Skip to content

Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.

fix: prevent shared LLM stop words mutation across agents

Greyson LaLonde184c228ae9f9
diff --git a/lib/crewai/src/crewai/agent/core.py b/lib/crewai/src/crewai/agent/core.py
index 10ebfd38c..561307680 100644
--- a/lib/crewai/src/crewai/agent/core.py
+++ b/lib/crewai/src/crewai/agent/core.py
@@ -1102,16 +1102,6 @@ class Agent(BaseAgent):
self.agent_executor.tools_handler = self.tools_handler
self.agent_executor.request_within_rpm_limit = rpm_limit_fn
- if isinstance(self.agent_executor.llm, BaseLLM):
- existing_stop = getattr(self.agent_executor.llm, "stop", [])
- self.agent_executor.llm.stop = list(
- set(
- existing_stop + stop_words
- if isinstance(existing_stop, list)
- else stop_words
- )
- )
-
def get_delegation_tools(self, agents: Sequence[BaseAgent]) -> list[BaseTool]:
agent_tools = AgentTools(agents=agents)
return agent_tools.tools()
diff --git a/lib/crewai/src/crewai/agents/crew_agent_executor.py b/lib/crewai/src/crewai/agents/crew_agent_executor.py
index 8bb27b395..f4d55fe80 100644
--- a/lib/crewai/src/crewai/agents/crew_agent_executor.py
+++ b/lib/crewai/src/crewai/agents/crew_agent_executor.py
@@ -49,6 +49,7 @@ from crewai.hooks.tool_hooks import (
)
from crewai.types.callback import SerializableCallable
from crewai.utilities.agent_utils import (
+ _llm_stop_words_applied,
aget_llm_response,
convert_tools_to_openai_schema,
enforce_rpm_limit,
@@ -141,15 +142,6 @@ class CrewAgentExecutor(BaseAgentExecutor):
self.before_llm_call_hooks.extend(get_before_llm_call_hooks())
if not self.after_llm_call_hooks:
self.after_llm_call_hooks.extend(get_after_llm_call_hooks())
- if self.llm and not isinstance(self.llm, str):
- existing_stop = getattr(self.llm, "stop", [])
- self.llm.stop = list(
- set(
- existing_stop + self.stop
- if isinstance(existing_stop, list)
- else self.stop
- )
- )
@property
def use_stop_words(self) -> bool:
@@ -210,21 +202,22 @@ class CrewAgentExecutor(BaseAgentExecutor):
self.ask_for_human_input = bool(inputs.get("ask_for_human_input", False))
- try:
- formatted_answer = self._invoke_loop()
- except AssertionError:
- if self.agent.verbose:
- PRINTER.print(
- content="Agent failed to reach a final answer. This is likely a bug - please report it.",
- color="red",
- )
- raise
- except Exception as e:
- handle_unknown_error(PRINTER, e, verbose=self.agent.verbose)
- raise
+ with _llm_stop_words_applied(self.llm, self):
+ try:
+ formatted_answer = self._invoke_loop()
+ except AssertionError:
+ if self.agent.verbose:
+ PRINTER.print(
+ content="Agent failed to reach a final answer. This is likely a bug - please report it.",
+ color="red",
+ )
+ raise
+ except Exception as e:
+ handle_unknown_error(PRINTER, e, verbose=self.agent.verbose)
+ raise
- if self.ask_for_human_input:
- formatted_answer = self._handle_human_feedback(formatted_answer)
+ if self.ask_for_human_input:
+ formatted_answer = self._handle_human_feedback(formatted_answer)
self._save_to_memory(formatted_answer)
return {"output": formatted_answer.output}
@@ -1082,21 +1075,22 @@ class CrewAgentExecutor(BaseAgentExecutor):
self.ask_for_human_input = bool(inputs.get("ask_for_human_input", False))
- try:
- formatted_answer = await self._ainvoke_loop()
- except AssertionError:
- if self.agent.verbose:
- PRINTER.print(
- content="Agent failed to reach a final answer. This is likely a bug - please report it.",
