Repository navigation
1.10.0 — remote control and --tmux on Windows, and a session list that folds away - #30
Merged
Merged
Conversation
Remote control was switched off on Windows because the hub "did not come up and did not go away" on CI. Measured on a real Windows machine, the hub came up in a second and stopped cleanly. What hung was whatever started it: `Command::spawn` hands the child every inheritable handle, and the caller's own stdout was one of them, so a test harness or a script reading `alc hub start` waited for an end of file that only came when the hub died. The hub is now started with the caller's std handles made non-inheritable for the length of the spawn, and with CREATE_NO_WINDOW rather than DETACHED_PROCESS, so the console programs it runs (a tmux probe, twice a second) share one hidden console instead of each opening a window. The gate is gone, and the hub tests that were unix-only run on Windows too. Three things a Windows pseudo-console does that a pty does not, each of which stopped a session outright: - portable-pty opens it with PSEUDOCONSOLE_INHERIT_CURSOR, so conhost asks for the cursor position and draws nothing until it hears. The hub now answers that one question itself, and keeps conhost's opening requests for win32 input and focus reports away from viewers - under --tmux the page's focus reports would otherwise have been typed into the agent. - Its output never ends when the process in it does; only closing it ends it. A watcher now closes it when the child exits, so a session ends when its agent does. - The local terminal needs virtual-terminal processing and input asked for explicitly, reads Ctrl+Z as end of file through std (so the keyboard is read with ReadConsoleW), refuses to write half a character on a non-UTF-8 code page (so whole characters are written), and is put back on Ctrl+Break or when its window closes. --tmux drives the native Windows port of tmux (arndawg.tmux-windows). It is real tmux and behaves the same for everything alc relies on, but it cannot start a pane given -c, re-splits the command line on spaces, hands the pane TMUX, and passes arguments, environment and working directory through the ANSI code page. So on Windows the pane runs `alc __tmux-pane <port> <token>`, which collects the exact launch from the hub over the loopback control port with a one-time token and runs the agent itself. Non-ASCII paths and values work, and the provider key never enters tmux's environment. psmux, which also installs a tmux.exe, is recognised and looked past; `alc doctor` says whether the native port was found. Found on the way, and not Windows-only: - Detaching (ctrl-\ then d) did nothing until the agent next drew: the relay was still blocked reading a connection nothing closed. The client now half-closes it, and the hub lets a departed terminal go and closes its end when the session ends, so the terminal is not left waiting for a key press either. - A reattached terminal could start blank: the attach reply was read through a buffer that also took the screen sent straight after it. - The control channel now caps and times out the request it reads before checking the secret, since on Windows it is a loopback port any local account can reach. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
On a wide screen with a session open, the list sat beside the terminal at 320px whether it was wanted or not. A button beside Back now folds it away, and the terminal slides over to take the width. For a --tmux session that width is the agent's - the page owns its size - so it gets the extra columns (measured: 239 to 284); a plain session's screen is drawn larger. The terminal is refitted once, where it comes to rest, rather than partway through the slide, so tmux is not told a width it keeps for a fifth of a second. The slide plays only in answer to the button, never when a folded page is widened or a session is attached, and not at all under reduced motion. The folded list is inert from the moment of the press, so a Tab straight afterwards cannot land on a card that is about to vanish. The choice is remembered per browser; with storage blocked the page still works, and reading `sessionStorage` at boot no longer takes the page down there either. There is no keyboard shortcut: every key on this page belongs to the terminal, and ctrl-b is tmux's prefix. The panes' landmark names are now translated too, since the new button points at one of them by name. The documentation site's left sidebar can be collapsed as well. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Both READMEs and both locales of the remote-control page said Windows refused `--share` and the hub. They now say it works, and the --tmux section says what Windows needs: `winget install arndawg.tmux-windows`, that psmux is looked past rather than used, that non-ASCII paths work and the provider key stays out of tmux, and that a session stopped there ends without an exit status because Windows has no hangup signal. The troubleshooting page gains the two refusals a Windows user can meet. The page's description mentions the button that folds the session list away. