Claude Sonnet 5 delegated to a lesser agent to perform a code review
- 主要語言
- Shell
- 星號
- 11.2k
- 分支
- 1.9k
- 平均合併
- 14 小時 16 分鐘
- 30 天內合併 PR
- 6
描述
### Describe the bug
I asked Claude Sonnet 5 to perform a code review. I specifically chose Claude Sonnet 5 for deep reasoning. The agent proceeded to review the structure of my project and then delegate to a lesser agent:
"Good, all directories exist. I'll launch a general-purpose agent to perform this code review, since it involves reading through before/after/project directories, comparing changes, and producing a structured markdown report — genuinely multi-step work."
"General-purpose(gpt-5.4) Perform code review for issue 44206"
The code review produced was of little use. I corrected the agent and requested Sonnet 5 to perform the review itself:
"I would like to start this code review over. I see that you delegated to gpt-5.4 and I do not want a weaker model doing the review. Please execute the review prompt in file 44206-review-prompt2.txt."
Sonnet replied "I'll perform this review myself directly (no sub-agent delegation this time). Let me examine the before/after directories to identify actual changed files."
Sonnet completed the review the second time and the quality was as expected.
### Affected version
GitHub Copilot CLI 1.0.75
### Steps to reproduce the behavior
_No response_
### Expected behavior
I expected that when selecting the agent that it perform the work itself.
### Additional context
_No response_
貢獻指南
研究方向
使用 review 情境和 44206-review-prompt2.txt prompt,在 GitHub Copilot CLI 1.0.75 中重現該行為,並比較 before、after 和 project 目錄。追蹤選取的 agent 在何處決定進行委派,然後確認 Sonnet 會自行執行要求的 review,而不會意外轉交給能力較低的 agent。
由索引模型根據 Issue 內容生成。
評估
- 技術堆疊
- github, shell
- 領域
- ai, cli
- Issue 類型
- 缺陷
- 難度
- 4/5
- 預估耗時
- 3-5 天
- 活躍度
- 冷清
- 描述清晰度
- 基本清楚
- 新手友好度
- 42/100