- color="red",
- )
- raise
- except Exception as e:
- handle_unknown_error(PRINTER, e, verbose=self.agent.verbose)
- raise
+ with _llm_stop_words_applied(self.llm, self):
+ try:
+ formatted_answer = await self._ainvoke_loop()
+ except AssertionError:
+ if self.agent.verbose:
+ PRINTER.print(
+ content="Agent failed to reach a final answer. This is likely a bug - please report it.",
+ color="red",
+ )
+ raise
+ except Exception as e:
+ handle_unknown_error(PRINTER, e, verbose=self.agent.verbose)
+ raise
- if self.ask_for_human_input:
- formatted_answer = await self._ahandle_human_feedback(formatted_answer)
+ if self.ask_for_human_input:
+ formatted_answer = await self._ahandle_human_feedback(formatted_answer)
self._save_to_memory(formatted_answer)
return {"output": formatted_answer.output}
diff --git a/lib/crewai/src/crewai/experimental/agent_executor.py b/lib/crewai/src/crewai/experimental/agent_executor.py
index b5e492c0b..de6f2ad62 100644
--- a/lib/crewai/src/crewai/experimental/agent_executor.py
+++ b/lib/crewai/src/crewai/experimental/agent_executor.py
@@ -71,6 +71,7 @@ from crewai.hooks.types import (
from crewai.tools.base_tool import BaseTool
from crewai.tools.structured_tool import CrewStructuredTool
from crewai.utilities.agent_utils import (
+ _llm_stop_words_applied,
check_native_tool_support,
enforce_rpm_limit,
extract_tool_call_info,
@@ -215,12 +216,6 @@ class AgentExecutor(Flow[AgentExecutorState], BaseAgentExecutor):
self.before_llm_call_hooks.extend(get_before_llm_call_hooks())
self.after_llm_call_hooks.extend(get_after_llm_call_hooks())
- if self.llm:
- existing_stop = getattr(self.llm, "stop", [])
- if not isinstance(existing_stop, list):
- existing_stop = []
- self.llm.stop = list(set(existing_stop + self.stop_words))
-
self._state = AgentExecutorState()
self.max_method_calls = self.max_iter * 10
@@ -2601,17 +2596,18 @@ class AgentExecutor(Flow[AgentExecutorState], BaseAgentExecutor):
inputs.get("ask_for_human_input", False)
)
- self.kickoff()
+ with _llm_stop_words_applied(self.llm, self):
+ self.kickoff()
- formatted_answer = self.state.current_answer
+ formatted_answer = self.state.current_answer
- if not isinstance(formatted_answer, AgentFinish):
- raise RuntimeError(
- "Agent execution ended without reaching a final answer."
- )
+ if not isinstance(formatted_answer, AgentFinish):
+ raise RuntimeError(
+ "Agent execution ended without reaching a final answer."
+ )
- if self.state.ask_for_human_input:
- formatted_answer = self._handle_human_feedback(formatted_answer)
+ if self.state.ask_for_human_input:
+ formatted_answer = self._handle_human_feedback(formatted_answer)
self._save_to_memory(formatted_answer)
@@ -2691,18 +2687,20 @@ class AgentExecutor(Flow[AgentExecutorState], BaseAgentExecutor):
inputs.get("ask_for_human_input", False)
)
- # Use async kickoff directly since we're already in an async context
- await self.kickoff_async()
+ with _llm_stop_words_applied(self.llm, self):
+ await self.kickoff_async()
- formatted_answer = self.state.current_answer
+ formatted_answer = self.state.current_answer
- if not isinstance(formatted_answer, AgentFinish):
- raise RuntimeError(
- "Agent execution ended without reaching a final answer."
- )
+ if not isinstance(formatted_answer, AgentFinish):
+ raise RuntimeError(
+ "Agent execution ended without reaching a final answer."
