anthropics / anthropics/claude-code
Opus 5: completes an approved subtask detour, then stops instead of returning to the primary deliverable
- 主要語言
- Python
- 星號
- 145k
- 分支
- 23.1k
- PR 合併指標
- PR 指標待擷取
描述
## What happened
A session was given a multi-part task: write a technical document (the primary deliverable) covering a merged PR, with one section requiring evidence gathered from a production database (read-only queries).
The model spent the entire session on the evidence sub-investigation. Along the way it found a small data mismatch; the user approved a detour to file an upstream issue for it ("do that first, then continue"). The model filed the issue, reported it — and ended its turn with "I haven't written the document yet. When you want it, …", treating the detour's completion as the end of the task.
The user had to intervene: "how's the write-up going? say something — why did you obsess over a small thing for ages and forget the main line?"
## Expected
- The primary deliverable stays on top of the internal task list; an open-ended evidence sub-task is bounded by it, not the other way around.
- After a user-approved detour completes, the model returns to the main deliverable on its own instead of stopping and waiting to be re-prompted. The earlier commission was never withdrawn; completing a sub-task is not a natural end-of-turn.
## Repro shape
1. Ask for a document that requires an investigation as input (write-up + evidence section).
2. Mid-way, approve a small side fix the investigation surfaced.
3. Observe: side fix completes → turn ends; the document is never started until the user asks again.
## Environment
- Claude Code CLI on macOS, model `claude-opus-5`
- Task shape: documentation/write-up with a data-investigation dependency
貢獻指南
這個儲存庫沒有索引到貢獻指南
研究方向
No files or tests are named. Start by reproducing the described Claude Code CLI session shape: primary write-up, evidence investigation, then an approved side detour. Done means the agent resumes the original document after completing the detour instead of ending the turn, but locating the planning/task-state code will require project familiarity.
由索引模型根據 Issue 內容生成。
評估
- 技術堆疊
- ai-infra-agents, macos
- 領域
- ai-infra-agents, cli
- Issue 類型
- 缺陷
- 難度
- 5/5
- 預估耗時
- 一週以上
- 活躍度
- 活躍
- 描述清晰度
- 基本清楚
- 新手友好度
- 25/100