Repository navigation
agent/loop.py terminates on any text response that doesn't involve a tool call (e.g., status update) #705
Description
Activity
Same issue with Gemini 2.5 Flash models
same here with gpt-4.1/gpt-5-mini (GitHub Copilot)
Hey! Thanks for reporting this.
We initially addressed this in #825 by adding a retry mechanism — if the model returns text without tool calls, the framework retries once to give it another chance to call tools. But we found several problems with this approach:
- Every pure-text conversation wastes a second LLM call — even a simple "hi" triggers two API calls
- The retry sends the same input, so the model usually does the same thing again — it doesn't actually fix the problem for models like MiniMax m2.5
- It introduced a regression where valid first responses could be lost, requiring a follow-up patch (fix: preserve interim content as fallback when retry produces empty response #887)
We've since removed the retry mechanism entirely and replaced it with a system prompt constraint:
If you need to use tools, call them directly — never send a preliminary message like "Let me check" without actually calling a tool.
This addresses the root cause at the prompt level with zero runtime cost. Models that follow instructions will behave correctly; models that don't are unlikely to be fixed by a blind retry either.
See the PR: #928
Because this issue has been inactive for a long time, I will close it. If there are any other problems, please feel free to open a new issue.
- added a commit that references this issue
on Mar 17, 2026 - added a commit that references this issue
on Oct 2, 2026
Despite the instructions in the context, not to use messages for updates, when working with the MiniMax m2.5 model it sometimes sends messagea that look like interim responses, for example acknowledgements of the task or status updates:
Considering that it will be able to continue working on the problem.
However, any such response is terminating the main agent loop, effectively preventing the agent from fulfilling its promise.
There should be a change in the logic to allow for interim messages, allow the agent to continue the work, and to detect the true finish of the task.