bug: [Auto-review] 自动审批偶发拒绝无害命令，报 litellm/Deepseek 接口格式错误

### 问题描述 / What happened

## 环境
- Cindy 桌面版
- sandbox_mode: workspace-write（文件系统受限，网络受限）
- 命令：`Get-ChildItem -Force`（仅列目录，无害）
- 出现时间：2026-08-02，Asia/Singapore

## 现象
执行无害命令时，被拒绝并提示：
`The action was rejected due to unacceptable risk.`

拒绝原因里包含：
`litellm.BadRequestError: DeepseekException - {"error":{"message":"This response_format type is unavailable now", ...}}`

## 推断
自动审批（auto-review）环节调用模型做评审时，模型接口（Deepseek）不支持该 response_format，评审失败，系统按"不可接受风险"默认拒绝。

## 期望行为
评审环节失败时应放行，或提示"自动评审暂不可用"，而不是直接拒绝无害命令。

## 其他
另一个会话/环境也出现过相同现象（疑似可复现）。

### 环境 / Environment

- Cindy 版本或 commit / version or commit: 41.2.0
- 平台与版本 / platform & OS version: Windows 11 Pro 10.0.26200 Build 26200
- 安装方式 / install method: 官方安装包（C:\Program Files\Cindy，桌面客户端）
- 沙箱模式 / sandbox mode: workspace-write（文件系统受限 + 网络受限）
- 出现时间 / occurred: 2026-08-02 (Asia/Singapore)

### 复现步骤 / Steps to reproduce

1. 在 workspace-write 沙箱模式下，让 Codex 执行一条无害命令，例如 `Get-ChildItem -Force`（仅列目录）。
2. 该命令未执行，被自动审批（auto-review）拦截，返回：
"The action was rejected due to unacceptable risk."
3. 拒绝原因内嵌报错：
litellm.BadRequestError: DeepseekException - {"error":{"message":"This response_format type is unavailable now","type":"invalid_request_error",...}}
4. 另一个会话/环境中也出现相同现象（疑似可复现）。

期望 / Expected:
自动审批的评审模型调用失败时，应放行无害命令或明确提示"自动评审暂不可用"，而不是直接按高风险拒绝。

### 日志与截图 / Logs & screenshots

Get-ChildItem -Force | Select-Object Name, Length, LastWriteTime
cwd: C:\Users\user\AppData\Roaming\CindyGlobal\owners\1106d4699367d440180b\dialogues\2026-08-02\4a5296b3-2b31-4f99-bdee-8c7c2bd196ac
This action was rejected due to unacceptable risk.
Reason: Automatic approval review failed: {"error":{"message":"litellm.BadRequestError: DeepseekException - {\"error\":{\"message\":\"This response_format type is unavailable now\",\"type\":\"invalid_request_error\",\"param\":null,\"code\":\"invalid_request_error\"}}","type":null,"param":null,"code":"400"}}
The agent must not attempt to achieve the same outcome via workaround, indirect execution, or policy circumvention. Proceed only with a materially safer alternative, or if the user explicitly approves the action after being informed of the risk. Otherwise, stop and request user input.

Start by tracing the auto-review flow and its LiteLLM/Deepseek response_format request, using the reproduction with Get-ChildItem -Force as the failure case. Determine how review-call errors are handled, then verify that an unavailable review model produces the intended user-facing outcome instead of an unconditional unacceptable-risk rejection.

Written by the indexing model from the issue text.