[devswha]

What

Settings > Terminal > Draw with libghostty (off by default, this device only). On a PC that runs herdr-core, as the desktop app does, a pane is drawn by libghostty-vt (Ghostty's terminal core, MIT) instead of xterm.js's own text layer, and the pictures a program sends with the Kitty graphics protocol are drawn: chafa -f kitty, timg -pk, kitten icat --transfer-mode=stream.

xterm stays the engine underneath: it still takes every key, IME syllable, paste, click, selection, link, search and scroll, and still answers the program's queries. Its text and cursor layer is hidden; libghostty draws the same bytes, in the same order, as plain DOM rows over xterm's screen at xterm's cell size, plus a canvas per Kitty placement (below or above the text, as Kitty's z says). libghostty answers nothing: a reply it made to a replayed stream would reach the program as typing (herdr-core already answers size queries itself).

This complements #2871 (sixel and iTerm2 through @xterm/addon-image, which has no Kitty graphics): libghostty has no sixel, the addon has no Kitty, and the two layers do not overlap.

How

• src/lib/ghostty/ghostty-vt.wasm: libghostty-vt at ghostty e95bed6 for wasm32, rebuilt by scripts/ghostty-vt/build.sh with kitty-on-wasm.patch. Upstream turns Kitty graphics off on freestanding wasm; the patch turns it on, refuses the file / temp-file / shared-memory media (no filesystem in a browser), and links ghostty's bundled wuffs PNG decoder. 1.11 MB (345 KB brotli). The build now writes .br/.gz copies beside it, but the server still serves precompressed copies for js/css only, so today the module goes out uncompressed (follow-up below). Notices for ghostty, wuffs and uucode in THIRD_PARTY_NOTICES.md.
• src/lib/ghostty/capture.ts: from the terminal's mount (setting on), every byte xterm is given is held; once the server lists its features and says terminal-history (herdr-core), the overlay module is loaded and starts from those bytes. herdr's own server, a module that fails to load, or more than 32 MB held leave the terminal to xterm and never fetch the module.
• src/lib/ghostty/vt.ts, overlay.ts: the wasm wrapper (rows as styled runs, the render state's Kitty image list, palette from the xterm theme, scroll sync with xterm's own herdr-core history) and the DOM layer. Each cell carries its own grapheme and width, so a combining mark never shifts a wide cell.
• CSP: script-src 'self' 'wasm-unsafe-eval' (the same line feat(terminal): draw sixel and iTerm2 inline pictures behind a Settings toggle #2871 adds), so the page may compile WebAssembly; no string eval.
• The choice applies to a pane the next time it is opened (the description says so): libghostty must see the stream from its first byte.

Design reference: Ghostty's own split of a terminal core (libghostty-vt) from the renderer an app draws with, rather than a second terminal widget; everything already built on xterm keeps working while the drawing moves. Clean-room from ghostty's public MIT source.

Validation

• bun run check fast: pass (exit 0) on 312eb94 (merged with origin/main 22a4b33, feat(terminal): draw sixel and iTerm2 inline pictures behind a Settings toggle #2871 and fix(test): read the header crumb apart from the session line in the reconnect regression #2937 included). check full not run here; CI runs the browser lane.
• bun scripts/terminal-ghostty-demo-regression.ts (new, added to scripts/ci-browser.sh): a herdr-core pane is drawn from its first byte (printed before the module loaded), a Kitty picture is drawn at its row and goes on a=d, a wide cell after a combining accent stays the CJK character (fails on the first head), ANSI red under a palette that names no ANSI colours is xterm's red (fails before 8189fde), typing still reaches the pane through xterm; on herdr's own server and with the setting off nothing is drawn by libghostty and the .wasm is never fetched.
• bun test src/lib/settings.test.ts server/static.test.ts server/static-encoding.test.ts scripts/compressed-assets.test.ts scripts/changelog-fold.test.ts: pass.
• Manual, isolated hc 0.1.17 core (own HC_SESSION, XDG dirs, socket) + this branch's server on loopback, headless Chromium at 1280x800 and 390x844: chafa / timg / icat (stream) pictures drawn; with the setting off xterm draws none; resize (rows match xterm), wheel through history (top 10 rows identical to xterm's), drag selection + copy + Quote, Korean text input, less alternate screen and back.
• Known, not in this PR: kitten icat in its default mode gives up because herdr-core answers the DA1 query on the server before the browser's Kitty reply arrives, and hc gives the pty no pixel size; those are herdr-core-side.

