不会真的有人手动整理 Chrome 标签页吧？
One click. Every tab in the window sorts itself into named, colored groups.
The LLM engine looked at 10 tabs and invented its own groups: AI 与开发 / 设计工具 / 网购 / 视频娱乐 / 社交社区.
You have 30 tabs open. Again. Manually dragging them into groups is a System-2 job — slow, boring, and you won't do it. Tab Sorter makes it a System-1 job:
| Rules engine (Jev) | LLM engine | |
|---|---|---|
| Group names | You define them, once | The model invents them, every run |
| Speed | ~1s for the whole window | 10–30s (model-dependent) |
| Cost | Jev: pennies per month of daily use (output tokens are ~free) | Standard chat pricing |
| Certainty | 100% structured answers, zero parsing failures | Strict-JSON prompt + lenient parser + fallback bucket |
| Best for | Daily cleanup with your fixed buckets | Discovering what your browsing actually clusters into |
Why Jev is the default: Jev is TypeSafe's "System One" model — state in, typed decisions out. Tab classification is literally one choice question per tab, answered in parallel, so a whole window resolves in a single sub-second call with zero completion tokens (Jev prices output at $0 — you only pay input). An LLM doing the same job burns 500+ output tokens per run and takes 20× longer. Rules for the daily habit, LLM for the exploration — both engines are one right-click away at all times.
- Toolbar click → your default engine groups the current window.
- Right-click the icon → pick either engine, or ungroup (mine / all).
- Manual groups are sacred — groups you built by hand are never touched.
Same tabs, rules engine: fixed groups, 456 ms end-to-end.
git clone https://github.com/AstonyCat/jev-tab-grouperchrome://extensions → Developer mode → Load unpacked → select the folder.
Grab tab-sorter-v1.1.0.zip from Releases, unzip, load unpacked.
Open the options page and configure either or both engines:
- Rules (Jev) — paste a typesafe.ai key, edit your group rules (
label | descriptionper line; the last line is the fallback bucket). - LLM — any OpenAI-compatible
/v1/chat/completionsendpoint: base URL, key, model, and optionally your own system prompt (it must still demand the strict-JSON contract below).
Both engines have a Test connection button that runs a real 3-tab classification.
- The model sees
#index [tab title] hostper tab and returns strict JSON:{"groups":[{"label":"…","description":"…","members":[0,2,5]}]} - Labels come back in the language of your tabs (Chinese tabs → Chinese labels).
- Parser is lenient (strips code fences); any index the model skips lands in an explicit
Ungroupedgroup — nothing is ever silently dropped. - Last LLM-generated groups are auto-saved and shown in the options page — one click adopts them as your permanent rules, so a good LLM run becomes tomorrow's 1-second default.
The built-in log viewer (options page, last 200 events) records for each run: engine, duration, token usage (prompt/completion/total), and resulting groups. Every API failure logs the HTTP status and the response body — when a gateway 422s you, the reason is already on your screen. Same lines go to the service-worker console ([tab-sorter] prefix).
Tab titles/domains are sent only to the endpoint you configured, only when you click. No analytics, no telemetry, no server of ours. Keys live in chrome.storage.local on your machine. Remote LLM endpoints require an explicit runtime permission grant. Full policy: PRIVACY.md.
manifest.json MV3 manifest (tabs, tabGroups, storage, contextMenus)
grouper.js shared core — both engines, logging, grouping pipeline
background.js service worker — toolbar click + right-click menu
options.html/js settings, connection tests, LLM-groups panel, log viewer
icons/ toolbar icons (regenerate: python3 tools/make_icons.py)
docs/screenshots real end-to-end captures