Support /v1/responses in complexity router

[krrish-berri-2]

Summary

• Adds get_structured_messages() to guardrail translation base + OpenAI chat, OpenAI Responses, and Anthropic handlers so non-chat formats can be converted to OpenAI-spec messages.
• Complexity router reuses the handler dispatch to classify requests that arrive via /v1/responses (or other non-chat formats).
• Auto-router: new _extract_text_from_messages helper handles tool-call turns (content=None) and multimodal content correctly.
• Widens async_pre_routing_hook messages type from Dict[str, str] to Dict[str, Any].

Fixes #25134

Clean re-spin of #25830 — excludes _experimental/out/*, repro_14633.py, and uv.lock.

Test plan

• make test-unit passes
• New tests in tests/test_litellm/router_strategy/test_auto_router.py and test_complexity_router.py pass
• New tests for get_structured_messages in OpenAI chat and Responses handlers pass

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov]

Codecov Report

❌ Patch coverage is 77.01284% with 197 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/llms/custom_httpx/llm_http_handler.py	12.80%	109 Missing ⚠️
...m/integrations/compression_interception/handler.py	83.43%	27 Missing ⚠️
...llm/integrations/websearch_interception/handler.py	50.94%	26 Missing ⚠️
litellm/compression/compress.py	92.77%	12 Missing ⚠️
...ass_through/messages/agentic_streaming_iterator.py	96.31%	6 Missing ⚠️
...lms/scaleway/audio_transcription/transformation.py	89.47%	6 Missing ⚠️
.../litellm_core_utils/get_supported_openai_params.py	0.00%	4 Missing ⚠️
...ms/anthropic/chat/guardrail_translation/handler.py	72.72%	3 Missing ⚠️
litellm/types/compression.py	66.66%	2 Missing ⚠️
litellm/proxy/proxy_cli.py	66.66%	1 Missing ⚠️
... and 1 more

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR wires /v1/responses (and other non-chat formats) into the complexity router by adding a get_structured_messages() method to each guardrail translation handler and a new _resolve_messages dispatcher in ComplexityRouter. The auto-router gains a _extract_text_from_messages helper to correctly handle tool-call turns and multimodal content, and async_pre_routing_hook's messages type annotation is widened to Dict[str, Any].

Confidence Score: 5/5

Safe to merge — no blocking bugs; all remaining findings are P2 style/robustness suggestions.

All findings are P2: a misleading comment and missing secondary fallback in _resolve_messages, ambiguous handler ordering in the fallback scan, and a redundant adapter instantiation. None affect correctness for the primary supported paths. New tests are well-structured mock-only unit tests.

litellm/router_strategy/complexity_router/complexity_router.py — _resolve_messages fallback logic

Important Files Changed

Filename	Overview
litellm/router_strategy/complexity_router/complexity_router.py	Adds _resolve_messages to convert non-chat formats into chat messages for complexity classification; two P2 issues with the fallback logic.
litellm/llms/base_llm/guardrail_translation/base_translation.py	Adds get_structured_messages as a safe non-abstract base method returning None.
litellm/llms/openai/responses/guardrail_translation/handler.py	Adds get_structured_messages delegating to existing transformation config.
litellm/llms/anthropic/chat/guardrail_translation/handler.py	Adds get_structured_messages; minor redundant adapter instantiation in _translate_to_openai.
litellm/router_strategy/auto_router/auto_router.py	Adds _extract_text_from_messages static helper; clean and well-tested.
litellm/integrations/custom_logger.py	Widens messages type annotation; backwards-compatible.

_{Reviews (2): Last reviewed commit: "fix: fallback to trying each handler whe..." | Re-trigger Greptile}

[greptile-apps]

Redundant unreachable guard

call_type is only ever assigned inside if ct in mappings (line 365), so the second if call_type not in mappings: return None block can never be reached. This is dead code that can be removed for clarity.

Simply delete lines 373-374 and go straight to handler = mappings[call_type]().

[greptile-apps]

New handler instance on every non-chat routing request

mappings[call_type]() instantiates a fresh handler object on each call to _resolve_messages. For AnthropicMessagesHandler this also creates a new LiteLLMAnthropicMessagesAdapter. While not a Router or DB object, it is unnecessary per-request allocation. Consider caching the handler instance (e.g. as an instance variable populated lazily) so it is created once and reused.