Skip to content

Commit f2410f7

Browse files
ccurmeccurme
authored
revert: Support for SystemMessage in create_agent (#33889)
Reverts #33640 Introduces lint errors into langchain-anthropic Should incorporate into 1.1 instead of patch release.
1 parent 91560b6 commit f2410f7

File tree

6 files changed

+31
-313
lines changed

6 files changed

+31
-313
lines changed

‎libs/langchain_v1/langchain/agents/factory.py‎

Lines changed: 9 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -516,7 +516,7 @@ def create_agent( # noqa: PLR0915
516516
model: str | BaseChatModel,
517517
tools: Sequence[BaseTool | Callable | dict[str, Any]] | None = None,
518518
*,
519-
system_prompt: str | SystemMessage | None = None,
519+
system_prompt: str | None = None,
520520
middleware: Sequence[AgentMiddleware[StateT_co, ContextT]] = (),
521521
response_format: ResponseFormat[ResponseT] | type[ResponseT] | None = None,
522522
state_schema: type[AgentState[ResponseT]] | None = None,
@@ -548,9 +548,11 @@ def create_agent( # noqa: PLR0915
548548
549549
If `None` or an empty list, the agent will consist of a model node without a
550550
tool calling loop.
551-
system_prompt: An optional system prompt for the LLM or
552-
can already be a [`SystemMessage`][langchain.messages.SystemMessage] object.
551+
system_prompt: An optional system prompt for the LLM.
553552
553+
Prompts are converted to a
554+
[`SystemMessage`][langchain.messages.SystemMessage] and added to the
555+
beginning of the message list.
554556
middleware: A sequence of middleware instances to apply to the agent.
555557
556558
Middleware can intercept and modify agent behavior at various stages. See
@@ -1038,10 +1040,8 @@ def _execute_model_sync(request: ModelRequest) -> ModelResponse:
10381040
# Get the bound model (with auto-detection if needed)
10391041
model_, effective_response_format = _get_bound_model(request)
10401042
messages = request.messages
1041-
if request.system_prompt and not isinstance(request.system_prompt, SystemMessage):
1042-
messages = [SystemMessage(content=request.system_prompt), *messages]
1043-
elif request.system_prompt and isinstance(request.system_prompt, SystemMessage):
1044-
messages = [request.system_prompt, *messages]
1043+
if request.system_prompt:
1044+
messages = [SystemMessage(request.system_prompt), *messages]
10451045

10461046
output = model_.invoke(messages)
10471047

@@ -1093,10 +1093,8 @@ async def _execute_model_async(request: ModelRequest) -> ModelResponse:
10931093
# Get the bound model (with auto-detection if needed)
10941094
model_, effective_response_format = _get_bound_model(request)
10951095
messages = request.messages
1096-
if request.system_prompt and not isinstance(request.system_prompt, SystemMessage):
1097-
messages = [SystemMessage(content=request.system_prompt), *messages]
1098-
elif request.system_prompt and isinstance(request.system_prompt, SystemMessage):
1099-
messages = [request.system_prompt, *messages]
1096+
if request.system_prompt:
1097+
messages = [SystemMessage(request.system_prompt), *messages]
11001098

11011099
output = await model_.ainvoke(messages)
11021100

‎libs/langchain_v1/langchain/agents/middleware/context_editing.py‎

Lines changed: 6 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -225,11 +225,9 @@ def wrap_model_call(
225225
def count_tokens(messages: Sequence[BaseMessage]) -> int:
226226
return count_tokens_approximately(messages)
227227
else:
228-
system_msg = []
229-
if request.system_prompt and not isinstance(request.system_prompt, SystemMessage):
230-
system_msg = [SystemMessage(content=request.system_prompt)]
231-
elif request.system_prompt and isinstance(request.system_prompt, SystemMessage):
232-
system_msg = [request.system_prompt]
228+
system_msg = (
229+
[SystemMessage(content=request.system_prompt)] if request.system_prompt else []
230+
)
233231

