stitchkit 0.100.1 → 0.100.3

package/llms/telegram-screens.txt

@@ -0,0 +1,365 @@
+# stitchkit/telegram/screens — stitchkit docs slice
+
+> Generated from the `docs/` tree by `bun run gen:llms`. The index of all slices is
+> `llms.txt` in this package.
+
+Contains: Guide: Telegram screens  (docs/guide/telegram-screens.md) · API reference: `stitchkit/telegram/screens`  (docs/api/reference.md)
+
+==============================================================================
+# Guide: Telegram screens  (docs/guide/telegram-screens.md)
+==============================================================================
+
+# Telegram screens
+
+`stitchkit/telegram/screens` turns a bot's menus into declared screens. A screen
+says what the chat should show — a message or several, with their buttons — and
+the runtime decides what to send, what to edit in place and what to delete. It
+is grammY middleware: it answers the presses and messages that belong to a
+screen and hands every other update to the next handler. Why it is shaped this
+way: [ADR 0211](../decisions/0211-telegram-screens-are-declared-views-over-grammy.md).
+
+```bash
+bun add grammy zod
+```
+
+grammY is the bot's own; screens use its `ctx.api` and name its types, and
+nothing imports grammY at run time.
+
+## A menu, end to end
+
+A bot that lets its owner manage projects: a list, a project's page, a
+two-message greeting screen and a screen that waits for the new greeting.
+
+```ts
+import { back, html, link, telegramScreens } from 'stitchkit/telegram/screens'
+import { z } from 'zod'
+
+export const tg = telegramScreens<BotContext>()
+
+export const home = tg.screen('/')
+  .load((c) => listProjects(c.ctx.from?.id))
+  .view(({ data: projects }) => ({
+    text: html`<b>Your projects</b>`,
+    keyboard: [
+      ...projects.map((project) => [link(project.name, projectPage, { projectId: project.id })]),
+      [link('➕ New project', newProject)],
+    ],
+  }))
+
+// A group shares a path prefix, a params schema and a load. The load runs
+// before every screen under it — its load, its actions, its input — and is
+// where access is checked: params arrive from a button, so they are input.
+const owned = tg.group('/project/:projectId', { params: z.object({ projectId: z.uuid() }) })
+  .load(async (c) => {
+    const project = await findProject(c.params.projectId)
+    return project?.ownerId === c.ctx.from?.id ? project : c.go(home).toast('Project not found')
+  })
+
+export const projectPage = owned.screen('/')
+  .action('archive', async (c) => {
+    await archiveProject(c.data.id)
+    return c.toast('Archived').go(home)
+  })
+  .view(({ data: project, params, act }) => ({
+    text: html`📁 <b>${project.name}</b>\n\nStatus: ${project.status}`,
+    keyboard: [
+      [link('👋 Greeting', greeting, params)],
+      project.status === 'active' && [act.archive('🗄 Archive')],
+      [back('« Back')],
+    ],
+  }))
+
+export const greeting = owned.screen('/greeting')
+  .action('clear', (c) => clearGreeting(c.data.id))       // nothing returned: stay, redrawn
+  .action('rule', z.object({ every: z.boolean() }), (c) => setRule(c.data.id, c.input.every))
+  .view(({ data: project, params, act }) => project.greeting
+    ? [
+        { key: 'header', text: html`<b>New members receive ↓</b>` },
+        {
+          text: project.greeting,                              // a plain string is escaped
+          keyboard: [
+            [link('✏️ Edit', greetingEdit, params), act.clear('🗑 Clear')],
+            [act.rule(project.every ? 'Every message' : 'First message only', { every: !project.every })],
+            [back('« Back')],
+          ],
+        },
+      ]
+    : { text: html`🗑 <b>No greeting.</b>`, keyboard: [[back('« Back')]] })
+
+export const greetingEdit = owned.screen('/greeting/edit')
+  .on('text', async (c) => {
+    if (c.text.length < 3) return c.notice(html`❌ Too short`, { expiresInMs: 5_000 })
+    await c.pending('⏳')
+    await saveGreeting(c.data.id, c.text)
+    return c.notice(html`<blockquote>✅ Greeting saved</blockquote>`).back()
+  })
+  .view(() => ({ text: html`📝 Send the new greeting ↓`, keyboard: [[back('« Back')]] }))
+```
+
+```ts
+// bot.ts
+const screens = tg.create({
+  screens: [home, projectPage, greeting, greetingEdit],
+  storage: myStorageAdapter,      // any grammY StorageAdapter
+})
+bot.use(screens)
+bot.command('start', (ctx) => screens.open(ctx, home))
+bot.catch((error) => logger.error('update failed', { error }))
+```
+
+## The model
+
+- **A screen** is a path, an optional `load`, any number of `action` and `on`,
+  and a `view`. `load` comes first or not at all — actions and input read what
+  it returned — and `view` ends the declaration.
