Ghostty update and OSC 7501 — 2026-10-09

Target and acceptance

Work on codex/ghostty-latest-osc-7501, committing tested increments. The clean starting point is ad857e6f. Advance the native dependency from b40acce58dcf77df52231c3798ea58e924647c89 to upstream main observed on October 9, 246f702876b924a1cb7cade1e99274d1470302fc. Audit every intervening first-parent change, port relevant behavior to the managed engine and shared host, update the public ABI, and validate native/managed behavior with focused xUnit tests.

Ghostling HEAD is still 63842bf8e5e481160f81d348da9ff6fd27986798 (August 9), matching the existing integration reference. GitHub's commits/main response was checked on October 9; there are no subsequent changes to port.

Reference decisions before implementation

• Ghostty's new public GHOSTTY_TERMINAL_OPT_PROGRAM_STATUS callback validates OSC 7501 reports and answers detection queries. Its public semantic-prompt and reset callbacks provide the lifecycle hooks. Use these APIs directly; no source overlay or separate native byte-stream parser for this protocol.
• Implement shared per-terminal bounded records, ancestry/app inheritance, full replacement, subtree clearing, least-recently-updated eviction, prompt and process-exit cleanup, and RIS reset. DECSTR and alternate-screen switching retain records. The managed parser follows the upstream parser and the protocol specification, revision 0.3. Reports never trigger commands or echo their text back.
• Windows Terminal src/terminal/adapter/adaptDispatch.cpp and xterm.js src/common/InputHandler.ts were inspected from upstream. Neither implements OSC 7501 in those dispatchers. Their soft-reset implementations preserve screen content, reset rendition/margins/charset and active saved cursor state. Prefer Ghostty's precise DECSTR contract for cross-engine parity and test differences.
• Windows Terminal's rectangular checksum is feature-gated; the Ghostty checksum extension is opt-in too. Preserve default privacy behavior and match Ghostty's algorithm/flags when enabled, with native differential cases.
• Native parser/library fixes arrive through the pin. Review existing guarded extensions against changed source before refreshing hashes. Keep exact fragment checks and existing regressions. Do not edit the upstream submodule checkout.
• Standalone platform UI/build changes (Swift/GTK, translations, packaging and Ghostty CLI) require a recorded applicability decision; they are not evidence of a new VT feature. Shared-host behavior changes must be evaluated separately.
• PowerShell 4c462e3d39478277bb8e89a7e01783efa375f25d, ConsoleHostUserInterfaceProgress.cs, writes OSC 9;4 progress only with VT, nonredirected stdout and UseOSCIndicator; completion writes OSC 9;4;0. Preserve that independent progress channel rather than replacing OSC 7501 records. No heuristic matching of PowerShell prompts or screen text is added.

Delivery checklist

• Refresh native pin, review overlays, build both native libraries (macOS arm64).
• Bind new public C types/options/callbacks/functions; regenerate ABI audit.
• OSC 7501 parser, bounded record store, native callbacks, lifecycle and host UI.
• Unknown OSC capture, terminators, cancellation and numeric validation.
• DECSTR and RIS palette behavior; charset single shift and Unicode handling.
• DECRQCRA/XTCHECKSUM and explicit host checksum policy.
• Mouse shape, UTF-8 mouse buttons, modified backspace and prompt-click fixes.
• Resize/reflow, saved cursor, line-selection and Kitty placeholder regressions.
• Snapshot decode compression and memory queries, with managed equivalents.
• Evaluate renderer/shader, font, image/zlib, tmux and other upstream deltas.
• Run ABI, unit, native integration, history and headless/build checks.
• Document every delivered behavior and any concrete remaining limitation.

Complete upstream delta audit

All 74 first-parent changes between the recorded pins are accounted for below. “Native pin” means the clean upstream implementation is compiled into libghostty-vt; managed or shared-host equivalents are stated separately. Standalone-app changes are recorded explicitly and are not represented as new RoyalTerminal features.

The input comparison also inspected Windows Terminal mouseInput.cpp and xterm.js MouseStateService.ts. xterm.js uses the same extended-button bits but removed UTF-8/urxvt encodings; Ghostty is the compatibility target for those encodings and X10's three-button limit.

