gemini-cli - 💡(How to fix) Fix [Bug] Agent gets stuck in infinite Tool-use Looping (echo commands) when resolving strict negative output constraints [1 participants]

What happened?

When using a custom skill (gh-address-cr) that implements a strict negative output contract (e.g., "Never output 'done', 'all resolved' or equivalent unless X condition is met"), the Agent correctly meets the condition but becomes "paralyzed" during the state transition from tool execution back to chat response.

Instead of breaking out of the ReAct loop and generating the final text summary, the agent repeatedly calls run_shell_command with redundant echo commands (e.g., echo "completed", echo "Ready to finish", echo "Done") as "noop for thought transitions". This creates an infinite loop that continues until the CLI's internal safety mechanism ("System: Potential loop detected.") forcefully interrupts the turn.

Steps to Reproduce:

1. Execute a skill that contains strict negative output constraints (e.g., Never output "done" without final_gate.sh pass).
2. Have the agent successfully complete all prerequisites and pass the validation gate.
3. Observe the agent's internal tool calls: it will successfully run the final validation tool but then fail to output the final text response, opting instead to repeatedly invoke run_shell_command with echo commands.
4. Wait until the system intervention triggers the Potential loop detected error.

Root Cause Analysis:

The ReAct model experiences "Reasoning Paralysis" due to the high penalty weight of the negative constraint in the prompt. Even when the condition is met, the model assigns a lower probability to directly outputting the completion text (fearing it might violate the constraint) and instead chooses the "safe" action of invoking another tool (echo), causing a loop.

[ACTION REQUIRED] 📎 PLEASE ATTACH THE EXPORTED CHAT HISTORY JSON FILE TO THIS ISSUE IF YOU FEEL COMFORTABLE SHARING IT.

Use this skill: https://github.com/RbBtSn0w/gh-address-cr-skill

What did you expect to happen?

Once the agent has successfully executed the necessary tools and met the criteria required to bypass the negative constraint, it should immediately break out of the tool-use loop and generate the final textual response (the "Completion Output Contract") to the user without generating redundant echo tool calls.

Client information

• CLI Version: 0.36.0
• Git Commit: 8b1e649c2
• Session ID: 3ca86e45-6fcf-487b-b7a2-43be60aac661
• Operating System: darwin v25.8.2
• Sandbox Environment: no sandbox
• Model Version: gemini-3.1-pro-preview
• Auth Type: oauth-personal
• Memory Usage: 198.2 MB
• Terminal Name: tmux 3.6a
• Terminal Background: #000000
• Kitty Keyboard Protocol: Unsupported

Login information

No response

Anything else we need to know?

This issue highlights a common failure mode in LLM Agent architectures dealing with hard negative constraints.

Potential Mitigation for the framework:

Consider adding logic to the ReAct parser that detects sequential, semantically identical, or zero-impact tool calls (like sequential echo commands that don't capture stdout) and prompts the model to break out of the loop and generate a text response.

Workaround for Skill Authors:

Adding explicit "State Transition Permissions" to the skill prompt (e.g., "Once the gate passes, you MUST immediately stop calling tools and output the final summary directly to the user in normal text") helps the model overcome this probabilistic hesitation.