+- **Params** come from the path literal: `/project/:projectId` gives
+  `params.projectId`. A group's or screen's `params` schema parses them (a
+  `z.coerce.number()` makes a number). The schema must accept what it produces,
+  because a button carries the parsed value and the schema parses it again when
+  it is pressed: `z.coerce.number()` does, `z.string().transform(Number)` does
+  not and is a compile error. A `link` to a screen whose path has params must
+  carry them; a link to a screen without params takes none — both are compile
+  errors.
+- **The parent** of a screen is the nearest declared screen whose path is a
+  prefix of its own, by segments. `back()` and the `back('« Back')` button go
+  there. There is no history stack: navigation is a tree of paths, so it
+  survives a restart and never grows.
+- **`load`** returns data, or an outcome instead of data — `c.go(home).toast('…')`
+  shows another screen. A chain of such redirects stops after four steps.
+- **`view`** is a pure function of `ctx`, `params`, `data` and `act`. It returns
+  one message or several. A message has exactly one content — `text`, `rich`
+  (a Bot API rich message), or a media file id or URL (`photo`, `video`,
+  `animation`, `document`, `audio`) with an optional `caption` — plus
+  `keyboard`, `key` and, for text, `linkPreview`. Rows and buttons may be
+  `false`, `null` or `undefined`, so a conditional button is `cond && [...]`.
+- **Text** is a string, which is always escaped, or a `TelegramHtml` from the
+  `html` tag, whose interpolations are escaped. `html.raw(markup)` marks markup
+  that is already safe and `html.join(parts, separator)` joins pieces. Every
+  message is sent with `parse_mode: 'HTML'` set explicitly, so an application's
+  own `parse_mode` transformer changes nothing.
+
+## Outcomes
+
+A load, an action and an input handler answer with an outcome, built by
+chaining; returning nothing is `stay()`.
+
+| Builder | Does |
+|---|---|
+| `c.stay()` | show this screen again, with fresh data |
+| `c.go(screen, params?)` | show another screen |
+| `c.back()` | show the parent screen |
+| `c.toast(text, { alert? })` | answer the press — up to 200 characters; the last toast wins |
+| `c.notice(content)` | a receipt that stays: posted, and the screen moves below it |
+| `c.notice(content, { expiresInMs })` | a remark under the screen that removes itself — `expiresInMs` up to 2³¹ − 1, a timer's limit |
+
+A remark is removed when its time comes, on the chat's next update after that,
+before anything new is sent into the chat and when the chat leaves its screen —
+so a screen is never split by a remark. A chat keeps at most 100 remarks; the
+oldest is removed to make room for the next. After a message there is no press
+to answer, so an input handler has no `toast`.
+
+## What the chat shows: reconciliation
+
+Each message of a view has a key: `key` when given, otherwise `'main'` for the
+last message and `s0`, `s1`, … for the ones before it. Between two views,
+messages with the same key are the same message, and only what changed is
+edited:
+
+| From → to | Call |
+|---|---|
+| nothing changed | none |
+| only the keyboard | `editMessageReplyMarkup` |
+| text → text, rich → rich | `editMessageText` |
+| text or rich → media, media → media | `editMessageMedia` |
+| a caption | `editMessageCaption` |
+| media → text or rich, text ↔ rich | delete and send |
+
+Order holds: once a message has to be sent, every message after it is sent
+too, and a message is deleted only after its replacement is in the chat. So
+"header + body" → "body" edits the body in place and deletes the header, and
+the reverse sends both and deletes the old body.
+
+When Telegram refuses a step:
+
+- **"message is not modified"** is success.
+- **A message gone or not editable** — deleted by the user, too old — is sent
+  again with every message after it.
+- **A deletion refused** (Telegram deletes only messages younger than 48 hours)
+  strips the message's buttons instead and reports it as `message-kept`. A
+  message the user already deleted counts as deleted.
+- **Anything else** — a rate limit, a network failure, HTML Telegram cannot
+  parse — stops the update, and the transition does not happen: the record stays
+  on the screen the chat was on, so its input still answers, and lists every
+  message the chat now shows, in the chat's order, with the buttons of both
+  renders. The next update repairs the rest. A receipt or a remark Telegram
+  refuses stops the update the same way, with the view already drawn recorded.
+
+A view is checked before anything is sent: at most 100 messages, keys of 1 to
+64 characters, at most 100 buttons whose address is kept in the record. A view
+past these fails in the handler's place, with the chat untouched.
+
+## Buttons and presses
+
+A screen button's `callback_data` is
+`‹prefix›‹screen id›[.‹action›](:‹param›)*(:‹key›=‹value›)*`. Values are typed,
+a UUID takes 24 bytes, and data that would pass Telegram's 64 bytes is kept in
+the chat's record behind a token derived from it — the same button renders the
+same token, so an unchanged keyboard is never edited. A screen's id is a
+short hash of its path, or its explicit `id` — give one to a screen whose path
+may change, so buttons already in chats keep working.
