Matrix conversations

A task conversation in Matrix is where you work with the AI tool assigned to a task: send follow-ups after a run, mention team members or agents, use Minds and slash commands, and review structured run summaries.

Send messages to an AI tool

Native tool tasks (Codex, Claude Code, Grok Build, OpenCode, Hermes Agent, Pi)

When a run on one of these tools has finished and the tool accepts follow-ups, a composer appears below the run transcript. Press Shift+Enter to send, or Enter for a new line. Your message is delivered as a new turn in the same tool session; earlier turns still count toward the context window.

The tool owner sends follow-ups directly. If you are not the tool owner, your follow-up is submitted as a request for the owner to approve. See Non-owner follow-up approval in Matrix tasks.

To continue in another tool instead, pick it in Continue in on the composer. The list shows every AI tool on a member's Mac that is ready in the task's Project. The new tool cannot open the earlier session, so it starts from that run's result: its reply, checklist, changed files, links and branch, followed by your message. Your own tool starts at once; another member's tool waits until they approve. The same picker appears when you continue one of your own sessions.

If the assigned AI tool cannot receive messages right now, make sure its Mac is online with the macOS app running, or use Handoff to route the task to another tool.

Tasks assigned to other agents

A task assigned to an agent that does not run on a member's Mac opens a discussion session with the full composer. Messages are sent in discussion mode, meaning the AI can read context but does not change files or approve actions. A Discussion badge in the composer confirms this mode. Until a tool is assigned, the composer shows: "Assign an AI tool to this task to send a message."

Mentions and commands

The composer supports @ mentions and / slash commands in both the native follow-up composer and the workshop composer.

@ mentions

Type @ to open the mention menu. The menu lists your team members and the AI agents available in this Project. Selecting a member or agent inserts their handle into the message.

/ slash commands

Type / to open the command menu. The menu lists available Minds and onemind slash commands. Commands are shown with a description based on the operation they run. Selecting a command inserts it into the message.

The menus the tool shows depend on what that tool reports. If a tool has not reported its available agents or commands yet, the menu shows a note.

Attachments, model, and effort

These options appear in the workshop composer (tasks not assigned to a managed native tool). They are not available in the native follow-up composer.

Attachments

Click Attach files in the composer toolbar to include files with your message. Limits:

• Up to 4 files per message
• Text files: up to 64 KiB each
• Images: up to 2 MiB each
• 4 MiB total per message, within the assigned tool's own limits

Not all tools or models accept files. If the selected model does not support attachments, a note appears in the toolbar.

Model and reasoning effort

Use the Model and Reasoning effort selectors in the toolbar to override the tool's default for a single message. The effort levels available are: None, Minimal, Low, Medium, High, Extra high, Maximum, and Ultra. Ultra is the highest level and is available only in task conversations, not in onemind agent conversations.

Settings apply per message. If the assigned tool no longer supports previously chosen settings, the toolbar shows a warning asking you to review the model and reasoning options before sending.

Discussion-only mode

In the workshop composer, a Discussion badge is always shown beside the send control. Messages sent from the workshop composer are always in discussion mode: the AI can read context but cannot change files or approve actions. This mode is fixed for workshop conversations and cannot be switched off.

Usage and context window

Usage details in the composer toolbar shows how much of the assigned tool's capacity the task is using. Only the tool's owner sees it: "Only you can see your plan usage and context window."

• Context window — how much of the tool's context the current session uses (used of total tokens) and when the tool last reported it. If the tool does not report its context window, the panel says so.
• Plan limits — each limit window of the tool's plan (for example 5-hour limit or Weekly limit) with the percentage used, used of limit, and when it resets. The toolbar summarizes this as "tool plan: N% of the window limit used".
• This session — the current execution's token counts and costs as reported by the device running the tool: Input tokens, Cached input tokens, Output tokens, Reasoning tokens, Total tokens, Model calls, API time, and Cost (in US dollars).

An onemind agent conversation has its own context breakdown; see Start a conversation.

Run summaries and states

Reply actions

Each reply in the task conversation has a Copy button and a thumbs-up / thumbs-down rating. Use the rating to give feedback on a response.

Run states

The conversation shows a state indicator while a run is in progress or after it ends:

Waiting for stop confirmation is not the same as Execution stopped: after you stop a run, onemind waits for the tool to confirm before it shows the final state.

Structured run summaries

When a run completes, the task detail may show a structured summary generated by the tool. The summary can include:

• Steps — a list of steps taken, each with a kind, target, and status
• Todos — remaining items with completion status
• Links — relevant URLs with kind and title

Composer warnings

The composer toolbar may show one of these warnings:

• Conversation limit reached — "This task conversation has reached its file or context limit. Remove files from this draft or shorten the message; previous records are preserved."
• Settings no longer supported — "The assigned tool no longer supports these execution settings. Review its model and reasoning options before sending again."
• Conversation changed — "This conversation changed. Refresh its latest records before sending again." This appears when another session updated the same conversation while your draft was open. Refresh the task and resend.
• Message too long — "Messages can contain up to 16 KiB of text." Shorten the message or split it into multiple sends.
• Tool runtime unavailable — "The assigned AI tool cannot receive messages right now. Connect its runtime or use Handoff."
• Request not confirmed — "The request could not be confirmed. Your draft is preserved; retry to check or resend the same request."

See Matrix tasks for creating tasks, assigning tools, and answering tool permission requests.