One setting boosted OpenClaw's score 44%: feed reasoning content back to the model

TheZachMueller · x · 2026-11-20

joelniklaus found that OpenClaw drops the model's earlier thinking when talking to custom OpenAI-compatible endpoints — no error, no warning — while reasoning models expect their prior thoughts in the conversation history. Flipping requiresReasoningContentOnAssistantMessages: true raised the return rate on GLM-5.3-Flash and DeepSeek-V4.1-Flash from 3-7% to nearly 100%. With its notes, OpenClaw needs about a third fewer model calls and hits time limits far less often; he's asked steipete and vincentkoc to make it the default.

Original post →

More from coding & agent

coding & agent channel →