234232
def count_tokens(messages: Sequence[BaseMessage]) -> int:
235233
return request.model.get_num_tokens_from_messages(
@@ -255,12 +253,9 @@ async def awrap_model_call(
255253
def count_tokens(messages: Sequence[BaseMessage]) -> int:
256254
return count_tokens_approximately(messages)
257255
else:
258-
system_msg = []
259-
260-
if request.system_prompt and not isinstance(request.system_prompt, SystemMessage):
261-
system_msg = [SystemMessage(content=request.system_prompt)]
262-
elif request.system_prompt and isinstance(request.system_prompt, SystemMessage):
263-
system_msg = [request.system_prompt]
256+
system_msg = (
257+
[SystemMessage(content=request.system_prompt)] if request.system_prompt else []
258+
)
264259

265260
def count_tokens(messages: Sequence[BaseMessage]) -> int:
266261
return request.model.get_num_tokens_from_messages(

‎libs/langchain_v1/langchain/agents/middleware/todo.py‎

Lines changed: 11 additions & 33 deletions
Original file line numberDiff line numberDiff line change
@@ -8,7 +8,7 @@
88
if TYPE_CHECKING:
99
from collections.abc import Awaitable, Callable
1010

11-
from langchain_core.messages import SystemMessage, ToolMessage
11+
from langchain_core.messages import ToolMessage
1212
from langchain_core.tools import tool
1313
from langgraph.types import Command
1414
from typing_extensions import NotRequired, TypedDict
@@ -199,22 +199,11 @@ def wrap_model_call(
199199
handler: Callable[[ModelRequest], ModelResponse],
200200
) -> ModelCallResult:
201201
"""Update the system prompt to include the todo system prompt."""
202-
if request.system_prompt is None:
203-
request.system_prompt = self.system_prompt
204-
elif isinstance(request.system_prompt, str):
205-
request.system_prompt = request.system_prompt + "\n\n" + self.system_prompt
206-
elif isinstance(request.system_prompt, SystemMessage) and isinstance(
207-
request.system_prompt.content, str
208-
):
209-
request.system_prompt = SystemMessage(
210-
content=request.system_prompt.content + self.system_prompt
211-
)
212-
elif isinstance(request.system_prompt, SystemMessage) and isinstance(
213-
request.system_prompt.content, list
214-
):
215-
request.system_prompt = SystemMessage(
216-
content=[*request.system_prompt.content, self.system_prompt]
217-
)
202+
request.system_prompt = (
203+
request.system_prompt + "\n\n" + self.system_prompt
204+
if request.system_prompt
205+
else self.system_prompt
206+
)
218207
return handler(request)
219208

220209
async def awrap_model_call(
@@ -223,20 +212,9 @@ async def awrap_model_call(
223212
handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
224213
) -> ModelCallResult:
225214
"""Update the system prompt to include the todo system prompt (async version)."""
226-
if request.system_prompt is None:
227-
request.system_prompt = self.system_prompt
228-
elif isinstance(request.system_prompt, str):
229-
request.system_prompt = request.system_prompt + "\n\n" + self.system_prompt
230-
elif isinstance(request.system_prompt, SystemMessage) and isinstance(
231-
request.system_prompt.content, str
232-
):
233-
request.system_prompt = SystemMessage(
234-
content=request.system_prompt.content + self.system_prompt
235-
)
236-
elif isinstance(request.system_prompt, SystemMessage) and isinstance(
237-
request.system_prompt.content, list
238-
):
239-
request.system_prompt = SystemMessage(
240-
content=[*request.system_prompt.content, self.system_prompt]
241-
)
215+
request.system_prompt = (
216+
request.system_prompt + "\n\n" + self.system_prompt
217+
if request.system_prompt
218+
else self.system_prompt
219+
)
242220
return await handler(request)

