[sumedh-aerram]

Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

1. Link to an existing issue (if applicable):

• Closes: LiteLlm sends no response_format for non-BaseModel output schemas (list, int, dict, Literal) #7946
• Related: LlmAgent output_schema silently becomes an empty Schema() for Literal and Annotated types #7927 / fix(agents): keep Literal and Annotated output schemas #7928

Problem:
_to_litellm_response_format only converted dicts and BaseModel classes or instances. For any other output_schema (list[Item], int, Literal[...], Annotated[...]) it logged Unsupported response_schema type ... for LiteLLM structured outputs and returned None, so no response_format was sent and the model answered unconstrained. Annotated[Model, Field(...)] actually raised a TypeError, because the alias passed the model_dump check.

Solution:

• Remaining schema types are converted with the existing schema_to_json_schema.
• For OpenAI-style strict json_schema targets, which require an object root, a non-object root is nested as {"type": "object", "properties": {"response": <schema>}, "required": ["response"], "additionalProperties": false}, with $defs kept at the document root. This mirrors the native Anthropic adapter's structured-output tool.
• The reply is unwrapped back to the value (json.dumps(reply["response"])), for non-streaming responses and for the final streaming aggregate.
• Unwrapping happens only when ADK did the wrapping. A BaseModel reply that contains a response key is untouched. So is a reply when the caller passed their own response_format.
• Partial streaming chunks keep the raw wrapped text.
• LiteLLM's Gemini path (json_object + response_schema) is unchanged and is never wrapped.
• Map schemas (dict[str, int], an object with only a schema in additionalProperties) can't be expressed in OpenAI strict mode. They keep the previous behaviour: a warning and no response_format. Types that schema_to_json_schema can't convert are handled the same way.
• When config.response_schema is unset, config.response_json_schema is used. That's how Literal/Annotated output schemas arrive after fix(agents): keep Literal and Annotated output schemas #7928, and how callers set a JSON schema directly.
• _to_litellm_response_format keeps its signature. _enforce_strict_openai_schema is not modified (it is touched by fix(openai): keep $defs of a recursive output_schema in strict mode #7806).

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally. (See below: pre-existing environment-dependent failures only.)

New tests in tests/unittests/models/test_litellm.py cover:

• the response_format for list[M], int, dict[str, int], Literal and Annotated on OpenAI and Gemini model strings;
• unwrapping of wrapped replies, streaming and non-streaming;
• the response_json_schema fallback;
• the unsupported-type warning;
• guard tests: a BaseModel reply with a response key, and a caller response_format override. Each fails if its guard is removed;
• thought parts and non-ASCII text in wrapped replies.

The new regression cases fail on main and pass with the change.

macOS:

• pytest tests/unittests/models: 1959 passed on each of Python 3.11, 3.12, 3.13 and 3.14.
• Full pytest tests/unittests -n auto on Python 3.11 (sandboxed runner, before the final test-only commit): 22672 passed, 46 failed, 3 errors. Every failure and error ID also occurs without this change: telemetry goldens, A2A, macOS VMaaS, and sandbox restrictions on sockets/file writes. No new failures.
• pre-commit on the changed files: clean. mypy on lite_llm.py: no new errors.

Manual End-to-End (E2E) Tests:

Live against OpenAI (LiteLlm(model="openai/gpt-4o-mini"), via LlmRequest.set_output_schema + generate_content_async):

output_schema	This PR	main
list[Item]	[{"name": "Vacuum Cleaner"}, {"name": "Toaster"}], validates	1. Toaster\n2. blender, fails validation
Literal["yes", "no"]	"yes", validates	Yes., fails validation
int	8, validates	A spider has eight legs., fails validation
Item (BaseModel)	{"name":"Coffee Mug"}, validates	{"name":"Toaster"}, validates (unchanged)

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes. (New and affected tests pass; see the pre-existing failures above.)
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

• Other LiteLLM providers that receive json_schema (for example anthropic/, ollama_chat/, azure/) get the same wrapping as OpenAI, since they also take the strict object-root format. They were checked by request inspection only, not live.
• Map schemas still get no response_format on strict targets, as before.