tty7

A terminal written in Rust that keeps its shells in a background server instead of the window. Quit the app and every shell keeps running; after a restart the panes return with their layout and the last of what was on screen, and supported coding agents resume the conversation they were in. Twenty-six coding CLIs are recognised by the command running in a pane and listed in one sidebar as working, waiting or done, a status that arrives through a hook installed into the agent’s own configuration. Because that status is something a program can ask for, an agent can ask it: split, send, wait and capture are the whole loop, and they work with the window closed. A remote workspace is the same server on another machine — tabs, panes, git and diffs stay over there and nothing is synced — reached through an SSH client the project wrote itself. Rendering is GPU-side on Zed’s gpui, the terminal core comes from Alacritty, and builds ship for macOS, Windows, Linux, iOS and Android.

Who built itThe account that owns the repository and wrote 1,302 of its 1,546 commits in twelve weeks; the other 244 come from thirty-three accounts, one with 89 commits and one with 30. Of the 316 commits carrying a co-author trailer, 92 name a Claude model. The report carries no profile data beyond the commit history.

1. 01

   Twelve weeks, 1,546 commits, and a version line restarted in July

   The first commit is dated 2026-07-06 and reads “a GPU-rendered, daemon-backed terminal in pure Rust”; the newest in the report is a link-parser fix on 2026-09-30, with 1,546 commits in between — 543 in July, 562 in August, 441 in September. One account wrote 1,302 of them; the other 244 come from thirty-three accounts, one with 89 commits and one with 30, and 1,544 of the 1,546 are linked to a GitHub account at all. Of the 316 commits that carry a co-author trailer, 92 name a Claude model, the largest group Opus variants, and the owner co-authors himself 199 times. The twenty releases in the report start at v0.16.0 on 2026-07-15 and pass through four versions in two days: v0.17.0 and v26.7.0 both land on 2026-07-16, after which the numbering is a year and a month — 26.7.x through 26.9.4 on 2026-09-29, whose own title claims an IDE-level code editor, a v5 redesign and a GitHub tab, while v26.8.3 calls itself a 151-fix polish pass. The nightly is a rolling prerelease tag rather than a version, the phone app has its own tags (mobile-v0.1.1 and mobile-v0.1.2, both on 2026-09-30), and the changelog committed to the repository is 376 KB.

2. 02

   The shells belong to the server, and the server must be replaceable

   The README states the design as one decision: the shells belong to a server running in the background and the window only shows them, so closing the window does nothing to them. The server is crates/tty7-core/src/daemon/ — server.rs at 68 KB, control.rs at 110 KB, pane.rs at 316 KB, spawn.rs at 94 KB, protocol.rs at 104 KB, router.rs at 52 KB — with the parts a long-lived process needs around it: singleton.rs, pidfile.rs, handoff.rs, update_guard.rs, scrollback.rs and history.rs. A separate 173 KB updater binary sits alongside it, with handoff.rs and update_guard.rs inside the daemon. What the daemon believes about the world is a machine tree in machine.json: pane records holding cwd, shell, SSH spec and agent facts, plus — added in pull request 1037 — a closed list of recently closed tabs with their layout and a closed_at time, so reopening a tab survives a GUI restart and is visible to every other window on that workspace instead of living in one window’s memory. A reboot is the case the design does not pretend to solve: the processes are gone, and what returns is the layout, the last of what was on screen and, for agents whose hook is installed, the same conversation.

3. 03

   gpui on a private fork, a VT core from Alacritty, and a screen that is a rectangle

   The subtitle is the stack: pure Rust, GPU rendering on Zed’s gpui, VT core from Alacritty. gpui paints each text line into a scene layer whose bounds decide its draw order against later primitives, and those bounds were computed from the unaligned origin, so right-aligned glyphs fell outside their own layer and drew through the buttons that appear when a row is hovered; the fix is a layer_bounds helper in a fork of Zed, l0ng-ai/zed@3d46ee70fb (1030). A second came from a component library: the title bar reserves an 80pt lead for the macOS traffic lights, so the centred Search Everywhere box sat 40pt right of the terminal column (1036). The repository carries kitty_graphics.rs (87 KB) and input.rs (54 KB), and on Windows it ships Microsoft’s conpty.dll and OpenConsole.exe, version 1.24.260710001, for x64 and arm64 rather than trusting the in-box pair (1043). One open issue reports the physical Escape key producing nothing in a TUI using the Kitty Keyboard Protocol on Windows; the author could not reproduce it and answered with probes rather than a fix — a unit test for the encoding, both ConPTY builds passing the sequence through, and a synthetic key event under Python and Node. Performance is measured by committed scripts: 95 ms for an 11 MB cat, DOOM-fire at 888 fps.

