Atlarix 15.13: Sessions That Answer Each Other, and Questions With Real Options

In 15.12, a message sent to a session that was sitting idle waited for you to type something there. The sender saw "waiting", even when nothing was running on the other side. Now, in the desktop app and the CLI, a plain message starts a turn in an idle session, the way a teammate picks up a ping. In editors (VS Code, Zed) and the Chrome panel it is read with your next prompt there. Two things deliberately do not wake anyone: a "that turn ended" notice, which is marked as a notice and needs no answer, and a session you stopped yourself. If you pressed stop, a message waits for you rather than restarting the work you just halted.

A message from another session used to arrive as a folded row with no text, so on your screen it looked blank. Now every surface shows who sent it and the start of what they said: Message from <name> (<surface>): <preview>. The full text is one keypress or click away: ctrl+o in the CLI, a click in the desktop app and the Chrome panel, Show all in VS Code, and the step's body in ACP editors such as Zed (its title carries the start). In the CLI, the sending side's row names the receiver instead of a session id. We also fixed the CLI bug behind the blank rows: a long transcript stuck to its top, so a turn that arrived without a keypress ran and was saved but never came into view.

When a session needs approval for a command or an edit, another session can now see exactly what it wants to do. list_sessions with pending_of shows the command or the diff. The reviewer then answers it: approve, ask for a change, or reject. Plans work the same way. High-risk requests, such as a command that deletes files, uses sudo or installs software, still go to you and never to a peer. Use it for a lead session that reviews its workers, or for a second model checking the first one's commands before they run.

When the agent needs a decision, ask_user now works like Claude Code's question tool. Each option can carry a one-line description and a Recommended mark. A question can carry a short topic chip and can allow several picks. You can still type your own answer. It has its own card in the desktop app, the CLI, VS Code and the Chrome panel; other ACP editors such as Zed show it as a list you answer by typing. The agent receives your picks as plain text, for example "Lint, Types, a11y audit".

While a turn runs, Enter steers it at its next step and Tab queues a message for after it. New in 15.13: ctrl+s stops the turn and sends everything you queued as one message. Esc is a plain stop, and anything you had queued comes back into your input box, images included, so nothing is sent behind your back or lost. The terminal bell also rings when a turn finishes or needs you.

Three changes target the same cost: re-sending a long conversation the provider has already seen. On Claude models, the main conversation keeps a one-hour prompt cache on its history, not only on its instructions. Things that change every turn (the active plan, finished background work, attachment notes) now travel in the per-turn tail instead of the cached header, so they no longer invalidate it. And a long conversation left idle until just before its cache would expire is compacted then, so coming back does not mean re-sending everything at full price. You can turn that off with ATLARIX_IDLE_COMPACTION=off.

We audited what Atlarix tells its own models and removed the claims the code did not back up: a search tool described as "relevance-ranked" when it is plain ripgrep, a context reminder that reported a different number from the one that triggered it, and memory-writing rules sent to modes that cannot write notes. A model that is told something false about its tools makes worse choices with them.