Alternate screen not restored after tssh reconnect

[awahlig]

This is somewhat edge-case behavior, but I can reproduce it consistently.

When using tssh (with the reconnect-patched tsshd), a reconnect triggered by rootshell’s eviction while a full-screen TUI application (e.g. vim or helix) is active, causes the alternate screen buffer to be lost. After the reconnect, exiting the editor no longer restores the previous terminal contents.

In addition, in case of helix, the cursor shape changes to underline (I use block normally) inside the editor after reconnect.

Both issues only occur for the editor surviving a reconnect. Closing and reopening the editor restores normal behavior.

rootshell 1.0.2 (26)
iPadOS 26.3

[awahlig]

On second thought, this bug report is somewhat mislabeled. The editor can redraw itself, which is why it becomes visible again after reconnecting. What’s actually lost is the scrollback buffer and whatever was on screen before the editor switched into alternate screen mode. Those contents disappear once the app is evicted.

So the real issue is about preserving that portion of the session state until reconnect. I realize that, depending on how libghosty works, that may not be straightforward.

[kitknox]

In the case of the mosh architecture there is a separate frame buffer with all of the tradeoffs involved. I've been thinking about a broader session restoration enhancement that would cover all session types scroll back history and then would help with this as well.

[kitknox]

Release notes snippet for next build. It does not save alternate screen mode data.

Terminal

- Scrollback Persistence: When a session reconnects or resumes after app
  termination, your previous scrollback history is now restored with full
  ANSI colors and styling intact.  The buffer is saved periodically, on new
  output, and when the app backgrounds.  Orphaned saves are cleaned up
  automatically.  Enabled by default; toggle in Settings.

[awahlig]

I’m testing the Scrollback Persistence feature in build 32 but have issues making it work. Part of the problem might be how I test it. Can you give me some pointers here?

It works in Local Shell if I follow these steps:

• type some commands generating output
• background the app
• wait 15-20 seconds
• kill the app
• reopen

SSH sessions started from local shell also work if this is followed. They reconnect with a fresh session but the previous output is preserved.

The same approach doesn’t seem to work for an SSH profile tab. I think I get a flash of the old screen (although that might be the OS showing the old state while the app launches) but then I end up with empty screen (just the “Last login at” line at the top).

TSSH profile tab is similar, except the state of bash is restored (cwd, current command) and there is no “Last login at”. That difference is expected as bash survives tssh reconnects.

Am I doing it wrong? How do I debug this?

Ubuntu Server 25.10 (no bash customizations, direct connection, no Tailscale)
rootshell 1.0.2 (32)
iPadOS 26.3

[kitknox]

If you scroll up you still have most of your scroll back history even for remote connections right? I think my little whimsical connect animations prevent a clean handoff because they render into the same terminal surface after the history was already restored. I can try moving those into alternate screen mode as one option. As it is the restore is not as seamless as I would like.

I also have another branch with a different experiment that is basically a DVR for all of your sessions where you can search and scrub in a timeline and play them back like a movie. This keeps 100% of the data including interactive sessions in text editors etc.

[kitknox]

I came up with a better solution. I defer doing the scroll back buffer restoration until we have a successful connection. The connecting animation etc renders and clears. On success sessions restores and then the connected starts flowing from there.

[awahlig]

If you scroll up you still have most of your scroll back history even for remote connections right?

I double checked and you’re correct. If the buffer is long enough to require scrolling, the content above is preserved. That would support the animation theory.

I came up with a better solution. I defer doing the scroll back buffer restoration until we have a successful connection.

That sounds like a sensible solution. I’ll report when it’s out.

[kitknox]

This worked out even better than I expected for tssh sessions. Hopefully your results match my testing so far in the latest build.

[awahlig]

The restore is working reliably on my end as well, nice job.

Some observations:

The output appears to only be saved after a scroll event. If the content doesn’t exceed the screen height, it doesn’t seem to be saved (at least not in the local shell where I tested this).

The fancy prompts in local shell are not restored correctly if they’re long:

Also, unrelated but with restored scrollback it feels like a bug that the local shell is losing the cwd on restart. Want me to create a FR for this (saving local shell session)?

[awahlig]

This worked out even better than I expected for tssh sessions.

Yeah I just tested with tssh. Seamless!

[kitknox]

The CWD thing ended up being worse, with there being some incorrect cross-tab contamination going on so I've prioritized fixing that for the next build.

[awahlig]

Nice, CWD restore is working in build 35. Thanks.

The issue of the color bar extending past the prompt after restore is still present (as in my last screenshot).

[kitknox]

The color bar issue required changes to the zig libghostty side to fix. It was technically saving things incorrectly so only newly saved state after running a fixed build will be correct.

[awahlig]

Works great now! 👍

By any chance, could you have a look at that Helix cursor issue I mentioned in the first post? It’s still there and I can’t figure out what’s causing it.

It’s easy to reproduce for me:

• start tssh session
• open hx
• kill rootshell
• restart rootshell

This restores the session and everything works except my normally block cursor changes to underline. It fixes itself as soon as I exit hx but it’s annoying. I tried changing the helix config while it’s still running to see if I can change the cursor shape. I can except every time I set it to back to block, it shows as underline. It’s very weird. Maybe a bug in helix?

[kitknox]

I can reproduce. It does seem that Helix is also doing some of its own cursor management by default. If I change the app cursor setting to underline and then launch Helix it switches me back to a block. The flip back to an underline after app restart happens even if the session scrollback code is disabled. It seems to be something with its internal state vs the default unit of a new ghostly surface. I might need to do something like store more of the internal surface’s state and then re-inject that again. This wouldn’t happen with mosh since its persisting the framebuffer etc and keeping that state on its own.