dsh-tokmeter

Live model decode speed (tok/s) pill for the DeepSeek Harness Web UI.

The stock status bar shows a session-average decode speed that only updates after an assistant step settles. dsh-tokmeter adds a small lightning pill that tracks the speed of the currently streaming response in real time, next to the native pills in the composer dock:

[ 1 turns 33 steps · 76 tok/s ] [ ⚡ 75.9 tok/s ]   ← live, updates every stream delta

The pill appears while the model is generating (text, reasoning, or tool-call arguments) and fades out ~2.5 s after the stream stops.

How it works

It is a standard DSH client plugin (slot entry in conversation.composer.dock, session scope). The chat client exposes the live partial assistant step as useChat(s => s.legacy.partial); while that is non-null, the pill samples the growing block content every 250 ms and converts character growth into an estimated token rate (~4 chars per token), smoothed with an exponential moving average:

rate = 0.7 * previous + 0.3 * (Δchars / 4) / Δt

The figure is an estimate derived from stream text growth — the client does not receive per-token timestamps, so long tool-call streams may show slightly different numbers than the provider-side metrics.

Installation

Requires DSH (DeepSeek Harness) with a Web UI profile.

# from a git checkout (or a published registry):
dsh plugin --profile web add <your-url-or-package>

The installer links the package into the profile and appends @local/dsh-tokmeter to dsh.profile.bundles in the profile's package.json. A restart of the Web UI server picks the bundle up (client module edits of an already-linked package are hot-reloaded without a restart).

Requirements

• DSH ≥ 0.2.1-rc with the Web UI bundle (@deepseek-ai/dsh-web-app).
• The chat client (@deepseek-ai/dsh-client-ui-chat) must be part of the composition so the session-scope slot receives the useChat hook.

Development

The plugin is a plain package — no build step.

File	Purpose
index.js	Host entry point (empty; all work happens client-side).
client.js	Client module: slot registration + the React pill.
cordis.patch.yml	Bundle composition patch (inserts the entry).
icon.svg	Plugin icon (plugin manager UI).

Local test loop:

dsh plugin --profile web add /path/to/dsh-tokmeter
dsh --profile web --dump-config   # must list '@local/dsh-tokmeter'

Then open the Web UI and ask the model for a long answer — the ⚡ pill in the bottom bar tracks the live decode speed.

Notes

• The pill renders only in sessions with a chat view; it is invisible when nothing is streaming.
• Colors and metrics use the standard --dsw-alias-* theme tokens, so it follows light/dark themes automatically.
• The pill text is intentionally language-neutral (tok/s).