Upstream commits	Change	Native / managed / host disposition
b115e4567	DECSTR	Native pin; managed reset contract and focused state-retention tests.
7ec9e26a2	Formatter origin coordinates	Both exporters restore cursor relative to margins only when margins and modes are exported; ordinary, horizontal-margin and pending-wrap replay tests.
b699ea79f, 4611e04ff	Single shift and Unicode charsets	Managed scalar dispatch consumes single shift once and bypasses legacy mapping above FF; native differential and grapheme-transfer tests.
a4aacd918	OSC 7501	Official native callbacks, full managed protocol parser, shared bounded store, lifecycle and status UI.
ca2356fa0	Prompt-bounded line selection	Shared selector already limits trimming to the semantic range; added native/managed blank-cell regression.
b094a6ba3	UTF-8 mouse button bytes	Shared encoder and native pin; extended back/forward buttons wired through adapters and Avalonia, with encoding and headless tests.
5dc28bb8e	Wide cell cut during resize	Managed resize clears split heads/tails and allocation metadata; both engines covered.
0e0ff282f	Initialize PNG callback output	Native pin; managed decoder initializes its output and uses the shared Skia decoder. PNG regression coverage.
04d685c6a	DECRQCRA and XTCHECKSUM	Native bindings and managed algorithm/flags; explicit shared opt-in policy and all-flag native differentials.
eee709f5d	Ctrl+Alt+Shift+Backspace	Both input paths emit the upstream legacy sequence; modifyOtherKeys coverage retained.
3425025e5	Allocator alignment documentation	C# vtable documentation identifies log2 alignment; allocator-failure paste test exercises callbacks.
33da6848d	OSC 105 color reset routing	Recognized by both parsers; unsupported special-color reset is a no-op, matching native.
8628db1a6, 54bada35d	Compressed snapshot history and memory queries	Native APIs and managed lossless row compression, lazy restoration, COW ownership and documented backend-specific memory estimates.
0081d4530, 364f8472a	Requested pointer shape and empty OSC 22	Public native getter and shared pointer path; empty OSC 22 restores text shape in both engines.
a806905ea	Reject paste after a refused writer call	Native pin plus failing-allocator/contract-violating-reader regression. Managed paste owns its complete encoded bytes before dispatch; it has no native MIME sink trampoline.
76895d97b	Same-size resize documentation	Native already updates cell pixels while preserving margins; managed NotifyResize corrected and both pixel/margin contracts tested.
daff2d8e1	Null OSC command access test	Official accessor returns false; integration regression included.
acf1209ee	RIS palette reset	Native pin and managed override reset; theme/default-color tests.
dc3f73a69, 26e64dfb4	Saved cursor and live pending wrap after resize	Managed pin mapping follows deferred hard breaks and wide/pending-wrap boundaries before clamping; repeated resize and anchor tests.
36ec90a07	Semantic prompt and reset effects	Public native callbacks and managed equivalents; owned command/error bytes, ordered reset and status lifecycle tests.
cc2f7ee4d	Prompt-relative click coordinates	Shared click capability reads public native grid/prompt APIs or managed row markers; history/page/viewport and passive-control interaction tests.
75e7a180d	Shader selection color layout	Shared Skia shader uniforms are bound by name; added missing cursor-text and selection foreground/background uniforms and distinct-color pixel tests.
f9e827093	Windows DLL global constructors	Arrives through native pin/build; no managed static initialization analogue. Windows CI builds the updated DLL.
0538f7535	CAN/SUB cancel OSC	Both protocol parsers and the existing native shell observer cancel partial commands; all-split regressions also cover C0 and ESC termination.
3a3047f6b	Unknown OSC callback	Native API and managed bounded capture through the optional host callback, owned bytes, terminators, frozen limits and cancellation tests.
02a3a049c	Clear Kitty placeholder row flag	Native pin; managed projections already derive from current cells. Full EL/ED clear and image-retention regressions in both engines.
f04a00c07	Strict OSC integers	Native pin; managed unsigned color/progress parsers hardened. Shared notification expiry and OSC 133 signed exit codes already reject separators. OSC 66/3008 are recognized no-ops in both terminal streams; their raw native parser improvements arrive through the pin.
d48c0372f	Wuffs zlib decoder	Native uses upstream Wuffs; managed already uses bounded .NET ZLibStream. New upstream compressed-PNG fixture decodes to identical RGBA in both engines.
c065cd3cd	Search tick after terminal free	Native pin and wrapper lifetime documentation; integration test checks InvalidValue after free. Managed search owns its captured rows and does not dereference freed native state.
35a81a980	Release GPU shaders	Standalone renderer fix is in the pin. Shared TerminalShaderPostProcessor already disposes every realized effect independently and idempotently; no equivalent delayed GPU release path.
ed350cbb4	Share GPU render-device state	Standalone Metal/OpenGL renderer internals are not linked by the renderer bridge. Shared Skia owns device/context resources; reproducing Ghostty device objects would not affect either VT backend.
b1d2b7ef1	Three-level Nerd Font constraints table	Same constraints in a smaller upstream lookup representation, not new glyph behavior. Shared Skia/HarfBuzz does not consume this Zig table; existing glyph constraint/rendering tests remain applicable.
4406075a0	CoreText BMP UTF-16 width	Standalone Ghostty shaper fix is in the pin. RoyalTerminal CoreText fallback already uses Rune.EncodeToUtf16 (one BMP unit, two supplementary units); no private shaper port required.
2fb0c9cac, a85cd2e16	tmux control parsing and pane mouse flags	Upstream control-mode fixes are in the pin. RoyalTerminal has no tmux control-mode viewer/session manager; ordinary tmux VT/mouse modes use the shared terminal path and its mouse-mode tests.
246f70287, c770410db	Theme browser paging/keypad key	Ghostty CLI-only key bindings. RoyalTerminal does not invoke this TUI; no terminal encoder change beyond the separately ported input fixes.
befcdfd2c	ssh-terminfo cache includes port	Ghostty CLI cache; RoyalTerminal transports do not run that command or share its cache.
ac2ae69e0	iTerm2 theme package refresh	Standalone Ghostty resource dependency; RoyalTerminal has its own theme catalog and does not import that package. No new palette protocol or renderer behavior.
f523504ea, bcfae289c, 6467b1dab, 068e15c2d	macOS restoration, Shortcuts, tab appearance/compiler check	Swift standalone app changes; RoyalTerminal uses Avalonia window/docking composition, not these NSWindow controllers or App Intents.
62fa23bfa	Kitty clipboard prompt program name	Standalone macOS prompt presentation. RoyalTerminal clipboard prompts do not display producer-supplied program names as trusted application identity.
4ddf1f79d, 59c2dc032	GTK DPI and EGL failure diagnostics	Standalone platform diagnostics; neither GTK monitor detection nor Ghostty EGL surfaceless display creation is used by the shared Avalonia/Skia host.
ce63fcba6, 8f0dd3709	Danish translations	Ghostty standalone localization assets; no shared terminal behavior or corresponding RoyalTerminal resource files.
da1afb561, d1eb46289	esctest reporting and zon2nix	Upstream CI/package tooling. RoyalTerminal keeps its xUnit/native differential suites and existing cross-platform Zig build pipeline.
9d479dcb1, bad5854f3, 0f171f650, 7551c5bad, a60e9e2a5, 34f39002c, 2febd0116, c3203ea4b, f96c9711b, 822e84272, f5a7f706f, 83edd491e, d5427745d, 1dc485f10, 4da7523fa, b0174d03f, 5b435d548, d67ab3213, 6f51f41d0, 12752b2ac	VOUCHED list updates	Upstream contributor metadata only.