+
+A press is answered exactly once, after the outcome is known and before the
+chat is edited. It is trusted only as far as Telegram vouches for it: the
+`callback_data` must be on the keyboard of the message it came with, otherwise
+the press was forged (a client can send any bytes) and goes to `onStale`.
+
+A press on a message the record does not know — an older menu higher up — is
+answered in place: a one-message view takes that message over, a longer view is
+sent below it and the pressed message is deleted. Either way the previously
+recorded view is removed: one live menu per chat.
+
+For a message the screens do not own — a notification from the backend —
+`screens.button(link('🔑 Refresh token', refreshToken, { botId }))` makes a
+button that opens the screen below and leaves the notification as it is.
+
+## Input
+
+`.on(kinds, handler, { keepMessage? })` takes a message while the screen is
+shown. Kinds: `text`, `rich`, `photo`, `video`, `animation`, `document`,
+`audio`, `voice`, `video_note`, `sticker`, `location`, `contact`. A screen may
+have several `.on`; a message of a kind no `.on` takes, and every command
+(`/start`), go to the next handler.
+
+The user's message is deleted **before** any load or the handler runs — input
+can be a secret, such as a bot token, and a load may redirect or fail — unless
+`keepMessage: true`; then the view is sent below the message. A message for a
+screen whose record cannot be read is not passed on to the next handler: the
+update fails, and `onError` decides. `c.pending(text)` shows a "working on it" message that is
+removed when the handler ends, however it ends. `c.message` is the message as
+Telegram sent it; it is never stored.
+
+## Opening a screen
+
+`screens.open(ctx, screen, params?, { previous })` shows a screen as a new
+message at the bottom of the chat, from a command, a deep link or a handler.
+The previous view stays (`'keep'`, the default) or is deleted (`'delete'`).
+Called from inside a handler of the same chat, it runs at once rather than
+waiting behind the update that called it, and the screen it opened stands: the
+handler may still answer the press with a toast, and returning a navigation or
+a notice as well is an error — drawn after it, the handler's own outcome would
+put the record back on the screen the chat left.
+
+## State and concurrency
+
+One record per chat, in any grammY `StorageAdapter<ScreenChatState>`, under
+`screens:<chat id>` by default (`chatKey` changes it; it never collides with a
+grammY session). It holds the shown screen, its params, message ids with
+fingerprints of their content, overflow tokens and expiring remarks — never
+message text or input. A record from another schema version is set aside
+(`state-discarded`); a storage failure is an update failure.
+
+Updates of one chat are handled one at a time within a process. Two processes
+serving one chat — releases overlapping, a webhook fan-out — are not
+serialised: the last write wins. The record is written after Telegram answers,
+so it never names a message that was not sent — and it is written when Telegram
+refused halfway too.
+
+## Configuration
+
+| Option | Default | Does |
+|---|---|---|
+| `screens` | — | the declared screens; ids, paths and access are checked at `create` |
+| `storage` | — | where each chat's record survives a restart |
+| `chatKey` | `screens:<chat id>` | the storage key of a chat's record |
+| `callbackPrefix` | `~` | what every screen button's `callback_data` starts with |
+| `chats` | `'private'` | `'all'` answers in groups too, where any member can press or type |
+| `onStale(c)` | answered silently | a press nothing can answer, with its `reason`; its outcome is shown |
+| `onError(error, c)` | none | a load, action or input that threw; its outcome is shown |
+| `onEvent(event)` | none | `transition`, `state-discarded`, `message-kept`, `sweep-failed` (a remark's timer could not delete it) — never message text |
+| `clock` | `Date.now` | the time expiring remarks are measured by |
+
+Without `onError` the press is answered silently and the error reaches
+`bot.catch` — which a bot must set. `screens.close()` cancels pending expiry
+timers on the way down; a remark left behind is removed on the chat's next
+update.
+
+## Security
+
+- **A group's `load` checks access** for every screen, action and input under
+  it. A param added below the last load is unchecked until a load of its own
+  runs: `create` refuses a screen whose params no load sees.
+- **Never `html.raw(c.text)`** — or anything else a user wrote. `html` escapes
+  interpolations; `raw` is for markup the application built.
+- Private chats only by default: in a group anyone could press an owner's menu.
+- A forged `callback_data` never reaches a handler; params and action input
+  still go through their schemas.
+
+## Testing
+
+`createScreenTestChat(bot)` takes the application's own `Bot` — its middleware,
+commands and error boundary in their real order — and answers its Bot API calls
+from memory. Updates go through `bot.handleUpdate`, and an error reaches
+`bot.catch` as it does under long polling.
