dsh-thinking-token-stat

English | 中文

A lightweight plugin that adds model thinking-token statistics to the bottom Dock and the end of each conversation. Client-only and zero-cost when idle: it reads the existing conversation snapshot and renders nothing at all when there is no reported or visible thinking output.

Thinking statistics depend on the model and provider. Some models do not expose reasoning tokens or reasoning text to the client; for those replies, the plugin has no thinking data to count and the readout is intentionally not shown.

It adds two readouts, both of which render nothing when the model has produced no thinking for the relevant scope:

Surface	Slot	Scope
Bottom composer dock	conversation.composer.dock	Whole session (cumulative)
Each assistant reply (timing row)	conversation.chat.assistant-actions	That single turn

The per-turn readout is appended to the reply's timing strip, alongside the built-in usage and timing pills, so it reads as part of that same unit. Click the thinking pill to open a native-style details panel. The panel follows the active DSH locale.

The details panel shows the thinking-token count, the all-token denominator and percentage, and the output-token denominator and percentage. Percentages use one decimal place; counts use compact units such as K and M when appropriate.

How thinking tokens are counted

For every finalized assistant message, in order:

1. Provider-reported — if usage.reasoningTokens is present (DeepSeek, OpenAI o-series, Anthropic, …), it is used as-is. This is exact.
2. Otherwise — the reasoning blocks are priced with the harness' fixed heuristic (four characters per token), so a provider that returns reasoning text but no token count still gets a figure.
3. Neither present → the message counts as zero thinking and contributes nothing.

The two denominators are derived from the same assistant messages, so the percentages always agree with the count on one consistent scope:

• share of all tokens = thinking ÷ (input + output + cache reads + cache writes)
• share of output = thinking ÷ output

Why it's lightweight

• Client-only, zero host behavior. The node half (lib/index.js) is an empty apply that exists only so the package appears in the host Loader. All work is done in the browser.
• Read-only. It consumes the existing conversation snapshot; it adds no service, projection, tool, or RPC.
• Self-contained. It depends on no other plugin.
• Zero-cost when idle. When nothing is being thought, both readouts render null — no element, no layout cost.
• Theme-aware. Colors come from the --dsw-alias-* tokens, so it follows light/dark automatically.

Install

The plugin follows the standard dsh.bundle + dsh.client convention, so it installs like any DSH plugin.

From GitHub:

dsh plugin add github:Six6stRINgs/dsh-thinking-token-stat

Or install the published npm package:

npm install dsh-thinking-token-stat

After installing, make sure the package is included in your DSH profile bundles, then restart dsh web and reload the page. The readouts appear only once the model starts thinking.

Testing

test/harness.mjs stubs the browser/React environment, runs the real factory and apply, and renders both entries against sample data (provider-reported and block-estimated paths, the one-decimal formatting, and the hidden-when-empty cases). Run with:

node test/harness.mjs

Known limitations

• Figures come from the in-window conversation snapshot. For very long, paged sessions the visible window is the counted scope; the shipped stats line and projections use whole-log folds instead.
• The character-per-token estimate is approximate, exactly as the harness' fixed heuristic is; provider-reported reasoningTokens is always preferred.