chat

Minor Changes

• 115cd39: Add @chat-adapter/gitlab for merge request and issue comment threads on GitLab.com, GitLab Self-Managed, and GitLab Dedicated. It verifies webhooks with a signing token or secret token, replies inside each comment's discussion, handles emoji reactions, uploads files, and resolves message.subject to the parent merge request or issue. Registers the adapter in @chat-adapter/catalog and the create-chat-sdk CLI scaffold, and adds GitLab emoji platform support.

• a32470e: Keep literal tildes, asterisks, underscores, and backslashes as written in outgoing WhatsApp messages. A backslash escape is kept only where the text would otherwise turn into WhatsApp formatting, such as \_word\_ or a line starting with \>. Render bold, italic, and strikethrough from their AST nodes so formatting markers inside inline and fenced code stay unchanged, and send hard breaks as plain newlines. Add a handlers option to stringifyMarkdown for platform-specific serialization.

• 228b30d: Pass the current message ID to Adapter.startTyping as TypingOptions.messageId.

Minor Changes

• daaa0ac: Add @chat-adapter/catalog, a zero-dependency catalog of every adapter on chat-sdk.dev. It lists official, vendor-official, and community adapters with their npm package, maintainer, README, and capability support, plus the factory export, peer dependencies, and environment variables for official and vendor-official adapters. New helpers include listAdapters(), getCatalogEntry(), and getFeatureSupport().

  The chat/adapters subpath is now deprecated. It re-exports the original API from @chat-adapter/catalog, so existing imports keep working. Its ADAPTERS keeps alphabetical key order, and its CatalogAdapter type accepts objects without the new features and readme fields. In @chat-adapter/catalog, ADAPTERS follows chat-sdk.dev listing order and CatalogAdapter requires both fields. create-chat-sdk reads adapter metadata from @chat-adapter/catalog.

  getAdapter() and listEnvVars() now return undefined and `` for inherited object keys such as "toString", instead of returning Object.prototype members.

• fd291b8: feat(teams): let a card carry a text input

  TextInput, which modals already use, is now also a CardChild, so a user can type a reply on the card instead of opening a dialog. Teams renders it as an Adaptive Card Input.Text with the same mapping the modal converter uses, and the @chat-adapter/teams/cards subpath emits the same element. Slack, Google Chat, Discord, GitHub, and Linear show its label as text. WhatsApp, Messenger, Twilio, and X omit it, so don't make an input the only content on a card that reaches them.

  ActionEvent gains values?: Record<string, string>, the card's own input values keyed by input ID. It has the same shape as ModalSubmitEvent.values. The Teams adapter fills it from the Action.Submit payload, and it is absent on platforms with no card inputs. A button's callbackUrl POST now carries values too, so a webhook-driven flow with no onAction handler still receives what was typed.

  An input is required unless optional is set, and Teams validates every input on the card before it lets a button through, so mark it optional unless the buttons beside it should be blocked until it is filled. That validation is client-side, so treat values as user input and validate it in the handler. A Teams button with actionType: "modal" now renders with associatedInputs: "none", so it opens its dialog without waiting on the card's inputs and doesn't send them.

  Teams merges card inputs into the button's own submit data, so a card TextInput with the ID actionId, value, or msteams throws a ValidationError when the card is rendered, and the @chat-adapter/teams/cards subpath throws an Error. Modal inputs can still use actionId and value.

  CardChild gained a member, so an adapter outside this repo that exhausts the union with a never check needs a text_input case. Every in-repo converter reaches it through a default: branch instead, so none needed changing.

• 19c7aa2: Add a handler-only onNewMessage(handler) overload for messages that reach pattern routing, including attachment-only messages without text. Invalid onNewMessage arguments now throw a ChatError at registration.

• 08081fa: Reject AI tool calls that have no conversation scope when they run inside a handler with no conversation, such as an action in App Home or inside a modal, an options-load handler, or a modal opened outside a thread or channel. Pass an explicit scope, or scope: false for intentionally workspace-wide access. Calls outside any handler, such as in a cron job or script, still run workspace-wide and log a warning. An empty scope id now throws when the tools are created.

