Webcore：单文件 Cloudflare Worker 自建聊天应用，跑在 Workers AI 上

Webcore

A single-file Cloudflare Worker that serves its own chat app and runs it on Workers AI: conversations saved in D1, five free models, optional web search, optional step-by-step thinking, and regenerated answers that stay reachable as switchable versions.

https://webcore.workers.dev/

• worker.js: the Worker, chat UI included. Deploy this file as-is; no build step, no dependencies.

Deploy

Paste the file into the Worker editor in the Cloudflare dashboard and deploy. The free tier is enough. No Wrangler is needed.

The Worker needs two bindings (Settings, Bindings in the dashboard) or it cannot start:

The variable names must match exactly. With a binding missing you get Database unavailable (503) or an AI error instead of a chat.

To update, paste the new file over the old one and deploy. The page is cached for up to 5 minutes, so a hard refresh may be needed before the new UI shows.

Optional: web search

The globe button only works once a search key exists. Add one as a secret (Settings, Variables and Secrets):

Unset, the app behaves exactly as before: turning search on shows an error toast explaining how to add a key, and the question is answered without search.

Optional: model license

Some models need their terms accepted once. If you see Model license not accepted, open the dashboard, AI, Models, open that model, and agree.

Using it

Open the Worker's URL. You land on the welcome page ("What can I help with?"); opening the site never opens an old chat and never creates one. Pick a conversation in the sidebar to continue it, or just type: the first message automatically creates a conversation called "New Chat". On a computer, Enter sends and Shift+Enter adds a new line. On a phone or tablet, Enter adds a new line and the Send button sends.

Versions: sending a message after switching to an older version continues from that version, the same as claude.ai.

Neuron bar: the bar at the bottom of the sidebar shows today's usage against the daily cap (DAILY_NEURON_LIMIT, default 10,000, UTC day). Requests that would exceed it are refused with the remaining amount.

HTML replies: a reply containing an HTML file gets a block with Code and Preview tabs, Copy and Download. The preview runs in a sandboxed iframe.

How it works

Branching. Every message row has a parent_id, and each parent stores which child is active (active_child_id). Regenerating inserts a new sibling under the same parent and never overwrites anything, so every version stays reachable after a reload or from another device.

Long-conversation memory. The last 30 messages go to the model word-for-word. Older ones are folded into a running summary stored on the conversation, refreshed after 20 more older messages pile up. Anything the summary does not yet cover is sent verbatim, so no message falls into a gap.

Server side.

• Tables are created and migrated once per isolate; a failed init is retried by the next request. Migrations only run when a column is actually missing, and indexes are created after them so older databases do not break.
• All system messages are merged into one, because several models reject more than one.
• Reasoning tags are removed from earlier replies before they are sent back to the model. Otherwise a reply made with thinking on teaches the model to keep "thinking" later, including when you regenerate with thinking off.
• One retry for transient AI errors. License, 4xx and quota errors are not retried. Empty model replies become a clear error instead of a blank message.
• Work is registered with ctx.waitUntil, so a reply is still saved if the browser disconnects.
• If an interrupted write leaves a message without an active child, the newest child is followed instead of cutting the conversation off.
• The neuron count comes from the model's usage report when it has one, otherwise from an estimate.
• Sources appear once: any "Sources" or "References" list the model writes itself is removed before the real list is appended, and the stored list is removed from history so the model does not copy it next time.
• Search results are used for that one call only and never stored. Each snippet is capped at 600 characters and the model is told to treat results as untrusted.
• /api/chat and /api/regenerate stream newline-delimited JSON: status events ("Searching the web...", "Thinking..."), then one result. The reply itself is not token-streamed; it arrives whole.

Client side (embedded UI). A markdown renderer written to cope with the different styles the models use: headings (# and underlined), bold, italic, bold-italic, strikethrough, highlight, inline code, links, bare URLs, nested bullet and numbered lists, task lists, tables with alignment (they scroll sideways on phones), blockquotes, horizontal rules, fenced code (triple backtick or tilde, including unclosed blocks), a few safe inline tags (<br>, <sub>, <sup>, <kbd>), and backslash escapes. Everything else is shown as plain text, never as live HTML. LaTeX is shown as code, not typeset. Reasoning blocks written as <thinking> or <think> become the collapsible "Thought process". The UI also has the version navigator, the model, search and thinking toggles, theme saved in localStorage, and swipe gestures for the sidebar on phones.

