fix: omit thinkingConfig for Gemini Flash Image models without thinking support

[aheritier]

Problem

google/gemini-2.5-flash-image rejects the no-thinking request shape used by Docker Agent in some paths. A minimal text request succeeds when thinkingConfig is omitted, but the same request with thinkingConfig: {"thinkingBudget": 0} returns HTTP 400 INVALID_ARGUMENT: Thinking is not enabled for this model.

Docker Agent's WithNoThinking / MCP sampling path can serialize this shape, causing requests to fail for Gemini Flash Image.

Evidence

Isolated gateway probes on 2026-09-10 reproduced the behavior:

• thinkingConfig omitted: accepted.
• thinkingConfig: {"thinkingBudget": 0}: rejected with the 400 above.

Sanitized reproduction

The following is illustrative and contains no credentials:

{
  "model": "google/gemini-2.5-flash-image",
  "contents": [{"role": "user", "parts": [{"text": "Reply with OK"}]}],
  "thinkingConfig": {"thinkingBudget": 0}
}

Send the JSON to the configured Gemini-compatible gateway with a redacted bearer credential, for example:

curl -sS -X POST "https://<gateway>/v1beta/models/google%2Fgemini-2.5-flash-image:generateContent" \
  -H "Authorization: Bearer <REDACTED>" \
  -H "Content-Type: application/json" \
  --data @request.json

Expected behavior

Omit thinkingConfig for models that do not support thinking. Retain the existing disable-thinking semantics for models that do support thinking.

Scope

This issue covers the confirmed Gemini Flash Image incompatibility only. The unconfirmed Gemini 3.8 failure is explicitly out of scope.

Related epic: #3996

[aheritier]

Update: confirmed direct Gemini API reproduction

We reproduced this independently of Docker Models Gateway using direct Gemini access via bypass_models_gateway.

For gemini-2.5-flash-image:

• A request with thinkingConfig omitted succeeds.
• The same request with thinkingConfig: {"thinkingBudget": 0} fails with HTTP 400 INVALID_ARGUMENT: Thinking is not enabled for this model.

Minimal request shape:

{
  "model": "gemini-2.5-flash-image",
  "contents": [{"role": "user", "parts": [{"text": "Reply with OK"}]}],
  "thinkingConfig": {"thinkingBudget": 0}
}

This confirms that the incompatibility is at the Gemini model/API level and is not specific to Docker Models Gateway.

The explicit-thinking CLI test confirms the model-level incompatibility. Separately, the affected Docker Agent path is WithNoThinking/MCP serialization, which emits a zero budget. The direct CLI test is not itself an exact MCP-path reproduction; it validates the same request-shape incompatibility at the provider API.

Expected fix: for models that do not support thinking, omit thinkingConfig entirely. Preserve existing disable-thinking behavior for models that do support thinking.

[gargsajal9]

@aheritier, would you be happy for me to pick this up? I'd keep it scoped to omitting thinkingConfig for non-thinking Gemini Flash Image models in the WithNoThinking/MCP sampling path, with regression tests preserving the behavior of thinking-capable models. I'd leave the unconfirmed Gemini 3.8 issue out of scope.

I see you're assigned, so I'll wait for your go-ahead before starting.

[aheritier]

I prefer to get some advices from @rumpl or @dgageot first. I am not sure about which path to follow to avoid a regression.