AI Agent Hub
Back to plugins
dsh-thinking-token-stat preview

dsh-thinking-token-stat

Client Updated 2026.09.13

Run the following command in DeepSeek Harness:

dsh plugin install Six6stRINgs/dsh-thinking-token-stat

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install Six6stRINgs/dsh-thinking-token-stat in your DeepSeek Harness terminal to install this plugin; the source code is available at https://github.com/Six6stRINgs/dsh-thinking-token-stat .

About this plugin

The DSH reply panel reports reasoning tokens per turn, but model providers differ widely in how they expose thinking tokens: some report a count, some return only the thinking text, and some provide neither. After a long multi-turn conversation there is no session-level view that answers how much thinking was spent overall, what share it represents of total output, or how many turns actually reasoned. dsh-thinking-token-stat fills that gap by adding a clickable thinking-token total to the statistics line under the message box, with an expandable detail panel that shows the reasoning-output share, the share against all billed tokens, the number of turns that thought, and which figures are estimates rather than reported values.

Clicking the statistic opens a detail panel with clearly labelled rows for every metric. At the bottom, a link leads to the turn-by-turn table, newest first, showing each turn thinking tokens, its full output, where the figure came from, the model that ran it, and the share of thinking within that turn output. Three filter switches isolate reported, estimated, or no-thinking turns; the table renders in pages of 100 rows and every row ends with a jump button that scrolls to the corresponding reply in the conversation. When a provider does not report a count, the plugin estimates from the thinking text using script-aware character density, and every estimated figure carries a tilde mark with an expandable explanation.

Under the hood the host side folds the session event log once and registers the result as a DSH session projection, so the numbers cover the entire conversation by construction, stay stable across windowing and compaction, and survive a reload. The browser side reads that single projection value and draws the UI; it makes no network requests, writes no local storage, and keeps no plugin-level state. If you want a complete, session-wide thinking-token ledger without modifying the conversation, issuing extra requests, or installing background processes, this plugin is the lightest path to get it.

Screenshots

Use Cases

  • Review session-wide thinking-token cost after a long conversation
  • Compare reasoning-output share of total output across models
  • Identify turns where the provider did not report a token count

Best For

  • Daily DSH users who want to track thinking-token spend
  • Developers benchmarking reasoning efficiency across models
  • Privacy-sensitive users requiring zero network calls and zero local storage