Fix DeepSeek tool-calling issues

The three failure modes we actually hit in runs — and the fix for each. Based on real AgentRunDB evidence.

Backed by 4 AgentRunDB runs · Success rate 75%

Goal: Diagnose and fix failing tool calls with DeepSeek.

agent
claude-code
model
deepseek
Last verified
2026-08-16

Failure mode 1 — malformed JSON arguments

Seen with: deepseek-reasoner. Fix: switch to deepseek-chat.

Failure mode 2 — tool not called at all

Cause: base URL typo or trailing slash. Fix: use exactly https://api.deepseek.com.

Failure mode 3 — hangs on parallel tool calls

Fix: reduce parallelism (max 1–2 concurrent tool calls) in agent settings.

Related: claude-code × deepseek Compatibility