Telegram Notifications Setup (Sentinel)

Sentinel pings the user on Telegram when a colony stalls and lets them reply to keep it going. It needs one thing: a Telegram bot token (123456:ABC-…), created via @BotFather and stored under credential id telegram. The same token sends alerts and reads the user's replies.

You do the whole thing end to end with the sentinel_setup tool (the same API the desktop connector uses): create the bot, store the token, detect the chat, point Sentinel at it, and send a test — without making the user paste anything.

Activate hive.browser-automation first — this skill assumes you know the lifecycle rules (the bridge attaches to the user's own Chrome; never launch or kill a browser), the screenshot + viewport-fraction coordinate workflow, and hive-browser interact. Telegram Web is a heavy SPA with custom rendering, so prefer the screenshot + coordinate path over selectors when the snapshot is unclear.

Telegram Web specifics (verified)

• Telegram Web defaults to the K version at https://web.telegram.org/k/. The selectors below are verified there; fall back to screenshot + coordinates if a selector doesn't match (e.g. the older /a/ version).
• Message composer: #editable-message-text — click it, type your text, send with the Enter key. Used for /newbot, /start, etc.
• Global search: the search box is top-left (around [0.13, 0.04]). Click it, type a name, then click the result row to open that chat.
• Read a bot's reply as text (don't OCR a screenshot): run hive-browser page text ".chat-message:last-child" --json in the terminal — it returns the latest message verbatim — this is how you capture the token without transposing characters.

Before you start — check what already exists

1. Run sentinel_setup({"action": "status"}). If it shows telegram already configured, the token's stored — skip to Step 3 (detect the chat) (or straight to Step 4 if the colony config is already set), don't make a new bot.
2. Confirm the user is signed into Telegram: run hive-browser open / hive-browser navigate https://web.telegram.org/k/ --json in the terminal. You cannot log in for them — you don't have their phone or SMS code. If it shows a QR / phone-number login, STOP and ask the user to log into Telegram Web in their Chrome, then continue. Don't fight the login screen or try to switch login methods yourself.

Step 1 — Create the bot with @BotFather

@BotFather is Telegram's official bot for managing bots.

1. Use in-app search to reach BotFather — not t.me/ URLs. t.me/BotFather (and hive-browser open-ing it) tends to land on the t.me web page or redirect without opening the chat. Instead: click the global search box (top-left), type BotFather, and click the verified result (blue check) to open the chat.
2. In the BotFather chat, click the composer (#editable-message-text) and send /newbot. It asks for:
   • a display name (e.g. Hive Sentinel), then
   • a username that must end in bot (e.g. hive_sentinel_<something>_bot). If the name is taken, BotFather says so — pick another until it's accepted.
3. On success BotFather replies with "Use this token to access the HTTP API:" followed by the token 123456789:AAE…. Read it by running hive-browser page text ".chat-message:last-child" --json in the terminal, not off a screenshot — the token is long and a single transposed character makes store_token fail. Copy the exact digits:rest string.

Step 2 — Store the token

sentinel_setup({"action": "store_token", "provider": "telegram", "token": "123456789:AAE…"})

This validates the token via getMe before storing and reports the bot's @username. If it's rejected, re-read the token from BotFather's message and retry — never invent or pad a token.

Prefer not to handle the token yourself? Fall back to credentials(action="collect", …) with credential_id telegram (field access_token) to pop a secure form. The default flow above is hands-off and is what the in-app "Set this up with the agent" button expects.

Step 3 — Detect the chat to notify

Telegram bots can only message a user/group after that chat has messaged the bot first. So trigger one message, then detect it:

1. Open the new bot's chat: search its @username (top-left search) and open the result — same as finding BotFather. (Clicking the t.me/<bot> link in BotFather's reply is less reliable; search is the sure path.)
2. Send /start (or any message): a fresh bot chat shows a Start button — click it; otherwise click the composer (#editable-message-text), type /start, and press Enter. Give it ~2–3s, then detect. That delivers the bot's first update.
   • Group target instead? Add the bot to the group, then send any message in that group. (You may need to disable BotFather's privacy mode via /setprivacy → Disable so the bot can read group messages.)
3. Run:

sentinel_setup({"action": "detect_chat"})

It returns the chat_id and the sender_id. Use the returned sender_id for the allowlist so only that user can drive the colony.

detect_chat is listener-aware: if this bot is already wired to another colony, Sentinel's background listener is long-polling Telegram and consuming updates — so a raw getUpdates would always come back empty. The tool handles that for you (it reads the chat the listener saw), so just call detect_chat. If it says PENDING, the message simply hasn't landed yet — make sure Start/your message actually sent, then retry. Do not start decrypting credentials or polling the Telegram API by hand to debug a PENDING — that fights the listener and gets you nowhere; messaging the bot and retrying is the whole fix.

Step 4 — Turn Sentinel on and verify

configure/test target the colony bound to your session (pass colony_id only to set up a different one):

sentinel_setup({"action": "configure", "channel": "telegram",
                "target": {"chat_id": "123456789"},
                "allowlist": ["123456789"], "enabled": true})
sentinel_setup({"action": "test"})

test sends a message to the chat — ask the user to confirm it arrived. If it didn't, the usual causes are: the chat hasn't messaged the bot yet (redo Step 3), the wrong chat_id, or (for groups) privacy mode still on. sentinel_setup({"action": "status"}) shows the stored token + the colony's current config any time.

When test succeeds you're done: tell the user Sentinel is live on Telegram and they can reply to its messages to keep the colony moving. (They can review or change this under Automations → Set up Sentinel.)

Guardrails

• Activate hive.browser-automation and obey its lifecycle rules. Never run google-chrome/chromium, --remote-debugging-port, or kill any browser/bridge process — you drive the user's existing Chrome.
• Reuse, don't duplicate. If sentinel_setup status shows telegram configured, the bot already exists — use it; don't spin up another via /newbot.
• Never invent, guess, or paraphrase a token. Read it by running hive-browser page text ".chat-message:last-child" --json in the terminal, not off a screenshot, so you don't transpose a character; if store_token rejects it, re-read — don't pad or guess.
• Don't try to log into Telegram for the user — you can't enter their phone/SMS code. If Telegram Web shows a login screen, ask them to sign in, then continue.
• Reach BotFather and the bot via in-app search, not t.me/ links (which often don't open the chat).
• The provider id is exact: telegram. The destination is {"chat_id": "…"} (a number as a string), not a username or link.
• A Telegram bot cannot initiate contact — if detect_chat stays PENDING, the user simply hasn't messaged the bot yet. Guide them to message it, then retry. Never decrypt the stored token, curl getUpdates, or otherwise debug the API by hand: the Sentinel listener owns the update stream, so manual polling just loses to it — detect_chat already reads what the listener saw.