[github-actions]

agentgateway/agentgateway#4151 — llm: send thinkingLevel instead of thinkingBudget for gemma-4 and gemini-4

Docs versions: standalone/main, kubernetes/main

What changes for users: OpenAI-compatible Chat Completions requests to Gemini and Vertex AI still use reasoning_effort, but Gemini 4 and Gemma 4 requests now send thinkingConfig.thinkingLevel instead of thinkingConfig.thinkingBudget. Gemma 4 maps none, minimal, and low to minimal, and Gemini 4 maps minimal and low to low.

What the docs now say: The Gemini and Vertex AI provider pages now list how each reasoning_effort value maps to thinkingBudget or thinkingLevel for Gemini 2.5 and older, Gemini 3, Gemini 4, and Gemma 4 models.

Not verified: The configuration was not applied to a cluster or standalone run; the docs are unrun.

How agentgateway-4151 was drafted, and what was not verified

• Code PR: llm: send thinkingLevel instead of thinkingBudget for gemma-4 and gemini-4 agentgateway#4151
• Code issue: [Bug] reasoning_effort on Gemma 4 models on Google AI Studio (Gemini Provider) translated to thinkingBudget instead thinkingLevel agentgateway#4148
• Product: agentgateway (upstream) → agentgateway/website
• Docs version: standalone/main and kubernetes/main
• Category: scoped-edit / pr

Plan, and what changed it

• Planned: Document the reasoning_effort conversion on the Gemini and Vertex AI provider pages for standalone/main and kubernetes/main.
• Learned: The native Gemini API pages only describe pass-through generationConfig.thinkingConfig, while the dossier diff changes Chat Completions provider conversion in crates/llm/src/conversion/vertex_gemini.rs.
• Did: Added reasoning-effort mapping tables to the standalone Gemini and Vertex AI provider pages, the Kubernetes main Gemini wrapper, and the Kubernetes shared Vertex AI asset with a main-only version gate.

Why a documentation change is necessary

The pull request changes how Chat Completions reasoning_effort is sent to Gemini-family providers. The provider pages explain how to configure Gemini and Vertex AI, but they did not say that reasoning_effort becomes thinkingBudget for older models and thinkingLevel for Gemini 3, Gemini 4, and Gemma 4. Without this update, a reader cannot tell why Gemma 4 and Gemini 4 need a different thinking configuration from Gemini 2.5.

What changed on disk

• assets/agw-docs/pages/agentgateway/llm/providers/vertex.md: Added the Kubernetes Vertex AI reasoning_effort mapping, gated to main and future versions.
• content/docs/kubernetes/main/integrations/llm/providers/gemini.md: Added the Kubernetes main Gemini reasoning_effort mapping.
• content/docs/standalone/main/integrations/llm/providers/gemini.md: Added the standalone Gemini reasoning_effort mapping.
• content/docs/standalone/main/integrations/llm/providers/vertex.md: Added the standalone Vertex AI reasoning_effort mapping.

How to verify this change

1. Configure a Gemini or Vertex AI provider as described on the edited page, using a model such as gemma-4-e2b-it.
2. Send a Chat Completions request with reasoning_effort: "low" through that provider. The request should not return Thinking budget is not supported for this model, because the provider request uses thinkingLevel: "minimal".
3. Repeat with a Gemini 4 model such as gemini-4-argon and reasoning_effort: "minimal". The request should not return Use thinkingLevel instead, because the provider request uses thinkingLevel: "low".

What was not verified

• The configuration was not applied to a cluster or standalone process; all edited examples and request flows are unrun.
• Internal links are unchecked because Hugo was not run.

How the pages were chosen (triage report)

Draft a documentation change

agentgateway/agentgateway#4151 — llm: send thinkingLevel instead of thinkingBudget for gemma-4 and gemini-4

• Product: agentgateway (upstream) → agentgateway/website
• Docs version: standalone/main, kubernetes/main
• Code issue: [Bug] reasoning_effort on Gemma 4 models on Google AI Studio (Gemini Provider) translated to thinkingBudget instead thinkingLevel agentgateway#4148
• Summary (read from the body): reasoning_effort was translated to thinkingConfig.thinkingBudget for every model except gemini-3*. But: Notes: - none sends minimal on gemma-4 instead of omitting the config, because Gemma 4 thinks by default when thinkingConfig is absent and minimal is its only off switch. - gemini-2.5 and older keep the existing budget behavior; an explicit thinking_config/thinkingConfig passth
• Not tested: no cluster was available, so every command in this diff is unrun

A surface this change does not reach

them again: they describe a field on a deployment mode the code change
never reached, so a reader following them would configure something that
does not exist yet. The rest of the draft is untouched.

• content/docs/kubernetes/main/integrations/llm/providers/gemini.md

Either the control-plane work for that mode is still owed, or the topic
belongs only to the other mode. The run summary names which surface and
which identifiers.

Release notes

This pull request also adds its release-note entries:

• content/docs/kubernetes/main/release-notes/release-notes.md: llm: send thinkingLevel instead of thinkingBudget for gemma-4 and gemini-4 (fix, under LLM, in a new 1.8 section)
• content/docs/standalone/main/release-notes/release-notes.md: llm: send thinkingLevel instead of thinkingBudget for gemma-4 and gemini-4 (fix, under LLM, in a new 1.8 section)

Check these release-note links and names. Each note is on its page, but a link may not reach the drafted content there, or a name may belong to the other mode:

• content/docs/kubernetes/main/release-notes/release-notes.md: llm: send thinkingLevel instead of thinkingBudget for gemma-4 and gemini-4 agentgateway#4151's note names thinkingLevel, which content/docs/standalone/main mentions and content/docs/kubernetes/main does not. Check that this tree's readers can set it, or word the note for this mode.

Draft branch: pr-tracker-draft-agentgateway-4151

[cloudflare-workers-and-pages]

Deploying agentproxy with    Cloudflare Pages

Latest commit:	661a940
Status:	✅  Deploy successful!
Preview URL:	https://d8a2e028.agentproxy.pages.dev
Branch Preview URL:	https://pr-tracker-draft-agentgatewa-415r.agentproxy.pages.dev

View logs

[kristin-kronstain-brown]

Moved this PR's release-notes entry to the combined release-notes PR, #1482, so the draft PRs stop conflicting on the release-notes pages. Review or change the entry there.