+ )
- if self.state.ask_for_human_input:
- formatted_answer = await self._ahandle_human_feedback(formatted_answer)
+ if self.state.ask_for_human_input:
+ formatted_answer = await self._ahandle_human_feedback(
+ formatted_answer
+ )
self._save_to_memory(formatted_answer)
diff --git a/lib/crewai/src/crewai/llm.py b/lib/crewai/src/crewai/llm.py
index 1c03eddc9..52e3b0b9f 100644
--- a/lib/crewai/src/crewai/llm.py
+++ b/lib/crewai/src/crewai/llm.py
@@ -688,7 +688,9 @@ class LLM(BaseLLM):
"temperature": self.temperature,
"top_p": self.top_p,
"n": self.n,
- "stop": (self.stop or None) if self.supports_stop_words() else None,
+ "stop": (self.stop_sequences or None)
+ if self.supports_stop_words()
+ else None,
"max_tokens": self.max_tokens or self.max_completion_tokens,
"presence_penalty": self.presence_penalty,
"frequency_penalty": self.frequency_penalty,
diff --git a/lib/crewai/src/crewai/llms/base_llm.py b/lib/crewai/src/crewai/llms/base_llm.py
index 4f45572ee..a9aa7dc01 100644
--- a/lib/crewai/src/crewai/llms/base_llm.py
+++ b/lib/crewai/src/crewai/llms/base_llm.py
@@ -72,6 +72,9 @@ _JSON_EXTRACTION_PATTERN: Final[re.Pattern[str]] = re.compile(r"\{.*}", re.DOTAL
_current_call_id: contextvars.ContextVar[str | None] = contextvars.ContextVar(
"_current_call_id", default=None
)
+_call_stop_override_var: contextvars.ContextVar[dict[int, list[str]] | None] = (
+ contextvars.ContextVar("_call_stop_override_var", default=None)
+)
@contextmanager
@@ -85,6 +88,31 @@ def llm_call_context() -> Generator[str, None, None]:
_current_call_id.reset(token)
+@contextmanager
+def call_stop_override(
+ llm: BaseLLM, stop: list[str] | None
+) -> Generator[None, None, None]:
+ """Override the stop list for ``llm`` within the current call scope.
+
+ Only ``llm``'s reads via :attr:`BaseLLM.stop_sequences` see ``stop``;
+ other LLM instances (e.g. an agent's ``function_calling_llm``) keep their
+ own ``stop`` field. Passing ``None`` clears any prior override for ``llm``
+ in the same scope. The instance-level ``stop`` field is never mutated,
+ so the override is safe under concurrent execution.
+ """
+ current = _call_stop_override_var.get()
+ new_overrides: dict[int, list[str]] = dict(current) if current else {}
+ if stop is None:
+ new_overrides.pop(id(llm), None)
+ else:
+ new_overrides[id(llm)] = stop
+ token = _call_stop_override_var.set(new_overrides)
+ try:
+ yield
+ finally:
+ _call_stop_override_var.reset(token)
+
+
def get_current_call_id() -> str:
"""Get current call_id from context"""
call_id = _current_call_id.get()
@@ -158,11 +186,18 @@ class BaseLLM(BaseModel, ABC):
@property
def stop_sequences(self) -> list[str]:
- """Alias for ``stop`` — kept for backward compatibility with provider APIs.
+ """Stop list active for the current call.
- Writes are handled by ``__setattr__``, which normalizes and redirects
- ``stop_sequences`` assignments to the ``stop`` field.
+ Returns the per-instance override set via :func:`call_stop_override`
+ when one is in effect for this LLM; otherwise the instance-level
+ ``stop`` field. Kept under this name for backward compatibility with
+ provider APIs that already read ``stop_sequences``.
"""
+ overrides = _call_stop_override_var.get()
+ if overrides is not None:
+ override = overrides.get(id(self))
+ if override is not None:
+ return override
return self.stop
_token_usage: dict[str, int] = PrivateAttr(
@@ -341,7 +376,7 @@ class BaseLLM(BaseModel, ABC):
Returns:
True if stop words are configured and can be applied
"""
- return bool(self.stop)
+ return bool(self.stop_sequences)
def _apply_stop_words(self, content: str) -> str:
"""Apply stop words to truncate response content.
