PR #4958: feat: add Responses API support for Azure, Bedrock, Gemini, Anthropic providers

• Repository: crewAIInc/crewAI
• Author: devin-ai-integration[bot]
• State: open | merged: False
• Link: https://github.com/crewAIInc/crewAI/pull/4958

Description (problem / solution / changelog)

Summary

Implements Responses API support for Azure OpenAI using the openai SDK's AzureOpenAI client, and adds an api parameter to Bedrock, Gemini, and Anthropic providers that raises NotImplementedError with clear guidance when api="responses" is used.

Azure (full implementation):

• New api parameter: "completions" (default, backward-compatible) or "responses"
• Uses openai.AzureOpenAI / AsyncAzureOpenAI for the Responses API (Azure AI Inference SDK doesn't support it yet)
• Supports: instructions, store, previous_response_id, include, builtin_tools, parse_tool_outputs, auto_chain, auto_chain_reasoning, seed, reasoning_effort, max_completion_tokens
• System messages extracted to instructions; tools converted to internally-tagged format; structured output via text.format
• Auto-chaining via response IDs and ZDR-compliant reasoning item tracking
• Sync/async streaming support

Bedrock, Gemini, Anthropic (error handling only):

• New api parameter with NotImplementedError when api="responses" is passed
• Error messages reference each provider's native API and suggest api='completions'

Closes #4957

Review & Testing Checklist for Human

• Verify _handle_responses / _ahandle_responses / streaming handler correctness — these ~400 lines of response handling, tool execution, and auto-chaining logic (azure/completion.py lines ~1300–1990) have zero integration test coverage. All tests mock away the actual API call. Manually test against a real Azure OpenAI endpoint with api="responses" to validate tool calling, streaming, structured output, and multi-turn chaining.
• Validate Azure API version "2025-03-01-preview" — this is hardcoded as the default for Responses API clients (line ~268). Confirm this version is available on your Azure deployments and supports the Responses API.
• Check openai package version compatibility — the implementation uses openai.AzureOpenAI(...).responses.create(). Verify that the openai version pinned in pyproject.toml (>=1.83.0) actually has responses.create() on the AzureOpenAI client.
• Review _convert_tools_for_responses and _prepare_responses_params — these transform Chat Completions-style tool dicts to Responses API format. Edge cases with nested schemas, function-wrapped tools, or tools with strict mode may not be handled.
• Test ResponsesAPIResult import from OpenAI provider — Azure imports ResponsesAPIResult from crewai.llms.providers.openai.completion. This cross-provider coupling should work but is worth verifying won't break if the OpenAI provider changes.

Notes

• The _call_responses path is only activated when api="responses" is explicitly set, so existing api="completions" (default) behavior is completely unchanged.
• Tests are all unit-level with mocks — no VCR cassettes or live API calls. The 37 Azure tests primarily cover _prepare_responses_params and initialization; the actual response handling code paths are untested.
• ~888 new lines in azure/completion.py — the bulk of the PR and highest review priority.

Link to Devin session: https://app.devin.ai/sessions/f7684e6b99e24a33a5303a862cb31d4d

Changed files

• lib/crewai/src/crewai/llms/providers/anthropic/completion.py (modified, +11/-0)
• lib/crewai/src/crewai/llms/providers/azure/completion.py (modified, +1012/-2)
• lib/crewai/src/crewai/llms/providers/bedrock/completion.py (modified, +12/-1)
• lib/crewai/src/crewai/llms/providers/gemini/completion.py (modified, +11/-0)
• lib/crewai/tests/llms/anthropic/test_anthropic.py (modified, +52/-0)
• lib/crewai/tests/llms/azure/test_azure.py (modified, +781/-0)
• lib/crewai/tests/llms/bedrock/test_bedrock.py (modified, +53/-0)
• lib/crewai/tests/llms/google/test_google.py (modified, +52/-0)