NoLetMe: DSH Session Page Inference Keyword Trajectory Panel

Preface

When observing sessions on the DeepSeek Harness (dsh) web interface, a common need is to view the reasoning traces within the model’s streaming output in real time without altering the existing UI. NoLetMe is a dsh web plugin designed for this scenario: it only counts reasoning blocks in the model’s streaming output and places keyword traces along with a set of local descriptive statistics beside the session page.

Below, we introduce its positioning, features, installation methods, and the boundaries to be aware of when using it.

What It Is

NoLetMe is a dsh web plugin maintained by Yuer6327, licensed under MIT.

It mounts a real-time reasoning keyword statistics panel on the session page, focusing only on reasoning blocks in the model’s streaming output. The plugin is mounted via the official shell.overlay slot, without modifying or patching the existing UI.

Core Features

NoLetMe’s capabilities are centered on real-time observation within the session page:

  1. Mounts a real-time reasoning keyword statistics panel.
  2. Counts only reasoning blocks in the model’s streaming output, not visible text.
  3. Identifies three types of keyword traces: Efficient·Direct Action, Hesitant·First-Person Probing, and Neutral·Task Restatement.
  4. Provides grayscale testing with per-turn determination: score ≥ 5 indicates a hit, score ≥ 2 indicates a potential hit, otherwise no hit.
  5. Displays local model-free statistics such as I'm doing count, list density, p50, TTR, word length, and TTFT.
  6. Mounts via the official shell.overlay slot, without modifying or patching the existing UI.
  7. Supports npm pre-built installation, direct GitHub installation, and local directory installation.

Prerequisites

Before using NoLetMe, the following conditions must be met:

  1. dsh CLI ≥ 0.1.0-rc.7 must be installed.
  2. A target profile must already be created.
  3. The runtime environment requires Node >=22.

NoLetMe is built according to the dsh 0.1.x client contract and is compatible with the following versions:

  • rc.7
  • rc.8
  • 0.1.1-rc.1
  • 0.1.1-rc.2
  • 0.1.2-alpha.1 (not yet published to npm, available on GitHub)

Earlier rc versions are not guaranteed to be compatible.

Installation and Activation

Three installation methods are described below.

Method 1: npm Pre-built Installation

npm installation is pre-built and does not require allowBuilds approval:

dsh plugin --profile demo add dsh-noletme

Method 2: Direct Installation from GitHub

dsh plugin --profile demo add github:Yuer6327/NoLetMe

If the environment uses pnpm ≥ 10, the prepare script will be intercepted by default. First, follow the pnpm prompt to add the package name to the pnpm-workspace.yaml of the respective profile, configure allowBuilds, and then re-run add.

Method 3: Local Directory Installation

cd /path/to/this/repo/..
dsh plugin --profile demo add ./NoLetMe
dsh web --profile demo

Typical Usage

After installation, the NoLetMe panel will be visible on the web interface.

To run calibration, use the following command:

pnpm calibrate

This command runs regression using positive and negative samples.

Use Cases and Considerations

NoLetMe is suitable for developers who want to observe reasoning keyword traces, grayscale test determinations, and local descriptive statistics alongside the dsh web interface.

Before use, please note:

  1. The plugin runs with the current dsh process permissions. Review the source code and license before installation.
  2. Lexical tags and grayscale test hits are only observational fingerprints; they cannot determine backend, routing, or checkpoint/model identity.
  3. V4 Flash may change its style while keeping scores unchanged.
  4. Panel numbers are local, model-free descriptive statistics and do not label sessions as Claude/Gemini/GPT.
  5. Grayscale testing only counts blocks with kind === 'reasoning'; visible text is not counted.

GitHub repository:

https://github.com/Yuer6327/NoLetMe
羽毛球分组比赛记分
小程序二维码

欢迎使用《羽毛球分组比赛记分》微信小程序

Xiaoye