@@ -363,14 +398,14 @@ class BaseLLM(BaseModel, ABC):
>>> llm._apply_stop_words(response)
"I need to search.\\n\\nAction: search"
"""
- if not self.stop or not content:
+ stops = self.stop_sequences
+ if not stops or not content:
return content
- # Find the earliest occurrence of any stop word
earliest_stop_pos = len(content)
found_stop_word = None
- for stop_word in self.stop:
+ for stop_word in stops:
stop_pos = content.find(stop_word)
if stop_pos != -1 and stop_pos < earliest_stop_pos:
earliest_stop_pos = stop_pos
diff --git a/lib/crewai/src/crewai/llms/providers/azure/completion.py b/lib/crewai/src/crewai/llms/providers/azure/completion.py
index 59fd5994e..dd18533e0 100644
--- a/lib/crewai/src/crewai/llms/providers/azure/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/azure/completion.py
@@ -679,8 +679,9 @@ class AzureCompletion(BaseLLM):
params["presence_penalty"] = self.presence_penalty
if self.max_tokens is not None:
params["max_tokens"] = self.max_tokens
- if self.stop and self.supports_stop_words():
- params["stop"] = self.stop
+ stops = self.stop_sequences
+ if stops and self.supports_stop_words():
+ params["stop"] = stops
# Handle tools/functions for Azure OpenAI models
if tools and self.is_openai_model:
diff --git a/lib/crewai/src/crewai/utilities/agent_utils.py b/lib/crewai/src/crewai/utilities/agent_utils.py
index 684fd9287..241bd6cd4 100644
--- a/lib/crewai/src/crewai/utilities/agent_utils.py
+++ b/lib/crewai/src/crewai/utilities/agent_utils.py
@@ -1,8 +1,9 @@
from __future__ import annotations
import asyncio
-from collections.abc import Callable, Sequence
+from collections.abc import Callable, Iterator, Sequence
import concurrent.futures
+import contextlib
import contextvars
from dataclasses import dataclass, field
from datetime import datetime
@@ -22,7 +23,7 @@ from crewai.agents.parser import (
parse,
)
from crewai.cli.config import Settings
-from crewai.llms.base_llm import BaseLLM
+from crewai.llms.base_llm import BaseLLM, call_stop_override
from crewai.tools import BaseTool as CrewAITool
from crewai.tools.base_tool import BaseTool
from crewai.tools.structured_tool import CrewStructuredTool
@@ -238,6 +239,38 @@ def extract_task_section(text: str) -> str:
return text
+def _executor_stop_words(
+ executor_context: CrewAgentExecutor | AgentExecutor | LiteAgent | None,
+) -> list[str]:
+ """Return the executor's stop words, regardless of which field name it uses."""
+ if executor_context is None:
+ return []
+ stops = getattr(executor_context, "stop", None)
+ if stops is None:
+ stops = getattr(executor_context, "stop_words", None)
+ return list(stops) if stops else []
+
+
+def _llm_stop_words_applied(
+ llm: LLM | BaseLLM,
+ executor_context: CrewAgentExecutor | AgentExecutor | LiteAgent | None,
+) -> Iterator[None]:
+ """Apply the executor's stop words to the LLM for the duration of one call.
+
+ Uses :func:`crewai.llms.base_llm.call_stop_override` so the LLM's stop
+ field is never mutated. Safe under concurrent execution: the override is
+ propagated via a :class:`contextvars.ContextVar` and is scoped to this
+ call's task / thread context.
+ """
+ extra = _executor_stop_words(executor_context)
+ if not extra or not isinstance(llm, BaseLLM) or set(extra).issubset(llm.stop):
+ yield
+ return
+ with call_stop_override(llm, list(set(llm.stop + extra))):
+ yield
+
+
def has_reached_max_iterations(iterations: int, max_iterations: int) -> bool:
"""Check if the maximum number of iterations has been reached.