+
+```ts
+import { Bot, MemorySessionStorage } from 'grammy'
+import { createScreenTestChat } from 'stitchkit/telegram/screens'
+
+const chat = createScreenTestChat(makeBot(new MemorySessionStorage()))
+await chat.send('/start')
+await chat.press('Alpha')
+expect(chat.messages.at(-1)?.text).toBe('📁 <b>Alpha</b>')
+expect(chat.calls.map((call) => call.method)).toEqual(['sendMessage', 'answerCallbackQuery', 'editMessageText'])
+```
+
+It keeps the Telegram rules screens depend on — an edit that changes nothing is
+"not modified", a deleted message cannot be edited, HTML it cannot parse, text
+past 4096 characters, a caption past 1024 and `callback_data` past 64 bytes are
+refused — and offers `failNext` to
+refuse a call, `deleteByUser`, `pressData` for a forged press and
+`restart(bot)` for a new process over the same storage.
+
+## Not included
+
+Reply keyboards, inline mode, media uploads (`InputFile`) in a view, showing a
+screen without an update (use `screens.button` in a notification), converting
+an incoming message to HTML, and a history stack.
+
+==============================================================================
+# API reference: `stitchkit/telegram/screens`  (docs/api/reference.md)
+==============================================================================
+
+## `stitchkit/telegram/screens`
+
+A bot's menus as declared screens over grammY: a screen says what the chat
+should show, and the runtime decides what to send, edit and delete. Server-only,
+Bun or Node; `grammy` and `zod` appear in the declarations, and nothing imports
+grammY at run time — the bot's own `ctx.api` and a grammY `StorageAdapter` are
+the whole dependency. → ADR 0211 — [guide](../guide/telegram-screens.md)
+
+| Export | Kind | Summary |
+|--------|------|---------|
+| `telegramScreens` | function | `telegramScreens<Ctx>()` — the root the bot's screens are declared from; the context type is named once |
+| `TelegramScreensRoot` / `ScreenScope` | class | `screen(path, options?)`, `group(prefix, { params?, id? })`, and on the root `create(config)` |
+| `ScreenGroupBuilder` / `ScreenBuilder` / `ScreenBodyBuilder` | class | the chain: `load` (first or never) → `action` / `on` → `view`, which ends the declaration |
+| `ScopeOptions` / `ScreenOptions` / `ParamsSchemaFor` / `ActionInputSchema` | _type_ | a group's `params` schema and `id`; a screen's explicit `id`; the schema a path's params accept; an action's flat input schema |
+| `TelegramScreen` / `AnyTelegramScreen` | class / _type_ | a declared screen, typed by its context, params, data and actions |
+| `ScreenLoadContext` / `ScreenActionContext` / `ScreenInputContext` / `ScreenViewContext` | _type_ | `ctx`, `params`, `data`; an action's `input`; an input's `message`, `text` and `pending(text)`; a view's `act` buttons |
+| `ScreenParams` / `ScreenParamValue` / `ScreenPathParams` / `ScreenLinkArgs` | _type_ | params parsed from the path literal (or its schema); the argument a `link` or `go` takes — required exactly when the path has params |
+| `ScreenActions` / `ScreenActionButtons` / `NoActionInput` | _type_ | action name → input; `act.<name>(label, input?)` typed by the action's schema |
+| `ScreenInputKind` / `ScreenInputOptions` | _type_ | `text` `rich` `photo` `video` `animation` `document` `audio` `voice` `video_note` `sticker` `location` `contact`; `keepMessage` leaves the user's message (default: deleted before the handler) |
+| `ScreenHandlerResult` / `ScreenOutcome` / `ScreenOutcomeBuilders` / `ScreenInputOutcomeBuilders` | _type_ / class | what a load, action or input answers: `stay()`, `go(screen, params?)`, `back()`, `toast(text, { alert? })` (≤ 200 characters), `notice(content, { expiresInMs? })`; `void` is `stay()`; after input there is no press, so no `toast` |
+| `ScreenNavigation` / `ScreenToast` / `ScreenNotice` / `NoticeOptions` | _type_ | the parts of an outcome; a notice without `expiresInMs` is a receipt the screen moves below, with it a remark under the screen that removes itself |
+| `ViewMessage` / `ScreenViewResult` / `TextViewMessage` / `RichViewMessage` / `PhotoViewMessage` / `VideoViewMessage` / `AnimationViewMessage` / `DocumentViewMessage` / `AudioViewMessage` / `ScreenMediaKind` | _type_ | one message or several; exactly one content (`text`, `rich`, or a media file id or URL with `caption`), plus `keyboard`, `key` and `linkPreview` |
+| `link` / `back` / `ScreenButton` | function / class | a button to a screen with its params; a button to the parent screen |
+| `Keyboard` / `KeyboardRow` / `KeyboardButton` / `PlainKeyboardButton` / `ButtonLabel` | _type_ | rows of screen buttons and plain Telegram buttons without `callback_data`; `false`, `null` and `undefined` are skipped; a label is text or `{ text, style?, icon_custom_emoji_id? }` |
