Hyperia v0.19.6 — the pane remembers what it was doing ✍️

Save a tab with vim open and get vim back, on the same file, when you restore it.

Running commands come back on restore

Saving a tab workspace already recorded each pane's working directory, but the resume-once checklist in the save confirm only ever listed nemesis8 sessions. Anything else running in a pane (vim, nano, npm run dev, an ssh session) was silently dropped, so a restored workspace was a row of bare shells in the right directories.

The cause was a gap between main and renderer: shell integration reported the running program on every preexec (the OSC 697 command line), but the field the save confirm looked at was never filled, and the renderer's idea of a busy pane (busy) never matched what main actually reports (running). The confirm now reads the shell-reported command line, exactly as you typed it and relative to the pane's directory, and treats a pane as running when either side says so.

Those rows are pre-checked, same as n8 resumes: what was running is what you expect back. Untick one to keep that pane as a plain shell. The safety rule is unchanged: only the shell-integration-reported command or an n8 session binding can ever be executed on restore, and the screen scrape is still display-only.

Bash panes report their first command again

With bash-preexec loaded from ~/.bashrc (common on Ubuntu and Pop!_OS setups), a fresh bash pane never reported the first command you typed. bash-preexec replaces the DEBUG trap at the first prompt and re-invokes Hyperia's hook as one of its own preexec functions, and Hyperia's history-number dedupe had already recorded that number while bash-preexec's install commands ran under the prompt, so the first real command matched the guard and was dropped. Open vim in a new pane, save, and the pane looked idle.

The bash integration now detects bash-preexec and registers with its preexec_functions / precmd_functions arrays instead of fighting it for the DEBUG trap. That also reports the full typed line (echo hi | cat, not just echo hi). Without bash-preexec, the DEBUG-trap path stays, deduped by a prompt marker instead of the history number, and it handles bash 5.1's array-valued PROMPT_COMMAND.

Whole-app restore runs resume-once too

A workspace file carrying resumeOnce behaved differently depending on how you brought it back: the + menu tab restore ran it, workspace_restore and the boot restore ignored it. Both paths now honor it the same way.

Web panes

Web panes already track in-page navigation into the saved URL, so a restored web pane loads the page it was last on. docs/workspace-format.md now says so explicitly.

Save confirm lists your saved workspaces

Under the name field, the Save Workspace… confirm now lists the tab workspaces you already have, newest first, with their pane counts. Click one to prefill its name; the button switches to Overwrite right away, so replacing a saved layout is one click plus Enter. Edit the name and it goes back to a plain save.

Sticky notes, rebuilt underneath

The two oversized sticky files — a 1,655-line main and a 2,175-line renderer — were split into cohesive modules with one owned show/hide path, which is what finally nailed the recurring "hidden stickies come back on start" symptom for good. Same notes, same files on disk; nothing moves on your screen.

Sticky windows also run locked down now: web security on, a strict content-security policy, no stray navigation or popups, and AI highlight rules validated and escaped before they touch the page — a malformed or hostile rule can't inject anything or break out of its attribute. Deeper renderer isolation is tracked separately in #208.

Closing the last window no longer takes four seconds

Closing a window saves the whole session first, then destroys the window. But the app's own close hook ran on that first, held close pass and tore the window down immediately: its rpc destroyed, its shells killed, and the window dropped from the capture set. The save then found zero windows, the sidecar refused the empty workspace, the fallback fired on a dead rpc, and only a 4-second failsafe finally closed the window. Every last-window close took 4 seconds and saved nothing, so a quit from the tray afterwards had nothing to write either.

Teardown now waits for the window to actually be gone (closed); only geometry is recorded on the close pass, while the window is still alive. Measured in the dev build: close is 70 ms including a successful save, and a tray quit with no windows left skips the pointless save and exits in under 100 ms. The quit and close paths also log an elapsed-time trail ([quit +Nms] …, [close +Nms] …) so a slow shutdown can be attributed to a stage.

The tray icon opens a window again

A side effect of that close rework: after the last window closed, the app kept a reference to the now-destroyed window, so clicking the tray icon (or asking for a new window) tried to show the dead one — nothing opened, and on Linux it surfaced an "Object has been destroyed" error. The window is now dropped the instant it's gone, and the tray and new-window paths never receive a destroyed window, so clicking the tray with nothing open reliably opens a fresh one.

Housekeeping

• The Hyperia Agent text link under the new-pane picker is replaced by configure, which opens the agent configuration view in that pane. Launching the agent stays in the picker's agent combobox (once configured), the tab-bar context menu, and the A hotkey.

• workspace-capture unit test caught up with the hidden-stickies restore change from v0.19.1 (its sticky stub lacked the new read).

Coming from further back? v0.19.1 made hidden stickies stay hidden; v0.19.0 added the Saved Sessions pulldown and a quieter close prompt.