What shipped in 1.13.0.

11 September 2026

Until now the panel’s Claude could read your buffers and propose edits, and that was all it could do: the build, the tests and every other command were simply not among its tools. An agent that cannot run the build cannot check what it just wrote, so it could only ever hand you something to try.
Preferences ▸ Claude now offers a second answer. Claude Code’s Own Rules gives the panel the full tool set and stops pinning a permission mode, so what is approved without asking is decided by your own settings.json — the same rules, and the same mode, as the claude you run in a terminal.
Linelark adds nothing on top of them. It pre-approves its own editor reads and never a built-in, so nothing here can approve Bash on your behalf; that decision stays in the one file where you can read it back, scope it to a project, or take it away. Anything your rules still leave to a person is asked in the panel exactly as a buffer edit is, with the command on the card.
Read-Only stays the default, and updating to 1.13.0 changes nothing about what the agent can already do. A window can also disagree with the app: the panel’s ⚡ menu sets the trust for that window alone, which is where the triggers and autostart already live.

It has listed your recent files and recent projects since it was written, and almost nobody has seen them. It was the empty window — and with session restore on by default, a window with nothing in it is a moment that after the first launch nearly never comes.
It now opens with the app, drawn over the pane with the tab strip left in place above it, so everything the last session restored is visible and one click away. Clicking a tab is what dismisses it, as is opening or creating anything.
Not when you opened the app at a file, though: double-clicking a document still goes straight to it, because a page in front of it would be answering a question nobody asked. There is a switch on the page itself, another in Preferences ▸ Files, and Help ▸ Welcome to Linelark brings it back after you have turned it off.

On Claude Code’s own rules the agent has the built-in Edit and Write, and those write to disk rather than into the buffer. Such an edit is not one ⌘Z and was not shown to you as a diff first. A tab with no unsaved changes reloads itself; a tab with unsaved changes is left alone and asks you the next time the app comes forward. The editor’s own tools still land in the buffer and the agent is told to prefer them, but it is free not to.
What gets approved without asking is whatever your settings.json already says, an auto-approving mode included. Linelark deliberately offers no second opinion on it — that is the whole of what the setting means, and it is why Read-Only is what you keep until you choose otherwise.
The panel still never sees the MCP servers in your own ~/.claude.json, under either trust: it gets this window’s editor tools and nothing else. That is a question about which servers a panel talks to rather than about who approves what, and it has not changed.
The Claude panel remains Studio only. It spawns a process and opens a socket, so the sandboxed App Store edition has none of it, and the trust setting with it.
The welcome page opens on every window a launch restores, not only on the one that ends up in front.
Code folding and recorded macros remain roadmap items with no promised date.

Artifact SHA-2566dbdfc9510adb9cc442eac7c5b0af499acc389357b06d1280ff2ce480a6a4341

Source commit 8f21a92a1ccc2fdbc60a5b14f4dad3b52fd2d60d

Claude can check its own work.