github / github/copilot-cli

Bug: Multiple unsolicited rubber-duck background agents spawned from a single task tool call

已關閉
#2,569 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視
area:agents
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

## Description

During a multi-step implementation session, **5 unexpected background agents** appeared with system notifications, despite only **1 rubber-duck agent being explicitly launched** via the `task` tool.

All 5 unexpected agents shared the same description ("Critiquing fix plan") and the same session owner ID as the one I launched, suggesting they were auto-spawned by the system rather than by me.

## Environment

- **Copilot CLI version:** 1.0.21
- **OS:** Ubuntu 24.04 (Linux 6.17.0-20-generic x86\_64)
- **Node.js:** v25.9.0
- **Mode:** Interactive (standard, not autopilot)
- **Session ID:** 10fe9180-e557-4c37-8efa-f294510de030

## What I Did

1. Received a `[[PLAN]]` user request to investigate and fix issues in a codebase
2. Investigated the codebase, identified root causes
3. Explicitly called the `task` tool **once** with these parameters:
```json
{
"agent_type": "rubber-duck",
"name": "rubber-duck-review",
"description": "Review fix plan for step_03",
"mode": "background",
"prompt": "..."
}
```
4. While waiting for `rubber-duck-review` to complete, system notifications arrived for **5 additional agents I never launched**

## Unexpected Agents (never explicitly launched)

```
💤 step3-fix-review rubber-duck "Critiquing fix plan" (824s)
💤 script-refs-review rubber-duck "Critiquing fix plan" (803s)
💤 step3-fix-critique rubber-duck "Critiquing fix plan" (747s)
💤 script-fix-critic rubber-duck "Critiquing fix plan" (695s)
💤 script-fix-critic-1 rubber-duck "Critiquing fix plan" (629s) [model: claude-haiku-4.5]
```

The agent I **did** launch:
```
💤 rubber-duck-review rubber-duck "Review fix plan for step_03" (902s)
```

All 6 were listed under the same session owner (`10fe9180-e557-4c37-8efa-f294510de030`), confirming they were spawned within the same session.

## Notable Details

- The unexpected agents had **different names** (step3-fix-review, script-refs-review, etc.) and slightly different descriptions despite all saying "Critiquing fix plan"
- One (`script-fix-critic-1`) used a **different model** (`claude-haiku-4.5`) — the `-1` suffix suggests a naming collision was detected and auto-incremented
- The agents arrived over a ~5 minute window, staggered, not all at once
- All were idle (completed) by the time the user asked "What is pending?"
- There is **no `/fleet` mode** active in this session
- The session was in standard interactive mode (not autopilot/experimental)

## Expected Behavior

Only the **one explicitly requested** rubber-duck agent (`rubber-duck-review`) should have been spawned. No other agents should appear unless the user or assistant explicitly calls the `task` tool.

## Actual Behavior

5 additional rubber-duck agents were spawned without any `task` tool call from the assistant. They ran in parallel, all critiquing the same plan, consuming quota unnecessarily and cluttering the `/tasks` view.

## Impact

- **Wasted quota:** ~5× the expected premium request consumption for agent spawning
- **Confusion:** The user noticed the anomaly and flagged it as a bug
- **Noise:** 5 redundant idle agents in `/tasks` list at session end
- **Potential loop risk:** If agents can self-spawn, a runaway loop scenario is possible in longer sessions

## Possible Hypotheses

1. A retry/fault-tolerance mechanism re-spawned the agent each time it failed to return a result within a timeout window
2. A skill or background automation triggered agent launches independent of the assistant's `task` tool calls
3. The `task` tool call was internally duplicated/retried at the orchestration layer
4. A race condition between agent completion notifications and internal state caused repeated spawning

貢獻指南

開啟貢獻指南

研究方向

Begin by reproducing the single background task call in Copilot CLI 1.0.21 on the reported Ubuntu and Node.js environment, then monitor the /tasks view and session notifications. Compare the task records, timing, names, models, and session owner against the six agents described. Done means one explicit call produces only one agent without duplicate quota use.

由索引模型根據 Issue 內容生成。

評估

技術堆疊
node.js, shell
領域
cli, devtools
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
48/100

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。