Automation messages: let an outside relay wake an agent without speaking as the user

[arthurcruzd]

Problem

I run an orchestrator in a T3 thread: one agent that supervises workers across several machines. An outside relay wakes it with t3_thread_send when something happens: a worker finished, a PR merged, a check failed.

We already filter at the source. A headless supervision session outside T3 handles routine events itself, and only events that may need the orchestrator reach the thread. Even so, every wake that does arrive lands in my conversation as if I had typed it, followed by the agent's acknowledgment. Each acknowledgment marks the thread unread and notifies me. Much of the timeline becomes "nothing to do" exchanges, and the replies that actually need me get lost among them.

What T3 already has

Delegated-task completions and PR-watch wakes are messages with a notification, rendered as a one-line, expandable item. A caller outside T3 cannot use that today.

Proposal

Expose it on t3_thread_send:

• automation: true stores the message with a notification of source automation. The first line is the summary and the full text is the expandable detail. The provider still receives the text, so the agent behaves exactly as before.
• The agent's reply to an automation message folds under it ("Replied after 2s", expandable) and does not count as unread or notify, unless the reply addresses the user. Failures, approvals and questions surface as they do today.
• A thread that was already unread stays unread.

Open question: how a reply "addresses the user"

We prototyped the simplest explicit version: T3 tells the model, in a one-line preamble, that its reply stays folded unless it starts with @user. A small model (Haiku) followed this reliably in a short test, but a marker in text is a convention, not a contract. Alternatives:

• A tool call or turn field.
• The inverse convention, where a quiet token folds and everything else shows. It fails safe but folds nothing by default.

Prototype

About 190 lines across contracts, server, web and mobile, tested live. It reuses the notification turn item and marks a quiet run as seen in the same commit that completes it, so every client's unread state stays consistent without new shell fields. Happy to share the branch.

Related

• Tell Claude and Codex when a turn wasn't written by the user #15456 asks T3 to tell the provider when a turn was not written by the user: the "agent knows" half of this.
• Optional compact incoming agent reports with source-thread navigation #20243 asks for compact incoming agent reports in an orchestrator thread: the "renders folded" half.
• Distinguish user prompts from automation runs in the thread minimap #19624 (automation runs in the minimap), Let an agent send a thread a message without starting a run #18104 (send without starting a run) and Let an orchestrator thread hear from its workers, and pass their questions to the user #17212 (an orchestrator hearing from its workers) touch the same workflow.

Questions for maintainers

• Do you want this shape?
• Which "addresses the user" signal do you prefer?
• Should automation sends be restricted to particular callers?