T34: review round 1: PowerShell reads a native program's code only when one ran in the command, by the new object each run stores, and 1 for any other failure; bash 3.2 runs the user's own DEBUG trap after its own and keeps $_, and a line that is only a subshell sends its D from the prompt by the command number; zsh marks the first prompt under a theme that sets PS1 in its own hook, through sched; each shell run gets its own folder

AGENTS.md

@@ -96,4 +96,4 @@ ELYTerminal. A terminal on GPUI and Ely components. `PRD.md` says what it does;
 - 2026-10-16: herdr 0.9 keeps its panes' OSC 7 and writes only a title, notifications, OSC 52 and bells to the terminal it runs in, so no terminal hears a folder from inside it; herdr's server knows each pane's `cwd`, the folder its shell reported, else the shell's own. On macOS and Linux each session asks twice a second which process group leads its pty's foreground (`tcgetpgrp` on a copy of the master) and what executable that group's leader runs: an exec keeps the group and swaps the executable, so `exec herdr` from a shell is a new program, and one that did not read is tried again; on Linux the ` (deleted)` mark an update leaves on a running executable is set aside. When that executable is `herdr` and its arguments start herdr's terminal client (no command, `session attach NAME` or `client`), ELYTerminal reads the client's arguments and variables (`/proc` on Linux; `KERN_PROCARGS2` on macOS, which hides only the variables of Apple's own programs) and finds its socket as herdr 0.9 does: `--session NAME`, then `HERDR_SOCKET_PATH`, then `HERDR_SESSION`, under `$XDG_CONFIG_HOME/herdr`, else `~/.config/herdr`; a relative socket lies in the client's folder, as herdr reads it (`/proc/PID/cwd`; `proc_pidinfo` on macOS). It asks that socket for `session.snapshot` twice a second, a connection each, since herdr closes one after its reply, and reports the focused pane's `cwd` as the session's folder when it changes and again each time herdr takes the foreground; an answer that lands after the shell reported a folder is dropped, since the shell reports only once herdr has left. Focus belongs to herdr's server, so clients of one server follow one focus; a client of another machine's server (`--remote`, `--machine`) is logged, not followed. A failure is logged once while it lasts. On Windows, which has no foreground group, a Toolhelp32 snapshot lists the `herdr.exe` processes at or under the session's shell, none under another herdr, since what a herdr starts, its server too, is its own; when that list changes, `sysinfo` reads their arguments and variables, nearest first, names in upper case, as Windows matches them, and the first terminal client is followed, a server or a command beside it passed by; its folder is found as herdr 0.9 finds it there, `XDG_CONFIG_HOME`, then `APPDATA`, then `USERPROFILE\AppData\Roaming`, then `HOME\.config`; and its server answers on the pipe herdr names `\\.\pipe\` and the socket's path as written, relative or not. A busy pipe is waited for, as herdr's client waits, and the pipe's server may identify ELYTerminal but never act as it. The pipe has no timeout: a herdr that never answers holds that session's follow until it answers or exits.
 - 2026-10-16: the main window opens at the size it was left at, which the owner asked for on macOS, Windows and Linux: `window_width` and `window_height` in the settings hold the size gpui's `window_bounds` gives 0.5 s after the bounds settle, saved through `SettingsStore::update`, which writes only a change, so a drag saves once and a move saves nothing. A maximized, full-screen or hidden window keeps the size from before; its state and size are read once they settle, since X11 marks a window maximized after its new size, gpui's X11 window then gives a maximized window's own size, and it calls a maximized window unmaximized while minimized. A window that shows again, uncovered, restored or back on its screen, is read again, so a size it took just before it hid is kept; a size it takes in the half second before it closes, or before it hides until the app quits, is lost. The window is no smaller than 100 by 100, the least the settings keep, or a smaller size the platform allows could not be kept. The window opens centered at whole points, since macOS grows a window by a point on a side whose edge falls on a half point; its place is not kept, and a kept side past the primary display's visible area shrinks to it, logged. A size outside 100 to 16384 stops the start, as any value out of range does.
 - 2026-10-16: the sessions and right columns resize by a drag on their border toward the terminal, which the owner asked for on macOS, Windows and Linux. Ely's `ShellColumn::resizable`, added to Ely for this, puts Ely's `resize_handle` on the border toward the filling column, the sessions column's right and the right column's left: the handle draws that border, a filled hairline, since gpui paints a border no wider than half its box, and the line it drew at first showed half a pixel on macOS and none on X11; it occludes the header under it, so a press there drags the border, not the window, on Windows too, where an unoccluded drag region takes the press as the caption. A drag sets off from the wider of the width drawn and the one kept: a header's parts may hold a column wider, and a window too narrow for both widths squeezes them, where a drag from the squeezed width jumped back against the pointer; a drag that has not moved sideways changes nothing, and one that comes back where it set off gives the width it had. A set width in Ely became a flex basis, so a column holds its header's own parts: at a set width under them, Windows' caption buttons ran out of the window. The widths, 160 to 960 pixels, show at once and are saved in `sessions_width` and `panel_width` through `SettingsStore::update` 0.5 s after a drag rests, as the window's size is; a drag in the half second before the window closes is lost, and a save the store refuses puts the columns back. A window too narrow for both yields the terminal its header's width, as Ely's set widths give way.
