feat(responses): strip custom_tool_call namespace for all providers

[Sameerlite]

Summary

• Add BaseResponsesAPIConfig.strip_custom_tool_call_namespace_from_responses_input and normalize_responses_api_request_dict so namespace is removed from custom_tool_call input items for every Responses API provider (OpenAI-compatible backends reject it).
• Call normalize_responses_api_request_dict on outbound bodies in llm_http_handler (sync/async responses + compact) so normalization always runs on the wire path.

Tests

• tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py: base strip + normalize, OpenAI transform+normalize, Azure transform+normalize.

function_call namespace is preserved for OpenAI deferred tool namespaces; only custom_tool_call is stripped.

[greptile-apps]

Greptile Summary

This PR adds strip_custom_tool_call_namespace_from_responses_input and normalize_responses_api_request_dict to BaseResponsesAPIConfig, then calls the normalizer across all four Responses API send paths (sync/async standard and sync/async compact) in llm_http_handler.py. The goal is to strip the namespace field from custom_tool_call input items before they reach any provider, since OpenAI-compatible backends reject it, while preserving namespace on function_call items.

Confidence Score: 5/5

Safe to merge — the change is minimal, targeted, and all four send paths are covered.

All four Responses API send paths have the normalizer applied. The stripping logic is type-guarded (only custom_tool_call), non-mutating, and idempotent. Tests are all-mock and cover the base, OpenAI, and Azure paths. No existing tests are modified. No P0/P1 issues found.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/base_llm/responses/transformation.py	Adds two new static methods for stripping namespace from custom_tool_call items; logic is correct and non-mutating.
litellm/llms/custom_httpx/llm_http_handler.py	Applies normalize_responses_api_request_dict at all four Responses API send sites (sync/async standard + compact); ordering relative to extra_body.update() is fine for typical usage.
tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py	Adds three new unit tests covering base strip, OpenAI transform+normalize, and Azure transform+normalize; no network calls, no existing test modifications.

Sequence Diagram

sequenceDiagram
    participant Caller
    participant llm_http_handler as BaseLLMHTTPHandler
    participant ProviderConfig as responses_api_provider_config
    participant Normalizer as BaseResponsesAPIConfig.normalize_*
    participant Provider as Remote Provider

    Caller->>llm_http_handler: async/sync responses_api_handler (input with custom_tool_call)
    llm_http_handler->>ProviderConfig: transform_responses_api_request(...)
    ProviderConfig-->>llm_http_handler: data (dict with input)
    llm_http_handler->>Normalizer: normalize_responses_api_request_dict(data)
    Note over Normalizer: strip namespace from custom_tool_call items<br/>preserve namespace on function_call items
    Normalizer-->>llm_http_handler: normalized data
    llm_http_handler->>Provider: POST /responses (no namespace on custom_tool_call)

Loading

_{Reviews (1): Last reviewed commit: "feat(responses): strip custom_tool_call ..." | Re-trigger Greptile}