dsh-plugin-thinking-size

A live reasoning-token badge for DSH web. It appends a compact Think(128) / Think(5.2K) marker to the title of every "Think" disclosure row in the conversation — ticking in real time while the model is still thinking, and staying put on history messages afterwards.

English | 中文

The badge

Format	Rule
Think(128)	Below 1000 tokens: plain integer
Think(5.2K)	1000 – 999 499: 1 decimal + K, fixed digits (5000 → 5.0K)
Think(1.2M)	A million or more: 1 decimal + M (1000000 → 1.0M)

• Live — the number ticks up while the reasoning is streaming.
• Persistent — every historical Think row keeps its badge too, across expand/collapse and session switches.
• Subtle — a 12px caption-colored span in tabular figures, appended inside the row title and marked aria-hidden, so the accessible title still reads just "Think".
• Adjustable precision — hover the badge to pick 1 or 2 decimal places (default 1); the choice applies to every badge instantly and is remembered in localStorage, surviving reloads and restarts.

雷霆大思考, measured

Some LLMs are famous for their 雷霆大思考 — thunder-length thinking that opens with several thousand tokens of pure reasoning. With this plugin you can watch the badge sprint in real time and see, at a glance, just how long each burst of thinking really is. No more wondering what the model was doing up there: the length is right on the row.

How it works

• Data comes from the current session snapshot (sessions service → chat.nodes): the full text of every reasoning block of each assistant-step node. The collapsed DOM never carries the full reasoning text, so counting must read the snapshot, not the page.
• Estimation follows DeepSeek's official approximation — 1 English character ≈ 0.3 tokens, 1 CJK character ≈ 0.6 tokens, the sum rounded up. Counts are cached per reasoning-block object (its reference is immutable), so a streaming chunk only re-counts the one block that changed.
• Writing appends the badge span into each row's title span. Think rows (data-variant="think") are matched to reasoning blocks by conversation anchor key plus in-anchor index. Snapshot subscription and a MutationObserver only set a dirty flag; a 100ms throttled sweep does the work — and if a React re-render wipes a badge, the observer flags it and the sweep puts it back.
• Silent by design — no host code is touched. If the sessions service is missing, the DOM contract changes, or a snapshot looks malformed, the plugin idles or skips rows quietly; it can never break the host.

Install

dsh plugin --profile web add dsh-plugin-thinking-size

Restart dsh web to take effect.

From source (local development)

pnpm install
pnpm run build
pnpm test

Then add this package to ~/.dsh/profiles/web/package.json as a link: dependency, register it in dsh.profile.bundles, and restart dsh web.

Keywords

dsh · deepseek harness · dsh plugin · web plugin · client plugin · chat ui · thinking · deep thinking · reasoning · reasoning tokens · reasoning length · token counter · token estimator · token usage · think badge · badge · streaming · 思考 · 深度思考 · 长思考 · 思考长度 · 思考时长 · 思考过程 · 推理 · 推理 token · 思维链 · token 统计 · token 估算 · 徽标 · 实时显示 · 聊天界面 · 雷霆大思考