Validation log

Initial environment: .NET SDK 10.0.201; Zig 0.16.0. Native runtime host: macOS. Both native libraries rebuilt successfully. The six changed overlay inputs retain all exact fragment matches; none of the upstream changes replaces the separate dirty-row, hyperlink ownership, clone rollback or notification/resize hooks. Reviewed changes add checksum/reset/parser effects, correct charset printing, reflow pins/wide-cell cuts and add memory/compression APIs. The three other overlay input files are byte-for-byte unchanged. All nine full-file hash guards and fragment-count checks remain enabled.

The native ABI now contains 172 types and 30 callbacks. Source audit and generated ABI checks pass. Release integration build: zero warnings/errors. Focused native API and ABI suite: 152 passed, zero skipped. Cases cover OSC 7501 opt-in queries, BEL/ST replies, validated report fields, prompt/reset callbacks, unknown OSC capture/cancellation, null command access, memory queries, incremental compressed snapshot restoration, pointer reset and opt-in rectangular checksums.

Full required-native integration suite: 274 passed, zero skipped. Initial OSC 7501 cross-engine suite: 12 passed, including every input split, hierarchy, replacement, eviction, prompt/exit/reset lifetimes and malformed-report limits. The adjacent regression suite found two cursor reflow differences from the new upstream pending-wrap correction; these belong to the resize parity work.