+| `html` / `TelegramHtml` / `HtmlTag` / `HtmlValue` / `TelegramText` | function / class / _type_ | Telegram HTML with every interpolation escaped; `html.raw()` for markup already safe, `html.join()`; a plain string is always escaped (`escapeTelegramHtml` is `stitchkit/telegram/html`'s) |
+| `ActionValue` | _type_ | what a button can carry: a string, a finite number, a boolean or `null` |
+| `TelegramScreensConfig` | _type_ | `screens`, `storage`, `chatKey` (default `screens:<chat id>`), `callbackPrefix` (default `~`), `chats` (`'private'` by default, or `'all'`), `onStale`, `onError`, `onEvent`, `clock` |
+| `TelegramScreens` / `OpenArgs` / `OpenOptions` | _type_ | `bot.use(screens)`; `open(ctx, screen, params?, { previous: 'keep' \| 'delete' })`; `button(link(…))` for a message the screens do not own; `close()` cancels expiry timers |
+| `ScreenStaleContext` / `ScreenStaleReason` / `ScreenErrorContext` / `ScreenTrigger` | _type_ | a press nothing can answer — `malformed` / `forged` / `inaccessible` / `unknown-token` / `unknown-screen` / `invalid-params` / `unknown-action` / `invalid-input`; an error with its trigger `open` / `press` / `input` |
+| `ScreenEvent` / `ExecutionReport` | _type_ | `transition` (trigger, path templates, duration, messages sent / edited / kept / deleted), `state-discarded`, `message-kept`, `sweep-failed`; never message text |
+| `ScreenChatState` / `ScreenChatStateSchema` / `TelegramScreenStorage` | _type_ / schema | one record per chat — the shown view and expiring remarks, no message text; any grammY `StorageAdapter<ScreenChatState>` |
+| `createScreenTestChat` / `ScreenTestChat` / `TestChatOptions` / `TestChatMessage` / `TestChatCall` / `TestMessageKind` | function / _type_ | a chat with the application's own `Bot` and no Telegram: `send`, `sendPhoto`, `press(label)`, `pressData`, `failNext`, `deleteByUser`, `restart(bot)`; `messages`, `calls`, `answers` |
+
+---

package/llms/telegram.txt

@@ -5,6 +5,9 @@
 
 Contains: Guide: Telegram bots  (docs/guide/telegram.md) · API reference: `stitchkit/telegram`  (docs/api/reference.md)
 
+Also read:
+- see `llms/telegram-screens.txt` for the Telegram screens guide
+
 ==============================================================================
 # Guide: Telegram bots  (docs/guide/telegram.md)
 ==============================================================================
@@ -21,6 +24,8 @@ way: [ADR 0201](../decisions/0201-a-telegram-bot-is-assembled-from-primitives.md
 | Job | Where |
 |---|---|
 | the bot as resources: command menu, then long polling with updates admitted by batch | `grammyBotResources` — [application kernel](application-kernel.md#optional-grammy-adapter) |
+| a bot on a webhook: set only where it is this process's, updates recorded before Telegram is answered | `claimTelegramWebhook`, `createTelegramUpdateIntake`, `receiveTelegramWebhook` — [below](#a-bot-on-a-webhook) |
+| text for Telegram's HTML: cleaned, cut to the limit, checked | `stitchkit/telegram/html` — [below](#message-markup) |
 | the process journal | `createJsonLogger` from `stitchkit/observability` |
 | the operator's chat: product events, apart from the journal | `createTelegramOperatorChannel` |
 | a broadcast that survives crashes, deploys and Ctrl-C | `runTelegramBroadcast` |
@@ -28,6 +33,7 @@ way: [ADR 0201](../decisions/0201-a-telegram-bot-is-assembled-from-primitives.md
 | why a send was refused | `classifyTelegramSendFailure` — [auth and errors](auth-and-errors.md#telegram-mini-apps) |
 | Mini App `initData` | `verifyTelegramInitData` — [auth and errors](auth-and-errors.md#telegram-mini-apps) |
 | one Bot API call without a bot library | `callTelegramBotApi` |
+| menus: screens, buttons, actions, input and message editing | `stitchkit/telegram/screens` — [Telegram screens](telegram-screens.md) |
 
 ## The journal
 
@@ -51,6 +57,24 @@ export const logger = createJsonLogger({
 logger.error('Application startup failed', { error })
 ```
 
+A provider's key, a webhook secret or a database password has no shape a
+pattern could guess, so the journal also masks exact values — and the
+environment already names them. `secretValuesFromEnv` collects every variable
+named as a token, secret, key, password or credential, the password and secret
+query values inside any URL variable (`DATABASE_URL`), and each URL-encoded:
+
+```ts
+import { createJsonLogger, secretValuesFromEnv } from 'stitchkit/observability'
+
+export const logger = createJsonLogger({
+  sensitiveUrlPatterns: [TELEGRAM_BOT_TOKEN_PATTERN],
+  sensitiveValues: secretValuesFromEnv(process.env),
+})
+```
+
+A key added to the environment is masked from its first line; a URL's user name
+is left alone, since it names a role that appears in ordinary text.
