Repository navigation
Agent loop runs extra step when client/server clocks are skewed #24183
Description
Description
The agent loop uses message IDs to detect whether a new user prompt arrived mid-response. User message IDs are stamped by the browser; assistant IDs are stamped by the server. When the client and server clocks differ, the browser-stamped user ID can compare as greater than the server-stamped assistant ID even with no new user input, causing the loop to run a spurious extra step. toModelMessagesEffect then produces a message array ending with an assistant message. On llama.cpp with Qwen3 and enable_thinking, this is rejected with: Assistant response prefill is incompatible with enable_thinking. Of course, this may have other effects beyond this, but it's the one that I ran into.
Plugins
none
OpenCode version
1.14.22
Steps to reproduce
- Run the OpenCode server and connect from a separate machine whose clock differs from the server's
- Send any prompt to a Qwen3 model via llama.cpp with
enable_thinkingenabled - Observe the error above after the model completes its response
Screenshot and/or share link
No response
Operating System
Arch
Terminal
Ghostty
Activity
github-actions commentedon Apr 24, 2026
This issue might be a duplicate of or closely related to an existing issue. Please check:
finish=stop, triggering prefill error on claude-opus-4-6 #17982: Same symptom (agent loop runs an extra step, last message ends with an assistant turn, triggering a prefill-incompatibility error). That issue attributes the root cause to the prompt loop not gating continuation onfinish=stop; your issue identifies clock skew between client and server as a trigger for the same behavior. These may be the same underlying bug or closely related manifestations worth investigating together.finish=stop, triggering prefill error on claude-opus-4-6 #17982