[ChatHelper] : a confused LLM could loop forever

[e-marchand]

• https://techcommunity.microsoft.com/blog/azuredevcommunityblog/understanding-agentic-function-calling-with-multi-modal-data-access/4504151

Every loop needs a safety valve. Without a maximum iteration count, a confused LLM could loop forever (calling tools that return errors, retrying, etc.)

agent loop

• https://www.inngest.com/docs/ai-patterns/agent-tool-loops
• https://www.zeroentropy.dev/concepts/agent-loop/

[e-marchand]

AI analyse

A loop guard — this is the real bug, not just a nice-to-have. _continueConversationAfterToolCalls re-enters _manageResponse, which can call _handleToolCalls again, with no step counter. A model that keeps emitting tool_calls (or loops on the same tool) recurses unbounded — growing the 4D call stack and burning tokens with no ceiling. You need a maxIterations (a.k.a. max steps) budget threaded through the loop, incremented per round, that forces a stop.

[e-marchand]

Possible PR https://github.com/4d/4D-AIKit/compare/feature/maxIterations?expand=1