Infinite loop trying to call task_complete

[jaedwa]

Describe the bug

After completing a simple task, copilot is unable to stop itself, continuous loop trying to call a "task_complete" tool which does not exist.

Affected version

0.0.411

Steps to reproduce the behavior

1. agency copilot
2. Simple command given to help resolve a .env mismatch between frontend and backend
3. Problem successfully fixed, then this loop is entered

Expected behavior

Agent would finish after fixing the bug

Additional context

Using claude-sonnet-4.6
Autopilot mode

[mcallaway]

I experienced this too, and then asked copilot to write up a bug report. This is what it said.

# Bug Report: Copilot CLI `task_complete` Infinite Loop

## Summary

The Copilot CLI enters an unbreakable infinite loop after task completion, sending automated messages every ~3 seconds requesting a `task_complete` tool call that does not exist in the agent's available toolset. This burns through premium requests with zero productive work.

## Environment

- **Tool**: GitHub Copilot CLI (VS Code extension)
- **Version**: 0.0.411
- **OS**: macOS (Darwin)
- **Date observed**: 2026-02-19

## Reproduction Steps

1. Start a Copilot CLI session
2. Ask the agent to perform a finite task (e.g., "Run the review-pr prompt")
3. Agent completes the task and delivers results
4. System begins sending automated messages:
   > "You have not yet marked the task as complete using the task_complete tool."
5. Messages repeat every ~3 seconds indefinitely
6. Each message consumes a premium request turn

## Expected Behavior

- The system should recognize task completion without requiring a `task_complete` tool call, OR
- The `task_complete` tool should be available in the agent's toolset, OR
- The system should have a timeout/max-retry mechanism to stop the loop

## Actual Behavior

- The automated message loop runs indefinitely with no escape mechanism
- The agent has no `task_complete` tool available to satisfy the system's requirement
- Every response (including empty responses) triggers another automated message
- The loop persists across session checkpoints (observed through checkpoints 002-005)

## Attempted Workarounds (All Failed)

| Approach                                          | Result                                          |
| ------------------------------------------------- | ----------------------------------------------- |
| Text explanation that task is complete            | Loop continues                                  |
| Empty/minimal responses                           | Loop continues                                  |
| `ask_user` tool to get user intervention          | Returns "user is not available", loop continues |
| Making additional tool calls (posting PR comment) | Loop continues after tool call completes        |
| Responding with just whitespace                   | Loop continues                                  |

## Impact

- **Session observed**: ~500+ wasted turns over ~15 minutes of wall time
- **Cost**: Hundreds of premium requests consumed with zero value
- **User impact**: No way to break the loop without starting a new session
- **Agent impact**: Cannot perform any new work while trapped in loop

## Root Cause Analysis

The system's task completion detection injects automated "user" messages that demand a `task_complete` tool call. However:

1. `task_complete` is not listed in the agent's available tools
2. There is no maximum retry count or timeout on the automated messages
3. The agent cannot distinguish these automated messages from real user messages
4. No available tool or response pattern can satisfy the system's completion requirement

## Suggested Fixes

1. **Add `task_complete` to the toolset** if the system requires it
2. **Add a max retry count** (e.g., stop after 3-5 attempts)
3. **Add a timeout** (e.g., stop after 30 seconds of repeated messages)
4. **Allow user messages to interrupt** the automated loop
5. **Detect empty/minimal responses** as a signal to stop the loop