Skip to content

Bug: Runaway garbled generation — 88k tokens of gibberish in one LLM step (step e6f3748b) #2597

Description

@kdp123

Summary / 问题概述

During a normal interactive session, the model produced a runaway generation: a single LLM step ran for 3214 seconds (~53 minutes) and emitted 88,114 output tokens of incoherent, repetitive gibberish (random multilingual fragments, broken Markdown, endless repetitions of words like "From", "Related", "Earlier", "https"). The content was completely unrelated to the task and wasted a large share of the user's rate-limit quota.

在一次正常交互会话中,模型出现失控生成:单个 LLM step 运行 3214 秒(约 53 分钟),输出 88,114 tokens 的无意义乱码(多语言碎片、重复词汇如 "From"/"Related"/"Earlier" 无限循环、破碎的 Markdown)。内容与原任务完全无关,白白消耗了用户大量频控额度。

Environment / 环境

  • OS: macOS (bash)
  • Install method: uv tool install kimi-cli
  • Model: default coding model via Kimi Code platform (kimi-for-coding)
  • Session/step ID: e6f3748b-f320-48c1-90f9-8fb12029902a
  • Time: 2026-08-08 ~10:17–11:11 (local, UTC+8)

Log evidence / 日志证据

From ~/.kimi/logs/kimi.log:

2026-08-08 10:17:48.393 | INFO | kimi_cli.ui.shell:run_soul_command:827 | - Running soul with user input: [TextPart(type='text', text='把阈值改成比例阈值(如 diff/source > 2% 才报),所有的ods表都改')]
2026-08-08 11:11:22.447 | INFO | kimi_cli.soul.kimisoul:_step:1256 | e6f3748b-f320-48c1-90f9-8fb12029902a - LLM step completed in 3214.0s (input=38892, output=88114)

Note the step ran 53 minutes and produced 88k output tokens — all gibberish.

Characteristics of the bad output / 异常输出特征

  • 无任何任务相关内容,开头即偏离
  • 大量重复 token:「From」「Related」「Earlier」「https」「/」「>>」等无限循环
  • 混杂中英文、代码片段、无效链接,无终止迹象,直到 88k tokens 才停止
  • 该 step 前一次正常输入仅 38,892 input tokens,任务本身是简单的两行配置修改

Expected behavior / 期望行为

  1. 模型不应产生失控重复生成;CLI 或平台侧应有 runaway generation 的检测与熔断(如重复 n-gram 检测、单 step 输出上限、超时中断)。
  2. 此类明显异常的请求消耗不应计入用户额度,或应有自动/人工补偿机制。

Request / 诉求

本次异常消耗了用户约 88k output tokens 的额度(约占当日 5 小时频控的 35%),请求官方返还/补偿该部分额度,并排查该 step(ID: e6f3748b-f320-48c1-90f9-8fb12029902a)的服务端日志定位根因。

Reproduction / 复现

无法稳定复现。前置上下文是一个较长的技术会话(约 38.9k input tokens),内容涉及 git、Flink CDC、StarRocks 监控脚本修改。用户输入后模型即开始失控生成。

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions