feat: Add OpenAI Responses API Integration

[devin-ai-integration]

feat: Add OpenAI Responses API Integration

Summary

Implements native support for OpenAI's Responses API (/v1/responses) as a new LLM provider in CrewAI, addressing feature request #4152.

The Responses API offers advantages for agent workflows including simpler input format, built-in conversation management via previous_response_id, and native support for o-series reasoning models.

Usage:

# Option 1: Using provider parameter
llm = LLM(model="gpt-4o", provider="openai_responses")

# Option 2: Using model prefix
llm = LLM(model="openai_responses/gpt-4o")

# With o-series reasoning models
llm = LLM(model="o3-mini", provider="openai_responses", reasoning_effort="high")

Key implementation details:

• New OpenAIResponsesCompletion class extending BaseLLM (~900 lines)
• Message conversion: system messages → instructions param, other messages → input param
• Tool calling support with Responses API format (sets strict: True by default)
• Streaming support (sync and async)
• Structured output via Pydantic models
• Support for stateful conversations via previous_response_id

Review & Testing Checklist for Human

• Verify message conversion logic: System messages become instructions, user/assistant messages become input array. This is fundamentally different from Chat Completions - confirm this matches Responses API expectations
• Test with real OpenAI API: All tests use mocks. Recommend testing basic call, streaming, and tool calling against actual API
• Verify tool format: The _convert_tools_for_responses method sets strict: True by default on all tools - verify this doesn't break tool schemas that aren't strict-compatible
• Test o-series model handling: Verify reasoning_effort parameter works correctly with o1/o3/o4 models
• Check streaming event types: Events like response.output_text.delta and response.function_call_arguments.delta need verification against actual streaming responses

Recommended test plan:

1. Create a simple agent with provider="openai_responses" and verify basic completion works
2. Test with a tool-using agent to verify function calling
3. Test streaming with stream=True
4. Test with an o-series model (o3-mini) with reasoning_effort="high"

Notes

• There's an unused import (ResponseOutputMessage) that could be cleaned up
• The context window size lookup uses ordered list matching with startswith - order matters for correct matching (e.g., gpt-4o before gpt-4)

Link to Devin run: https://app.devin.ai/sessions/5344f7b180844f958605133c3772c492
Requested by: João (joao@crewai.com)

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring

[devin-ai-integration]

Closing due to inactivity for more than 7 days. Configure here.