+
 The error line contract, for whatever recognises a failure from the journal: a
 line whose `level` is `50` or more is an error. There is no separate `fatal`;
 an unrecoverable end is an `error` line followed by a non-zero exit.
@@ -79,8 +103,32 @@ operators.post(`New user ${user.id}`, 'users')
 
 Every message is masked before it leaves: a bot token always — bots post their
 errors here, and an error about a local Bot API file carries the token in its
-path — plus any `sensitivePatterns` the bot adds; `telegramOperatorSender` also
-masks its own exact token. What to post stays the bot's. On the way down, drain it within the grace period
+path — plus any `sensitivePatterns` and exact `sensitiveValues` the bot adds;
+`telegramOperatorSender` also masks its own exact token, and with `parseMode:
+'HTML'` repairs the markup — a text cut at Telegram's limit never leaves a tag
+open.
+
+A failure that repeats a hundred times, or a storm of a hundred different ones
+when a database goes down, is a line with a count for the person reading, not a
+hundred messages. `dedupe` sends one message per fingerprint per window — the
+topic and the text with numbers, UUIDs and long identifiers blanked — and the
+next one of that kind carries how many were held back; past `maxPerWindow`
+messages in a window the rest are counted onto the next message sent:
+
+```ts
+const operators = createTelegramOperatorChannel<'errors'>({
+  chatId: env.OPERATOR_CHAT_ID,
+  send: telegramOperatorSender({ token: env.BOT_TOKEN }),
+  sensitiveValues: secretValuesFromEnv(process.env),
+  dedupe: {
+    windowMs: 10 * 60_000,
+    maxPerWindow: 20,
+    repeatedLine: (count) => `Same again: ${count}`,
+  },
+})
+```
+
+Held-back messages reach `onDropped` as `repeated` or `over-budget`. What to post stays the bot's. On the way down, drain it within the grace period
 and close it: `drain: (context) => operators.drain(context.signal)`,
 `close: () => operators.close()` in a resource.
 
@@ -124,6 +172,102 @@ application (the application's signal, with admission held by the caller).
 The recipient state contains Telegram ids: keep it in the state directory, out
 of git.
 
+## Message markup
+
+`stitchkit/telegram/html` is Telegram's HTML parse mode as a tree, with no peer
+and no DOM, so a server sending a message and a browser previewing it share one
+parser. Text written by someone else — a model, an editor, an owner's greeting —
+becomes markup Telegram accepts: synonyms renamed (`strong` → `b`), block tags
+turned into line breaks, a link kept only with an `href` Telegram opens,
+anything Telegram forbids (a tag it does not know, formatting inside `code`, a
+quote inside a quote) left as its words.
+
+```ts
+import { sanitizeTelegramHtml, splitTelegramHtml } from 'stitchkit/telegram/html'
+
+for (const part of splitTelegramHtml(summary)) {
+  await ctx.reply(part, { parse_mode: 'HTML' })
+}
+```
+
+Telegram's limits — 4096 for a message, 1024 for a caption (`{ limit:
+TELEGRAM_CAPTION_LIMIT }`) — count the text after the markup is parsed, so a
+text is cut by what the reader sees, not by the length of its tags. A cut
+prefers a paragraph, then a line, then a space; it never splits a character or
+a custom emoji, and an element open across it is closed in one part and opened
+again, attributes and all, in the next. `truncateTelegramHtml` keeps one part
+with an ellipsis; `telegramHtmlText` is what the reader sees; `parseTelegramHtml`
+gives the nodes to render a preview or an email from. `checkTelegramHtml` says,
+without repairing anything, what Telegram would refuse — for an editor that
+should warn a person rather than change their text.
+
+## A bot on a webhook
+
+A webhook is one place in all of Telegram, and `setWebhook` takes it without
+asking whose it was: a staging host or a developer's machine started with the
+production token takes the production bot with one call. `claimTelegramWebhook`
+sets it only where it already points here — address *and* secret — or where the
+bot names the current owner's host as the one it takes over from (`'none'` when
+there is no webhook: a bot on long polling is somebody's bot too). Telegram never
+returns the secret, so the address carries an `owner` tag derived from it; a
+copy of the production address with another secret is refused as
+`other-secret` instead of quietly installing its own.
