutException",
}
)
def _is_fallback_eligible(exc: BaseException) -> bool:
"""Eligible if the exception or any base in its MRO matches by class name."""
return any(cls.__name__ in _FALLBACK_ELIGIBLE_NAMES for cls in type(exc).__mro__)
class ScopedModelFallbackMiddleware(ModelFallbackMiddleware):
"""``ModelFallbackMiddleware`` that re-raises non-provider exceptions."""
def wrap_model_call( # type: ignore[override]
self,
request: ModelRequest[Any],
handler: Callable[[ModelRequest[Any]], ModelResponse[Any]],
) -> ModelResponse[Any] | AIMessage:
last_exception: Exception
try:
return handler(request)
except Exception as e:
if not _is_fallback_eligible(e):
raise
last_exception = e
for fallback_model in self.models:
try:
return handler(request.override(model=fallback_model))
except Exception as e:
if not _is_fallback_eligible(e):
raise
last_exception = e
continue
raise last_exception
async def awrap_model_call( # type: ignore[override]
self,
request: ModelRequest[Any],
handler: Callable[[ModelRequest[Any]], Awaitable[ModelResponse[Any]]],
) -> ModelResponse[Any] | AIMessage:
last_exception: Exception
try:
return await handler(request)
except Exception as e:
if not _is_fallback_eligible(e):
raise
last_exception = e
for fallback_model in self.models:
try:
return await handler(request.override(model=fallback_model))
except Exception as e:
if not _is_fallback_eligible(e):
raise
last_exception = e
continue
raise last_exception

View file

@ -31,12 +31,12 @@ from app.agents.multi_agent_chat.subagents.shared.subagent_builder import (
) )
class _AlwaysFailingChatModel(BaseChatModel): class RateLimitError(Exception):
"""Mimics a provider hard-failing on every call (rate limit / empty stream). """Provider-style 429; matches the scoped-fallback eligibility allowlist by name."""
``ModelFallbackMiddleware`` triggers on any ``Exception``, so the exact
error type doesn't matter for the contract under test. class _AlwaysFailingChatModel(BaseChatModel):
""" """Mimics a provider hard-failing on every call (rate limit / empty stream)."""
@property @property
def _llm_type(self) -> str: def _llm_type(self) -> str:
@ -50,7 +50,7 @@ class _AlwaysFailingChatModel(BaseChatModel):
**kwargs: Any, **kwargs: Any,
) -> ChatResult: ) -> ChatResult:
msg = "primary llm exploded" msg = "primary llm exploded"
raise RuntimeError(msg) raise RateLimitError(msg)
async def _agenerate( async def _agenerate(
self, self,
@ -60,17 +60,17 @@ class _AlwaysFailingChatModel(BaseChatModel):
**kwargs: Any, **kwargs: Any,
) -> ChatResult: ) -> ChatResult:
msg = "primary llm exploded" msg = "primary llm exploded"
raise RuntimeError(msg) raise RateLimitError(msg)
def _stream(self, *args: Any, **kwargs: Any) -> Iterator[ChatGeneration]: def _stream(self, *args: Any, **kwargs: Any) -> Iterator[ChatGeneration]:
msg = "primary llm exploded" msg = "primary llm exploded"
raise RuntimeError(msg) raise RateLimitError(msg)
async def _astream( async def _astream(
self, *args: Any, **kwargs: Any self, *args: Any, **kwargs: Any
) -> AsyncIterator[ChatGeneration]: ) -> AsyncIterator[ChatGeneration]:
msg = "primary llm exploded" msg = "primary llm exploded"
raise RuntimeError(msg) raise RateLimitError(msg)
yield # pragma: no cover - unreachable, satisfies async generator typing yield # pragma: no cover - unreachable, satisfies async generator typing