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Remote control on Windows, --tmux there too, and a session list that folds away. A feature, so a minor release. On Windows, `alc <agent> --share`, the hub, `alc sessions`, `alc attach` and `alc kill` now work instead of refusing. What kept them off turned out to be the hub holding open the output of whatever started it, which made CI wait for ever; the hub itself was fine. Getting a session to actually run under a Windows pseudo-console took three more fixes - conhost waits for an answer to its cursor query before drawing anything, never reports the end of a session on its own, and the local console needs asking for the modes a unix terminal has in raw mode. `--tmux` on Windows uses the native port of tmux: `winget install arndawg.tmux-windows`, then a new terminal. The pane runs a small launcher that collects the agent's launch from the hub, because the port mangles anything that is not ASCII on the way to a pane; non-ASCII folders and values work, and the provider key never enters tmux there. Everywhere: detaching lets go straight away rather than when the agent next draws, a reattached terminal gets its screen back, and the terminal returns to its shell when the agent exits without waiting for a key. On the page, a button beside Back folds the session list away, and a --tmux session gets those columns. Nothing to do when upgrading. On Windows, `alc doctor` says whether --tmux has what it needs. A hub still running the previous version is named by `alc update`; stop it with `alc hub stop` when its sessions are done. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
--tmuxcould not be used on Windows. It turned out that nothing about remote control could:--share, the hub,sessionsandattachwere all switched off there in 7890d6d, after Windows CI hung on the hub tests. This lifts that, makes--tmuxwork on Windows through the native port of tmux, and lets the page's session list fold away so the terminal - and under--tmux, the agent - gets the width.Why remote control was off, and what it actually was
The note on the gate said the hub "does not come up and does not go away" on Windows. On a real Windows machine it came up in a second and stopped cleanly. What hung was the caller: Rust's
Command::spawnhands a child every inheritable handle, and the caller's own stdout is one of them whenever something captured it, soassert_cmd- or any script readingalc hub start- waited for an end of file that came only when the hub died. Reproduced with a pipe-capturing caller, fixed, and the formerly unix-only hub tests now run on Windows.CREATE_NO_WINDOWreplacesDETACHED_PROCESSwhile at it: with no console at all, every console program the hub runs (a tmux probe, twice a second) would open a window of its own.What a Windows pseudo-console needed
Each of these stopped a session outright, and each was measured:
INHERIT_CURSOR; conhost printsESC[6nand draws nothing until answered. A--tmuxmirror never attached.?9001h,?1004h) from viewers - under--tmuxthe page's focus reports would have been typed into the agent.ReadConsoleW, whole characters written.--tmuxon Windowstmux has a native Windows port,
arndawg/tmux-windows- real tmux 3.6 over ConPTY and named pipes. Measured, it behaves exactly as alc needs for the option sequence,remain-on-exitand the pane's exit status,send-keys -H(the prefix opens no prompt),ignore-size, andkill-server. It also:-c, whatever the path;TMUX, which-e TMUX=cannot undo;So on Windows the pane runs
alc __tmux-pane <port> <token>, all ASCII. It collects the exact launch from the hub over the loopback control port with a one-time 256-bit token, removesTMUX, runs the agent in a job tied to itself, and exits with its status. The provider key never enters tmux's environment on Windows. If alc itself lives under a non-ASCII path, its 8.3 short path is used.psmux also installs a
tmux.exeand reports itself as tmux 3.3 on the first line of-V, but cannot run theset-option … ; new-sessionsequence a session is created with.findrecognises it and looks further down PATH, so the two can coexist.alc doctorsays whether the native port was found.Not Windows-only
BufReader, which took the screen the hub sent straight after it.The session list folds away
On a wide screen with a session open, a button beside Back hides the list and the terminal slides over. A
--tmuxsession gets the columns (239 → 284 in Chrome at 1400px); a plain one is drawn larger. It refits once where it comes to rest, slides only when pressed and never under reduced motion, makes the hidden list inert immediately, remembers the choice per browser, and has no keyboard shortcut - every key belongs to the terminal. The docs site's sidebar is now collapsible too.Verified
cargo test --all-targets574 + 67, including 7 new live tests against tmux-windows (sizing vote, exit status,kill-server, prefix key, and the launcher end to end with non-ASCII args/cwd/env).--share --tmuxand plain--share- typing, arrows, Ctrl+Z, detach, reattach with the screen intact, exit statuses,alc kill, no processes left behind.cargo clippy -D warningsfor Windows and forx86_64-unknown-linux-gnu;node --test web/*.test.js64/64; the docs site builds in both locales with broken links and anchors fatal.Not verified here: the unix runtime tests (Linux was type-checked, not run - CI covers it), and typing by hand in Windows Terminal itself.
🤖 Generated with Claude Code