Phones. Larger text and touch targets, safe-area padding for notches, a two-row input bar (model and toggles above, message and Send below), the layout follows the on-screen keyboard, the keyboard is not opened automatically, and the browser theme colour follows light/dark. A failed send restores what you typed and reloads the conversation, so no phantom message is left behind. Code blocks cut off by the token limit are still shown.

Editing and verifying worker.js

The UI is a long string of concatenated single-quoted pieces, and its JavaScript lives inside that string, so node --check worker.js does not validate it. A syntax error there breaks the whole page in the browser while the outer file still checks clean. Always run both:

cp worker.js worker.mjs
node --check worker.mjs

node -e "
import('./worker.mjs').then(async m => {
  const r = await m.default.fetch(new Request('https://x/'), {}, { waitUntil() {} });
  const html = await r.text();
  const s = html.match(/<script>([\s\S]*)<\/script>/)[1];
  new Function(s.replace(/document\.getElementById\([^)]*\)/g, '({})'));
  console.log('UI script parses');
});
"

Inside the UI string, use single quotes only for the string pieces (escape any ' in UI code), and double every backslash (\\n, \\u0000). In UI code, do not pass code text as a string to .replace(); use a function (.replace(x, function () { return y })), because $& and similar patterns in a replacement string corrupt code.

To add a model, change all three places together: FREE_MODELS, the <select> options, and the MODEL_NAMES map in the UI. If they disagree, the dropdown offers a model the server rejects.

Quick test checklist

1. Homepage opens on the welcome page ("What can I help with?"), not on an old chat, and the sidebar neuron bar fills in.
2. Send a message: a conversation called "New Chat" is created and the reply appears.
3. Reload: the conversation and reply are still there.
4. Regenerate: a 2/2 counter appears and < returns to the original.
5. Switch to version 1 and send another message: it continues from version 1.
6. Ask for a long HTML page: a Code / Preview block appears even if the reply hit the token limit.
7. Lightbulb on: a "Thought process" block appears above the answer.
8. Regenerate that reply with the lightbulb off: the old "Thought process" block disappears at once and the new reply has none.
9. Globe on with a key set: one Sources list is appended (not two). Without a key: an error toast explains how to add one.
10. Remove the AI or DB binding temporarily: you get a clear error, not a blank page.
11. Click the pencil icon on a conversation: a "Rename Chat" box opens with the current name filled in and a "Rename" button.
12. On a phone: the keyboard does not open by itself, the input bar stays above the keyboard, Enter adds a line, tables scroll sideways, and nothing is hidden under a notch.

Limitations

• No login. Anyone with the URL can read, rename and delete every conversation and use up the daily neurons, and CORS_ORIGIN is *. Put Cloudflare Access in front of the Worker, or at least set CORS_ORIGIN to your own domain, before sharing the URL.
• One shared history. There are no accounts; every visitor sees the same conversation list.
• Neuron counts are estimates. The pre-check counts only the new prompt and search text, not the history, and summary calls are not added to the daily total.
• Summaries are per conversation, not per branch. A summary made on one branch can be reused on another of similar length. One that covers more messages than the new path has is discarded.
• Pre-branching data. Messages saved before branching support all have no parent, so they show as several roots and only the first is displayed.
• Length caps. 2,000 characters in, and 800 to 1,000 tokens out depending on the model. Long answers are cut off.
• Search uses your raw message text, not a rewritten query, so follow-ups like "what about the second one?" search poorly.
• Thinking mode is prompt-based. Weaker models sometimes ignore the <thinking> format and the reasoning then shows as part of the answer.
• MAX_HISTORY in CONFIG is not used by anything.
• Unconfirmed on a live deployment (checked for syntax and tested in isolation only, not against real D1 and Workers AI): merged system messages, the AI retry, the memory gap fix, DB init and migration, waitUntil, and the three UI fixes. Use the checklist above after deploying.