feat(models): wire maxOutputTokens for native Gemini and Vertex

[anthonytwh]

What

Wire maxOutputTokens end to end for the native Gemini and Gemini Vertex AI model providers.

Previously neither provider could cap a response's output tokens via ModelConfig: GeminiConfig was an empty struct and GeminiVertexAIConfig.MaxOutputTokens was declared but never read by the translator or applied at runtime. Both providers subclass the ADK native Gemini model, which takes generation config from the per-request LlmRequest.config rather than the model definition, so nothing plumbed the ModelConfig-level setting through to the API.

Changes

• Add MaxOutputTokens to GeminiConfig (v1alpha2) and regenerate the CRDs (both go/api/config/crd/bases and helm/kagent-crds).
• Carry it on the adk.Gemini / adk.GeminiVertexAI serialization structs.
• Copy Spec.Gemini / Spec.GeminiVertexAI.MaxOutputTokens in the translator.
• Parse it on the Gemini / GeminiVertexAI pydantic models and build the wrappers with it.
• Apply it in a shared _GeminiGenerationConfigMixin that injects max_output_tokens into GenerateContentConfig without overriding a value the agent already set on the request.

A per-request value always wins over the model-level default.

Tests

• Go translator tests: Test_AdkApiTranslator_GeminiMaxOutputTokens (native Gemini + Vertex AI) — pass; full translator package green.
• Python unit tests for the mixin: applies the cap when unset, does not override a per-request value, and is a no-op when uncapped — 3 passed.

[Copilot]

Pull request overview

This PR wires maxOutputTokens end-to-end for both the native Gemini and Gemini Vertex AI providers, carrying a model-level default from the Kubernetes ModelConfig CR through the Go translator into the Python ADK runtime, and applying it per-request without overriding any explicitly set per-request value.

Changes:

• Adds maxOutputTokens to the Gemini model config API/CRD and propagates it through the Go ADK config serialization.
• Introduces a shared Python _GeminiGenerationConfigMixin to apply model-level max_output_tokens only when the request doesn’t already specify it.
• Adds Go translator tests and Python unit tests to verify correct precedence and no-op behavior when uncapped.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
python/packages/kagent-adk/tests/unittests/models/test_gemini.py	Adds unit tests validating model-level max_output_tokens behavior and precedence vs per-request config.
python/packages/kagent-adk/src/kagent/adk/types.py	Extends Gemini/Gemini Vertex AI model config schemas and passes max_output_tokens into runtime wrappers.
python/packages/kagent-adk/src/kagent/adk/models/_gemini.py	Adds generation-config mixin and a Vertex AI wrapper to apply model-level defaults per request.
python/packages/kagent-adk/src/kagent/adk/models/init.py	Exports the new Gemini Vertex AI wrapper.
helm/kagent-crds/templates/kagent.dev_modelconfigs.yaml	Updates Helm CRDs to expose spec.gemini.maxOutputTokens.
go/core/internal/controller/translator/agent/adk_api_translator.go	Copies MaxOutputTokens from ModelConfig into serialized ADK model config (Gemini + Vertex AI).
go/core/internal/controller/translator/agent/adk_api_translator_test.go	Adds translator tests verifying MaxOutputTokens is set for Gemini and Vertex AI.
go/api/v1alpha2/modelconfig_types.go	Adds MaxOutputTokens to the Gemini API type.
go/api/config/crd/bases/kagent.dev_modelconfigs.yaml	Updates generated CRD bases to include spec.gemini.maxOutputTokens.
go/api/adk/types.go	Adds max_output_tokens to serialized ADK Gemini / GeminiVertexAI model structs.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[anthonytwh]

Thanks @copilot — addressed all five comments in ad27cd0. They were the same point (maxOutputTokens accepted 0/negative even though the translator treats <= 0 as "unset"), so I enforced a positive value at every layer:

• +kubebuilder:validation:Minimum=1 on GeminiConfig and GeminiVertexAIConfig.MaxOutputTokens, CRDs regenerated (minimum: 1 now in both the bases and Helm copies) — Kubernetes rejects invalid values on apply.
• pydantic ge=1 on the Gemini / GeminiVertexAI max_output_tokens fields so an invalid ModelConfig fails fast at parse time.
• Added unit tests for the parse-time validation.

Go translator tests + Python model tests pass.

[anthonytwh]

Thanks @copilot — all five comments make sense and are the same point (reject 0/negative maxOutputTokens since the translator treats <= 0 as "unset"). They're already addressed in commit 75d90bfd ("validate: reject non-positive Gemini maxOutputTokens"), which landed after the review was posted:

• Go: +kubebuilder:validation:Minimum=1 on GeminiConfig and GeminiVertexAIConfig.MaxOutputTokens (modelconfig_types.go:77,248)
• CRDs regenerated with minimum: 1 in both api/config/crd/bases and helm/kagent-crds (lines 540 & 554)
• Python: Field(default=None, ge=1) on Gemini and GeminiVertexAI.max_output_tokens (types.py:297,311) — fails fast at parse time

Added parse-time validation tests as well. Kubernetes now rejects invalid values on apply, and the pydantic models reject them on parse.