dsh-client-ui-thinking-stats

1. 安装并启动 DeepSeek Harness：npx @deepseek-ai/dsh web
2. 在终端执行上面的安装命令（CLI 会解析插件并核验来源）
3. 用 dsh plugins list 确认已安装，必要时重启 Harness 生效

插件以当前 dsh 进程的权限运行，安装时可能执行代码。请先通读仓库源码与许可证，确认无破坏性命令与越权访问；本站只做索引，不对第三方插件安全性作担保。

事实信息来自公开插件目录快照（2026-10-01），介绍文案由本站再加工。

以下为插件仓库 README 全文（原始内容，由公开目录抓取整理）。

# dsh-thinking-token-stat

![license](https://img.shields.io/badge/license-MIT-green)
![dsh](https://img.shields.io/badge/dsh-plugin-4B32C3)
[![repo](https://img.shields.io/badge/repo-github-181717?logo=github)](https://github.com/Six6stRINgs/dsh-thinking-token-stat)

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.

[Chinese version / 中文版](./README.zh.md)

![thinking-token statistics](./assets/screenshot.png)

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 — the same hover
row that shows `20:55 · 用时 11秒 · 首token 10秒 · 28 tok/s` — so it reads as part
of that same unit.

Each readout shows, left to right, a **brain glyph**, the thinking-token count,
its **share of all tokens**, and its **share of output tokens**, each with one
decimal place, for example:

```
💭 15 · 0.2% · 60.0%
```

Hover for the exact numbers.

## 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 this repository:

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

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.