Summary
π― Found It: glm-4.7-flash Outputs reasoning Field β OpenClaw Expects content
Your tests reveal the critical difference:
Model | Streaming Chunk Format
-- | --
β
llama3.2 | "delta":{"role":"assistant","content":"Hello"}
β glm-4.7-flash | "delta":{"role":"assistant","content":"","reasoning":"Let"}
The problem: OpenClaw expects the standard OpenAI format where delta.content contains the response text. When glm-4.7-flash sends chunks with empty content and a reasoning field instead, OpenClaw likely:
- π Waits for
content that never arrives in early chunks β timeout - π§© Fails to parse the unexpected
reasoning field β JSON/schema error - π« Treats empty
content as "no response" β fetch failed
This is why curl works (it just prints raw SSE) but OpenClaw fails (it tries to parse and act on the response).
Steps to reproduce
Use Ollama reasoning models like glm-4.7-flash
OpenClaw never responds
If llama3.2 is used, OpenClaw responds
Expected behavior
OpenClaw responds
Actual behavior
OpenClaw fails to respond
OpenClaw version
2026.2.25
Operating system
Ubuntu 24.04
Install method
No response
Logs, screenshots, and evidence
Impact and severity
Cannot use Ollama reasoning local models
Additional information
No response
Summary
π― Found It:
glm-4.7-flashOutputsreasoningField β OpenClaw Expectscontentdelta.contentcontains the response text. Whenglm-4.7-flashsends chunks with emptycontentand areasoningfield instead, OpenClaw likely:contentthat never arrives in early chunks β timeoutreasoningfield β JSON/schema errorcontentas "no response" βfetch failedcurlworks (it just prints raw SSE) but OpenClaw fails (it tries to parse and act on the response).Steps to reproduce
Use Ollama reasoning models like glm-4.7-flash
OpenClaw never responds
If llama3.2 is used, OpenClaw responds
Expected behavior
OpenClaw responds
Actual behavior
OpenClaw fails to respond
OpenClaw version
2026.2.25
Operating system
Ubuntu 24.04
Install method
No response
Logs, screenshots, and evidence
Impact and severity
Cannot use Ollama reasoning local models
Additional information
No response