-- 2026-10-16: prompts and commands carry OSC 133 marks, which the owner asked for, read as Ghostty reads them: A or N starts a prompt, B or I ends it, C starts a command's output, and D ends the command with its status, D's first option, none when that is no number. The tap reads any program's marks before the parser, so it never sees the cursor: A's fresh line and L, which need it, are not done, and P adds nothing to A. Each session keeps the status its last command ended with. The integration marks what the shell leaves unmarked, and the user's prompt still sees the command's status. zsh: precmd reads `$?` and sends D after a command ran, preexec sends C, and a last precmd hook wraps PS1 in A and B under `prompt_percent`; it moves to the end of `precmd_functions` at each prompt, since a theme that sets PS1 in a hook of its own loads after it, so such a theme's first prompt goes unmarked. bash: PROMPT_COMMAND's first entry reads `$?`, which each entry resets, and its last wraps PS1; from bash 4.4 C rides PS0, whose subscript notes that a command ran, since bash prints no PS0 for an empty line; macOS's bash 3.2 has no PS0, so a DEBUG trap armed at the prompt sends C, unless the user set one. fish 4.3 marks its own prompts and commands, fish 4.0 all but B, older ones none: the script adds what the running fish lacks, around the prompt shown, and wraps it again when the user's changes. PowerShell: the prompt reads `$?` first and sets it back for the user's prompt through `Write-Error '' -ErrorAction Ignore`, which adds nothing to `$Error`; a command that left a new error ended with 1, any other failure with `$LASTEXITCODE`, read through `Get-Variable`, since strict mode refuses an unset variable and a prompt that fails falls back to `PS>`; C goes out when PSReadLine hands over a line that is not blank. herdr passes no OSC 133 on, so a session running herdr keeps the status from before it.
+- 2026-10-16: prompts and commands carry OSC 133 marks, which the owner asked for, read as Ghostty reads them: A or N starts a prompt, B or I ends it, C starts a command's output, and D ends the command with its status, D's first option, none when that is no number. The tap reads any program's marks before the parser, so it never sees the cursor: A's fresh line and L, which need it, are not done, and P adds nothing to A. Each session keeps the status its last command ended with. The integration marks what the shell leaves unmarked, and the user's prompt still sees the command's status. zsh: precmd reads `$?` and sends D after a command ran, preexec sends C, and a last precmd hook wraps PS1 in A and B under `prompt_percent`; it moves to the end of `precmd_functions` at each prompt, since a theme that sets PS1 in a hook of its own loads after it, and at the first prompt `sched +0` wraps it once more, since zsh runs due events after every precmd hook; that call returns the command's status, which the prompt shows, but it would reset `$pipestatus`, so it serves the first prompt alone. bash: PROMPT_COMMAND's first entry reads `$?`, which each entry resets, and sends D once a command ran, and its last wraps PS1. From bash 4.4 PS0 sends C and its subscript notes that a command ran, since bash prints no PS0 for an empty line; both need `promptvars`, on by default. macOS's bash 3.2 has no PS0: a DEBUG trap armed at the prompt sends C and notes it, then runs any trap the user set, with `$?` kept, and `$_` too, since the call is passed `$_` and bash sets `$_` from a call's last word. bash 3.2 runs no trap for a line that is only a subshell, so such a line sends no C, and its D comes from PS1 once `\#`, the command number, has moved; readline sends a prompt's escapes again at each redraw, so D rides PS1 only then. fish 4.3 marks its own prompts and commands, fish 4.0 all but B, older ones none: the script adds what the running fish lacks, around the prompt shown, and wraps it again when the user's changes. PowerShell: the prompt reads `$?` first and sets it back for the user's prompt through `Write-Error '' -ErrorAction Ignore`, which adds nothing to `$Error`; a failure ends with `$LASTEXITCODE` when a native program set it during the command, since each run stores a new object, equal or not, which `[object]::ReferenceEquals` tells apart, and with 1 otherwise, a cmdlet's under `-ErrorAction Ignore` too; it is read through `Get-Variable`, since strict mode refuses an unset variable and a prompt that fails falls back to `PS>`. C goes out when PSReadLine hands over a line that is not blank. herdr passes no OSC 133 on, so a session running herdr keeps the status from before it.