### What happened?

Sending Anthropic-style `thinking` (or `reasoning_effort`) to a **DeepSeek R1** model on **AWS Bedrock** (Converse API) causes AWS to reject the request with a generic 400 `Validation Error`, because litellm forwards `thinking` verbatim into `additionalModelRequestFields` without any model-family gating — DeepSeek R1 on Bedrock doesn't support that field; it reasons via a `<think>` prompt-completion convention, not a Converse-API-level toggle.

Without `thinking`/`reasoning_effort` in the request, calls to the same model succeed (including via `/v1/messages`, which defaults reasoning on for DeepSeek R1 through the model's native `<think>` behavior — no `thinking` field needed).

This is the same class of bug reported in #13959 (gpt-oss/qwen/deepseek-v3.1 on Bedrock), which was auto-closed as stale/not-planned without a fix. Root cause is the same: `AmazonConverseConfig.map_openai_params()` sets `optional_params["thinking"] = value` unconditionally for any non-adaptive-thinking model, and `_handle_reasoning_effort_parameter()` only special-cases `gpt-oss` and Nova 2 — every other model (including `deepseek.r1`) falls through to the Anthropic-shaped `thinking` map…