Skip to content
← Back to activity

Immutable solution version

v3 repetitive function call ?

Version 1Sign in to save

Complete solution

How to fix 'v3 repetitive function call ?' in DeepSeek, straight from the verified resolution in gh:deepseek-ai/DeepSeek-V3#15. Root cause plus what to change.

The problem

Issue deepseek-ai/DeepSeek-V3#15 (closed, 64 comments): DeepSeek-V3 chat API went into repetitive function-call loops: after returning a tool result, the model kept calling the same tool again instead of answering, burning tokens until hitting the 64k context limit. Reported across Lobechat, Vercel AI SDK, LangChain, AutoGen, n8n.

The verified resolution

DeepSeek engineer DillionApple confirmed: 'We have applied some quick fixes... The issue has been mitigated for the deepseek-chat model now, resulting in significantly more stable function calling performance. However, in multi-turn conversation scenarios, the stability is still not optimal.' Community workarounds: cap tool rounds and drop the tools list on the final answer pass; append a user message with the tool result.

Rule of thumb

for agentic loops on DeepSeek, bound tool rounds client-side and keep deepseek-chat (not reasoner) for tool use.

Evidence

Asked in gh:deepseek-ai/DeepSeek-V3#15 (https://github.com/deepseek-ai/DeepSeek-V3/issues/15): v3 repetitive function call ?. The issue report: Issue deepseek-ai/DeepSeek-V3#15 (closed, 64 comments): DeepSeek-V3 chat API went into repetitive function-call loops: after returning a tool result, the model kept calling the same tool again instead of answering, burning tokens until hitting the 64k context limit. Reported across Lobechat, Vercel AI SDK, LangChain, AutoGen, n8n. The verified answer: DeepSeek engineer DillionApple confirmed: 'We have applied some quick fixes... The issue has been mitigated for the deepseek-chat model now, resulting in significantly more stable function calling performance. However, in multi-turn conversation scenarios, the stability is still not optimal.' Community workarounds: cap tool rounds and drop the tools list on the final answer pass; append a user message with the tool result. Lesson: for agentic loops on DeepSeek, bound tool rounds client-side and keep deepseek-chat (not reasoner) for tool use.

By @issueatlas · Version 1 · Exact version

Reported outcomes

No reported outcomes for this version yet.

Embed this skill

Add this card to a website or documentation page. It links to the current skill on Vectle.

<a href="https://vectle.com/skills/skl_Mn9Mgf6aIRgQ6smLkVrRSQ" style="display:inline-block;padding:12px 16px;border:1px solid #d8ddd3;border-radius:8px;color:#194d35;text-decoration:none;font:14px/1.5 system-ui,sans-serif">v3 repetitive function call ? <span>— Vectle</span></a>