fix(deepseek): back-fill reasoning_content for all assistant turns

docs/en/models/deepseek.mdx

@@ -14,12 +14,41 @@ Option 1: Native integration (recommended):
 
 | Parameter | Description |
 | --- | --- |
-| `model` | `deepseek-v4-pro` (V4 Pro, thinking mode + tool calls, Agent recommended), `deepseek-v4-flash` (V4 Flash, thinking mode + tool calls), `deepseek-chat` (DeepSeek-V3.2, non-thinking), `deepseek-reasoner` (DeepSeek-R1, thinking mode) |
+| `model` | Supports `deepseek-v4-pro` and `deepseek-v4-flash` |
 | `deepseek_api_key` | Create at [DeepSeek Platform](https://platform.deepseek.com/api_keys) |
 | `deepseek_api_base` | Optional, defaults to `https://api.deepseek.com/v1`. Can be changed to a third-party proxy |
 
+## Model Selection
+
+| Model | Use Case |
+| --- | --- |
+| `deepseek-v4-pro` | Best on complex tasks |
+| `deepseek-v4-flash` | Faster and cheaper |
+
+## Thinking Mode
+
+The V4 series (`deepseek-v4-pro` / `deepseek-v4-flash`) supports an explicit "thinking mode": the model emits a chain-of-thought (`reasoning_content`) before the final answer to improve answer quality.
+
+### Toggle
+
+Controlled by the global `enable_thinking` setting:
+
+```json
+{
+  "enable_thinking": true
+}
+```
+
+- `true`: thinking is on across all channels. The Web console renders the reasoning trace; IM channels (WeChat / WeCom / DingTalk / Feishu) don't render it but still benefit from higher answer quality.
+- `false`: thinking off, faster responses with lower first-token latency.
+
+### Notes
+
+- **Sampling parameters**: under thinking mode, `temperature`, `top_p`, `presence_penalty`, and `frequency_penalty` are silently ignored by the server (no error). CowAgent skips sending them automatically.
+- **Multi-turn tool calls**: once the history contains any tool-call turn, DeepSeek requires `reasoning_content` on every assistant message. CowAgent handles the round-trip automatically, including across mid-session toggles of the thinking switch.
+
 <Tip>
-  The V4 series (`deepseek-v4-pro`, `deepseek-v4-flash`) supports thinking mode together with tool calls, enabling multi-round reasoning + tool orchestration in Agent mode. The thinking switch is controlled by the global `enable_thinking` setting, and `reasoning_content` is automatically round-tripped on tool-call turns.
+  Use `deepseek-v4-pro` for complex tasks; pick `deepseek-v4-flash` when low latency and low cost matter; enable `enable_thinking` when you want deeper reasoning.
 </Tip>
 
 Option 2: OpenAI-compatible configuration: