Feature: Conditionally Include attempt_completion Tool Based on Task Context

[VooDisss]

Problem Description

Currently, the attempt_completion tool is always available and its automatic invocation can prematurely end a user's session, especially in a standard, non-orchestrated task. This forces a workflow where the user's task is considered "complete" even if they intended to continue.

However, for subtasks created by the orchestrator mode, attempt_completion is a necessary mechanism to signal completion and return results to the parent task. A global toggle would break this critical workflow.

This proposal suggests a more intelligent, context-aware approach to making the attempt_completion tool available.

Proposed Solution

Instead of a simple on/off switch, the attempt_completion tool's availability in the system prompt should be conditional based on the task's context:

1. Standard Tasks (User-initiated):

   • By default, the attempt_completion tool should NOT be included in the system prompt.
   • This prevents the LLM from prematurely ending the task and allows for a more continuous, conversational workflow until the user is truly finished.

2. Orchestrated Subtasks (Created via new_task by orchestrator mode):

   • The attempt_completion tool MUST be included in the system prompt for these subtasks.
   • This is essential for the subtask to signal that it has finished its work and to pass the results back to the parent orchestrator task, allowing the main workflow to proceed.

A user setting could still be provided in the AutoApproveMenu, but it would only control the behavior for standard tasks. For example, a setting like "Allow Assistant to End Session" could be off by default but could be enabled by users who prefer the current behavior. Orchestrated tasks would ignore this setting.

Implementation Details

This implementation requires the system to be aware of the context in which a task is running.

1. Task Context Awareness: The logic that creates a new_task from the orchestrator mode needs to flag the subtask as "orchestrated". The Task object in src/core/task/Task.ts should be updated to track this state.
2. Conditional Prompt Generation: The system prompt generation logic, likely in files like src/core/prompts/tools/index.ts, must be modified. Before adding the attempt_completion tool to the prompt, it should check:
   • Is this an orchestrated subtask? If yes, include the tool.
   • If not, is the "Allow Assistant to End Session" user setting enabled? If yes, include the tool.
   • Otherwise, exclude the tool.
3. UI (Optional but Recommended): A new setting toggle can be added to webview-ui/src/components/chat/AutoApproveMenu.tsx to control the behavior for standard tasks.

Impact

This change makes the assistant's behavior much more intuitive. It preserves the essential attempt_completion mechanism for automated workflows (orchestrator) while providing a more flexible and user-friendly experience for standard, conversational tasks.

[roomote]

I'm exploring a potential fix for this issue and will open a PR to share progress. The proposal to conditionally include the attempt_completion tool based on task context makes sense - I'll work on implementing the context-aware approach as described.

[roomote]

I've successfully implemented the solution and created PR #7532: #7532

Implementation Summary

The PR makes the attempt_completion tool conditionally available based on task context:

Key Changes:

• Orchestrated subtasks: Always have attempt_completion available (preserves existing orchestrator functionality)
• Standard user tasks: attempt_completion is excluded by default (addresses the issue)
• User control: New allowAttemptCompletion setting allows users to enable it if desired

Technical Details:

• Added isOrchestrated flag to track subtask context
• Modified system prompt generation to conditionally include the tool
• Added UI controls in the Auto-Approve settings menu
• Comprehensive test coverage included

Note:

There's currently a failing translation check in the CI. This appears to be related to missing translations for the new allowAttemptCompletion setting in non-English locales. The core functionality is complete and tested.

This implementation provides the requested behavior while maintaining backward compatibility for orchestrated workflows.

[VooDisss]

It seems that such functionality already exists in RooCode 3.26.1. But maybe Roomote figured something better or maybe there are some new insights to be gained... At the time of writing, I just did not notice that functionality of "Subtasks" toggle button in the autocompletion menu.

[daniel-lxs]

Thank you for raising this. The problem is that without the attempt_completion tool there’s no reliable way for Roo Code to determine when a task is complete. Multiple mechanisms depend on it to signal the final state.

This can already be addressed with custom instructions, or by enabling the setting that prevents Roo Code from completing a task until all TODO items are resolved (see screenshot below). That avoids premature completion without introducing breaking changes and workflows which critically rely on attempt_completion.

If the default feels unintuitive, we can look at improving the visibility of that setting, but conditionally removing the tool would introduce regressions in orchestrated workflows.