Introduction

When observing session text morphology on the DSH web interface, a specific requirement is: which category the current reasoning text resembles more, and what are the tendencies of the paragraph opening styles. DSH’s plugin extensibility is suitable for turning such observations into in-page widgets: open the session page first, then check the readings, without needing to copy the text out for offline statistics. dsh-grayprint is a web plugin designed specifically for this scenario.

What is this

dsh-grayprint is a web extension for DeepSeek Harness, with the package name dsh-grayprint, maintained by HongzhongL, and licensed under MIT.

By default, it opens two independent real-time fingerprints from the bottom right corner of the session page:

  • Whether the current thinking text resembles a grayscale sample or the current version sample more.
  • Whether the paragraph opening leans more towards Let me or We need.

The readings here only describe textual morphology and do not represent actual model version, capability, backend, routing, or checkpoint determination.

Core Features

Two Independent Fingerprints

The first one focuses on whether the current thinking text resembles a grayscale sample or the current version sample. The plugin counts the first-person pronouns and styles like I'm/I'll at the start of internal paragraphs.

The second one independently tracks the quantity and ratio of openings starting with Let me and We need. It only counts paragraph opening styles, used to observe paragraph opening tendencies.

Six Guards

The plugin refuses to judge or degrades when evidence is insufficient or conditions are not met:

  1. Refuse to judge when the proportion of Chinese characters exceeds 15%.
  2. Refuse to judge when Latin letters account for less than 40% of all letters.
  3. Refuse to judge when English stylistic signals are insufficient.
  4. Directly refuse to judge if there are fewer than 5 internal paragraphs in a normal sample.
  5. Judging for “grayscale fingerprint” requires an additional >=48 internal paragraphs; otherwise, refuse to judge or degrade.
  6. Refuse to judge or degrade if there is no thinking text.

Interface and Data Boundaries

The panel is initially located in the bottom right corner of the session page. Both the expanded state and the collapsed capsule can be dragged; a short press on the sticky header bar can retract it.

The interface follows the DSH Chinese/English settings, and panel state is saved in localStorage.

The plugin only counts the reasoning block within the current browser session snapshot; data does not leave the browser.

Installation and Enablement

Prerequisites are dsh CLI >= 0.1.0-rc.7 and a target profile created.

Installation command:

dsh plugin --profile web add github:HongzhongL/dsh-grayprint

After installation, restart the dsh web host, then refresh the page.

Uninstallation command:

dsh plugin --profile web remove dsh-grayprint

Typical Usage

  1. Open a DSH web session page with the plugin enabled.
  2. By default, view the panel in the bottom right: grayscale sample fingerprint and independent Let me / We need paragraph opening style tendencies.
  3. Drag the panel to a suitable position; short press the sticky header bar to retract it.
  4. Expand details to view the two scoring criteria and raw statistics of paragraph openings.

By default, only the main fingerprint is displayed; the two scoring criteria and raw paragraph opening statistics are contained within the expandable details.

Suitable Scenarios and Notes

Suitable for the following scenarios:

  • Observing the textual style of the current session’s reasoning block within the DSH web host.
  • Need to count the quantity and proportion of Let me and We need at the start of paragraphs.
  • Wish for statistics to only occur within the current browser session snapshot.

Note the following boundaries:

  • Only compares textual styles, not representing actual model version, capability, backend, routing, or checkpoint determination.
  • Polarized samples come from different rendering channels; some differences may come from the channel rather than the model version, and this is not excluded.
  • The grayscale side only has positive examples; neither side has task scoring or rubrics.
  • Preset influence is greater than version influence: for the same model, switching presets can change the score from 10% to 86%.
  • The plugin runs with the permissions of the current dsh process; before installation, check the source code and license; the license is MIT.

Conclusion

The value of dsh-grayprint is to transform the reasoning paragraph opening style into a draggable, rejectable, and detailed web-based reading metric. It does not perform model identity determination and does not send data out of the browser.

GitHub repository: https://github.com/HongzhongL/dsh-grayprint

The community directory is an independent site and does not form an official subordinate relationship with DeepSeek / Hypersphere; the directory page address is subject to actual retrieval on the independent community directory site; this draft does not include unconfirmed directory URLs.