OSC 7501 plus presentation/headless suite: 18 passed, zero skipped. Both engines keep 256 records, inherit application names dynamically, preserve done/error/idle on prompt and process exit, retain all records on DECSTR/screen switches, and clear on RIS/new sessions. Native semantic prompt-start events (A/N/P) and their managed equivalents drive transient cleanup. OSC 9;4 remains independent. TerminalControl coalesces immutable snapshots onto the UI thread, handles real transport exit/stop, and discards old processor state on replacement. The app's compiled-XAML status bar displays the most urgent record and all records in a tooltip, labels the originating terminal, and escapes invisible direction text.

Reset/charset regression batch: 440 passed. Follow-up reset, pointer and input tracker batch: 87 passed. DECSTR now resets only Ghostty's defined mode subset, active saved cursor, pen, protection, charset, modifyOtherKeys and palette; it preserves dynamic colors, text/cursor, pending wrap, tabs, links, keyboard stacks and other modes. RIS also restores palette overrides. Single shift is consumed by exactly one input scalar, including combining characters; internal spacers are raw writes and non-8-bit Unicode bypasses legacy charset mapping. Live pending-wrap cursors advance after resize when no longer at the right edge. Empty OSC 22 restores the text pointer in both engines.

Protocol/input/lifecycle batch: 539 passed, zero skipped. Unknown OSC selectors have bounded, owned byte payloads, captured BEL/ST terminators and fixed limits per command; supported-but-malformed payloads and CAN/SUB never reach that callback. OSC 105 is recognized without changing unsupported special colors. Color/progress integer fields reject signs on unsigned fields, separators and whitespace. Ctrl+Alt+Shift+Backspace matches the new native legacy encoding. DECRQCRA/XTCHECKSUM uses explicit opt-in and preserves host defaults across RIS/DECSTR; every flag combination was compared against native using Unicode, combining marks, attributes, margins, empty rectangles and held output. Public lifecycle callbacks expose owned OSC 133 command/error bytes, prompt roles and exit codes, plus ordered protocol-only RIS effects. Native uses the new official callbacks. Managed decoding matches first-option/quoting/percent decoding rules and the 2048-byte capture, including malformed and split input. The shared line selector already bounds whitespace trimming to the semantic range; the new upstream unwritten-gap regression passes in both engines.

October input/resize batch: 241 passed, zero skipped, including six new headless cases. Saved pins account for deferred hard breaks and pending wrap before clamping; shrinking without reflow clears both halves of a split wide glyph and their snapshot metadata. Same-size resize retains margins and updates pixel reports. OSC 133 click_events=1/2 now reach both engines through a focused host capability; prompt-relative rows use absolute coordinates across page boundaries and scrolled viewports. Native reads public cursor/grid APIs. The control excludes modified clicks, drags, selections and application mouse reporting; hosts can disable cursor-click-to-move. Extended back/forward buttons flow through Avalonia, both adapters and every mouse encoding, including UTF-8 button bytes; X10 remains limited to the three standard buttons. Native exposes button IDs 10/11 but its encoder does not emit them, so the shared host does not claim support for them.

Compression/storage regression suite: 2,097 cases, with one anchor expectation updated for the upstream saved-pin correction. Focused completion suite: 78 passed, including compression, COW, incremental admission, reflow/export, same-screen memory queries, shader pixels, origin-relative formatter replay, compressed PNGs, Kitty placeholder erasure and shell-observer cancellation. A warmed local 1,201-row fixture used 4,616,640 bytes of plain row payload versus 304,091 resident bytes with compression (211,839 compressed bytes), about 93% less. Restore took 26.86 ms plain versus 62.21 ms compressed in that sample. These are diagnostic payload/timing measurements, not heap-size or general performance guarantees. Compression is opt-in and does not change history quotas.

Release solution build: zero warnings/errors. Final native integration suite: 276 passed, zero skipped, including search-after-free rejection and a paste reader that incorrectly reports success after allocation refuses its write. History contract suite: 45 passed, zero skipped. Native test-history passed; all 172 ABI types, 30 callbacks and history exports validated. CI also built both native libraries for macOS/Linux/Windows, each x64 and arm64, at 29d4f383 (the native source/pin is unchanged by subsequent managed completion commits).