  The standalone tool factories exported from chat/ai, such as fetchMessages and postMessage, now accept scope and strictScope and follow the same rules as createChatTools. Inside a handler they inherit the conversation being handled, so a call that targets another channel is now rejected. Outside a handler they log the same unscoped warning once per tool. Passing a guard function still works but is deprecated in favor of scope.

• 7b7d73c: Add a visibleTo option to Button and LinkButton that lists the user IDs allowed to see the button. The Slack adapter renders it as visible_to_user_ids. Other adapters ignore it and show the button to everyone. It only hides the button, so action handlers still need to check who clicked.

• f8527b2: Add an optional aiGenerated field to PostableRaw, PostableMarkdown, PostableAst, PostableCard, StreamingPlanOptions, and StreamOptions. When it's true, adapters with a native AI label show it. The Teams adapter shows Teams' "AI generated" label on new, edited, channel, targeted, and streamed messages. An edit must set the flag again to keep the label. The built-in AI tools now set aiGenerated: true on the markdown and raw messages they post or edit.

• 9a83b22: Isolate Slack message concurrency and turn cancellation by installation to prevent cross-workspace bot token reuse. Adapters can now namespace dedupe, lock, queue, and turn keys with the optional Adapter.getConcurrencyScope() hook. In multi-workspace mode, messages queued under the old keys aren't migrated, and old and new workers don't exclude each other, so upgrade chat and @chat-adapter/slack together and restart all workers.

• f635dbe: Add a Twitch adapter. It reads channel chat through EventSub webhooks, replies with the Helix Chat API as a bot account, deletes messages as a moderator, and sends and receives whispers. The chat/adapters catalog and create-chat-sdk now include Twitch.

Patch Changes

• 2f1cfa2: Allow SlackAdapter and TelegramAdapter to be assigned to Adapter when exactOptionalPropertyTypes is enabled.

• 19c7aa2: Fix onNewMessage patterns with the global or sticky flag so they match on every message instead of skipping alternate ones.

• 80e7bae: Stop describing the ioredis state adapter as supporting Redis Cluster. Its client option accepts an ioredis Redis instance, which a Cluster doesn't satisfy; Sentinel is still supported.

• 08081fa: Keep modal context until the modal closes, so relatedThread and relatedChannel are still available when a user resubmits after a submit handler returns validation errors, or after an update or push response.

• 8178fdb: Fix Workflow DevKit workflows failing to start with ReferenceError: AbortController is not defined. The default thread signal is now created on first use, and threads revived inside the workflow VM get an inert signal that never aborts.

Minor Changes

• ad90432: Add opt-in selection change events for Select and RadioSelect inputs in Slack modals

• dc2a777: Add TanStack AI support. thread.post(), thread.reply(), and channel.post() accept the stream returned by TanStack's chat(), and the new chat/ai/tanstack subpath provides toTanStackMessages for thread history and createTanStackTools for the Chat SDK toolset. Neither adds a runtime dependency on @tanstack/ai; the tools require zod 4.2 or newer.

• 139d337: Support Vercel Connect in the Teams adapter with lazy app identity resolution and custom webhook verification. Ensure custom Bot Framework and Graph token factories take precedence over client-secret environment variables.

  Add Teams to create-chat-sdk --connect and interactive Connect scaffolding, including the helper, connector environment variable, and setup documentation.

• 2e2426d: Add onInstalled and onUninstalled handlers for installation lifecycle events. The Teams adapter emits them for installationUpdate activities, including the add-upgrade and remove-upgrade actions, with a persistable channelId for later proactive messages.

  The Teams adapter now sends, edits, deletes, reacts, and types against the service URL encoded in the thread ID instead of the app-wide default, so conversations hosted on regional or sovereign Bot Framework endpoints are reached correctly. An explicit apiUrl still pins every call to that endpoint.