Follow-ups (this PR is step 0)

The end state is one rendering path: libghostty draws every terminal, on herdr-core and on herdr, and xterm's drawing code goes.

1. Default on for herdr-core once: the two renderers match row text and cell colours across the themes (a two-renderer browser regression); light-theme contrast correction matches xterm's minimumContrastRatio 4.5; cursor and selection positions match for emoji, CJK, combining marks, Nerd icons and box glyphs; link hover underline and cursor blink are drawn.
2. herdr mode too, then drop xterm's drawing, gated on:
   • serve the module's precompressed copy (server/static.ts compresses js/css only today);
   • phone first load: the extra 345 KB brotli adds at most 0.5 s to the first terminal text on an emulated 4G phone with a 4x CPU slowdown (median of 5, off vs on);
   • repeat loads: 0 wasm bytes on the second visit and across a build that keeps the module (the service worker serves /assets/ cache-first and carries a same-named file into the new build's cache);
   • iOS long-press select and copy on the libghostty rows (iOS Simulator script);
   • screen readers: VoiceOver and NVDA read new output and walk the rows (the layer is aria-hidden today);
   • low-end phone frame time under runaway output at a 6x CPU slowdown: no more 50 ms stalls than xterm, p95 frame gap at most 33 ms.
     Then input, selection and search move to libghostty's key encoder, selection and search APIs, and xterm's parser and renderer are removed.

[coderabbitai]

Warning

Review limit reached

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Next included review available in 4 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used all 4 included reviews currently available.

Learn how review limits work.

Review configuration:

Run configuration

• Source mode: Auto
• Repository: devswha/ovis
• Branch: feat/terminal-ghostty-renderer
• File: .coderabbit.yaml
• Commit: 312eb94
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Advanced
• Run ID: 8c2465e6-a662-4304-8994-7ca36292ce8b

Commits

Reviewing files that changed from the base of the PR and between 22a4b33 and 312eb94.

Files ignored due to path filters: 1

• src/lib/ghostty/ghostty-vt.wasm is excluded by !**/*.wasm

Files selected for processing: 24

• DESIGN.md
• THIRD_PARTY_NOTICES.md
• changelog.d/terminal-ghostty-renderer.md
• docs/development.md
• scripts/ci-browser.sh
• scripts/compressed-assets.ts
• scripts/ghostty-vt/build.sh
• scripts/ghostty-vt/kitty-on-wasm.patch
• scripts/terminal-ghostty-demo-regression.ts
• server/static.ts
• src/components/PaneTerminal.css
• src/components/PaneTerminal.tsx
• src/components/SettingsDialog.tsx
• src/lib/ghostty/capture.ts
• src/lib/ghostty/overlay.ts
• src/lib/ghostty/vt.ts
• src/lib/i18n/settings.ja.ts
• src/lib/i18n/settings.ko.ts
• src/lib/i18n/settings.zh.ts
• src/lib/settings.legacy-fixture.ts
• src/lib/settings.test.ts
• src/lib/settings.ts
• src/lib/useTerminalMount.ts
• src/lib/ws.ts

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn

[devswha]

Review round (omo + Codex Terra 6.1 fast, xhigh) on 0486e36: no HIGH from either. Resolved in the next commit:

• MEDIUM, reproduced (both): a wide cell after a combining mark was placed one code point off. frame() now hands each cell's grapheme with its width; the browser regression covers e\u0301界X and fails on 0486e36.
• MEDIUM (Codex): a cell-size-only change (font size, zoom) rebuilt the rows while libghostty redrew none. Rows are refilled from the last drawn HTML.
• MEDIUM (Codex, omo unreproduced): live queries during module load, replayed queries answered again, Kitty APC replies not recognised as automatic by the server, CSI 6;h;w t read by the attach parser. libghostty now answers nothing at all; chafa, timg and kitten icat --transfer-mode=stream still draw on a live herdr-core 0.1.17 pane (herdr-core answers size queries itself).
• LOW/MEDIUM (both): dispose freed only the terminal and render state. It now frees the row, cell and image iterators and the scratch buffer; no function-table entries are added any more.
• omo unreproduced, taken: the layer hides under a folded mirror (data-folded); a throw from the module during a write returns the pane to xterm instead of stalling its output queue; the per-terminal picture limit is 64 MB instead of 320 MB (one wasm memory for every pane).
• Noted, not changed: 'wasm-unsafe-eval' in script-src (no string eval; the same line as feat(terminal): draw sixel and iTerm2 inline pictures behind a Settings toggle #2871).

[devswha]

Second review round (omo, with its Codex Terra pass) on 14f0036, resolved in 8189fde and the merge of origin/main (d9bd543):

• HIGH, reproduced: under a palette that names no ANSI colours (the dark ones: amber, charcoal, tokyo-night...) the 16 colours were set to black. Colours a theme leaves unnamed are now xterm.js 6's own defaults, as xterm draws them; the browser regression prints ANSI red and fails without the fix.
• MEDIUM, reproduced: the regression waited for xterm's rows to be visible, which the overlay hides on purpose; it now waits for them attached (and paints the fixture after the features, as a herdr-core server does).
• LOW: a throw from the module inside a frame now returns the pane to xterm instead of erroring every frame.
• Merged origin/main with feat(terminal): draw sixel and iTerm2 inline pictures behind a Settings toggle #2871 (Pictures in the terminal): both settings sit side by side, the CSP comment names both WebAssembly users, and both picture regressions pass together.
• Follow-ups, not in this PR (gate 1-C of the default-on plan): ZWJ emoji sequences are three wide cells in libghostty and one two-cell grapheme in xterm, so later cells shift; Kitty layer-1 pictures sit under the whole text layer, cell backgrounds included; a resize while the module loads replays the held bytes at the new grid (unreproduced); one 64 MB picture budget per terminal in one shared wasm memory.

[cloudflare-workers-and-pages]

Deploying ovis with    Cloudflare Pages

Latest commit:	d9bd543
Status:	✅  Deploy successful!
Preview URL:	https://2ef89c75.ovis.pages.dev
Branch Preview URL:	https://feat-terminal-ghostty-render.ovis.pages.dev

View logs

[devswha]

Review at d9bd543: omo CLEAR (with its Codex Terra xhigh pass), no HIGH. Remaining findings, left for follow-ups after two fix rounds:

• MEDIUM (code reading, not reproduced): if libghostty throws while replaying the held bytes, the write/reset wrappers are already installed and are not restored; a later write could then reach the freed module. Only on a module trap or out-of-memory. Fix: install the wrappers after the replay succeeds.
• MEDIUM (reproduced by Codex earlier): ZWJ emoji sequences are wider in libghostty than in xterm, so later cells shift. Tracked as a follow-up issue; it is gate 1-C before default-on.
• LOW: a folded mirror hides the whole layer, pictures included.
• LOW: cell backgrounds libghostty draws cover xterm's selection and search highlights on coloured text.
• LOW: the .wasm is served uncompressed (1.11 MB) because the server serves precompressed copies for js/css only; the PR body now says so, and serving it compressed is on the follow-up list.

[devswha]

Review at 312eb94: omo CLEAR (with its Codex Terra xhigh pass), no HIGH. Non-blocking notes for the follow-ups:

• MEDIUM, reproduced engine-to-engine only (not seen in the app): after a resize the two engines can reflow differently (libghostty resizes on xterm's resize event; held bytes replay at the current grid). Part of parity gate 1-A before default-on.
• MEDIUM, code reading: the replay-failure path (wrappers installed before the backlog replay) — install them after the replay succeeds.
• MEDIUM, code reading: run rendering in runHtml — to be covered by the two-renderer comparison regression of gate 1-A.

[devswha]

Merge note: GitHub required checks green at this head, omo CLEAR at this head, threads resolved. Opt-in renderer (Settings > Terminal > Draw with libghostty), off by default, herdr-core mode only; xterm stays the input, selection and search engine. The wasm build patch enabling kitty graphics is documented in the PR.