Full unit/headless suite completed in 16 isolated batches: 7,889 passed, 16 pre-existing skips, zero failures (7,905 distinct executed/discovered runtime test IDs; some theories expand beyond list-tests output). The existing exclusions cover ncurses/mouse PTY harness cleanup, macOS telnet execution and JSON atomic file-store roundtrips. None was added by this update. Total passing .NET cases across unit/headless, native integration and history suites: 8,210. Generated X11 colors and whitespace checks also pass. Final cross-platform CI remains a separate PR check; local runtime validation is macOS arm64.

PR review follow-up

Reviewed the complete change set, including public API compatibility, native bindings, parser/lifecycle behavior, snapshot ownership, input, shaders and host presentation. Fixed two issues found during review:

• Explicit StopPty() detaches the transport's exit callback before stopping it, leaving working/blocked OSC 7501 records behind. The stop cleanup now drops transient records directly, preserving completed records. Headless tests cover both managed and native processors without relying on an exit callback.
• Adding optional constructor parameters removed the existing CLR signatures for TerminalUnknownSequence and TerminalShaderFrameContext, and the sequence record lost its three-value deconstruction. Explicit overloads preserve those signatures alongside the new metadata and shader colors. A consumer compiled against the pre-update API sources runs unchanged with the updated assemblies.

The review regression suite passes 142 cases with no skips, including seven new tests for explicit stop, public compatibility, strict Base64 padding and active pane status routing. Background-pane updates cannot replace the active pane's status; changing focus shows the corresponding snapshot. All six native CI builds and ABI checks also passed at 64456587, whose native code is unchanged by these review fixes.

OSC 133 line-click follow-up (2026-10-11)

Preview 1 parsed cl=line but the two host encoders accepted only click_events=1/2. The follow-up uses Ghostty's pinned Screen.promptClickMove / promptClickLine and Surface.maybePromptClick as the behavior contract: count only input-marked cells, traverse soft wraps, stop at hard line boundaries, clamp clicks beyond input, and use DECCKM to choose CSI or SS3 Left/Right sequences. The other cl values (m, v, w) currently use the same single-line fallback in Ghostty and RoyalTerminal. click_events=1/2 retain priority. A zero-distance click emits no input.

Compared with Windows Terminal's ControlCore::_repositionCursorWithMouse, which clamps to command marks and counts cell distance, Ghostty provides the explicit OSC 133 policy and per-cell input classifications needed here. xterm.js MoveToCell also uses horizontal arrows on the normal buffer and honors application cursor keys, but does not filter movement by OSC 133 input metadata. RoyalTerminal follows Ghostty's cell-based contract, including its current wide-cell counting, rather than inferring shell editing behavior from text or whitespace.

The implementation shares an allocation-free movement calculation between the engines, with managed row access and native public grid/cell/row accessors. Only the emitted byte array is allocated; native references are borrowed under the existing processor lock. No native ABI extension or shell-specific handler is needed. Focused processor and headless input tests cover both engines.

Validation: 186 focused tests pass with zero skips on macOS, including clean Bash and Zsh PTYs with each backend. The tests edit abcdef into abcXdeYf using actual encoded left/right clicks and assert the executed command's result; neither shell installs a mouse handler. Allocation checks confirm only the returned byte array is allocated, and no-op clicks allocate zero bytes. This fix follows the published v0.7.0-preview.1 tag.

Line-click review

Resize/reflow can replace the initial prompt row marker with a continuation. When that continuation reaches the beginning of retained history, Ghostty's PageList.PromptIterator.nextLeftUp returns the cursor's continuation row. Using that as the earliest permitted click incorrectly rejects earlier visible input after reflow, or after the original prompt scrolls out of history. Both RoyalTerminal backends now admit line clicks through the first retained continuation row; the movement calculator still stops at hard line boundaries and counts only input cells. This is a deliberate correction to the upstream host boundary check, using existing public native row accessors. Explicit click_events retain Ghostty's original shell-visible coordinate origin.

Regression tests cover reflow in both directions, truncated prompt history, unchanged explicit mouse-event origins, live cursor-key mode/input during synchronized output, alternate-screen restoration, and reset. The Bash/Zsh test now waits for both repainted text and the final cursor position before sending another click, avoiding a race when a repaint crosses PTY reads. All 196 focused processor, headless, allocation and live-shell checks pass on macOS arm64 with no skipped tests.