+
+```ts
+import {
+  checkTelegramWebhook,
+  claimTelegramWebhook,
+  createTelegramUpdateIntake,
+  receiveTelegramWebhook,
+  sqliteTelegramUpdateStore,
+} from 'stitchkit/telegram'
+import { Database } from 'bun:sqlite'
+import type { Update } from 'grammy/types'
+
+const webhook = { token: env.BOT_TOKEN, url: env.WEBHOOK_URL, secret: env.WEBHOOK_SECRET }
+
+const intake = createTelegramUpdateIntake<Update>({
+  store: sqliteTelegramUpdateStore({ database: new Database(env.UPDATES_DB) }),
+  handle: (update) => bot.handleUpdate(update),
+  onFailure: (failure) => logger.warn('Update failed', { ...failure }),
+})
+
+// start: set the webhook, then handle what was recorded before the restart
+await bot.init()
+await claimTelegramWebhook({ ...webhook, takeoverFrom: env.WEBHOOK_TAKEOVER_FROM })
+await intake.start()
+
+// the route
+const route = (request: Request) =>
+  receiveTelegramWebhook(request, { secret: webhook.secret, accept: intake.accept })
+
+// stop admission: `await intake.close()` waits for the handlers in flight
+```
+
+A refusal names the owner's host, never its path — a foreign webhook's path may
+itself be a secret — and says what to pass: set the takeover for one start and
+remove it. `checkTelegramWebhook` on a timer notices a webhook that moved
+elsewhere, which this process cannot see any other way.
+
+**The update is recorded before Telegram is answered.** A handler that runs
+longer than Telegram waits — a large file from a local Bot API server — gets the
+same update again while it still works, and a process that answers and then
+dies loses the update for good. `receiveTelegramWebhook` answers 200 once the
+intake recorded the body; the handler runs after, outside the request. Updates
+of one chat run in order, other chats alongside (`maxConcurrent`, 32). An
+attempt holds a lease it renews while the handler lives; a sweep takes over what
+was left — `pending` whose process never got to it (`pendingGraceMs`), `failed`
+whose retry is due, `processing` whose process died. A failure is retried with
+backoff or after Telegram's `retry_after`; one Telegram said repeating cannot
+fix (a refused message, a blocked user) is abandoned; `maxAttempts` (5) ends
+the rest. Delivery is at least once: a handler that finished but could not be
+recorded as finished runs again.
+
+The store is an interface of six atomic steps (`TelegramUpdateStore`), so the
+bot's own database can hold it; `sqliteTelegramUpdateStore` takes a
+`bun:sqlite` or `node:sqlite` handle as it is, and `memoryTelegramUpdateStore`
+is for tests.
+
 ## Files from a local Bot API server
 
 A local `telegram-bot-api` keeps downloaded files under `<dir>/<bot token>/`,
@@ -179,11 +323,23 @@ injected grammY bot. → ADR 0143, ADR 0201 — [guide](../guide/telegram.md)
 | `TelegramBroadcastConfig` / `TelegramBroadcastSend` / `TelegramBroadcastReport` / `TelegramBroadcastRunOutcome` / `TelegramBroadcastOutcome` / `TelegramBroadcastRecipient` | _type_ | `name`, `directory`, `recipients()`, `send({ recipient, attempt })`, `maxAttempts`, `signal`, `onProgress`; counts `delivered` / `unreachable` / `failed` / `uncertain` / `pending` and `finished` / `stopped` / `halted` / `dry-run` |
 | `telegramBroadcastSender` / `TelegramBroadcastSenderConfig` / `TelegramBroadcastMessage` | function / _type_ | the standard `send`: `sendMessage` with a text or `copyMessage` of a prepared message, through the Bot API |
 | `createTelegramOperatorChannel` | function | the operator chat, apart from the journal: `post(text, topic?)` returns at once and never throws; paced sends into forum topics, 429 honoured, oldest dropped on overflow, `drain(signal)` and `close()` for the way down |
-| `TelegramOperatorChannelConfig` / `TelegramOperatorChannel` / `TelegramOperatorMessage` / `TelegramOperatorDrop` / `TelegramChatId` | _type_ | `chatId`, `topics`, `send`, `maxQueued`, `minIntervalMs`, `maxAttempts`, `onDropped`, `sensitivePatterns` (a bot token is always masked); a drop is `overflow` / `refused` / `attempts` / `closed` with its classification |
-| `telegramOperatorSender` / `TelegramOperatorSenderConfig` | function / _type_ | the standard `send`: `sendMessage` into the chat and topic, plain text unless `parseMode` |
+| `TelegramOperatorChannelConfig` / `TelegramOperatorChannel` / `TelegramOperatorMessage` / `TelegramOperatorDrop` / `TelegramChatId` | _type_ | `chatId`, `topics`, `send`, `maxQueued`, `minIntervalMs`, `maxAttempts`, `onDropped`, `sensitivePatterns` and `sensitiveValues` (a bot token is always masked), `dedupe`; a drop is `overflow` / `refused` / `attempts` / `closed` / `repeated` / `over-budget` with its classification |
