Gemini thinking_config malformed request causing 400 errors

Fixes on record

Merged to main. A stamp counts the issues a fix is linked to.

• #17904 fix(transport): omit thinking_config for Gemma on the gemini provider (#17426)

  merged to main

  files: agent/transports/chat_completions.py, tests/agent/transports/test_chat_completions.py

  Show the diff

  • agent/transports/chat_completions.py

    @@ -24,6 +24,18 @@ def _build_gemini_thinking_config(model: str, reasoning_config: dict | None) ->
         if reasoning_config is None or not isinstance(reasoning_config, dict):
             return None
     
    +    normalized_model = (model or "").strip().lower()
    +    if normalized_model.startswith("google/"):
    +        normalized_model = normalized_model.split("/", 1)[1]
    +
    +    # ``thinking_config`` is a Gemini-only request parameter. The same
    +    # ``gemini`` provider also serves Gemma (and historically PaLM/Bard);
    +    # those reject the field with HTTP 400 "Unknown name 'thinking_config':
    +    # Cannot find field" — including the polite ``{"includeThoughts": False}``
    +    # form. Omit the field entirely on non-Gemini models. (#17426)
    +    if not normalized_model.startswith("gemini"):
    +        return None
    +
         if reasoning_config.get("enabled") is False:
             # Gemini can hide thought parts even when internal thinking still
             # happens; omit thinkingLevel to avoid model-specific validation quirks.
    @@ -34,9 +46,6 @@ def _build_gemini_thinking_config(model: str, reasoning_config: dict | None) ->
             return {"includeThoughts": False}
     
         thinking_config: Dict[str, Any] = {"includeThoughts": True}
    -    normalized_model = (model or "").strip().lower()
    -    if normalized_model.startswith("google/"):
    -        normalized_model = normalized_model.split("/", 1)[1]
     
         # Gemini 2.5 accepts thinkingBudget; don't guess a budget from Hermes'
         # coarse effort levels. ``includeThoughts`` alone is enough to surface

  • tests/agent/transports/test_chat_completions.py

    @@ -244,6 +244,46 @@ def test_gemini_flash_minimal_clamps_to_low(self, transport):
                 "thinking_level": "low",
             }
     
    +    def test_gemma_does_not_receive_thinking_config(self, transport):
    +        # The `gemini` provider also serves Gemma (e.g. `gemma-4-31b-it`),
    +        # but Gemma rejects `thinking_config` with HTTP 400 (#17426). Even
    +        # when Hermes has reasoning enabled, the field must be omitted for
    +        # non-Gemini models on this provider.
    +        msgs = [{"role": "user", "content": "Hi"}]
    +        kw = transport.build_kwargs(
    +            model="gemma-4-31b-it",
    +            messages=msgs,
    +            provider_name="gemini",
    +            reasoning_config={"enabled": True, "effort": "high"},
    +        )
    +        assert "thinking_config" not in kw.get("extra_body", {})
    +
    +    def test_gemma_disabled_reasoning_still_omits_thinking_config(self, transport):
    +        # The `Unknown name 'thinking_config': Cannot find field` rejection
    +        # fires even on `{"includeThoughts": False}` — the entire field must
    +        # be absent, not just disabled. (#17426)
    +        msgs = [{"role": "user", "content": "Hi"}]
    +        kw = transport.build_kwargs(
    +            model="gemma-4-31b-it",
    +            messages=msgs,
    +            provider_name="gemini",
    +            reasoning_config={"enabled": False},
    +        )
    +        assert "thinking_config" not in kw.get("extra_body", {})
    +
    +    def test_google_prefixed_gemma_also_omits_thinking_config(self, transport):
    +        # OpenRouter-style `google/gemma-...` IDs hit the same provider path
    +        # and must also omit `thinking_config`. The existing `google/`
    +        # prefix-stripping must not accidentally classify Gemma as Gemini.
    +        msgs = [{"role": "user", "content": "Hi"}]
    +        kw = transport.build_kwargs(
    +            model="google/gemma-4-31b-it",
    +            messages=msgs,
    +            provider_name="gemini",
    +            reasoning_config={"enabled": True, "effort": "medium"},
    +        )
    +        assert "thinking_config" not in kw.get("extra_body", {})
    +
         def test_max_tokens_with_fn(self, transport):
             msgs = [{"role": "user", "content": "Hi"}]
             kw = transport.build_kwargs(