Agent Continues Running After 100% Project Completion (Token Burn)

[amoshalava]

Thank you for this tool - it's incredibly useful for structured development. Just want to make sure I'm getting the most efficient use out of it!

After a project reaches 100% completion with all features passing, the agent continues to run indefinitely - performing regression tests, attempting builds, exploring the file system, and re-verifying already-completed features. This results in significant unnecessary token consumption.
Environment

Host: Proxmox VE
Container OS: Rocky Linux 9 (LXC container)
Node.js Version: v20.19.5
Claude Model: claude-sonnet-4-20250514

Observed Behavior

Project reaches 100% completion (55/55 features passing in my case)
Agent displays: "The codebase is in excellent shape. All core functionality appears to be implemented correctly..."
Agent then proceeds to:

Run regression tests on randomly selected features
Attempt to build the project (sometimes with incorrect paths)
Execute exploratory commands (pwd && ls -la)
Loop back to re-check features

Agent shows [Done] but then initiates new tasks
This cycle continues until manually paused

Expected Behavior

Once all features are marked Complete and the agent confirms the codebase is in good shape, the agent should:

• Stop automatically, or
• Enter an idle/waiting state, or
• Prompt the user before performing additional verification cycles

Console Output Example
19:28:41 Status: ✅ PASSING** - History is automatically limited to 20 entries after each insertion
19:28:41 ## Regression Test Summary
19:28:41 All three randomly selected features are verified and working correctly:
19:28:41 ✅ Feature #46: Case-insensitive wake word matching - PASSING
19:28:41 ✅ Feature #8: Notification channel created - PASSING
19:28:41 ✅ Feature #21: History limited to 20 entries - PASSING
19:28:42 The codebase is in excellent shape. All core functionality appears to be implemented correctly. Let me check if there are any outstanding issues or improvements needed:
19:28:42 [Tool: Bash]
Input: {'command': 'cd android-app && ./gradlew assembleDebug --no-daemon 2>&1 | tail -30', 'description': 'Build the app to verify compilation', 'timeout': 120000}
19:28:43 [Error] Exit code 1
19:28:43 /bin/bash: line 1: cd: android-app: No such file or directory
19:28:45 [Tool: Bash]
Input: {'command': 'pwd && ls -la', 'description': 'Check current directory'}
19:28:45 [Done]

The agent continues exploring despite the error being unrelated to feature completion.

Questions for Maintainer

Is there a configuration setting to enable automatic stopping when all features pass? I may have missed this in setup.
Is this a known issue with certain environments? I'm running on Rocky Linux 9 in a Proxmox LXC container - wondering if there are dependencies or configurations specific to this setup that could cause the loop behavior.
Would a different OS be recommended? Should I try:

Ubuntu LTS instead of Rocky Linux?
A full VM instead of an LXC container?
Different Node.js version?

Are there any additional packages or dependencies that should be installed to support proper completion detection?
Is there a way to configure a "completion hook" or callback that triggers when all features pass, to prevent continued execution?

Workaround
Currently using the Stop button manually when I see 100% completion, but this requires monitoring the agent and doesn't prevent token usage before I notice.
Suggestions
Would it be possible to add:

A configurable "auto-stop on completion" flag
A maximum iteration limit after 100% is reached
A confirmation prompt before running post-completion verification cycles
Token usage warnings/limits per session

Additional Context
This has happened across multiple projects. Happy to provide additional logs, configuration files, or test on a fresh instance if that would help diagnose the issue.