Introduction

The built-in panels of DeepSeek Harness (DSH) usually focus on statistics for a single-turn response, while developers often need to understand the inference cost and structure of an entire session. When the model provider does not report thinking tokens for each turn, manual counting can become tedious and inaccurate. The dsh-thinking-token-stat plugin solves this problem by calculating session-level totals through collapsed aggregation and providing an intuitive statistics view in the Composer Dock.

Plugin Overview

This is a client-side plugin maintained by Six6stRINgs and released under the MIT license. It performs a one-time persistent computation through DSH’s session projection mechanism, then reads the results and displays them in the UI Dock. Its core value lies in aggregating inference token statistics scattered across multiple turns of conversation and providing filterable details on a per-turn basis.

Core Features

  1. Session-level statistics view: Displays the total number of thinking tokens for the session, the proportion of inference output, and the number of turns that actually produced thinking in the statistics row below the message box.
  2. Per-turn detail view: Provides a turn-by-turn table listing the thinking tokens, output content, source (reported/derived), model, and proportion for each turn.
  3. Filtering and selection: In the per-turn view, toggles allow filtering to show only turns with “reported counts,” “derived counts,” or “no thinking counts.”
  4. Estimation handling: Numbers that cannot be derived from text or are unreported are marked with ~, and the source of counts is distinguished between “reported” and “derived.”
  5. Persistent calculation: Uses DSH’s session projection mechanism to ensure statistics are based on the complete log of the entire session, rather than only the current window, and keeps data consistent after page refreshes.

Installation

The installation command is not explicitly listed in the provided source text. The plugin source code or documentation can be accessed through the following methods:
- GitHub repository: Six6stRINgs/dsh-thinking-token-stat
- Skill library catalog: dsh-thinking-token-stat

Typical Usage

  1. View session overview: Open a conversation that contains multi-turn thinking. A statistics row in the Composer Dock will display information such as 💭 37.2K · 70.7% · 4 / 6 turns. These values represent the total number of thinking tokens, the proportion of inference output, and the number of turns that actually thought versus the total number of turns, respectively.
  2. Click to view details: Click the statistics row to open the session details panel. Jump buttons at the bottom of the panel allow navigation to the per-turn view.
  3. Filter turns: In the per-turn view, use the toggles at the top to filter specific types of turns (for example, only view turns that require derived tokens).
  4. Jump to response: In the per-turn table, click the jump button at the end of a row to quickly locate the specific response content for that turn.

Notes

  • Does not override single-turn statistics: The plugin does not display thinking token values for individual responses, because DSH’s official panel already provides this functionality.
  • Data precision and format: Percentages are kept to one decimal place, and large numbers use K format (for example, 24.8K).
  • Estimation marker: All numbers based on text collapsing or derivation are prefixed with ~, indicating estimated values.
  • Pagination limit: The per-turn view renders only 100 turns per page; entries beyond this limit must be accessed via “Newer / Older” pagination.
  • Includes failed turns: The list includes all turns, including failed turns or turns that did not produce a response. These turns are displayed as — and None in the table.

Conclusion

dsh-thinking-token-stat provides DeepSeek Harness users with a reliable session-level inference token analysis tool. Through its projection mechanism, it ensures data accuracy, and through a clear interface, it helps developers understand model inference behavior at the session level. For developers who need to finely control inference costs or analyze model thinking patterns, this plugin provides the necessary visualization support.