@@ -459,18 +492,15 @@ def get_llm_response(
"""
messages = _prepare_llm_call(executor_context, messages, printer, verbose=verbose)
- try:
- answer = llm.call(
- messages,
- tools=tools,
- callbacks=callbacks,
- available_functions=available_functions,
- from_task=from_task,
- from_agent=from_agent,
- response_model=response_model,
- )
- except Exception as e:
- raise e
+ answer = llm.call(
+ messages,
+ tools=tools,
+ callbacks=callbacks,
+ available_functions=available_functions,
+ from_task=from_task,
+ from_agent=from_agent,
+ response_model=response_model,
+ )
return _validate_and_finalize_llm_response(
answer, executor_context, printer, verbose=verbose
@@ -515,18 +545,15 @@ async def aget_llm_response(
"""
messages = _prepare_llm_call(executor_context, messages, printer, verbose=verbose)
- try:
- answer = await llm.acall(
- messages,
- tools=tools,
- callbacks=callbacks,
- available_functions=available_functions,
- from_task=from_task,
- from_agent=from_agent,
- response_model=response_model,
- )
- except Exception as e:
- raise e
+ answer = await llm.acall(
+ messages,
+ tools=tools,
+ callbacks=callbacks,
+ available_functions=available_functions,
+ from_task=from_task,
+ from_agent=from_agent,
+ response_model=response_model,
+ )
return _validate_and_finalize_llm_response(
answer, executor_context, printer, verbose=verbose
diff --git a/lib/crewai/tests/agents/test_agent.py b/lib/crewai/tests/agents/test_agent.py
index f6101a9e0..037f70af4 100644
--- a/lib/crewai/tests/agents/test_agent.py
+++ b/lib/crewai/tests/agents/test_agent.py
@@ -2452,3 +2452,167 @@ def test_agent_mcps_accepts_legacy_prefix_with_tool():
mcps=["crewai-amp:notion#get_page"],
)
assert agent.mcps == ["crewai-amp:notion#get_page"]
+
+
+class TestSharedLLMStopWords:
+ """Regression tests for shared LLM stop words mutation (issue #5141).
+
+ Stop words from one executor must not leak into the shared LLM permanently
+ or pollute other agents sharing that LLM.
+ """
+
+ @staticmethod
+ def _make_executor(llm: LLM, stop_words: list[str]) -> CrewAgentExecutor:
+ """Build a CrewAgentExecutor with minimal deps."""
+ from crewai.agents.tools_handler import ToolsHandler
+
+ agent = Agent(role="r", goal="g", backstory="b")
+ task = Task(description="d", expected_output="o", agent=agent)
+ return CrewAgentExecutor(
+ agent=agent,
+ task=task,
+ llm=llm,
+ crew=None,
+ prompt={"prompt": "p {input} {tool_names} {tools}"},
+ max_iter=5,
+ tools=[],
+ tools_names="",
+ stop_words=stop_words,
+ tools_description="",
+ tools_handler=ToolsHandler(),
+ )
+
+ def test_executor_init_does_not_mutate_shared_llm(self) -> None:
+ """Constructing executors must not touch the shared LLM's stop list."""
+ shared = LLM(model="gpt-4", stop=["Original:"])
+ original = list(shared.stop)
+
+ a = self._make_executor(shared, stop_words=["StopA:"])
+ b = self._make_executor(shared, stop_words=["StopB:"])
+
+ assert shared.stop == original
+ assert a.llm is shared
+ assert b.llm is shared
+
+ def test_effective_stop_reflects_override_inside_context(self) -> None:
+ """Inside the helper, the effective stop list includes the executor's words."""
+ from crewai.utilities.agent_utils import _llm_stop_words_applied
+
+ shared = LLM(model="gpt-4", stop=["Original:"])
+ executor = self._make_executor(shared, stop_words=["Observation:"])
+
+ with _llm_stop_words_applied(shared, executor):
+ assert set(shared.stop_sequences) == {"Original:", "Observation:"}
+ assert shared.stop == ["Original:"]
+
+ assert shared.stop == ["Original:"]
+ assert shared.stop_sequences == ["Original:"]
+
+ def test_override_cleared_when_context_raises(self) -> None:
+ """A failed call must still clear the per-call stop override."""
+ from crewai.utilities.agent_utils import _llm_stop_words_applied
+
+ shared = LLM(model="gpt-4", stop=["Original:"])
+ executor = self._make_executor(shared, stop_words=["Observation:"])
+
+ try:
+ with _llm_stop_words_applied(shared, executor):
+ raise RuntimeError("boom")
+ except RuntimeError:
+ pass
+
+ assert shared.stop == ["Original:"]
+ assert shared.stop_sequences == ["Original:"]
+
+ def test_override_applies_for_post_processing_when_api_lacks_stop_support(
+ self,
+ ) -> None:
+ """Models that lack API-level stop support still need the override.
