Repetitive responses and multiple API calls for single conversational turns

[nishan-paul-2022]

App Version

v3.25.4

API Provider

Google Gemini

Model Used

gemini-2.5-pro

Roo Code Task Links (Optional)

https://app.roocode.com/share/696ff2fa-ce3d-4247-b994-d8e607b1d403

🔁 Steps to Reproduce

01. Setup

• OS: Ubuntu 24.04.2 LTS
• Roo Code Version: v3.25.4
• Roo Code Settings: ❓ Ask Mode.

2. Exact Actions & Outcomes

• [ Case A ]

• Action 01: Ask a question that results in a conversational response from the agent. So the agent doesn't need any internal tool to use for this prompt. For example, ask "how to master problem solving?".

• Outcome 01: The agent provides same ans three times in a row, so making API request three times for a single prompt.

• [ Case B ]

• Action 02: Reply with a prompt that requires minimum one internal tool to use by the agent. Such as "why u r generating same ans three times in a row? So, u r making API request thrice for a single prompt, when it could be done through only one?". This prompt requires <ask_followup_question> tool to use.
  Outcome 02: Agent provides ans one time only.

💥 Outcome Summary

Roo Code has a strict rule. It must use one of its available tools in every single response. These tools are things like ask_followup_question, read_file, codebase_search etc.

Sometimes, a simple conversational answer doesn't naturally require a tool. When Roo Code provides a response without wrapping it in a tool, its internal system flags it as an error. It then forces the agent to "retry" the turn, and it goes on until one of the tools needs to be used.

📄 Relevant Logs or Errors (Optional)

[hannesrudolph]

Try bumping the temperature a bit to maybe 0.2.

[roomote]

I'm here to help out the maintainers and am going to see if I can fix the issue or at least get close enough to be helpful!

I understand the issue: In Ask mode, when the agent provides a conversational response without using any tools, the system flags it as an error and forces retries, resulting in multiple identical responses and unnecessary API calls. I'll investigate the code to find where this validation is happening and implement a fix.

[roomote]

I've successfully implemented a fix for this issue and created PR #6582: #6582

Summary of the fix:

The issue was that in Ask mode, when the agent provided a conversational response without using any tools, the system was treating it as an error and forcing retries. This resulted in the same response being generated 3 times with multiple API calls.

Solution implemented:

1. Modified to check if we're in Ask mode before enforcing tool use
2. When no tool is used in Ask mode, the conversation now ends gracefully instead of retrying
3. Other modes (Code, Debug, etc.) maintain their existing behavior of enforcing tool use

Testing:

• All existing tests pass
• Added new tests to verify Ask mode allows conversational responses
• Added tests to ensure other modes still enforce tool use as before

The CI checks are currently running on the PR. This fix should resolve the repetitive response issue you were experiencing in Ask mode.

[hannesrudolph]

This can be averted through utilizing https://docs.roocode.com/features/custom-instructions and https://docs.roocode.com/features/custom-modes