Telegram

Telegram Channel Skill

Operate the joelclaw Telegram channel — the primary mobile interface between Joel and the gateway agent. Built on grammy (Bot API wrapper), supports text, media, reactions, replies, inline buttons, callbacks, and streaming.

Architecture

Joel (Telegram app)
  → Bot API (long polling via grammy)
    → telegram.ts channel adapter
      → enrichPromptWithVaultContext()
        → command-queue → pi session
          → outbound router → telegram.ts send → Bot API → Joel

Key files:

• packages/gateway/src/channels/telegram.ts — channel adapter (inbound + outbound)
• packages/gateway/src/telegram-stream.ts — streaming UX (progressive text updates)
• packages/gateway/src/outbound/router.ts — response routing
• packages/gateway/src/channels/types.ts — Channel interface

SDK: grammy@1.40.0 — Bot instance at module scope, exposed via getBot().

Multi-instance poll ownership (2026-03-05): Telegram long polling now uses a Redis lease per bot token hash.

• Owner key: joelclaw:gateway:telegram:poll-owner:<tokenHash>
• Status key: joelclaw:gateway:telegram:poll-status:<tokenHash>
• Only owner polls getUpdates; non-owners stay passive/send-only and retry lease acquisition with backoff.

Conflict guard still applies for non-cooperative pollers: telegram.channel.start_failed (with conflict metadata) + telegram.channel.retry_scheduled + telegram.channel.polling_recovered.

Delivery audit

Telegram has three distinct delivery lanes:

1. Human conversation: Telegram update → command queue → Pi session → stream/router → Bot API.
2. Canonical CLI notification: joelclaw notify send → gateway Redis event queue → immediate send or Pi session → Bot API.
3. Direct system notification: producer → gateway/send.message → Redis joelclaw:outbound:messages → gateway drain → Bot API.

Both lanes emit privacy-safe OTEL receipts under one flowId. Message bodies and previews are not stored. Receipts keep only SHA-256, character count, byte count, producer, origin system, route, queue timing, reply/message IDs, and terminal status.

Canonical actions:

• telegram.inbound.accepted
• channel.delivery.queued
• channel.delivery.dequeued
• telegram.delivery.attempted
• telegram.delivery.confirmed
• telegram.delivery.failed

Query the whole lane, then narrow to one flow:

joelclaw otel search "telegram.delivery" --hours 24
joelclaw otel search "channel.delivery" --hours 24
joelclaw otel search "<flowId>" --hours 24

If the default ClickHouse adapter is unavailable, use the working Typesense mirror explicitly:

joelclaw otel search "telegram.delivery" --hours 24 --adapter typesense-otel
joelclaw otel search "<flowId>" --hours 24 --adapter typesense-otel

telegram.delivery.confirmed means Telegram's Bot API accepted the send and returned message ID(s). It does not mean Joel read the message.

The queue still uses LRANGE then DEL before delivery. A queued flow with no dequeue/terminal receipt exposes that gap, but does not recover the lost item. Reliable claim/ack plus a dead-letter queue is separate follow-up work.

Capabilities

Sending Messages

// Via channel adapter
await telegramChannel.send("telegram:7718912466", "Hello", { format: "html" });

// Direct grammy API (from telegram-stream or daemon)
const bot = getBot();
await bot.api.sendMessage(chatId, text, { parse_mode: "HTML" });

• Max message length: 4096 chars (Telegram API limit)
• Chunking: TelegramConverter.chunk() for HTML-aware splitting, chunkMessage() for raw text
• Format: markdown→HTML via TelegramConverter.convert(), with plain text fallback on validation failure
• Buttons: InlineButton[][] → inline_keyboard reply markup

Reactions (ADR-0162)

// grammy API
await bot.api.setMessageReaction(chatId, messageId, [
  { type: "emoji", emoji: "👍" }
]);

Telegram supports a fixed set of emoji reactions. Common ones: 👍 👎 ❤️ 🔥 🎉 🤔 👀 ✅ ❌ 🤯 💯

Agent convention: Include <<react:EMOJI>> at the start of a response. The outbound router strips it and calls setMessageReaction before sending text.

