T40: a press whose record went, a twin Linux stacked or one past the 64 kept, still shows its session; on Linux the body goes as markup's text; only a command C started posts its end, so a prompt bash 3.2 draws again posts nothing; the posting conditions are pure and tested; the decision says what macOS, Wayland and gpui's log do at a press and outside the bundle

AGENTS.md

@@ -102,4 +102,4 @@ ELYTerminal. A terminal on GPUI and Ely components. `PRD.md` says what it does;
 - 2026-10-17: programs ask for notifications through OSC 9, 777 and 99 and report progress through OSC 9;4, which the owner asked for. OSC 9 is ConEmu's command where Ghostty reads one, 9;1 to 9;12 each by its own shape, else iTerm2's notification, its text the title; OSC 777 is rxvt's `notify;title;body`, another extension logged and passed by. OSC 99 is kitty's: `k=v` pairs before the payload, a malformed pair passed by; `i`, of `[A-Za-z0-9_-+.]` alone or the sequence drops, gathers chunks while `d=0`, 32 notifications and 16384 bytes of text at most; a payload holds 2048 bytes plain or 4096 encoded at most; plain text is UTF-8 without a control character; `e=1` reads it as base64 group by group, so chunks split before encoding or after join, and it carries any UTF-8, newlines too, checked once the text is whole; a chunk that breaks a rule drops what its notification gathered; metadata holds 1024 bytes at most, and the tap keeps an OSC 99 body to 5124 bytes, room for metadata and a full payload, where other bodies stop at 4096; `p` takes title, body, close, ? and alive, and other payloads pass by; `a`, `c` and `w` are kept for posting, and `o` passes by; a lone body is the title, and one with neither passes by. The query and alive are answered as the notifications below say; an answer names the asker's id, or 0. OSC 9;4's states, remove, set, error, indeterminate and pause, read as Ghostty reads them, the progress clamped to 100, stand as the root record: set and pause working with their progress, indeterminate working without, error an error, remove none; this lasts until a program reports through OSC 7501, and again after a full reset, as the spec asks, so a mapped 9;4 never wipes a real report. The terminal posts the notifications, as below.
 - 2026-10-17: inside herdr, which passes no OSC 7501 on, herdr's agents are the session's records. The snapshot the follower asks for lists them; one in a state herdr knows, idle, working, blocked or done, is the record `herdr:PANE`, whose colon no OSC 7501 id holds, so a program's record of a name like it, `herdr/build` among them, stays its own; unknown and any state herdr adds later pass by. herdr names no kind and no progress; the app is the agent, the title its name, else the name herdr shows for it, and the message its terminal title less the spinner herdr strips, else the title its metadata sets; herdr's done is idle and not yet seen. A snapshot that lists other agents replaces them whole, so a full reset or a clear that dropped them sees them back at the next; they go when herdr leaves the foreground, and a snapshot that lands after the shell reported a folder is dropped, its agents with it. A change of their states is logged; a herdr whose snapshot lacks `agents` lists none.
 - 2026-10-17: a session's row shows its status, which the owner asked for: one mark before its title, the most pressing state first. Blocked shows its kind, a shield for permission, a circled question mark for a question, a key for a login and a hand for none, in the theme's warning color; error a circled cross; working Ely's `ProgressRing` at its progress with Ely's loader icon inside, since the ring's track takes the theme's border color, which a selected row's ground matches in Ely Dark, where 0% showed nothing, 24 pixels wide, since at an icon's 16 its stroke filled it, and every mark takes that box, so the title stays put as the state changes; else the loader alone, held still, since a turning spinner would draw the window at every frame for as long as a program works; done a circled check; then a failed last command, a circled alert; then idle, a hollow circle. Among the records in that state, the one reported last speaks. Hover over the mark lists each record, ten at most: its state, its app, its own or its nearest ancestor's as the spec asks, its title and its message, each on one line and cut at 120 characters, with C1 controls, direction controls and invisible formatting taken out, as the spec asks; joiners stay, since scripts and emoji need them. A failed last command's status follows. Hover reads the records as they stand at each frame, so an open one follows a change; it stays while the pointer is over it and scrolls within the window's height, so ten long records all show. A key the user types, a paste too, sees the terminal's done and error records, which then go, as the spec suggests; herdr's stay, since herdr marks them seen. The view notes the state its row shows after each change it hears, not once an output, and tells the workspace of each, and a hidden session whose mark turns to blocked sets its title strong until it is shown, though it went on within the same output, so one output marks as two would. A change of herdr's records draws the view again, so an open hover follows them though the state holds. The view logs each change of that state, and keeps its last command's status.
