> ## Documentation Index
> Fetch the complete documentation index at: https://docs.opencomputer.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Threads

> Each Slack thread is one session; its runner exits after 2 hours idle and resumes on the next message

![A finished turn: two step cards that ended as Ran N steps, the text between them, and the footer with model, duration and context size](https://mintcdn.com/opensandbox/wc3ROkMcBR7DifaQ/images/slack/turn-finished.png?fit=max\&auto=format\&n=wc3ROkMcBR7DifaQ\&q=85\&s=f60d55b67af95b64b1cdd7d49461199f)

## Starting a session

| Where | What starts a session | Later messages |
| - | - | - |
| Channel | A message that mentions `@computer`. Its thread is the session. | Every reply is delivered; the agent answers the ones meant for it (see [Shared threads](https://docs.opencomputer.dev/slack/shared-threads.md)). |
| Existing thread | A reply that mentions `@computer`. The session also receives the replies above it. | Same. |
| Direct message | Any message. No mention needed. | Every message in that DM thread is for the agent. |

Only people on the people list get a session; anyone else gets an ephemeral "not enabled for you yet" reply. One-machine mode has no people list: anyone who can reach `@computer` can start a session.

Edited messages and messages from bots are ignored.

## Turn output

1. **Status.** Slack shows "is thinking…" on the thread, or an :eyes: reaction if the status call fails.
2. **Step cards.** Each tool call is a step with its arguments and an output tail. Consecutive steps share one card, which ends as "Ran N steps" with a count of failures.
3. **Text.** The agent's text streams into the same message between the cards.
4. **Footer.** Model, reasoning effort (once set with `!effort`), permission mode (when not `bypassPermissions`), duration and context size. A stopped turn's footer starts with the reason. Cost is in `!status`.

![A running turn: a finished step card, streamed text, a live step with its timer, and the thinking status under the composer](https://mintcdn.com/opensandbox/wc3ROkMcBR7DifaQ/images/slack/turn-running.png?fit=max\&auto=format\&n=wc3ROkMcBR7DifaQ\&q=85\&s=7a035c66cddba76f03ac6cdc2f82b93d)

A long reply continues in a new message: Slack closes a stream after about 10 minutes and caps a message's length. A turn waiting on an answer does not open a new message, so it sends one notification.

| What you see | Cause |
| - | - |
| ":warning: agent process ended: …" | The harness process died. The next message starts it again. |
| ":warning: couldn't reach the session: …" | The gateway could not deliver the message to the runner. |
| "The agent host is offline; your message will be delivered when it reconnects." | The host is not connected to the gateway. The message is queued for up to 24 hours. |

## Commands

A message that starts with `!` is a command to the session, not a message to the agent: `!status`, `!stop`, `!diff`, `!restart`. See Slack commands.

## Idle exit

The host stops a session's runner 2 hours after the last message (`SP_IDLE_MS`). It waits while a turn is running or waiting on an answer, a subagent is running, or an API terminal is open.

The next message starts a new runner that resumes the same conversation in the same directory; the first reply is slower. See Session durability.

One-machine mode has no idle exit.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.