• 8421953: feat(teams): render tables as the Adaptive Card 1.5 Table element

  The Teams adapter now renders Table as the native Adaptive Cards Table element instead of a Container of ColumnSets. Teams draws grid lines between cells, sizes columns by relative weight and marks the header row for accessibility. The @chat-adapter/teams/cards subpath emits the same element.

  Table gains optional rendering options that only Teams reads: widths (positive integer column weights), verticalAlign (vertical alignment of cell content), gridLines (default true) and gridStyle. Other adapters ignore them. Pass gridLines: false to keep a borderless table.

  The chat JSX runtime now forwards align on <Table>, matching fromReactElement.

• 056d883: Allow history.user.maxPerUser and legacy transcripts.maxPerUser to be false to disable count-based history eviction. The default remains 200 entries per user.

• fcdc1c9: Honor an adapter's definitive isMention: false instead of re-detecting mentions from the message text.

  An adapter that reads structured platform content reports true or false, and the SDK falls back to matching @username in the text only when the adapter reports nothing. A definitive false now wins, so an adapter can suppress a mention that exists only in text its platform renders literally, such as a code sample. Adapters that previously returned false to mean "not detected" should return undefined instead. Direct messages are unchanged: with no onDirectMessage handler registered they still route to onNewMention.

  Adapter changes that follow from the new contract:

  • Linear: ordinary comments are left undetermined, so SDK text detection still applies to them.

  • Notion: keyword mode leaves comments without a keyword match undetermined instead of reporting false, so @userName in the text still counts.

  • X: posts rebuilt from raw or fetched by id are left undetermined instead of reporting false. Only post.mention.create events report true.

  • Discord: a real ping, role mention, @everyone, or allowlisted channel reports true; anything else is left undetermined instead of false, so a literal @botname typed in the text still counts as before.

Patch Changes

• f233ffe: Retry initialization after a failed state connection. Once the state adapter (for example Redis) recovers, the next initialize() call or webhook tries again instead of rejecting with the original connection error. Concurrent callers still share one attempt. Adapter initialization failures remain cached until shutdown() so retries cannot restart adapters that are already running or still starting.

• 91683e5: Wait for Telegram polling handlers to settle before advancing update offsets.

• c21ccbc: Add an opt-in webhook option to propagate message, action, and slash command handler errors through waitUntil while preserving default behavior.

Minor Changes

• 2ce2be0: Add Slack Agent Sessions lifecycle support, native stop cancellation through thread.signal, automatic session titles, and session stop/title-change events while preserving the legacy assistant_view compatibility path.

• 169788b: Introduce the unified History API (bot.history) with user, thread, and channel scopes.

  bot.history.user replaces bot.transcripts for cross-platform per-user message persistence. The API surface is identical — migrate by changing the transcripts config key to history.user and updating call sites from bot.transcripts.* to bot.history.user.*. bot.transcripts remains available as a deprecated alias.

  bot.history.thread and bot.history.channel expose promise-based helpers for per-thread and per-channel message access, aligned with the existing thread.messages and channel.threads() iterators. Reads delegate to the adapter; adapters that persist history in the SDK-side store (persistThreadHistory: true) are served from that cache. An unregistered adapter prefix or an unsupported capability throws instead of returning an empty result.

  The new toPromptEntries helper converts history.user.list() entries into { role, content } messages ready for LLM prompts.

  The TranscriptEntry type is deprecated in favour of HistoryEntry. Both are exported from chat.

• 5b538f6: Keep thread locks alive while message handlers run so queue, burst, and debounce strategies remain serialized beyond the lock TTL. Renewal is capped by the new concurrency.maxLockLifetimeMs option (default 10 minutes) so a hung handler cannot block a thread forever. When the heartbeat detects that lock ownership was lost, the queue drain and debounce loops stop instead of competing with the new lock holder, and the debounce loop now keeps draining messages that arrive while a handler is running instead of stranding them until the next webhook.

