feat(models): add LlmRequest.to_generate_content_body()

[cnpierrepapi]

Link to Issue or Description of Change

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

• Closes: A custom BaseLlm has no documented mapping from LlmRequest.config to a Vertex request body, and getting it wrong drops your tools with no error #6880

Opening this at @surajksharma07's go-ahead on that thread, which also settled that the helper belongs on LlmRequest in ADK rather than in genai.

Problem:

LlmRequest.config is a single flat GenerateContentConfig with 35 fields. On the wire those fields go to three different destinations, and nothing on the type says which is which:

top level (8):         cachedContent, labels, modelArmorConfig, safetySettings,
                       serviceTier, systemInstruction, toolConfig, tools
generationConfig (22): audioTimestamp, audioTranscriptionConfig, candidateCount,
                       frequencyPenalty, imageConfig, logprobs, maxOutputTokens,
                       mediaResolution, modelConfig, presencePenalty, responseLogprobs,
                       responseMimeType, responseModalities, responseSchema, routingConfig,
                       seed, speechConfig, stopSequences, temperature, thinkingConfig,
                       topK, topP
dropped (3):           httpOptions, automaticFunctionCalling, shouldReturnHttpResponse

Getting the split wrong does not raise. A custom BaseLlm that flattens config by hand and buries tools inside generationConfig still gets a 200 with a well formed response in it. The tools are simply absent, so the model answers from memory instead of calling anything, and it reads as a model that chose not to call the tool.

Solution:

LlmRequest.to_generate_content_body(), delegating to the same converters google-genai uses for its own built-in models (_GenerateContentParameters_to_vertex / _to_mldev) rather than keeping a hand written field list.

The delegation is the point. A documented table or a hardcoded list is a second copy of the mapping, and it goes stale the moment GenerateContentConfig gains a field — every custom BaseLlm in the wild then silently drops that field. I can vouch for the rot personally: the field table I first posted on #6880 said 6 fields go top level. It is 8. I had missed modelArmorConfig and serviceTier, and that list was less than a day old when I wrote it.

Three edge cases are handled, since each would bite anyone calling the converters directly:

1. The converters return the model under a private _url key for the genai client to build the path from. It is routing information, not body content, and sending it is a 400. Stripped.
2. model_selection_config lands in generationConfig as modelConfig. It is the only field that changes name in transit, so grepping your own config keys against the emitted body gives a false miss on exactly that one. Covered by a test.
3. enable_enhanced_civic_answers raises ValueError in Vertex mode (Developer API only). Loud, which is right, but it surfaces from inside the genai converter and a BaseLlm author reading that traceback has no reason to connect it to a field they set several layers up. Re-raised with the LlmRequest.config field name and the target API attached, chained via from exc so the original is intact.

The helper takes an api_client when the caller has one, and falls back to a vertexai flag otherwise, so it stays usable from a BaseLlm that talks REST directly and never builds a genai client. Vertex and the Developer API differ in more than the URL — some fields exist on one and not the other — so that choice is explicit at the call site rather than inferred.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

13 new tests in tests/unittests/models/test_generate_content_body.py, run together with the existing test_llm_request.py:

$ pytest tests/unittests/models/test_generate_content_body.py \
         tests/unittests/models/test_llm_request.py -q
46 passed, 5 warnings in 1.81s

Environment: Python 3.12, google-genai 2.18.1, against main at 2c9c00d.

Coverage: tools landing top level rather than in generationConfig; temperature staying in generationConfig; the three client-only fields never reaching the wire; the _url strip; the full 33-field set splitting 8/22/3 with nothing unaccounted for; the model_selection_config → modelConfig rename; the Developer-API-only field being named in the error message; a field Vertex rejects being accepted on the Developer API; a string system_instruction becoming a content block; a request with no config at all; the method matching the free function; and the API mode being passed through.

The one that matters most for the rot argument: test_every_config_field_is_classified walks GenerateContentConfig.model_fields and fails if a field appears that lands in none of the three buckets. If genai adds a field and this helper cannot place it, CI says so instead of a user finding out through a silently dropped parameter.

Manual End-to-End (E2E) Tests:

Verified against live Vertex AI (gemini-2.5-flash, ADC, us-central1) before this was written up — a config with 33 of the 35 fields populated (omitting response_json_schema, which is mutually exclusive with response_schema, and enable_enhanced_civic_answers, which is Developer-API only) pushed through the converter produced exactly the 8/22/3 split above, with tools at the top level and no field silently lost. The unit tests reproduce that split offline, so no credentials are needed to check it.

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.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

Formatted with pyink==25.12 (the version pinned in pyproject.toml) using the repo's own
[tool.pyink] config, run in a python:3.12-slim container. All three files come back
unchanged.

The Google CLA is signed.