-- 2026-10-17: a session posts system notifications through gpui's, which the owner asked for. At start `set_app_identity` names the app: on Windows that sets the process's AppUserModelID, which the taskbar groups its windows by too, so a pin made before may show apart, and the first notification writes the app's name under `HKCU\Software\Classes\AppUserModelId`, as gpui needs for an app without a package; on macOS gpui posts only from the app bundle, asks for leave at the first, and logs once that it posts nothing outside one; on Linux it posts over D-Bus through notify-rust. Each terminal view posts its own, in the order its output asks, so OSC 99's answers keep their place before a DA1 sent after them: when its row's mark turns to waiting, done or error, at each turn, as the row's waiting mark is set, the hover's line for the record its body; when a command ends, `Command finished`, or `Command exited with N`; and when a program asks through OSC 9, 777 or 99, its title and body the body, each line shown as text. It posts only while its window is not active or it is not the session shown, which the workspace tells it, one every 3 seconds at most, a later one logged and dropped; the title is the program's title, else the app's name, as the row names the session. The session's own share a tag, so a new one takes the old one's place where the system allows, as an OSC 99 id does; OSC 9, 777 and OSC 99 without an id each take their own; a session keeps its last 64. A press goes to the window whose session posted it: it shows the window and the session, unless OSC 99 said `a=-focus`, and tells a program that asked `a=report`, `OSC 99 ; i=ID ;`. gpui hears of no close on any system, so `c=1` hears `p=close;untracked` once posted, and alive lists the ids posted and not yet pressed, closed or expired. On macOS and Windows a close takes the notification back, and `w` over 0 takes it back once it runs out, its session gone or not, unless a later post holds its tag; gpui on Linux closes and replaces none, so there a close and `w` pass by, logged, a closed one stays alive, and an update posts beside the old one. `o` passes by: each notification posts only while unseen, which is kitty's invisible, so the query answers `a=focus,report:o=invisible:p=title,body,close,?,alive:w=1`, on Linux `a=focus,report:o=invisible:p=title,body,?,alive`, never with the `c` that close events would need. Windows' toasts are silent, as macOS's are; on Linux the notification daemon decides.
+- 2026-10-17: a session posts system notifications through gpui's, which the owner asked for. At start `set_app_identity` names the app: on Windows that sets the process's AppUserModelID, which the taskbar groups its windows by too, so a pin made before may show apart, and the first notification writes the app's name under `HKCU\Software\Classes\AppUserModelId`, as gpui needs for an app without a package; on macOS gpui posts only from the app bundle, asks for leave at the first, and logs once, under the `gpui_macos` target, that it posts nothing outside one; on Linux it posts over D-Bus through notify-rust. Each terminal view posts its own, in the order its output asks, so OSC 99's answers keep their place before a DA1 sent after them: when its row's mark turns to waiting, done or error, at each turn, as the row's waiting mark is set, the hover's line for the record its body; when a command C started ends, `Command finished`, or `Command exited with N`, so a prompt drawn again, as macOS's bash 3.2 draws one with its D at a resize, posts nothing, and bash 3.2, which sends no C, posts no command's end, as Ghostty's integration, which skips it, posts none; and when a program asks through OSC 9, 777 or 99, its title and body the body, each line shown as text. It posts only while its window is not active or it is not the session shown, which the workspace tells it, one every 3 seconds at most, a later one logged and dropped; the title is the program's title, else the app's name, as the row names the session. The session's own share a tag, so a new one takes the old one's place where the system allows, as an OSC 99 id does; OSC 9, 777 and OSC 99 without an id each take their own; a session keeps its last 64. A press goes to the window whose session posted it: it shows the window and the session, unless OSC 99 said `a=-focus`, and tells a program that asked `a=report`, `OSC 99 ; i=ID ;`; a press whose record went, a twin Linux stacked and pressed before or one past the 64 kept, still shows its session and tells no program. macOS brings the app forward at any press, `a=-focus` or not, and on Wayland gpui asks for the window with a serial the compositor likely refuses, so KWin and Mutter mark it as asking for attention instead. gpui hears of no close on any system, so `c=1` hears `p=close;untracked` once posted, and alive lists the ids posted and not yet pressed, closed or expired. On macOS and Windows a close takes the notification back, and `w` over 0 takes it back once it runs out, its session gone or not, unless a later post holds its tag; gpui on Linux closes and replaces none, so there a close and `w` pass by, logged, a closed one stays alive, and an update posts beside the old one. `o` passes by: each notification posts only while unseen, which is kitty's invisible, so the query answers `a=focus,report:o=invisible:p=title,body,close,?,alive:w=1`, on Linux `a=focus,report:o=invisible:p=title,body,?,alive`, never with the `c` that close events would need. Windows' toasts are silent, as macOS's are; on Linux the notification daemon decides, and reads the body as markup, so its `&`, `<` and `>` go as entities.