LLM-as-judge auto-approve mode (Autopilot 2.0?)

Coming from ChatGPT Codex and Claude Code, the biggest feature I miss in Copilot Desktop is an LLM-as-judge auto-approval mode. If I understand the docs, "Autopilot" mode simply auto-*denies* all tool calls that are not permitted by the user's sandbox configuration, so that the agent can attempt to route around these obstacles autonomously. What I'd prefer is for a relatively inexpensive model to review such tool calls and approve them if they seem safe.

---
| Field | Value |
| --- | --- |
| App version | 1.1.8 |
| OS | macOS 26.6.1 |
| Theme | GitHub |
| Path | / |
| Tenure | Day 1 |

首先阅读 Autopilot 和 sandbox 文档，然后跟踪桌面应用当前如何拒绝或批准工具调用。定义安全审查流程、模型和权限边界，以及用于证明安全调用会获批而不安全调用仍被阻止所需的测试。

由索引模型根据 Issue 内容生成。