@intentface/latch-channels

Bring a Latch agent to the places people talk — Slack, Telegram, WhatsApp. Reuses the core runtime, storage and durability (no separate state store); identity (platform user → Principal) and channel → agent binding stay in user-land, per Latch's isolation model.

What it does

• Generic core (.) — the platform-agnostic half:
  • consumeRun(response) → AsyncIterable<RenderEvent> folded from the runtime's SSE Response; pendingApprovals / pendingQuestion read what a paused turn is waiting on from persisted history. Adding a platform is "translate RenderEvents", never "re-parse the AI stream."
  • createTurnKernel(config, surface) — everything a direct-message surface does between "a webhook arrived" and "the platform API was called", written once: dedupe, the identity gate (with stash + post-link replay), thread epochs (/new), commands, questionnaires, media folding (files inline, voice transcribed), the answer posted once at finish (split past the cap; a typing hint meanwhile), approvals and button taps back into the run. A Surface is pure protocol: send text / link / approval / question, edit or retire a card, download a file.
  • AuthResult, the lta: / ltq: button-payload codecs, splitText.
• Telegram (./telegram) and WhatsApp (./whatsapp) — a webhook parser plus a Surface each, over the kernel. telegramChannel(config) / whatsappChannel(config) verify the webhook, ACK fast, and run the turn in the background (waitUntil).
• Slack (./slack) — slackChannel(config): one agent's dedicated app; native message streaming with tool calls as task cards, initiator-owned threads, signed actor-bound approval buttons, atomic admission (claimOnce). One live run per chat comes from core's storage invariant on the shipped adapters; on a custom adapter without it, wire the optional claimTurn / releaseTurn lease — every adapter (kernel and Slack) honours it. Plus the per-agent app manifest generator, a minimal Web API client and hand-rolled Block Kit (no Chat SDK runtime dependency). Slack keeps its own engine because its surface is threads, not DMs.

Usage

import { telegramChannel } from "@intentface/latch-channels/telegram";

const channel = telegramChannel<Principal>({
  agent, botRef, botToken, webhookSecret,
  runtime: { handleChat, applyApproval, applyToolResult, compactChat, loadHistory },
  authorizeUser, seen, stashPending, askState, thread, waitUntil,
});
return channel.handleRequest(request); // one webhook route per bot

A new DM platform: implement Surface<User, Ref> and a parser from your webhook body to Inbound<User, Ref>, then createTurnKernel(config, surface).handleInbound(event).