Patch Changes

• 16ea171: preserve adapter-returned thread ids when editing channel messages

• eddcd7e: Return Telegram file downloads as portable ArrayBuffer data while preserving Buffer support in the shared attachment contract.

• 929878b: Allow JSX link buttons to include an explicit action ID.

• 500b7e6: enforce the conversation scope on write tools and stop trusting client-supplied message history in the web adapter

  createChatTools now runs the same scope guard on write tools that read tools already used, so a thread or channel id the model supplies that resolves outside the scoped conversation is rejected before the write executes. sendDirectMessage targets a user id rather than a conversation and stays gated by approval alone.

  The web adapter no longer treats the request body's messages array as a source of conversation state. Only the latest user message is consumed, and tool parts are stripped from it so a browser cannot inject forged tool-call or approval state. Text, file, and custom data parts pass through unchanged; a message left with no parts after stripping is rejected with HTTP 400. Prior turns come from the state adapter when persistMessageHistory is enabled.

Minor Changes

• 0f24cc3: Preserve normalized replied-to message context and populate it from Telegram replies.

• bdeb2bf: Add a workflow-safe chat/serialization entrypoint and isolate automatic Chat class serializers from Node-only runtime dependencies.

• a0cba02: Add Vercel Connect credential resolvers and custom webhook verification to the Discord adapter, with create-chat-sdk --connect scaffolding for Discord bots.

• 83ede7e: add native message replies with WhatsApp contextual reply support

• 18d4a23: Add a shared thread API for marking messages as read across WhatsApp, Messenger, and XChat.

  Note for anyone calling XchatAdapter.markAsRead() directly: it now rejects when a receipt fails instead of logging a warning and resolving. Automatic read receipts are unaffected, since the adapter still catches and logs those internally. If you call the method yourself without awaiting it, add a .catch() so a failed receipt does not surface as an unhandled rejection.

Minor Changes

• 2a2b2c5: Add a native Instagram Direct Messages adapter with signed webhooks, media, quick replies, story context, reactions, and typed Meta API errors.

• 4ac0455: Add message update and delete lifecycle callbacks, with Slack message_changed and message_deleted dispatch support.

• 0ec6a73: Add @chat-adapter/notion for Notion page and block comment discussions: webhook HMAC verification, Post+Edit streaming, conversation history, message.subject page metadata, plain-text @userName/@botUserId mention detection, and File Uploads (up to 3 native attachments). Registers the adapter in the chat/adapters catalog and create-chat-sdk CLI scaffold, and adds Notion emoji platform support.

• 85e3d22: Close residual gaps in agent read-tool scoping. createChatTools's read guard now wraps modal, assistant-thread, assistant-context, app-home, app-context, and member-joined dispatch so tools built in those handlers inherit the active conversation, and it logs a warning (instead of failing open silently) when a read runs with no resolvable scope. Scoping stays channel-level by default, so a thread scope still permits sibling threads in its channel. Pass the new strictScope: true to confine a thread scope to that thread alone, rejecting both sibling threads and the parent channel, which matters on platforms where a channel is the widest read available (a GitHub channel is an entire repo).

  Note that reads inside those newly wrapped handlers were previously unscoped. An agent built in an onModalSubmit, onAppHomeOpened, or onMemberJoinedChannel handler that reads another channel will now be rejected. Pass an explicit scope, or scope: false for intentionally workspace-wide reads.

Minor Changes

• c5d86b1: confine built-in agent read tools to the conversation being handled, with an optional scope override

• 0153a39: Add DateInput and NumberInput modal children. The Slack adapter renders them as a datepicker and a number_input, the Teams adapter as Input.Date and Input.Number, and both submitted values arrive in event.values as strings.

  Teams submit values that arrive as JSON numbers are now stringified into event.values instead of being dropped. This fixes Input.Number, but applies to any numeric value a Teams dialog submits — a key that was previously absent from event.values will now be present as a string.

Patch Changes

• 257a32d: Route Teams personal and group conversations using their explicit conversation type so group chats use buffered fallback even when their IDs resemble direct messages.

