TL;DR—[cordless]is a CLI-first tool that manages your remote terminal / coding-agent sessions and puts them on your phone.v0.8is aboutorganisingthe swarm:Chrome-mobile-style tab groups,custom launchers(bring any command as a profile), a built-inGitHub Copilot CLIprofile,renameable tabs, andpersisted scrollback historyso a reopened session after a reboot shows what it was doing. There's also a properinstaller. Six features, each on its own branch — designed withcordless setup
GPT-5.6 Sol, built withGitHub Copilot CLI.
v0.7 taught cordless to tell me which of my eight agent sessions needs me. v0.8 is the next problem: once you're running that many, they stop fitting on the screen. You need to organise them — like browser tabs, because that was the whole metaphor from day one. So this release is six features that make a big session swarm manageable.
The headline. Sessions can now live in named, colored groups — think "API migration", "Website", "Flaky tests" — exactly like Chrome mobile's tab groups.
In the terminal dashboard, groups render as collapsible headers with a live per-group waiting/session count:
── Sessions (6) ──
▼ API migration 2 waiting · 3 sessions
! claude build pipeline waiting
● codex api tests working
○ shell api ~/src/api idle
▶ Website 2 sessions
▼ Ungrouped 1 session
○ copilot scratch idle
Press g
for the group menu (new / assign / ungroup / collapse / rename / delete), and f
to cycle a smart-view filter — All · Attention · Claude · Codex · Copilot · Shell
. Those filters are views, not groups; collapse state is per-device. On the phone you get the same idea as a chip strip above the tabs (All · Unread · one chip per group, with counts and color dots) plus rename/move in the details sheet.
The design decision I kept going back and forth on — and where Sol was firm — was manual groups over continuous auto-grouping. Auto-sorting your tabs by directory sounds clever until a session changes cwd and silently jumps groups. So groups are something you choose; deleting one never kills its sessions, it just ungroups them.
Until now cordless had three profiles: shell
, claude
, codex
. v0.8 makes profiles yours. Drop this in ~/.cordless/config.json
:
"profiles": {
"api": { "command": "pwsh", "args": ["-NoLogo"], "cwd": "C:/src/api" },
"notes": { "command": "nvim", "args": ["~/notes.md"] }
}
…and cordless new api
spawns it directly (resolved against the daemon's PATH
, not a shell string — no sh -c
injection surface). A missing executable doesn't crash anything; it shows as unavailable with a reason, and cordless profiles
lists every launcher with its source and availability.
And because the agent world moved, v0.8 ships a built-in ** copilot** profile — the standalone
claude
and codex
. The nice part: attention detection is now copilot
(and any custom profile you tag attentionPreset: "agent"
) gets the same Small feature, big daily quality-of-life. cordless rename <id> "API migration"
, or the dashboard e
key, or a long-press on the phone. Titles are Unicode-normalized, control-stripped and length-capped, an empty title restores the generated default, and the change broadcasts live to every connected client (with a monotonic revision so a stale update can't clobber a newer one).
This is the one that turned into a proper engineering story. A cordless session already survives a daemon restart (it reopens the same shell in the same directory). But it came back blank — you lost the scrollback. v0.8 fixes that: the daemon persists a capped, plain-text copy of each session's scrollback (2000 lines / 512 KB, gzipped, user-only) and, on restore, shows it as frozen context above the reopened session with a ── session reopened after system restart ──
banner.
A couple of decisions worth calling out:
Here's the part I'm glad CI caught. My persisted-history tests were green on Windows and I shipped v0.8.0. The Linux and macOS CLI builds went red on the exact same test.
The cause is a lovely cross-platform trap. On a graceful stop, the daemon catches SIGTERM
and runs shutdown()
: it saves each session's history, then kills the PTYs. Killing a PTY fires its exit handler — and that handler deleted the just-saved history and rewrote the "reopen these on restart" manifest without the now-exited session. So on Linux the session came back with no history; on macOS it didn't come back at all.
Why did Windows pass? Because on Windows SIGTERM
from another process is an uncatchable hard-kill — shutdown()
never runs, the exit handler never fires, and the periodically-flushed history file just… survives. The platform that "worked" only worked by accident.
The fix is a _shuttingDown
flag so the exit handler preserves history and the manifest during a graceful stop. What I like about it: I reproduced the Linux-only failure on Windows by driving shutdown()
directly in a unit test (the harness's SIGTERM
can't), so the regression is now locked shut on every platform. Re-tagged, rebuilt, and v0.8.0 went out clean with the full asset set.
Same loop, tightened further: me on GitHub Copilot CLI, GPT-5.6 Sol as the design partner — and this time with a rule I'm keeping for good: every feature on its own branch, merged --no-ff, with clean well-documented code. The v0.8 program was six branches (
feature/persisted-history
, feature/custom-profiles
, feature/copilot-profile
, feature/session-rename
, feature/session-groups
, feature/group-ui-phone
) plus a fix/history-shutdown
. The daemon's test harness grew to The feeling I was chasing: open cordless with a dozen agents running and not feel underwater. The waiting ones surface (v0.7); the rest are filed into groups you named, launched from profiles you defined, titled how you think about them — and if the box reboots overnight, each one comes back with its history intact. It's the browser-tabs promise, finally delivered for a swarm of terminals.
cordless setup
Part of the #AI4Good series. Built one day at a time. — @naveenneog