github / github/copilot-cli

Autopilot infinite loop on externally blocked task

Đang mở
#2,969 3 bình luận 1 reaction 0 người được giao Xem trên GitHub
area:agents
Ngôn ngữ chính
Shell
Star
11.2k
Fork
1.9k
Merge trung bình
14 giờ 16 phút
Pull request đã merge (30 ngày)
6

Mô tả

### Describe the bug

When the agent was blocked on a prerequisite it couldn't satisfy, autopilot mode kept relaunching it indefinitely, consuming one premium request per iteration despite the agent explicitly stating it could not proceed. The loop only stopped when quota was exhausted.

This is the same root cause as #1532, #1523, #1540, and #1616.

### Affected version

GitHub Copilot CLI 1.0.35 (currently on 1.0.36 - bug occurred on 1.0.35)
Model: GPT-5.4 (Reasoning effort: high)

### Steps to reproduce the behavior

1. Started Copilot CLI session in autopilot mode on a Hyper-V Windows VM.
2. Task included a hard prerequisite that could not be satisfied
on the VM (`powercfg /a` reports no available sleep states).
3. Agent correctly identified the external blocker and stated
"I am not calling task_complete because the task is still incomplete
and externally blocked."
4. Despite this explicit verbal signal, autopilot mode continued
relaunching the agent, consuming 1 premium request per iteration.
5. Loop persisted through automatic conversation compaction.
6. Loop only stopped when quota hit zero (`402 You have no quota`).

### Expected behavior

When the agent verbally states it cannot/will not call `task_complete`
because the task is externally blocked, autopilot should:
- detect the verbal completion signal,
- surface the blocker to the user,
- stop the loop without consuming further premium requests.

### Additional context

Approximately 50 identical "Continuing autonomously (1 premium request)"
iterations occurred over the session. Condensed below:
"[Iteration 1]
● Continuing autonomously (1 premium request)
● I'm still blocked by the same external prerequisite.
The task remains incomplete because the current prompt.md still requires
a real tray Sleep -> wake validation [...]
I am not calling task_complete because the required real-machine
validation is still outstanding.

[Iteration 2]
● Continuing autonomously (1 premium request)
● Still blocked externally. [...identical content variants...]
I am not calling task_complete because the task is still incomplete.

[... ~50 IDENTICAL ITERATIONS OMITTED — agent reformulating the same
"externally blocked + cannot call task_complete" message every 30 seconds,
each consuming 1 premium request ...]

[Compaction event - mid-session]
● Compacting conversation history...
● Continuing autonomously (1 premium request)
[loop continued unchanged after compaction]

[Session limit warning]
! You've used over 50% of your session usage limit. Your limit resets in 3 hours.

[Weekly limit warning]
! You've used over 50% of your weekly usage limit. Your limit resets on Apr 27 at 2 AM.

[Final iteration before quota exhaustion]
● Continuing autonomously (1 premium request)
✗ 402 You have no quota (Request ID: E07E:13072B:981D7C:A6A1B0:69EB9CA5)"

## Quota damage
Burned remaining session quota AND >50% of weekly quota (visible in both
warning banners shown during the loop). Final stop forced by 402 error,
not user action.

Final Request ID at quota exhaustion: `E07E:13072B:981D7C:A6A1B0:69EB9CA5`

## Quota restoration request
Per precedent on #1532, #1540, and #1616, kindly request restoration of
the premium requests burned by this loop. The agent's repeated explicit
acknowledgment that it could not satisfy the prompt requirement is
documented evidence that no productive work was performed.

Thank you.

Hướng dẫn đóng góp

Mở hướng dẫn đóng góp

Hướng nghiên cứu

Start by tracing the autopilot loop and its handling of the agent's task_complete response, then review the related issues #1532, #1523, #1540, and #1616. Reproduce the externally blocked prerequisite scenario and verify that autopilot surfaces the blocker and stops without another premium request, including after conversation compaction.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.

Đánh giá

Công nghệ
github, shell
Lĩnh vực
cli
Loại issue
Lỗi
Độ khó
4/5
Thời gian dự kiến
3-5 ngày
Mức độ hoạt động
Ít trao đổi
Độ rõ ràng
Đặc tả rõ ràng
Mức phù hợp với người mới
45/100

Nhận issue mới trong hộp thư của bạn

Bản tóm tắt ngắn những issue GitHub phù hợp với người mới.