4. 04

   Agents that run agents, without wrapping them

   Twenty-six coding CLIs are recognised from the command running in a pane, and detection sees through a full path, a .cmd or .exe on Windows, leading environment assignments and an interpreter in front (node .../claude/cli.js); a wrapper of your own is mapped to an agent in config.json. Status is not: it comes from the agent itself over a channel tty7 installs into that agent’s own configuration, one click per agent, powering the status dots, notifications, the tray icon, tty7 wait and resume after a reboot. The asker need not be a person: tty7 split, send, wait and capture are the whole loop. The documentation is clear about where this breaks — a screen is a rectangle, and an exit code was never on it. The README says the agents are not wrapped or proxied, in one line: “The agent you start is the real one, running in an ordinary PTY. Simple things are more likely to keep working.” Behind that sit agent_hooks.rs (163 KB), cli_agent.rs (109 KB) and agent_history.rs (72 KB); it also ships an agent skill (skills/tty7/SKILL.md 18 KB, a 25 KB command reference, a 13 KB delegation reference) installed with npx skills add l0ng-ai/tty7. The CLI is a thin non-interactive client: a pane address such as %42 is stable for the life of the pane, while a tab address such as @7 shifts whenever anything is created or closed.

5. 05

   Remote is the same server, which is where the hard bugs live

   A remote workspace is the same daemon on the other machine: tabs, panes, git and diffs live over there and nothing is copied, reached through the project’s own Rust SSH client — keychain profiles, jump hosts, SFTP, port forwarding, and a tty7-server that installs once without root. The boundary is real enough to have its own script: daemon/install/ is 70 KB plus wsl.rs at 63 KB, and .github/scripts/check-host-boundary.sh is kept for exactly that line. The bugs are where the interest is. A 17-pane remote workspace hung for over ten seconds because sshd allows ten sessions per connection and every remote pane is one: the eleventh open was refused, and the connection was retired while ten panes still used it. Reattach made it worse — any failure other than silence was read as the pane being gone, so it spawned a fresh shell that auto-ran claude --resume while the original was still alive (1034). Elsewhere, the window on the daemon’s own machine skipped workspace takeover, so a remote client and the local window drove the same panes at once, the last resize won and TUIs reflowed; the rule written down afterwards is one workspace to one window at a time (1042). Windows hosts added their own run of POSIX assumptions, from sh -c to setsid and kill -TERM (1041).

6. 06

   Nightly tables, UTC build numbers, and notes that say what was not checked

   Four workflows — ci at 19 KB, nightly at 26 KB, release at 23 KB, mobile at 8 KB — and eleven scripts under .github/scripts/, among them a macOS bundler, an installer script and assertions about the binaries that come out. The nightly notes open with a download table for all four platforms, built from the files that night actually produced, so a failed build is left out rather than linked dead (1054, 1055). The Windows ARM64 leg is cross-compiled and marked experimental until it has shipped once, so its failure drops the ARM64 packages instead of blocking the release (1043). The recorded failures are ordinary: TestFlight build numbers were the local time, so a machine in UTC+8 numbered builds eight hours ahead of the UTC runner and a CI upload of the same version within eight hours came out lower and was refused (1050); a certificate check matched one fixed string while newer apksigners name the signer by SDK range (1047). Two notes are worth more than the fixes: a mobile change was checked in the iOS simulator in dark theme only — “Not yet checked in the light theme or on a device” — and a stuck-relay fix skipped the pull-request path: “shipping a TestFlight build from the branch first” (1039, 1057). The README itself was rewritten in a pull request to lead with three pillars and the line “The terminal that outlives its window” instead of a feature list (1046).