GLM-5.3 Flash: Stream output mixed with random English fragments and shortened reasoning in Kelivo client

オープン
#1,039 コメント 3 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
25/100
issue の種類
バグ
明瞭さ
おおむね明確
活発さ
活発

調査の方向性

The report identifies GLM-5.3 Flash through the CommandCode API gateway, used from Kelivo on iOS, with failures especially after web search in long role-play conversations. Start by tracing the API gateway’s streaming and non-streaming response paths and how web search results are passed to the model. Done means reproducing the mixed-language output and instruction-following failure, then identifying a fix that addresses both modes.

索引モデルが issue の本文から書いたものです。

説明

Summary

When using GLM-5.3 Flash model via CommandCode API gateway on Kelivo client (version 1.3.1 / 80, iOS), the output randomly inserts unexpected fragments of English, Russian, Thai and other languages into Chinese text.
The issue is prominent in streaming mode, and can still happen occasionally after streaming is disabled. Disabling streaming also causes the model to fail to follow instructions and generate uncontrollable content.
I have tested temperature=1.0 and temperature=0.8 for comparison. The problem becomes much worse with web search enabled, mostly appearing in role-play long context conversations after search finishes.

Expected Behavior

The model should output clean Chinese text without unexpected mixed foreign language fragments. The model should faithfully follow user instructions regardless of streaming mode and temperature settings.

Actual Behavior

In streaming mode, GLM-5.3 Flash randomly inserts fragments of English, Russian, Thai and other foreign languages into Chinese sentences. This phenomenon is very obvious when stream is enabled.
Even if streaming is disabled, this bug still occurs occasionally. Moreover, when streaming is turned off, the model cannot fully follow user instructions and produces uncontrollable content.
The problem is much worse when web search is enabled. It mainly happens during long text generation after web search completes in role-play conversations.
Tested temperature: 1.0 makes the bug more obvious; temperature=0.8 reduces the frequency but cannot eliminate it completely.

Steps to reproduce the issue
  1. Use Kelivo client v1.3.1 / 80 on iOS, set API endpoint to CommandCode, select GLM-5.3 Flash model and enable stream mode.

  2. Turn on the web search function in the prompt.

  3. The model completes web search and retrieves reference materials.

  4. After the search phase finishes, the model starts generating long text for role-play.

  5. Observe the streaming output: random fragments of English, Russian, Thai and other foreign languages are inserted inside Chinese sentences.

  6. Additional test: Disable streaming. The foreign fragment bug still occurs occasionally, and the model may not follow instructions properly, producing uncontrollable output.

  7. Compare test: temperature=1.0 makes the bug more obvious; temperature=0.8 reduces the frequency but cannot eliminate it completely.

  8. When web search is disabled, this bug appears far less often.

Command Code Version

API Gateway (web API, not local CLI)

Operating System

Other

Terminal/IDE

Kelivo Mobile Client

Shell

Not applicable

Session file (optional)

No response

Fix prompt (optional)

No response

Additional context

The issue started happening 2-3 days ago. Client: Kelivo version 1.3.1 / 80, iOS system. When web search function is enabled, the mixed-language problem becomes much worse. If web search is disabled, the bug occurs less frequently. I can provide screenshots of mixed-language output if needed. No API keys or account secrets included.

Note: The following content is for testing purposes only and does not represent any position. The user takes no responsibility for any output generated here. It is only used as a bug sample.

Sample broken output example:

Image Image Image
主要言語
言語のデータがありません
スター
4.2k
フォーク
377
平均マージ
1日 17分
マージ済み PR(30日)
7

環境構築

このプロジェクトには開発コンテナ、Dockerfile、コントリビューションガイドがありません。まず README を読み、一般的な手順ははじめてのコントリビューションガイドを参照してください。

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

CommandCodeAI/command-code のほかの issue

CommandCodeAI/command-code の issue をすべて見る

似ている issue

Backend & API Design の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。