Gemini 3.6 Flash / 3.5 Flash lite sends deprecated sampling parameters through OpenRouter

[anandpant]

Description

When using google/gemini-3.6-flash or google/gemini-3.5-flash-lite through OpenRouter, OpenCode automatically sends:

  {
    "temperature": 1,
    "top_p": 0.95,
    "top_k": 64
  }

Google documents temperature, top_p, and top_k as deprecated for these models and instructs clients to remove them:

https://ai.google.dev/gemini-api/docs/generate-content/latest-model#migrate-to-gemini-3-6-flash

This causes OpenRouter requests using provider.require_parameters: true to fail before inference because the Google AI Studio and Vertex endpoints do not advertise top_k:

HTTP 404
No endpoints found that can handle the requested parameters.

note... a controlled test succeeds top_k is omitted so i guess technically that would fix my problem, but it's misleading when tuning behavior to hardcode values that dont do anything

Expected: OpenCode should not automatically add temperature, top_p, or top_k for Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and future Gemini models covered by Google’s sampling-parameter deprecation

Plugins

no

OpenCode version

1.18.4

Steps to reproduce

1. Configure the OpenRouter model:

  {
    "$schema": "https://opencode.ai/config.json",
    "model": "openrouter/google/gemini-3.6-flash",
    "provider": {
      "openrouter": {
        "models": {
          "google/gemini-3.6-flash": {
            "name": "Gemini 3.6 Flash",
            "options": {
              "provider": {
                "only": ["google-ai-studio", "google-vertex"],
                "require_parameters": true
              },
              "reasoning": {
                "effort": "low",
                "exclude": true
              }
            }
          }
        }
      }
    }
  }

2. Run:

opencode run --pure \
  --model openrouter/google/gemini-3.6-flash \
  "Reply with the single word done."

3. Observe the pre-inference 404.
4. Capture the outgoing /chat/completions request w/ mitmproxy or add a breakpoint. It contains:

{
"temperature": 1,
"top_p": 0.95,
"top_k": 64
}

5. Send the same request without top_k. It routes successfully with
   HTTP 200.

same thing w/ gemini-3.5-flash-lite

Screenshot and/or share link

no

Operating System

macOS 15.7.7

Terminal

ghostty

[kagura-agent]

I'd like to work on this. The root cause is in packages/opencode/src/provider/transform.ts — the temperature(), topP(), and topK() functions use a blanket id.includes("gemini") match that applies defaults to all Gemini models, including 3.6 Flash and 3.5 Flash-Lite which have deprecated these sampling parameters. Fix: add a version-aware guard that returns undefined for Gemini 3.5-flash-lite and 3.6+ models.