dsh-thinking-token-stat
English | 中文
A lightweight thinking/reasoning token statistics plugin for DeepSeek Harness (DSH). Displays whole-session thinking token totals and ratios right in the composer dock, with official frosted-glass detail popovers and turn-by-turn model breakdowns.
Composer dock readout and overview popover
Turn-by-turn details
Features
1. Composer Dock Readout
Always visible in the status line below the message input box:
💭 37.2K · 70.7% · 4 / 6 turns
- Total Thinking Tokens: The cumulative count of reasoning tokens across the entire conversation.
- Thinking / Output Ratio: The proportion of thinking tokens relative to total reasoning output.
- Reasoning Turns: Active thinking turns out of total turns (displayed when non-thinking turns exist).
2. Session Overview Panel
Clicking the dock readout opens the overview panel, fully styled according to DSH official popover design and frosted-glass vibrancy:
- Thinking / Reasoning Output: Share of thinking tokens within reasoning-capable turns.
- Thinking / All Tokens: Share of thinking tokens out of total billed session tokens (input + output + cache).
- Turn Counts: Total turns vs. active reasoning turns.
- Estimate Transparency: When a model does not report token counts, the estimated turn count and token volume are clearly broken out with a
? explaining the formula.
- Scope Audit: Statically proves what the figures cover (empty turns, fork-excluded prefixes, etc.).
3. Turn-by-Turn Breakdown
Click "Detailed turn statistics" in the overview to open the per-turn table (newest turns first):
- Turn Number & Thinking Tokens: Per-turn count (estimated figures prefixed with
~ and highlighted in business accent).
- Turn Total Output: Total tokens written in that turn.
- Thinking Source: Indicates whether the count was provider-reported, text-estimated, or mixed.
- Model Attribution: Exact Provider and Model names (e.g.
deepseek-official/deepseek-reasoner), tracking retries and model switches accurately.
- Output Share: Percentage of the turn's output spent on reasoning.
- Quick Filtering: Filter by Reported, Estimated, or None (collapsed by default).
- Jump to Reply: Jump button on each row pages DSH history and navigates straight to the message.
Token Counting & Estimation
Not all model providers report reasoning tokens via API usage metadata. To ensure accurate and meaningful statistics in every environment, the plugin employs a transparent hybrid strategy:
| Provider Response | Method | Display |
|---|
| Reports reasoning tokens | Uses exact provider count | Stated exactly; marked as Reported |
| Provides reasoning text only | Multi-script density estimation | Prefixed with ~ (e.g. ~5.1K); marked as Estimated |
| Provides neither | Counted as non-thinking | Collapsed by default in the turn list |
Multi-Script Character Density Rules
Estimating tokens using a flat 4 characters/token English heuristic undercounts dense scripts by 2–4×. This plugin segments text by script family:
| Script | Characters / Token | Rationale |
|---|
| CJK (Chinese, Japanese kana, Korean Hangul, fullwidth) | 1 | Measured tokenizers range ~0.6–1.7 tokens/char (DeepSeek/Qwen tokenizers sit at 0.6–0.8). 1 char/token is conservative and avoids overestimating. |
| Cyrillic, Greek, Arabic, Hebrew, Indic, Thai, etc. | 2.5 | Typically measured at 2–3 characters per token. |
| Latin, digits, punctuation, whitespace | 4 | Standard empirical benchmark for English and Western prose. |
All estimated numbers are explicitly prefixed with ~ and never misrepresented as exact provider telemetry.
Architecture & Performance
- Durable Session Projection:
- Folded incrementally on the host Node process as an event-driven Reducer over
assistant/message settlements.
- Persisted as a session projection checkpoint (
<root>/session_projcache/sessions/<id>.json), requiring only a few KB per session.
- Covers the complete session event log, immune to client-side window paging, virtual scrolling, or compaction.
- Zero Frontend Overhead:
- The browser half simply reads a single projection value via
useProjection("thinkingStats"). No message parsing, no background API requests, and no localStorage caching.
- Native DSH Alignment:
- Strictly adheres to DSH
stat-dialog.module.css specs, incorporating var(--dsw-menu-backdrop-filter) (40px blur) and var(--dsw-radius-lg) corner radius.
- Automatically adapts to Light and Dark themes.
- Matches the client UI language automatically (Chinese & English).
Installation
Method 1: DSH Plugin Market (Recommended)
Open the Market in DSH Desktop or Web GUI, search for dsh-thinking-token-stat, and click Install.
Method 2: Command Line
Run in your terminal:
dsh plugin add dsh-thinking-token-stat
Or install the latest development branch from GitHub:
dsh plugin add github:Six6stRINgs/dsh-thinking-token-stat
After installation, refresh your browser or desktop window (Ctrl+R). The indicator appears automatically once the model begins thinking.
Development
The plugin consists of two self-contained modules:
lib/index.js: Host half; registers the thinkingStats session projection and fold logic.lib/client.js: Client half; injects the readout, popover, and table into the composer dock slot.
Run tests:
npm test
(Executes test/projection.mjs for host fold semantics and test/harness.mjs for client UI rendering and degradation paths).
License
MIT