+
+ Native providers (e.g. Azure on gpt-5/o-series) read ``stop_sequences``
+ in ``_apply_stop_words`` to truncate the response post-hoc even when
+ ``supports_stop_words()`` returns False, so the override must be set
+ regardless of API-level support. (Issue raised by Cursor Bugbot.)
+ """
+ from unittest.mock import patch
+ from crewai.utilities.agent_utils import _llm_stop_words_applied
+
+ shared = LLM(model="gpt-4", stop=["Original:"])
+ executor = self._make_executor(shared, stop_words=["Observation:"])
+
+ with patch.object(shared, "supports_stop_words", return_value=False):
+ with _llm_stop_words_applied(shared, executor):
+ assert set(shared.stop_sequences) == {"Original:", "Observation:"}
+
+ assert shared.stop == ["Original:"]
+ assert shared.stop_sequences == ["Original:"]
+
+ def test_concurrent_overrides_do_not_collide(self) -> None:
+ """Concurrent agents on a shared LLM must each see their own effective stop."""
+ import asyncio
+ from crewai.utilities.agent_utils import _llm_stop_words_applied
+
+ shared = LLM(model="gpt-4", stop=["Original:"])
+ exec_a = self._make_executor(shared, stop_words=["StopA:"])
+ exec_b = self._make_executor(shared, stop_words=["StopB:"])
+
+ async def run(executor: CrewAgentExecutor, expected: str) -> set[str]:
+ with _llm_stop_words_applied(shared, executor):
+ await asyncio.sleep(0)
+ seen = set(shared.stop_sequences)
+ assert expected in seen
+ return seen
+
+ async def main() -> tuple[set[str], set[str]]:
+ return await asyncio.gather(
+ run(exec_a, "StopA:"), run(exec_b, "StopB:")
+ )
+
+ a_seen, b_seen = asyncio.run(main())
+ assert a_seen == {"Original:", "StopA:"}
+ assert b_seen == {"Original:", "StopB:"}
+ assert shared.stop == ["Original:"]
+ assert shared.stop_sequences == ["Original:"]
+
+ def test_override_does_not_leak_to_other_llm_instances(self) -> None:
+ """Override for one LLM must not affect another LLM (e.g. function_calling_llm).
+
+ Regression for Cursor Bugbot: a global ContextVar would leak the
+ override to every BaseLLM that reads stop_sequences during the scope.
+ """
+ from crewai.utilities.agent_utils import _llm_stop_words_applied
+
+ target = LLM(model="gpt-4", stop=["TargetStop:"])
+ other = LLM(model="gpt-4", stop=["OtherStop:"])
+ executor = self._make_executor(target, stop_words=["Observation:"])
+
+ with _llm_stop_words_applied(target, executor):
+ assert set(target.stop_sequences) == {"TargetStop:", "Observation:"}
+ assert other.stop_sequences == ["OtherStop:"]
+
+ assert target.stop_sequences == ["TargetStop:"]
+ assert other.stop_sequences == ["OtherStop:"]
+
+ def test_override_propagates_to_nested_direct_llm_calls(self) -> None:
+ """Once invoke wraps with the override, nested direct llm.call sites
+ (StepExecutor, handle_max_iterations_exceeded) see the merged stops.
+
+ Regression for Cursor Bugbot: those direct call sites bypass
+ get_llm_response, so the override must be set at executor entry, not
+ only around get_llm_response.
+ """
+ from crewai.utilities.agent_utils import _llm_stop_words_applied
+
+ shared = LLM(model="gpt-4", stop=["Original:"])
+ executor = self._make_executor(shared, stop_words=["Observation:"])
+
+ seen: list[set[str]] = []
+
+ def nested_direct_call() -> None:
+ seen.append(set(shared.stop_sequences))
+
+ with _llm_stop_words_applied(shared, executor):
+ nested_direct_call()
+
+ assert seen == [{"Original:", "Observation:"}]
+ assert shared.stop == ["Original:"]