Bug: Model generates identical response twice

[Q-Peppa]

Description

The model outputs the exact same response twice in a row.

Plugins

No response

OpenCode version

1.14.31

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• UI bug: Assistant responses rendered twice/duplicated #14935: UI bug where assistant responses were rendered twice/duplicated
• Some AI text is printed twice or more times #14623: AI text printed twice or more times
• Streaming content duplication during LLM streaming in desktop/web UI #14822: Streaming content duplication during LLM streaming in desktop/web UI
• [Bug] Web UI: Duplicate text rendering during streaming due to dual render path race condition in session-turn.tsx #14280: Web UI duplicate text rendering during streaming

These are all closed, but if you're seeing this on a recent version (1.14.31), it may be a regression. Could you clarify whether the duplication is a rendering artifact (same response shown twice in the UI) or if the model is actually generating two separate identical completions? That will help narrow down the root cause.

[nutriandrea]

I’m seeing the same behavior on 1.14.31, and I think it’s likely a frontend streaming/rendering duplication rather than the model producing two outputs.

From what I can observe, the response appears twice in sequence with identical content, which suggests a possible double-commit in the UI layer (e.g. both streaming updates and final completion event being appended without replacing the intermediate state).

It would be helpful to confirm:

whether the duplication also appears in raw logs / websocket stream, or only in the rendered UI
whether the session handler might be registering multiple listeners or triggering both onToken and onComplete as append operations

[wjr050430]

This issue can be easily reproduced. Using the DeepSeek official API with the models DeepSeek V4 Pro or DeepSeek V4 Flash, if you ask it to "Now you do nothing, just output a piece of text" it will call the model twice and output the result twice.
Just like this:

The user only made one request, but the model was called twice, even though only one was needed. From the raw messages, it can be seen that it is indeed two calls, not repeated output.

Additional: only reproducible under webui, environment: Arch Linux, installed opencode via npm. Not reproducible on Windows client.

After restarting the system and launching opencode again, the issue cannot be reproduced. It is suspected to be caused by opencode running for a long time. When I reproduced this bug, opencode had been running continuously for over 12 hours, and the operating system's uptime may have been even longer.

[crd2333]

It's very annoying, as one question is quoted twice. This problem makes the whole webui not suitable to use.

[wjr050430]

Confirmed issue: Installing opencode via npm on Arch Linux and running web mode continuously for over 12 hours consistently triggers this problem.
Model: DeepSeek V4 Flash/Pro, thinking mode: Max.

[wjr050430]

The issue could not be reproduced after enabling systemd-timesyncd and updating to 1.14.35.