Replies

// grammy API — reply to a specific message
await bot.api.sendMessage(chatId, text, {
  reply_parameters: { message_id: targetMessageId }
});

Already wired in the adapter via RichSendOptions.replyTo. The agent uses <<reply:MSG_ID>> directive.

Media

Supports photo, video, audio, voice, and document sending/receiving:

// Send
await telegramChannel.sendMedia(chatId, "/path/to/file.jpg", { caption: "Look at this" });

// Receive — handled by bot.on("message:photo") etc.
// Downloads via Bot API getFile → local /tmp/joelclaw-media/
// Emits media/received Inngest event for pipeline processing

File size limit: 20MB download via Bot API (larger files need direct Telegram API).

Streaming (ADR-0160)

Progressive text updates with cursor:

import { begin, pushDelta, finish, abort } from "./telegram-stream";

// On prompt dispatch
begin({ chatId, bot, replyTo });

// On each text_delta event
pushDelta(delta);

// On message_end
await finish(fullText);

• Plain text during streaming (no parse_mode) — avoids broken HTML on partial content
• HTML formatting only on finish() — final edit with parse_mode: "HTML"
• Throttled edits: 800ms minimum between API calls
• Cursor: ▌ appended during streaming, removed on finish
• initialSendPromise awaited in finish() to prevent race conditions

Inline Buttons & Callbacks (ADR-0070)

// Send message with buttons
await sendTelegramMessage(chatId, "Choose:", {
  buttons: [
    [{ text: "✅ Approve", action: "approve:item123" }],
    [{ text: "❌ Reject", action: "reject:item123" }],
  ]
});

// Callback handler fires telegram/callback.received Inngest event
// Then edits message to show action taken + removes buttons

Callback data max: 64 bytes. Format: action:context.

Commands

• /stop — abort current turn without killing the daemon.
• /esc — alias for /stop.
• /kill — hard stop: disables launchd service + kills process. Emergency use only.

Configuration

Currently via environment variables (migrating to ~/.joelclaw/channels.toml per ADR-0162):

Security

• Single-user lockdown — middleware drops all messages from users other than TELEGRAM_USER_ID
• No token in config — channels.toml references agent-secrets key names, not raw tokens
• Media downloads to /tmp/joelclaw-media/ with UUID filenames (no path traversal)

Troubleshooting

Bot not receiving messages

1. Check gateway is running: cat /tmp/joelclaw/gateway.pid && ps aux | grep daemon.ts
2. Check Telegram polling started: grep "telegram.*started" /tmp/joelclaw/gateway.log
3. Verify token: curl https://api.telegram.org/bot<TOKEN>/getMe
4. Check polling errors in stderr: rg "telegram.channel.start_failed|failed to start polling|getUpdates" /tmp/joelclaw/gateway.err
5. Check ownership lifecycle telemetry:
   • joelclaw otel search "telegram.channel.poll_owner" --hours 1
   • joelclaw otel search "telegram.channel.retry_scheduled" --hours 1

If you see repeated 409 conflicts, another bot process is polling the same token. Telegram phone/desktop apps are not Bot API pollers and do not cause getUpdates contention.

Messages arriving but no response

1. Check command queue: grep "command-queue\|enqueue" /tmp/joelclaw/gateway.log | tail -10
2. Check pi session health: grep "session\|prompt" /tmp/joelclaw/gateway.log | tail -10
3. Check outbound routing: grep "outbound\|response ready" /tmp/joelclaw/gateway.log | tail -10

Streaming not working

1. Verify text_delta events: grep "text_delta" /tmp/joelclaw/gateway.log | tail -5
2. Check telegram-stream lifecycle: grep "telegram-stream" /tmp/joelclaw/gateway.log | tail -10
3. Common issue: model does tool calls before text → no deltas until after tools complete
4. Race condition fix: initialSendPromise in finish() (commit 175c6ca)

HTML formatting broken

1. Check converter output: TelegramConverter.convert(text) + .validate(result)
2. Fallback: adapter auto-strips HTML and sends plain text if validation fails
3. Streaming path sends plain text (no parse_mode), only finish() adds HTML

Related ADRs

• ADR-0042 — Media download pipeline
• ADR-0070 — Inline buttons and callbacks
• ADR-0160 — Telegram streaming UX
• ADR-0162 — Reactions, replies, and channel configuration