PR #4443: fix: handle streaming tool calls when available_functions is None (fixes #4442)

• Repository: crewAIInc/crewAI
• Author: devin-ai-integration[bot]
• State: closed | merged: False
• Link: https://github.com/crewAIInc/crewAI/pull/4443

Description (problem / solution / changelog)

fix: handle streaming tool calls when available_functions is None (#4442)

Summary

When using crew.akickoff() with stream=True and agents that have tools, the streaming path in the OpenAI provider fails with ValueError: Invalid response from LLM call - None or empty. This happens because 4 streaming methods accumulate tool call chunks but only handle the case where available_functions is provided (execute tools inline). When available_functions is None (the executor handles tool execution), the methods fall through and return empty/None.

This PR adds an early-return path in each of the 4 affected streaming methods to format and return accumulated tool calls when available_functions is None, matching the behavior already present in:

• The non-streaming counterparts of the same methods
• The Azure provider's streaming implementation

Methods fixed in lib/crewai/src/crewai/llms/providers/openai/completion.py:

• _handle_streaming_completion (sync Chat Completions)
• _ahandle_streaming_completion (async Chat Completions)
• _handle_streaming_responses (sync Responses API)
• _ahandle_streaming_responses (async Responses API)

9 unit tests added covering single/multiple tool calls, sync/async, both API styles, and a regression check that available_functions being provided still executes tools as before.

Review & Testing Checklist for Human

• Verify the returned tool call format matches what CrewAgentExecutor._handle_native_tool_calls() expects. The Chat Completions path returns {"id", "type", "function": {"name", "arguments"}} dicts, while the Responses API path returns {"id", "name", "arguments"} dicts — confirm both are consumed correctly downstream.
• Check if other providers (Anthropic, Gemini, Bedrock) have the same streaming gap. Only OpenAI was patched here; Azure already had the fix. The same bug pattern may exist in other providers' streaming methods.
• End-to-end test: Run the reproduction code from issue #4442 — specifically crew.akickoff(inputs={}) with stream=True and tool-equipped agents — to confirm the fix resolves the actual user-reported scenario.

Notes

• Link to Devin run: https://app.devin.ai/sessions/48f4ae0f16a74a349983e6c225963c6c
• Requested by: João

Changed files

• lib/crewai/src/crewai/llms/providers/openai/completion.py (modified, +62/-0)
• lib/crewai/tests/llms/openai/test_openai_streaming_tool_calls_no_available_functions.py (added, +449/-0)

PR #4444: Fix async/sync streaming to return tool calls when available_functions is None

• Repository: crewAIInc/crewAI
• Author: Chase-Xuu
• State: open | merged: False
• Link: https://github.com/crewAIInc/crewAI/pull/4444

Description (problem / solution / changelog)

Summary

Fixes #4442 - Async stream does not work with function calls

When streaming is enabled (crew.stream=True) and the agent has tools registered, the executor receives ValueError: Invalid response from LLM call - None or empty because streaming methods return an empty string instead of the accumulated tool calls.

Root Cause

The CrewAgentExecutor calls the LLM with available_functions=None when it wants to handle tool execution externally (via _handle_native_tool_calls).

Non-streaming methods correctly handle this case:

# In _ahandle_completion (lines 1918-1932)
if message.tool_calls and not available_functions:
    return list(message.tool_calls)  # ✅ Returns tool calls for external handling

Streaming methods only checked if tool_calls and available_functions: and would return an empty full_response string when tool calls existed but available_functions was None:

# Before fix - streaming methods
if tool_calls and available_functions:  # ❌ Misses case where available_functions=None
    # execute tools internally
# ...
return full_response  # Returns "" when tool calls exist but weren't handled

Solution

1. Added import for ChatCompletionMessageToolCall and Function types

2. Added _build_tool_call_list() helper to convert accumulated streaming tool call data into proper ChatCompletionMessageToolCall objects

3. Added the check for tool_calls and not available_functions to both sync (_handle_streaming_completion) and async (_ahandle_streaming_completion) methods:

if tool_calls and not available_functions:
    tool_call_list = self._build_tool_call_list(tool_calls)
    if tool_call_list:
        return tool_call_list

1. Added regression tests in test_streaming_tool_calls_without_functions.py

Checklist

• Code follows existing patterns in the codebase
• Added regression tests
• Syntax verified with py_compile
• Consistent with how non-streaming methods handle this case

Changed files

• lib/crewai/src/crewai/llms/providers/openai/completion.py (modified, +71/-0)
• lib/crewai/tests/llms/test_streaming_tool_calls_without_functions.py (added, +161/-0)

PR #4456: fix: ensure openai tool call stream is finalized

• Repository: crewAIInc/crewAI
• Author: greysonlalonde
• State: closed | merged: True
• Link: https://github.com/crewAIInc/crewAI/pull/4456

Description (problem / solution / changelog)

fixes #4442

Changed files

• lib/crewai/src/crewai/llms/providers/openai/completion.py (modified, +112/-88)
• lib/crewai/tests/llms/openai/test_openai.py (modified, +165/-0)