Multi-pane ChatGPT workspace as a Chrome extension: several ChatGPT chats side by side in one extension page, with bottom tabs and a chat-picker pane.
Each pane is a live https://chatgpt.com iframe, so all of ChatGPT keeps
working (model picker, deep research, composer, streaming). Speed comes from
not paying for what you don't look at:
- Lazy panes — only the focused pane mounts its iframe. Other panes show a
lightweight placeholder until clicked. On reload only the focused pane loads
eagerly. Per-pane "sleep" (☾) unloads an iframe to reclaim memory.
State lives in the pure, tested
src/lib/pane-store.js. - Conversation trim — a MAIN-world content script patches
fetchsoGET /backend-api/conversation/{id}is trimmed to the last N messages on the active branch. Long chats stop freezing. Logic:src/lib/conversation-trim.js. - Settings bridge — an ISOLATED-world content script mirrors settings into
<html>dataset so the MAIN-world patch can read them; both are injected into chatgpt.com withall_frames: trueso they run inside the pane iframes (standalone extensions don't, which is why they appeared not to work here).
| Area | File |
|---|---|
| Settings schema/defaults | src/lib/settings.js |
| Pane state machine | src/lib/pane-store.js |
| Conversation trim (pure) | src/lib/conversation-trim.js |
| MAIN-world fetch patch | src/content/perf-fetch-trim.js |
| ISOLATED settings bridge | src/content/settings-bridge.js |
| Workspace view | app.html, src/app.js, src/app.css |
| Settings page | options.html, options.js |
All toggleable from the settings page; injected into the pane iframes via
all_frames content scripts, with pure logic factored into src/lib/* and
unit-tested.
- Lazy panes + conversation trim (see above).
- Native chat picker — a
+pane lists your chats (fetched from/backend-api/conversations) without booting the full app; click to open one in that pane.src/content/picker.js,src/lib/picker-model.js. - Pane title / URL sync — each pane tracks the chat it shows across SPA
navigation, so reloads restore the right chat per pane and the tab shows the
real chat name.
src/content/title-report.js. - Auto-confirm Custom GPT actions (+ auto-expand tool calls).
src/content/features/auto-confirm.js,src/lib/confirm-match.js. - Prompt queue with auto-send when ChatGPT goes idle.
src/content/features/queue.js,src/lib/queue-model.js. - Auto-collapse old messages in the live pane.
src/content/features/collapse.js,src/lib/collapse-model.js.
A single shared MutationObserver + rAF scheduler (features-bundle.js) drives
all features and only runs them while the tab is visible.
- Goal Agent (🎯 button) — give a final goal; the focused pane is the
executor, a second pane is the agent (evaluator). After each executor
turn the controller extracts only the final answer (no tool calls/reasoning)
and asks the agent — in a fresh, memory-disabled chat — whether the goal
is reached. The agent either lists what's missing (fed back to the executor)
or emits the exact marker
GOAL REACHED GOALto finish. Memory is disabled by PATCHing account settings before each agent turn. Pure logic:src/lib/goal-agent.js+src/lib/goal-loop.js; orchestration:src/goal-controller.js↔src/content/agent-runtime.js. - Telegram bridge — set a bot token + user id in settings. The background
service worker (
service-worker.js) long-polls Telegram on a 1-min alarm (survives restarts) and sends messages chunked to 4096 chars. 📤 mirrors a pane's replies to Telegram; inbound Telegram messages are routed into the executor. On goal completion it posts a report (goal, duration, message count). Granular settings: what to forward (executor/agent, all/last, tool calls, hidden, agent opinion). Pure logic:src/lib/telegram.js.
Open via the ⚙ button in the tab bar or the extension's options page. Toggle lazy panes, conversation trim (+ how many messages to keep), and every ported feature. Changes apply live.
chrome://extensions→ enable Developer mode- Load unpacked → select this folder
- Click the extension icon or open
chrome-extension://<id>/app.html
If ChatGPT changes CSP/frame behaviour, update src/rules/*.json.
npm test # unit + perf (node:test, no deps)
npm run test:perf # perf only
Covers the trim algorithm (incl. a 4000→20 message perf budget) and the pane state machine (lazy loading, focus/load transitions, persistence shaping).
- Lazy-load panes + conversation trim
- Port auto-confirm + auto-expand tool calls
- Port prompt queue with auto-send
- Port auto-collapse of old messages in the live pane
- Native chat list in the picker pane via
/backend-api/conversations - Sync pane title/url with the real chat (content-script postMessage)
Possible next steps: smart reveal of collapsed messages on scroll-up; queue keyboard shortcut; per-pane model preselect.