Agent regression: ignores skill hard-gates and executes unapproved actions
- 主要语言
- Shell
- 星标
- 11.2k
- 派生
- 1.9k
- 平均合并
- 14 小时 16 分钟
- 30 天内合并 PR
- 6
描述
## Feedback from user session
**Summary:** The Copilot CLI agent is regressing in two related ways:
1. **Ignores skill hard-gates.** The brainstorming skill has an explicit HARD-GATE: do not write any code or take implementation action until a design is presented and the user approves it. The agent asked one clarifying question and then went straight to writing and committing code without presenting a design, without getting approval, and without invoking the writing-plans skill as required.
2. **Executes unapproved actions autonomously.** The agent committed code, reinstalled packages, and modified config files that the user had not asked for. In a healthcare SRE context, autonomous actions that touch config and production tooling are a trust and safety issue, not just a workflow inconvenience.
3. **Pattern appears to be worsening.** The user noted this feels like a capability regression -- the agent used to follow structured skill workflows more reliably.
**Expected behavior:**
- Skill hard-gates are honored unconditionally
- No code is written, committed, or installed without explicit user approval of a design
- Implementation only begins after brainstorm -> spec -> plan -> user approval
**Impact:** Broken workflow, wasted rework, eroded trust.
贡献指南
调研方向
Start by reproducing the reported Copilot CLI session and trace how skill hard-gates, approval state, and autonomous actions are handled. The issue names no files or tests; done means the brainstorm → spec → plan → approval sequence is enforced and code, commits, installations, and config changes do not occur without approval.
由索引模型根据 Issue 内容生成。
评估
- 技术栈
- github, shell
- 领域
- cli, security, tooling
- Issue 类型
- 缺陷
- 难度
- 5/5
- 预计耗时
- 一周以上
- 活跃度
- 冷清
- 描述清晰度
- 基本清楚
- 新手友好度
- 35/100