@linxin666/dsh-live-stats

English | 中文

Live input/output token estimates and generation throughput for DSH Web. It feeds the built-in session status row: input and output token totals update while a response streams, and the generation throughput group (TPS 31.4 tok/s) renders right after the step counts, ahead of the billing groups:

1 turns · 3 steps TPS 31.4 tok/s Input ~7.9K tok · Output ~12 tok

~ marks a heuristic estimate. Provider usage replaces the estimate when it arrives; exact cache accounting continues to come from DSH's durable token-usage projection. A retry replaces the prior estimate for that step, and an aborted turn removes its unsettled estimate.

What it does

• Host half: registers the replayable liveTokenUsage session projection (ctx.sessionProjections). The fold estimates input tokens from the surface log plus header/tool framing, estimates output tokens from streaming chunks, and replaces estimates with provider usage as soon as a usage chunk or final message lands. TPS is derived from output tokens over wall-clock time of the active step, and the rate is resident: once any step measured one, the projection keeps reporting it (falling back to the last measured value while a new step has not produced output yet, or after a rate-less step), so the row never flickers out.
• Client half: kept for roster compatibility only. The TPS group lives inside the conversation stats line — ui-conversation reads the liveTokenUsage projection directly — so the client mounts nothing.

Installation

Install the family aggregate package @linxin666/dsh-web-ui-all (all plugins and skins in one) or this plugin alone:

# Recommended: install directly from npm
dsh plugin --profile web add @linxin666/dsh-live-stats

# Or from the repository (development loop)
git clone https://github.com/zhu1090093659/dsh-web-ui.git
cd dsh-web-ui
pnpm install && pnpm -r build
dsh plugin --profile web add link:$(pwd)/packages/dsh-live-stats

Restart dsh web, and the TPS group appears in the session status line.

Alternatively, as a plain overlay row in the personal DSH overlay (~/.dsh/config.yaml), hot-reloaded on save:

- insert:
    - id: live-stats
      name: '@linxin666/dsh-live-stats'
      config:
        charsPerToken: 4
        blockOverhead: 4
        roleOverhead: 4

All three estimator values are optional (defaults shown).

Configuration

| Key | Type | Default | Meaning | |---|---|---|---| | charsPerToken | number | 4 | Approximate text characters represented by one token | | blockOverhead | number | 4 | Fixed framing tokens assigned to each content block | | roleOverhead | number | 4 | Fixed framing tokens assigned to each message or assistant response |

Export shape

A function/namespace plugin: inject / Config / apply, no default export. The estimator (./estimator) and the projection fold (./projection) are pure and unit-tested; the client TpsLine renders through the runtime's projection hook. The invariant companion registers under ./invariant.

Model Experience

Prompt and tool surface

What the model sees

Nothing. The plugin injects no prompt sections, registers no tools, and emits no session events of its own — it only consumes the durable stream and the projection carrier's wire path.

Token effect

Zero per request.

KV Cache effect

No system-prompt contribution, so no cache-stability effect.

Known Limitations and Deferred Work

• Heuristic estimates: input/output totals are character-count heuristics (~) until provider usage arrives; exact cache accounting always comes from DSH's durable token-usage projection.
• Web only: the TPS row renders in DSH Web's composer dock; there is no TUI equivalent yet.
• Single active step: the projection tracks one active step per session and the dock row shows that session's view; concurrent sessions each get their own projection.
• Density assumption: charsPerToken defaults to 4 characters, which undercounts CJK text and overcounts pure ASCII; tune it per deployment if estimates drift.