[Bug]: gemini-3.7-flash unusable — hardcoded thinking_level "minimal" is rejected by the model (only low/medium/high supported)

[GlikIta]

Describe the bug

Selecting gemini-3.7-flash as the LLM (google / google_vertex provider — the model field allows custom input) makes every generation fail with an INVALID_ARGUMENT error about the thinking level (observed live on app.dograh.com, 2026-08-21: 400 … THINKING_LEVEL_MINIMAL). The agent is completely broken with that model.

Root cause

The bundled pipecat fork applies a blanket low-latency thinking default to every Gemini 3 Flash model when no explicit thinking config is set (and the app never sets one — there is no thinking field anywhere in the configuration schema):

# dograh-hq/pipecat, src/pipecat/services/google/llm.py — _maybe_unset_thinking_budget()
elif model.startswith("gemini-3") and "flash" in model:
    generation_params["thinking_config"] = {"thinking_level": "minimal"}

But per Google's documentation, Gemini 3.7 Flash does not support minimal: supported values are LOW, MEDIUM (default) and HIGH, and "explicitly setting thinking_level to MINIMAL will return an API validation error" (What's new in Gemini 3.7 Flash, Vertex model page).

Fix already exists upstream

Upstream pipecat has already fixed exactly this, with a per-model lowest-level map:

# pipecat-ai/pipecat, src/pipecat/services/google/llm.py
_LOWEST_MODEL_THINKING_LEVELS = {
    "gemini-3.7-flash": "low",
}
...
elif model.startswith("gemini-3") and "flash" in model:
    level = next(
        (lowest for prefix, lowest in _LOWEST_MODEL_THINKING_LEVELS.items()
         if model.startswith(prefix)),
        "minimal",
    )
    generation_params["thinking_config"] = {"thinking_level": level}

Its docstring states the contract explicitly: "Gemini 3 Flash accepts 'minimal', 'low', 'medium', and 'high', except Gemini 3.7 Flash, which accepts only 'low', 'medium', and 'high'."

Cherry-picking that mechanism into the fork fixes the bug with no behavior change for any other model.

Workarounds today

• Use gemini-3.6-flash (still accepts minimal), or
• none otherwise: there is no thinking-level field in the org configuration, the workflow model_configuration_v2_override, or any env var, so users cannot express a valid level for 3.7.

Same class of bug tracked elsewhere for reference: BerriAI/litellm #18245 (reasoning_effort mapping vs Gemini 3 thinkingLevel).

(Optionally, exposing thinking_level as an advanced LLM config field would also future-proof this — happy to split that into a separate feature request.)