+| `TelegramOperatorDedupe` | _type_ | `windowMs` (10 min): one message per `fingerprint` per window, the next one sent carries `repeatedLine(count)`; `maxPerWindow` (20) across fingerprints, the overflow counted onto the next one sent with `overBudgetLine(count)`; `now` |
+| `telegramOperatorSender` / `TelegramOperatorSenderConfig` | function / _type_ | the standard `send`: `sendMessage` into the chat and topic, plain text unless `parseMode`; with `'HTML'` the markup is cleaned and cut by what it shows, so the channel's cut never leaves a tag open |
 | `createTelegramLocalFiles` | function | files the local Bot API server wrote under `<root>/<token>/`: `resolve(file_path)` and `remove(file_path)` only inside the bot's directory, judged on real paths so a link is judged by where it leads; `check()` for readiness. Recommended root variable: `BOT_API_FILES_ROOT` |
 | `TelegramLocalFilesConfig` / `TelegramLocalFiles` / `TelegramLocalFilesCheck` / `TelegramLocalFileError` / `TelegramLocalFileRefusal` | _type_ / class | `root`, `token`; refusals `root-not-absolute` / `bot-directory-unavailable` / `outside-bot-directory` / `missing` / `not-a-file`, never carrying the path or the token |
 | `callTelegramBotApi` / `TelegramBotApiCall` / `TelegramBotApiError` | function / _type_ / class | one Bot API method over `fetch`; a refusal keeps `error_code`, `description` and `parameters` for the classifier, and no error names the URL |
 | `TELEGRAM_BOT_TOKEN_PATTERN` | constant | the shape of a bot token, for a logger's `sensitiveUrlPatterns` |
+| `claimTelegramWebhook` | function | set the webhook only where it already points here, or by naming the host it is taken from; confirm Telegram holds it; pending updates kept unless `dropPendingUpdates` |
+| `ClaimTelegramWebhookConfig` / `TelegramWebhookConfig` / `TelegramWebhookClaim` | _type_ | `token`, `url`, `secret`, `apiRoot`, `fetch`, `signal`; `takeoverFrom` (a host, or `'none'`), `allowedUpdates`, `maxConnections`, `dropPendingUpdates`; the claim's tagged `url`, `takenFrom`, `pendingUpdates` |
+| `TelegramWebhookClaimError` / `TelegramWebhookRefusal` / `TELEGRAM_WEBHOOK_NONE` | class / _type_ / constant | `owned-elsewhere` / `other-secret` / `not-confirmed` with `ownerHost` — a host, never a path; `'none'` names no webhook |
+| `telegramWebhookUrl` | function | the address with an `owner` tag derived from the secret: equal for one secret, different for another, the path untouched |
+| `checkTelegramWebhook` / `TelegramWebhookState` | function / _type_ | whether the webhook still points here: `owned`, `ownerHost`, `pendingUpdates`, Telegram's `lastError` |
+| `receiveTelegramWebhook` / `ReceiveTelegramWebhookOptions` / `TelegramUpdateAcceptance` | function / _type_ | one webhook `Request`: 403 without the secret (compared in constant time), 405 for another method, 400 for no update, 200 once `accept` recorded it |
+| `createTelegramUpdateIntake` | function | webhook updates recorded before Telegram is answered and handled after: chat order kept, other chats alongside up to `maxConcurrent`, a renewed lease per attempt, a sweep that takes over what a restart or a dead process left, retries, `maxAttempts`, finished updates remembered `retainMs` to refuse repeats; at least once |
+| `TelegramUpdateIntakeConfig` / `TelegramUpdateIntake` / `TelegramUpdateEnvelope` / `TelegramUpdateFailure` / `TelegramUpdateStoreStep` | _type_ | `store`, `handle`, `retry`, `maxAttempts` (5), `leaseMs` (120 000), `pendingGraceMs` (60 000), `sweepEveryMs` (30 000), `retainMs` (24 h), `maxConcurrent` (32), `onFailure`, `onStoreError`, `now`; `accept(body)`, `start()`, `sweep()`, `idle()`, `close()` |
+| `TelegramUpdateStore` / `StoredTelegramUpdate` / `TelegramUpdateState` / `TelegramUpdateSettlement` / `TelegramUpdateClaimOptions` / `TelegramUpdateDueQuery` | _type_ | the store an intake runs on: `add`, `claim`, `renew`, `settle`, `due`, `prune`, each atomic on one record — a claim takes `pending`, a due `failed` or a lapsed `processing`; renew and settle only by the attempt holding it |
+| `memoryTelegramUpdateStore` | function | the store in this process's memory — tests, or a bot that accepts losing unhandled updates to a restart |
+| `sqliteTelegramUpdateStore` / `SqliteTelegramUpdateStoreConfig` / `TelegramSqliteDatabase` | function / _type_ | the store in SQLite over the handle the application opened — `bun:sqlite` `Database` or `node:sqlite` `DatabaseSync`; `table` (default `telegram_updates`) created when missing |
 
 ---