repetition-guard

中文

LM Studio 实时重复输出守卫：本地模型正文 / 思考流一开始"复读"，立刻自动截停，已生成内容完整保留。

解决什么问题

量化模型（尤其是小模型）长生成时容易陷入循环：wait wait wait wait …、思考里反复念同一句话、工具调用参数刷同一段 JSON——白烧 token 还占着上下文。目前 LM Studio 没有内置的"检测到循环就停"机制，本插件补上这一环。

工作原理

插件注册一个 Guarded Generator，包裹本地已加载的模型，逐片段转发输出的同时做两层检测：

连续单元重复：尾部出现同一单元（1–24 字符）连续重复 ≥ 阈值（默认 10）次。覆盖单字符循环、词语循环、短语循环，忽略空白差异。
窗口吟唱：同一 32 字符窗口在尾部反复出现 ≥ 阈值（默认 4）次。覆盖超过单元长度的整句循环。

检测同时作用于三路流：正文、思考（reasoning）、工具调用参数，可分别开关。

触发后：底层预测立即 abort（服务端停止生成）、已输出片段保留、本轮正常收尾（等同用户点停止），并在末尾附一行提示（可关闭）。不删除任何内容。

防误杀

Markdown 表格分隔行、水平线（- | 默认忽略字符）不会误判。
纯装饰/空白片段直接跳过。
围栏代码块内阈值自动 ×3（可调），代码里合法的重复（分隔线、ASCII 表格、重复数据）不受影响。

配置项

项	默认	说明
Model Key	（空=当前模型）	要守护的本地模型
Temperature	0.8	透传给模型
Repeat Threshold	10	连续重复触发阈值，越小截停越早
Chant Threshold	4	整句循环触发阈值
Guard Thinking Stream	开	检测思考流
Guard Answer Stream	开	检测正文流
Notify On Stop	开	截停时附提示行

使用

在聊天的模型选择器里选本插件提供的 guarded 生成器即可生效。

English

Real-time repetition guard for LM Studio: the moment a local model starts looping on repeated text — in the answer, in the thinking stream, or in tool-call arguments — generation is stopped automatically. Everything generated before the loop is kept intact.

Problem

Quantized local models (especially small ones) can degenerate into wait wait wait wait … loops mid-generation, chant the same sentence in reasoning, or spam identical tool-call JSON — burning tokens and polluting the context. LM Studio ships no built-in loop breaker; this plugin adds one.

How it works

The plugin registers a Guarded Generator that wraps a locally loaded model. Every fragment is forwarded to the UI and inspected by two detectors:

Consecutive-unit repetition: the tail ends with the same unit (1–24 chars) repeated ≥ threshold (default 10) times — catches char loops, word loops and phrase loops, whitespace-insensitive.
Window chanting: the same 32-char window recurs ≥ threshold (default 4) times — catches long sentence-level loops.

All three streams are guarded independently: answer text, thinking/reasoning, and tool-call arguments, each toggleable.

On trigger: the underlying prediction is aborted immediately (the server stops generating), emitted fragments are preserved, the turn finishes cleanly (exactly like a user stop), and an optional one-line note is appended. Nothing is deleted.

False-positive protection

Markdown table separators / horizontal rules (- | are ignored chars) never trigger.
Purely decorative or whitespace fragments are skipped.
Threshold auto-scales ×3 inside fenced code blocks (configurable) — legitimate repetition in code (separators, ASCII tables, generated data) is unaffected.

Configuration

Setting	Default	Meaning
Model Key	(empty = current)	Local model to guard
Temperature	0.8	Passed through to the model
Repeat Threshold	10	Consecutive-repeat trigger; lower cuts earlier
Chant Threshold	4	Sentence-loop trigger
Guard Thinking Stream	on	Detect loops in reasoning
Guard Answer Stream	on	Detect loops in the answer
Notify On Stop	on	Append a note when cut off

Usage

Pick the guarded generator provided by this plugin in the chat model selector.

Credits

Detection strategy inspired by dsh-repetition-guard / dsh-dupguard (DeepSeek Harness) and Qwen Code's LoopDetectionService.

License

MIT