‎libs/langchain_v1/langchain/agents/middleware/types.py‎

Lines changed: 4 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -26,7 +26,6 @@
2626
AIMessage,
2727
AnyMessage,
2828
BaseMessage,
29-
SystemMessage,
3029
ToolMessage,
3130
)
3231
from langgraph.channels.ephemeral_value import EphemeralValue
@@ -86,7 +85,7 @@ class ModelRequest:
8685
"""Model request information for the agent."""
8786

8887
model: BaseChatModel
89-
system_prompt: str | SystemMessage | None
88+
system_prompt: str | None
9089
messages: list[AnyMessage] # excluding system prompt
9190
tool_choice: Any | None
9291
tools: list[BaseTool | dict]
@@ -104,7 +103,7 @@ def override(self, **overrides: Unpack[_ModelRequestOverrides]) -> ModelRequest:
104103
Args:
105104
**overrides: Keyword arguments for attributes to override. Supported keys:
106105
- model: BaseChatModel instance
107-
- system_prompt: Optional system prompt string or SystemMessage object
106+
- system_prompt: Optional system prompt string
108107
- messages: List of messages
109108
- tool_choice: Tool choice configuration
110109
- tools: List of available tools
@@ -1257,7 +1256,7 @@ def wrapped(
12571256
request: ModelRequest,
12581257
handler: Callable[[ModelRequest], ModelResponse],
12591258
) -> ModelCallResult:
1260-
prompt = cast("str | SystemMessage", func(request))
1259+
prompt = cast("str", func(request))
12611260
request.system_prompt = prompt
12621261
return handler(request)
12631262

@@ -1267,7 +1266,7 @@ async def async_wrapped_from_sync(
12671266
handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
12681267
) -> ModelCallResult:
12691268
# Delegate to sync function
1270-
prompt = cast("str | SystemMessage", func(request))
1269+
prompt = cast("str", func(request))
12711270
request.system_prompt = prompt
12721271
return await handler(request)
12731272

‎libs/langchain_v1/tests/unit_tests/agents/test_context_editing_middleware.py‎

Lines changed: 0 additions & 124 deletions
Original file line numberDiff line numberDiff line change
@@ -13,7 +13,6 @@
1313
from langchain_core.messages import (
1414
AIMessage,
1515
MessageLikeRepresentation,
16-
SystemMessage,
1716
ToolMessage,
1817
)
1918
from langgraph.runtime import Runtime
@@ -400,126 +399,3 @@ async def mock_handler(req: ModelRequest) -> AIMessage:
400399

401400
assert isinstance(calc_tool, ToolMessage)
402401
assert calc_tool.content == "[cleared]"
403-
404-
405-
# ==============================================================================
406-
# SystemMessage Tests
407-
# ==============================================================================
408-
409-
410-
def test_handles_system_message_prompt() -> None:
411-
"""Test that middleware handles SystemMessage as system_prompt correctly."""
412-
tool_call_id = "call-1"
413-
ai_message = AIMessage(
414-
content="",
415-
tool_calls=[{"id": tool_call_id, "name": "search", "args": {}}],
416-
)
417-
tool_message = ToolMessage(content="12345", tool_call_id=tool_call_id)
418-
419-
system_prompt = SystemMessage(content="You are a helpful assistant.")
420-
state, request = _make_state_and_request([ai_message, tool_message], system_prompt=None)
421-
# Manually set SystemMessage as system_prompt
422-
request.system_prompt = system_prompt
423-
424-
middleware = ContextEditingMiddleware(
425-
edits=[ClearToolUsesEdit(trigger=50)],
426-
token_count_method="model",
427-
)
428-
429-
def mock_handler(req: ModelRequest) -> AIMessage:
430-
return AIMessage(content="mock response")
431-
432-
# Call wrap_model_call - should not fail with SystemMessage
433-
middleware.wrap_model_call(request, mock_handler)
434-
435-
# Request should have processed without errors
436-
assert request.system_prompt == system_prompt
437-
assert isinstance(request.system_prompt, SystemMessage)
438-
439-
440-
def test_does_not_double_wrap_system_message() -> None:
441-
"""Test that middleware doesn't wrap SystemMessage in another SystemMessage."""
442-
tool_call_id = "call-1"
443-
ai_message = AIMessage(
444-
content="",
445-
tool_calls=[{"id": tool_call_id, "name": "search", "args": {}}],
446-
)
447-
tool_message = ToolMessage(content="x" * 100, tool_call_id=tool_call_id)
448-
449-
system_prompt = SystemMessage(content="Original system prompt")
450-
state, request = _make_state_and_request([ai_message, tool_message], system_prompt=None)
451-
request.system_prompt = system_prompt
452-
453-
middleware = ContextEditingMiddleware(
454-
edits=[ClearToolUsesEdit(trigger=50)],
455-
token_count_method="model",
456-
)
457-
458-
def mock_handler(req: ModelRequest) -> AIMessage:
459-
return AIMessage(content="mock response")
460-
461-
middleware.wrap_model_call(request, mock_handler)
462-
463-
# System prompt should still be the same SystemMessage, not wrapped
464-
assert request.system_prompt == system_prompt
465-
assert isinstance(request.system_prompt, SystemMessage)
466-
assert request.system_prompt.content == "Original system prompt"
467-
468-
469-
async def test_handles_system_message_prompt_async() -> None:
470-
"""Test async version - middleware handles SystemMessage as system_prompt correctly."""
471-
tool_call_id = "call-1"
472-
ai_message = AIMessage(
473-
content="",
474-
tool_calls=[{"id": tool_call_id, "name": "search", "args": {}}],
475-
)
476-
tool_message = ToolMessage(content="12345", tool_call_id=tool_call_id)
477-
478-
system_prompt = SystemMessage(content="You are a helpful assistant.")
479-
state, request = _make_state_and_request([ai_message, tool_message], system_prompt=None)
480-
# Manually set SystemMessage as system_prompt
481-
request.system_prompt = system_prompt
482-
483-
middleware = ContextEditingMiddleware(
484-
edits=[ClearToolUsesEdit(trigger=50)],
485-
token_count_method="model",
486-
)
487-
488-
async def mock_handler(req: ModelRequest) -> AIMessage:
489-
return AIMessage(content="mock response")
490-
491-
# Call awrap_model_call - should not fail with SystemMessage
492-
await middleware.awrap_model_call(request, mock_handler)
493-
494-
# Request should have processed without errors
495-
assert request.system_prompt == system_prompt
496-
assert isinstance(request.system_prompt, SystemMessage)
497-
498-
499-
async def test_does_not_double_wrap_system_message_async() -> None:
500-
"""Test async version - middleware doesn't wrap SystemMessage in another SystemMessage."""
501-
tool_call_id = "call-1"
502-
ai_message = AIMessage(
503-
content="",
504-
tool_calls=[{"id": tool_call_id, "name": "search", "args": {}}],
505-
)
506-
tool_message = ToolMessage(content="x" * 100, tool_call_id=tool_call_id)
507-
508-
system_prompt = SystemMessage(content="Original system prompt")
509-
state, request = _make_state_and_request([ai_message, tool_message], system_prompt=None)
510-
request.system_prompt = system_prompt
511-
512-
middleware = ContextEditingMiddleware(
513-
edits=[ClearToolUsesEdit(trigger=50)],
514-
token_count_method="model",
515-
)
516-
517-
async def mock_handler(req: ModelRequest) -> AIMessage:
518-
return AIMessage(content="mock response")
519-
520-
await middleware.awrap_model_call(request, mock_handler)
521-
522-
# System prompt should still be the same SystemMessage, not wrapped
523-
assert request.system_prompt == system_prompt
524-
assert isinstance(request.system_prompt, SystemMessage)
525-
assert request.system_prompt.content == "Original system prompt"

0 commit comments

Comments
 (0)