Skip to content

[Bug]: CugaLite stops mid-execution with verbal intent instead of writing code on some models #110

@sami-marreed

Description

@sami-marreed

What happened

CugaLite sometimes halts mid-execution on certain models (e.g. gpt-4.1) and produces a natural-language announcement such as:

"I am going to perform X."

without writing any Python code. The agent does not proceed autonomously — it waits for the user to explicitly instruct it to continue, at which point it resumes normally and completes the task.

How to reproduce

  1. Run CugaLite with a gpt-4.1 (or similarly affected) model backend.
  2. Trigger a multi-step task that requires tool/code execution.
  3. Observe that at some step the agent outputs a verbal intent message instead of generating a code block.
  4. Instruct the agent to continue — it will proceed correctly.

Environment

  • Model: gpt-4.1 (and potentially other models)
  • Component: CugaLite agent loop

Logs, screenshots, or config

No code block generated at the stopping step — the agent outputs plain text intent instead of a python fenced block. Execution resumes normally after user prompt.

Metadata

Metadata

Labels

bugSomething isn't workingcomponent: agentCore agent loop, DynamicAgentGraph, LLM node, tool execution, CugaLitecomponent: evaluationBenchmarks and evaluationneeds-triageNewly filed, not yet reviewed

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions