tmux-ide 2.9.23

2.9.23 is a CLI/daemon release. It fixes dropped terminal connections, makes the daemon cheaper under heavy output and at rest, and fixes five problems on tmux servers that are not tmux-ide's own. It adds no commands and needs no Mac app update.

Terminals no longer drop while the daemon stays up

In 2.9.22 the daemon could close every terminal stream of a tmux server while the daemon itself kept running. The Mac app then lost its terminals.

The cause: each time a client listed the tmux servers, the daemon asked every server a question with a 500 ms limit. A slow answer, or a tmux that could not be started, counted as "server gone", so the daemon dropped its connection to that server, closed its streams and built a new connection on the next list. The Mac app lists servers whenever it reads the catalog, opens a session or creates one, so it was hit far more often than the TUI, which lists only when it reconnects.

The reproduction, on a separate daemon with tmux answering in 0.7 s, one list request and one open stream:

A tmux that cannot be started because the process limit is reached gave the same drop in 2.9.22 and gives none in 2.9.23. A busy tmux server, many list requests at once and plain CPU load did not reproduce the drop. Which of the two causes hit a given machine can't be told from 2.9.22's log; 2.9.23 logs it.

What changed:

• Read requests can no longer retire or replace a tmux server connection. They wait at most 1 s for work in flight, then answer with what is known.
• A background check every 2 s watches the servers. It starts no tmux process while their sockets and processes are unchanged.
• The daemon retires a connection only on evidence that the server changed: its socket was replaced, its process is gone, or another process answers. A second probe must confirm it.
• A connection that fails to build is logged and retried, from 500 ms up to every 30 s.

One case is left open: when a tmux server has died, another process has taken its process id, and its old socket file is still there, the server is listed as online until one of those changes.

A stop that ends

After SIGTERM or SIGINT the daemon now stops within a limit. If its teardown makes no progress for 10 s, or takes 25 s in all, it logs the steps still open, removes daemon.json and exits with status 1. Two waits on the stop path that had no limit now have one. A restart in place whose stop hangs exits the same way.

We have not found why a daemon sometimes did not stop after its tmux server was replaced. It did not happen in 74 local runs. The limit and the new log line are there to end the hang and to name the step if it happens again.

More connection fixes

• Two daemons on one tmux server no longer remove each other's hooks. A daemon's hook install removed every tmux-ide daemon's hooks, and each gap lost send-keys and capture-pane observations and made clients re-read the terminal inventory. A daemon now removes only its own hooks and those no living daemon owns. A daemon older than 2.9.23 on the same server still removes everyone's.
• Terminal streams and the events socket have a heartbeat. The daemon pings every 10 s and drops a connection after 3 unanswered pings, so a half-open stream no longer looks like a frozen terminal.
• Every synchronous tmux call has a 5 s limit. One hung tmux call could freeze the whole daemon, including its signal handling.
• An absent or stale tmux server costs a file check or a connect, not a tmux process. A refused connection never retires a server connection, and on macOS touching the socket file no longer does either.
• A daemon restart in place keeps its port.
• A terminal inventory read that was overtaken twice by changes answered 503. It now keeps reading. This was one cause of a click on an agent that had just appeared doing nothing.
• With --json, all log records go to stderr, so stdout stays one JSON document.
• tmux-ide doctor adds a "tmux server" row with the socket the daemon uses. It warns when this shell's tmux (TMUX_TMPDIR, or a tmux started with -S or -L) means another server. See Troubleshooting.

A cheaper daemon

Measured on a separate daemon, with runs of the old and the new build alternated.

Heavy output in a viewed pane no longer slows tmux itself. The daemon committed after every write, which made it the bottleneck for tmux. It now skips a commit while more output is queued. With a 2 MB flood and one viewer:

A pane nobody has watched for 10 s is parked. tmux pauses its output to the daemon, and the daemon keeps the last frame. Agent status is not affected. The cost: when you return to a parked pane, the last frame appears as fast as before, and the current content follows after about 250 to 290 ms instead of 100 to 130 ms. On a heavily loaded machine it takes longer.

The daemon uses less memory at rest. Its process footprint fell from 327–333 MiB to 283–288 MiB.

The daemon starts fewer tmux processes while idle. The check of its healthy hooks runs every 5 minutes instead of every 30 s: with three tmux servers, 5.8 tmux processes a minute become 0.6. The fleet poll slows from every 500 ms to every 1 s after ten unchanged reads, which halves its tmux processes while idle. A CPU saving from the fleet poll was not measured.

tmux servers that are not tmux-ide's own

Five fixes for a server started by Homebrew's or another stock tmux (3.7c and 3.8 were tested):

• On tmux 3.8, a window with pane-border-status set to top-floating or bottom-floating can be opened. Before, its session had no pane to attach to.
• Opening a session on an agent works. Before, it answered 503.
• Browser panes and the Markdown reader work. Before, both answered 503. The Mac app was not run against a stock server for this fix.
• When the tmux socket cannot be read, the workspace catalog answers 503 tmux-unavailable instead of an empty session list.
• The daemon starts on a tmux socket whose path is too long once /tmp is written as /private/tmp. Before, it exited at startup.

Panes whose content holds tab characters still do not load on a stock server.

Also in 2.9.23

• Clients keep re-reading a terminal inventory the daemon refuses: after 250 ms, 500 ms, then every 1 s for up to 15 s. After that the client reconnects, with 1, 2, 4 and up to 30 s between attempts. Before, it stopped after two quick retries and kept showing a stale terminal.
• The daemon announces when a session's terminal inventory is readable again, so an idle client re-reads it.
• A terminal stream closes with code 1012 and reason tmux-server-retired when its tmux server is gone or replaced, and 1012 daemon-restarting when the daemon restarts in place. 1001 daemon-shutdown now means the daemon stops.
• The daemon announces on the events socket when a tmux server comes online or is retired.

New log lines

The daemon log gains:

• tmux-server-owners: a connection kept after a failed probe, retired, built or failed to build, with the trigger, the probe's duration and its error.
• daemon-lifecycle: a restart in place with its trigger, and a slow or failed stop with its slowest steps. A stop that misses its limit logs stop missed the signal deadline; exiting anyway.
• pane-stream: a shutdown that closes streams, with the code, the reason and the number of streams.

Updating

Update with tmux-ide update or rerun the installer. A running older daemon restarts on 2.9.23 by itself; your tmux sessions keep running. Update remote machines too: the fix is in the daemon on the machine that runs tmux.

Full list of changes: CHANGELOG.md.