An agent can start other chats, make worktrees and message other chats
through Tori’s MCP server or the
tori CLI. Tori keeps track of who did what, shows it
on both sides, and only lets an agent reach the chats it is tied to.
A worktree an agent made
When an agent makes a worktree with worktree_new, or starts a chat in a new
worktree with session_spawn, the call shows as a worktree card instead of a
plain tool row. It reads Making worktree while it runs, then Worktree
with the branch, and:
- the project’s name, when the worktree is in another project than the chat;
- with a chat, when the agent started a chat there in the same call;
- the pull request’s number and state, like PR #482 open, once the branch has one;
- Removed once the worktree is gone;
- Setup failed when the project’s setup command failed in it.
Open goes to the worktree, or to the chat the agent started there. A call that waits for your approval, was refused or failed stays a plain tool row, which is where the prompt or the error is.
From and To
The strip above the composer, beside main and the subagents, also shows the chats tied to this one:
- From is what led to this chat: the chat that started it or made the worktree it runs in, and chats that sent it a prompt.
- To is what came out of it: chats it started, chats in worktrees it made, chats it prompted, and chats you linked to it.
Each chip has a dot for that chat’s state and a mark for how it is tied. Hover it for the chat’s name, how it is tied (like Started this chat or Runs in a worktree made here), whether it is Working, Needs you, Idle or Done, and its project and branch. Click it to open that chat.
Each side shows two chips, the chat that started this one always first. The rest are behind a +3 chip, which opens the whole list grouped as Started this chat, Prompted this chat, Started from here, Prompted from here and Linked.
Prompts from another chat
A message another chat sent shows on the right like one of yours, with a from line naming that chat. Click the name to open it. The first prompt of a chat another chat started shows its sender the same way.
Who an agent can reach
An agent can steer or read a chat only when the two are related:
- one started the other, or made the worktree the other runs in;
- the Autopilot and the chats it runs for its items;
- a chat and the one it was handed over to;
- chats in the same Topic;
- chats you linked.
Reading is refused outright for any other chat, and the agent is told to ask you. To let a chat read one session without linking them, reference that session with @ in the chat: the reference is your permission.
A message goes straight in when it goes down the tree (to a chat the sender started or whose worktree it made, or the Autopilot to its items), to a linked chat, or to a related chat that is in the middle of a turn. Anything else waits for you, including a message to a chat that is not related at all, and a worker writing to its idle parent: a chat never starts a turn only because another agent asked.
Terminal tabs are not held to this. A tab’s token cannot tell you at the shell
from an agent running in it, so tori steer from a terminal tab works as
before. A worker still cannot steer anyone.
Answering, waiting on and stopping take more than being related: a chat answers another session’s question, waits on it, lists what it is waiting on or stops its turn only when it owns that session. It owns a session it started, one running in a worktree it made, and, for the Autopilot, its items. A chat that took over by a handover owns the sessions its predecessor started. A terminal tab answers no session’s question, since an agent in it would be answering for you; the question reaches it without the id an answer needs.
Messages waiting for you
A message that waits shows above the receiving chat’s composer as From Retry worker: and the text, with how many files came with it:
| Button | Does |
|---|---|
| Accept | Delivers this message, once |
| Link | Delivers it and links the two chats |
| Discard | Drops it |
The sender is told its message is waiting for a person, so it does not try again.
Link two chats
Linked chats can message and read each other from then on, both ways. The linked chat shows as a To chip in both. Unlink sits beside it in the list behind the +3 chip, and removes the link on both sides. Only you can link or unlink chats; there is no tool for an agent to do it.