Omniscio

Team Chat reply watch — an agent hears back from a person

An agent that sends a Team Chat message for you keeps listening for the answer. When a person replies in that conversation or thread, the reply lands in the session that wrote the message, and nothing is ever sent to the person unless the agent chooses to.

On this page

What it is

Agents can already send on your behalf — a direct message to a teammate, or a post in a channel. What was missing was the other half: when the person wrote back, nobody told the agent. The reply sat in Team Chat until somebody happened to look.

The reply watch closes that. Sending a message arms it; the person's answer wakes the agent that asked. It is not a subscription you have to set up or remember — the act of writing is the ask.

It also means an agent can read the pictures a person attaches. Team Chat messages carry their attachments as cloud file locations, and an agent reading a message could see that screenshots were there and had no way to open them. An agent can now download one and open it.

Where to find it

Nothing to switch on. It happens in the background once an agent has sent a Team Chat message through the local control server — either the direct-message route or the channel message route.

An agent can see its own active watches, and stop one, with three routes:

  • GET /team-chat/reply-watches — every conversation this session is listening to, whether it is live yet, and the last reason nothing was delivered.
  • DELETE /team-chat/reply-watches/:channelId — stop listening to one conversation. A conversation this session is not watching answers 404 NOT_FOUND, so a caller can tell "stopped" from "there was nothing to stop".
  • GET /team-chat/channels/:channelId/messages/:messageId/attachments/:index/download — fetch one attachment to a temporary file and return its path. A video or audio attachment (a clip) answers 200 with { skipped: true, note: 'skipped: voice/video clip' } and no file instead, so check skipped before reading path.

How it behaves

What wakes the agent. Only a person. A message another agent wrote never wakes it, an ordinary automated row from another part of the product never wakes it, and a message that predates the one the agent sent does not wake it either.

Inside a conversation or a thread, your own typing does not wake it either — there you are not "a different person", and the agent already speaks for you. There is one exception, and it is the one that matters most day to day: when an agent posts at the top level of a channel, a message typed straight under that post wakes the session that wrote it when it names you — and that holds even when you are the one typing it, because a top-level answer has to say who it is for. That is the natural way to answer a post in a channel — there is no thread to reply into — and beyond the name the line held is who WROTE it, not who is signed in: anything an agent wrote still never wakes an agent.

A channel is a shared room, so being in it is not being addressed. A top-level message under the post counts only when it @mentions the person the agent posted for — the signed-in person. An ordinary status update in that channel, with nobody mentioned, is a new subject rather than an answer, and it no longer wakes the agent or gets labelled as a reply. @everyone does not count: a broadcast is addressed to the room, not to you, so it is not an answer to one post. Replying inside the post's thread always counts, mention or not, which is what the reply button in Team Chat does. If the app cannot tell who is signed in, a top-level message counts for nobody: an identity it cannot read is a reason to stay quiet, not to guess. That answer counts for a day after the post, and only while nothing newer in that channel has been read, because by then a new top-level message is far more likely to be a new subject than an answer.

What arrives. The person's own words, in a clearly marked block, followed by a note from the app naming who wrote it, saying plainly that they are not you and hold no authority, and giving the exact request to answer with. If a person's text tries to close that block early and write its own note, the attempt is disarmed.

How often. Once per message. A reply is never delivered twice, even if the app is interrupted between handing it over and writing down that it did.

If the agent is busy. The reply is held, and delivered the moment the agent is free — never dropped. A few replies arriving together are read in the order they were written.

If a message cannot be delivered. The agent is not told it arrived when it did not. Nothing advances until a delivery actually succeeds.

When it stops. A watch is an exchange, not a subscription: it is there for the answer to one message, and it ends when that answer has arrived. So it stops once the person's reply has been delivered and the conversation has gone quiet again — the agent's next message there starts a fresh watch, so a back-and-forth keeps working — and a watch that never hears anything ends 72 hours after it was armed. It also ends when the agent's session is archived, ended or has failed — a session you have closed is never reopened to hand it a message — when you stop it by name, and in one further case: if a conversation grew so long that the app can no longer read back to where the agent last looked, the replies it can still see are delivered first and then the watch ends rather than sitting there silently. GET /team-chat/reply-watches names which ending it was (reply-delivered, idle-expired, backlog-too-deep) and the reason it happened, and the agent session itself is told in words when its question timed out. No ending ever throws an answer away: a watch retiring on the 72-hour mark hands over anything already waiting first, and only then stops. A paused session is not ended: its watch waits until you resume it, and a waiting watch still checks its conversation once a minute — from the app's own memory while nothing has changed, so a quiet conversation costs no cloud calls.

One channel, one agent. Only one session watches a conversation at a time, so a session that has already been answered stops receiving everything the person writes there afterwards — and a conversation is never taken away from a watch that already has it. When an agent asks in a conversation another session is watching, the new ask is refused: the agent is told who holds the conversation, that a reply there reaches the holder rather than itself — so it reads the conversation instead of sending the person a second message — and that the watch frees when that one ends, either because the person replies or after three idle days, so it knows when to try again. Nothing is ever taken, because a watch part-way through an answer must not have it pulled out from under it. An ask that fails outright — the watch could not be started at all — is not answered with silence either: the send still counts as sent, and the agent is told its replies will not reach it and to read the conversation itself. The watch itself still covers the whole conversation, because a direct message's answer arrives as a new message there and a narrower watch would miss the second line of a two-line reply. "Another session is watching" means the two watches necessarily pick up the SAME message: a conversation watch covers its whole channel, and two watches on one thread root cover that thread. Two agents answering inside two different threads do NOT collide, and one may watch each. Two agents that each posted at the TOP LEVEL of one channel also do NOT block each other — but only one of them hears the person's answer, the one that asked most recently, and the other keeps watching for its own answer — it is not told anything, and if nothing comes it simply reaches its own three-day limit like any unanswered question.

The only limit is one you set. Every reply that wakes an agent is a turn on your account, so Settings → Team Chat notifications has a Replies per conversation per day box: the most replies from one conversation your agent will be woken for in a day, blank — the default — for no limit at all. It is your number and nothing else applies: leave it blank and every reply arrives. When a limit you set is reached, the replies are not dropped — they wait, and the agent session is told so in words, once, naming the limit and that they arrive after 00:00 UTC.

A watch never goes quiet on its own. Whatever happens — a reply held while the agent is busy, a conversation too long to read all the way back — the state is on the watch you can list, and the reply itself is never thrown away for a reason the app kept to itself.

Nothing goes back to the person. This feature writes nothing to Team Chat. Nothing is acknowledged, no read receipt is sent, and no reply is posted on your behalf. The person sees only what the agent deliberately sends.

Reading an attachment. The file is written to a temporary folder and its path is returned. The caller can name an attachment only by its position on a message it is allowed to read — never a file location of its own choosing — and an attachment larger than the ordinary download limit is refused: checked against its stated size before the download and against its real size after, so a file whose stated size was wrong is refused too.

Agent lanes covers your agents talking to a teammate's agents, with no person in the loop. Cross-device agent messages is the same idea between your own computers. Agent messages is the general picture of how a session receives a message from anywhere. Team Chat API platform is the full set of routes an agent can drive.

Last verified 2026-11-18