• b547f45: Stop treating email addresses as bot mentions. A message containing jane@acme.com no longer triggers a bot named acme, because the @ in detectMention must not follow a word character. Real mentions are unaffected, including at the start of a message, after punctuation, and suffixed names such as GitHub's mybot[bot].

• caa6325: Add XChat support to @chat-adapter/x, shipped from the new @chat-adapter/x/chat subpath so it sits alongside the existing X adapter. The XChat crypto stack (@xdevplatform/chat-xdk, @xdevplatform/xdk, juicebox-sdk) is an optional peer dependency, so existing @chat-adapter/x users are unaffected. All cryptography is handled inside the adapter via @xdevplatform/chat-xdk (wasm) and all REST goes through the typed @xdevplatform/xdk client. Only a bot token and a Juicebox PIN are required: the bot's identity (user id and @handle) is resolved from GET /2/users/me at startup.

  • Encrypted send/receive in DMs and groups (webhook push + polling), signature verification on by default; undecryptable or unverified events are dropped

  • Webhook POSTs must carry a valid x-twitter-webhooks-signature, which X sends on every delivery. Set consumerSecret (or X_CONSUMER_SECRET) to receive webhooks, or disableWebhookVerification when an upstream layer already verifies them. Polling deployments are unaffected

  • Mention detection from structured mention entities, swipe-replies to the bot, and a plain-text @handle fallback; group replies sent as quoted replies

  • openDM(userId) starts (or reuses) an encrypted 1:1, running a full key exchange when needed so the bot can message first

  • Media both ways: inbound attachments with lazy download+decrypt, outbound encrypted uploads

  • Edit and delete of the bot's own messages; the first edit of a fresh message is age-gated by editSafetyDelayMs (default 5000ms) so receiving clients have stored the original

  • Reactions in and out, read receipts (sendReadReceipts, default on), typing keep-alive, configurable group welcome message

  • Cards degrade to text with tappable URL/mention entities plus a URL preview attachment

  • Requests carry a chat-sdk-xchat/<version> User-Agent product token so Chat SDK traffic is identifiable in X API request logs (a User-Agent set via apiHeaders takes precedence)

  • Registered in the chat/adapters catalog and the create-chat-sdk CLI scaffold, with a new optional importPath catalog field for adapters that ship on a subpath

Minor Changes

• 4cb7e5d: Add a chat/workflow subpath with requestApproval(): durable human-in-the-loop approvals built on Workflow SDK. It posts an approval card with Approve/Deny buttons, suspends the workflow until a user decides (or an optional timeout elapses), validates approvers, finalizes the card with the outcome, and returns { approved, timedOut, user }. Also exports the buildApprovalCard and buildResolvedCard builders. Requires the new optional workflow peer dependency.

• 46681f5: Expose Microsoft Graph email addresses on normalized incoming Teams message authors. Resolved user profiles are cached in the state adapter (1 hour, failed lookups 5 minutes) so the lookup doesn't add a Graph call per message.

Patch Changes

• 80def3a: Add optional isSystem field to the normalized message Author type to distinguish platform-generated messages from humans and bots. The Slack adapter now sets isSystem: true for messages authored by Slack's reserved USLACK user (e.g. "@user archived the channel" notifications in DMs), so consumers no longer need to hard-code Slack-specific user IDs.

• 93a58af: Show explicitly configured progress as a native Teams DM status while preserving native streaming.

• 25f3099: toAiMessages no longer drops messages that have no text. A message with an empty text body is now kept when it has links or attachments the converter can include: images and text files (text/*, JSON, XML, YAML, etc.) with a working fetchData(). Messages whose only attachments are unsupported (video, audio, other file types, or attachments without fetchData()) are still skipped, and onUnsupportedAttachment now fires for video/audio attachments on these previously filtered messages.

  Note: multipart content no longer always starts with a text part. When a kept message had no text, its content array contains only attachment parts.