* fix: switch prefix ASCII input source on the foreground client
The macOS prefix ASCII-switch ran in the headless server, whose
TISCopyCurrentKeyboardInputSource read is a stale per-process cache
(refreshed only via a main-run-loop notification the server never pumps),
so prefix could leave the terminal on the wrong input source.
Forward it to the foreground client, like clipboard: sync_prefix_input_source
emits AppEvent::PrefixInputSource as an intent; the server forwards it to the
foreground client and the monolithic app applies it in-process via
RealPrefixInputSource, which pumps its main run loop before the stale TIS read.
Key it on Mode::wants_ascii_input so multi-level prefix commands keep ASCII
until returning to the terminal, and restore the IME for rename text entry.
Adds ServerMessage::PrefixInputSource to the already-unreleased protocol 15.
refs #774
* fix: keep prefix input-source switch out of the headless server process
An App-internal drain (the exhaustive drain at the top of
handle_api_request, reachable from prefix-key runtime mutations) can
consume a queued PrefixInputSource intent before the server's
forwarding drain sees it, applying the TIS switch in the headless
server process and stranding the restore state. Gate the in-process
switch on App.local_input_source_switch, disabled by the headless
server alongside the other local side effects, so a swallowed intent
degrades to a skipped switch. Widen the forwarding-bypass maintenance
test to the handle_internal_event_with_prefix_sync wrapper.
refs #774
* docs: clarify prefix input-source comments and flag docs
refs #774
* fix: bump protocol for prefix input source
refs #774
---------
Co-authored-by: Ogulcan Celik <ogulcancelik@gmail.com>
* fix: request a render on kitty graphics writes
pty output containing a kitty graphics sequence only scheduled the
deferred settle render. that render's wake-up notifies the render loop
only when the dirty flag was not already set, so under a continuous
stream of image updates the notify can be dropped and the pane keeps
showing an old frame until something else forces a full render.
request an immediate render like any other output. the dirty flag and
the minimum render interval still coalesce chunked writes, and the
settle render stays as a backstop.
refs #947
* fix: fingerprint kitty images exactly across retransmissions
the image fingerprint hashed only the first, middle, and last 4096
bytes of pixel data. a program that streams updates by retransmitting
one image id keeps the length and dimensions constant, so a frame whose
changed pixels all fall outside those three windows gets the same
fingerprint as the frame before it. the encoder then treats the image
as already uploaded, emits no graphics bytes, and the frame is dropped
as identical to the previous render. the pane stays stuck on the last
uploaded image until a resize resets the graphics surface.
hash the full payload instead. to keep that off the render hot path,
cache the fingerprint per image id and recompute it only when the
image's transmit time changes. libghostty-vt already refreshes the
transmit time on every transmission; vendor patch 0002 exposes it
through a new GHOSTTY_KITTY_IMAGE_DATA_TRANSMIT_TIME_NS accessor. a
static image costs one map lookup per render and a streaming image one
hash per transmitted frame.
refs #947
* fix: delete superseded host kitty images
host image ids derive from image content, so a pane image whose pixels
change maps to a new host id on every update. nothing deleted the host
image it replaced, and a streaming pane left one full-size image behind
in the host terminal per changed frame until the graphics surface was
reset. track which host image backs each pane image and delete the
replaced one once no source references it.
refs #947
* fix: prune stale kitty graphics source entries
a source entry whose pane closed or whose placement disappeared stayed
in the cache forever. through the shared-reference check it could keep
an old host image alive that nothing visible used anymore, and the
entries themselves accumulated without bound. prune sources against
the visible set at the start of every encode pass.
refs #947
* docs: link issue and upstream discussion for vendor patch 0002
the transmit time accessor now has a removal path: the ghostty
discussion proposing it upstream is linked in the patch ledger.
refs #947
---------
Co-authored-by: Can Celik <ogulcancelik@gmail.com>
Ogulcan Celik
akbash
kangal-bot
Wuyue (Tony) Sun
Mirza Arnaut
ppggff
Gabriel
Ward Peeters
Dillon Mulroy