Skip to content

fix: prevent response_model from being passed in ReAct flow when LLM lacks function calling #4698

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
devin-ai-integration wants to merge 3 commits into
base: main
Choose a base branch
from
+555 −73
Open

fix: prevent response_model from being passed in ReAct flow when LLM lacks function calling #4698

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
20be4ae
fix: prevent response_model from being passed in ReAct flow when LLM …
devin-ai-integration[bot] Mar 4, 2026
7e60321
fix: address review comments - unused import, FC+no-tools routing
devin-ai-integration[bot] Mar 4, 2026
c5d4384
fix: use self.tools instead of self.original_tools for no-tools routing
devin-ai-integration[bot] Mar 4, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
118 changes: 45 additions & 73 deletions lib/crewai/src/crewai/agents/crew_agent_executor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -311,16 +311,22 @@ def _invoke_loop(self) -> AgentFinish:
Final answer from the agent.
"""
# Check if model supports native function calling
use_native_tools = (
supports_fc = (
hasattr(self.llm, "supports_function_calling")
and callable(getattr(self.llm, "supports_function_calling", None))
and self.llm.supports_function_calling()
and self.original_tools
)

if use_native_tools:
if supports_fc and self.original_tools:
return self._invoke_loop_native_tools()

# FC-capable LLM with no user-defined tools but with response_model
# and no internal tools (delegation, human input, etc.): use simple
# native call path which correctly passes response_model for structured
# output instead of dropping it in the ReAct path.
if supports_fc and not self.tools and self.response_model:
return self._invoke_loop_native_no_tools()

# Fall back to ReAct text-based pattern
return self._invoke_loop_react()

Expand Down Expand Up @@ -351,50 +357,30 @@ def _invoke_loop_react(self) -> AgentFinish:

enforce_rpm_limit(self.request_within_rpm_limit)

# In the ReAct flow, do NOT pass response_model to the LLM call.
# When the LLM doesn't support function calling, passing response_model
# forces structured output (via instructor/tools mode) before the agent
# can reason through the Action/Observation loop. The output schema is
# already embedded in the prompt text for guidance, and the final
# conversion to pydantic/json happens in task._export_output().
# See: https://github.com/crewAIInc/crewAI/issues/4695
answer = get_llm_response(
llm=self.llm,
messages=self.messages,
callbacks=self.callbacks,
printer=self._printer,
from_task=self.task,
from_agent=self.agent,
response_model=self.response_model,
response_model=None,
executor_context=self,
verbose=self.agent.verbose,
)
# breakpoint()
if self.response_model is not None:
try:
if isinstance(answer, BaseModel):
output_json = answer.model_dump_json()
formatted_answer = AgentFinish(
thought="",
output=answer,
text=output_json,
)
else:
self.response_model.model_validate_json(answer)
formatted_answer = AgentFinish(
thought="",
output=answer,
text=answer,
)
except ValidationError:
# If validation fails, convert BaseModel to JSON string for parsing
answer_str = (
answer.model_dump_json()
if isinstance(answer, BaseModel)
else str(answer)
)
formatted_answer = process_llm_response(
answer_str, self.use_stop_words
) # type: ignore[assignment]
else:
# When no response_model, answer should be a string
answer_str = str(answer) if not isinstance(answer, str) else answer
formatted_answer = process_llm_response(
answer_str, self.use_stop_words
) # type: ignore[assignment]

# When no response_model is passed, answer should be a string
answer_str = str(answer) if not isinstance(answer, str) else answer
formatted_answer = process_llm_response(
answer_str, self.use_stop_words
) # type: ignore[assignment]

if isinstance(formatted_answer, AgentAction):
# Extract agent fingerprint if available
Expand Down Expand Up @@ -1152,16 +1138,22 @@ async def _ainvoke_loop(self) -> AgentFinish:
Final answer from the agent.
"""
# Check if model supports native function calling
use_native_tools = (
supports_fc = (
hasattr(self.llm, "supports_function_calling")
and callable(getattr(self.llm, "supports_function_calling", None))
and self.llm.supports_function_calling()
and self.original_tools
)

if use_native_tools:
if supports_fc and self.original_tools:
return await self._ainvoke_loop_native_tools()

# FC-capable LLM with no user-defined tools but with response_model
# and no internal tools (delegation, human input, etc.): use simple
# native call path which correctly passes response_model for structured
# output instead of dropping it in the ReAct path.
if supports_fc and not self.tools and self.response_model:
return await self._ainvoke_loop_native_no_tools()

# Fall back to ReAct text-based pattern
return await self._ainvoke_loop_react()

Expand All @@ -1188,50 +1180,30 @@ async def _ainvoke_loop_react(self) -> AgentFinish:

enforce_rpm_limit(self.request_within_rpm_limit)

# In the ReAct flow, do NOT pass response_model to the LLM call.
# When the LLM doesn't support function calling, passing response_model
# forces structured output (via instructor/tools mode) before the agent
# can reason through the Action/Observation loop. The output schema is
# already embedded in the prompt text for guidance, and the final
# conversion to pydantic/json happens in task._export_output().
# See: https://github.com/crewAIInc/crewAI/issues/4695
answer = await aget_llm_response(
llm=self.llm,
messages=self.messages,
callbacks=self.callbacks,
printer=self._printer,
from_task=self.task,
from_agent=self.agent,
response_model=self.response_model,
response_model=None,
executor_context=self,
verbose=self.agent.verbose,
)

if self.response_model is not None:
try:
if isinstance(answer, BaseModel):
output_json = answer.model_dump_json()
formatted_answer = AgentFinish(
thought="",
output=answer,
text=output_json,
)
else:
self.response_model.model_validate_json(answer)
formatted_answer = AgentFinish(
thought="",
output=answer,
text=answer,
)
except ValidationError:
# If validation fails, convert BaseModel to JSON string for parsing
answer_str = (
answer.model_dump_json()
if isinstance(answer, BaseModel)
else str(answer)
)
formatted_answer = process_llm_response(
answer_str, self.use_stop_words
) # type: ignore[assignment]
else:
# When no response_model, answer should be a string
answer_str = str(answer) if not isinstance(answer, str) else answer
formatted_answer = process_llm_response(
answer_str, self.use_stop_words
) # type: ignore[assignment]
# When no response_model is passed, answer should be a string
answer_str = str(answer) if not isinstance(answer, str) else answer
formatted_answer = process_llm_response(
answer_str, self.use_stop_words
) # type: ignore[assignment]

if isinstance(formatted_answer, AgentAction):
fingerprint_context = {}
Expand Down
Loading
Loading