View file

@ -0,0 +1,148 @@
"""Exception-scope contract for ``ScopedModelFallbackMiddleware``.
Upstream ``ModelFallbackMiddleware`` catches every ``Exception`` and walks
the fallback chain. That means a programming bug (``KeyError`` from a
botched tool config, ``TypeError`` from middleware, ...) burns 1+N model
round-trips and ~Nx tokens before its real cause surfaces. The scoped
variant only falls back on provider/network exception types so bugs fail
fast, with clean tracebacks.
"""
from __future__ import annotations
from collections.abc import AsyncIterator, Iterator
from typing import Any
import pytest
from langchain_core.callbacks import (
AsyncCallbackManagerForLLMRun,
CallbackManagerForLLMRun,
)
from langchain_core.language_models.chat_models import BaseChatModel
from langchain_core.messages import AIMessage, BaseMessage
from langchain_core.outputs import ChatGeneration, ChatResult
class _RaisingChatModel(BaseChatModel):
"""LLM that raises a configurable exception on every invocation."""
exc_to_raise: Any
@property
def _llm_type(self) -> str:
return "raising-test-model"
def _generate(
self,
messages: list[BaseMessage],
stop: list[str] | None = None,
run_manager: CallbackManagerForLLMRun | None = None,
**kwargs: Any,
) -> ChatResult:
raise self.exc_to_raise
async def _agenerate(
self,
messages: list[BaseMessage],
stop: list[str] | None = None,
run_manager: AsyncCallbackManagerForLLMRun | None = None,
**kwargs: Any,
) -> ChatResult:
raise self.exc_to_raise
def _stream(self, *args: Any, **kwargs: Any) -> Iterator[ChatGeneration]:
raise self.exc_to_raise
async def _astream(
self, *args: Any, **kwargs: Any
) -> AsyncIterator[ChatGeneration]:
raise self.exc_to_raise
yield # pragma: no cover - unreachable
class _RecordingChatModel(BaseChatModel):
"""Returns a fixed message and counts how often it was called."""
response_text: str = "fallback-ok"
call_count: int = 0
@property
def _llm_type(self) -> str:
return "recording-test-model"
def _generate(
self,
messages: list[BaseMessage],
stop: list[str] | None = None,
run_manager: CallbackManagerForLLMRun | None = None,
**kwargs: Any,
) -> ChatResult:
self.call_count += 1
return ChatResult(
generations=[
ChatGeneration(message=AIMessage(content=self.response_text))
]
)
async def _agenerate(
self,
messages: list[BaseMessage],
stop: list[str] | None = None,
run_manager: AsyncCallbackManagerForLLMRun | None = None,
**kwargs: Any,
) -> ChatResult:
return self._generate(messages, stop, None, **kwargs)
# Locally defined provider-style error: importing openai/anthropic/etc.
# would couple the test to provider SDKs the contract intentionally avoids.
class RateLimitError(Exception):
"""Mimics ``openai.RateLimitError`` for name-based eligibility."""
def _build_agent(primary: BaseChatModel, fallback: BaseChatModel):
"""Compile a no-tools agent with the scoped fallback wired in."""
from langchain.agents import create_agent
from app.agents.new_chat.middleware.scoped_model_fallback import (
ScopedModelFallbackMiddleware,
)
return create_agent(
model=primary,
tools=[],
middleware=[ScopedModelFallbackMiddleware(fallback)],
system_prompt="be helpful",
)
@pytest.mark.asyncio
async def test_provider_errors_trigger_fallback():
"""Class names matching the provider allowlist drive the fallback chain."""
primary = _RaisingChatModel(exc_to_raise=RateLimitError("429 from provider"))
fallback = _RecordingChatModel(response_text="recovered")
agent = _build_agent(primary, fallback)
result = await agent.ainvoke({"messages": [("user", "hi")]})
final = result["messages"][-1]
assert isinstance(final, AIMessage)
assert final.content == "recovered"
assert fallback.call_count == 1
@pytest.mark.asyncio
async def test_programming_errors_propagate_without_invoking_fallback():
"""``KeyError`` from agent-side bugs must surface immediately, no fallback retry."""
primary = _RaisingChatModel(exc_to_raise=KeyError("missing_state_field"))
fallback = _RecordingChatModel(response_text="should-never-arrive")
agent = _build_agent(primary, fallback)
with pytest.raises(KeyError, match="missing_state_field"):
await agent.ainvoke({"messages": [("user", "hi")]})
assert fallback.call_count == 0, (
"fallback was invoked for a programming error; "
"scoping rule is broken"
)