tracker: add Linux support

[wey-gu]

Linux support — status tracker

Rolling tracker for the Linux port. Live design plan: docs/impl/linux-port.md. Related Windows tracker: #34.

TL;DR

• Linux preview has shipped beyond the original placeholder plan: Unix PTY + shared libghostty-vt + GPUI-owned per-row StyledText renderer.
• The preview renderer supports SGR colors, bold, italic, underline, strikethrough, inverse, block cursor, theme palette sync, IoskeleyMono font normalization, client-side decorations, transparent rounded windows, and KWin Wayland blur where the compositor exposes it.
• PR linux: cache styled rows in preview renderer #65 landed row-level StyledText caching so the current preview path only rebuilds changed rows/cursor rows instead of the whole grid.
• PR perf: reduce non-macOS terminal command latency #68 landed direct PTY output wake and shared VT profiling, removing the old intended 8 ms idle-poll wake floor for terminal output.
• PR [codex] Add Windows and Linux CJK IME input #99 landed CJK IME commit/preedit input for the Linux preview via GPUI InputHandler; remaining Linux IME work is real-desktop validation across Wayland/X11 and IME engines.
• PR Fix:Linux WSL illegal-instruction startup && inline LaTex #100 hardened Linux release portability by building libghostty-vt with an explicit generic Zig target derived from Cargo's target triple, fixing/mitigating the WSL Ubuntu 22.04 illegal-instruction startup report in crash on WSL ubuntu 22.04.5 #97 and reducing host-CPU feature leakage risk.
• The post-perf: reduce non-macOS terminal command latency #68 Linux stale-screen regression after quitting htop / vim was fixed before merge: visible rows are refreshed on each VT generation update and row containers clear across full width. User-side testing confirmed the stale-screen issue was fixed on the branch before merge.
• Release v0 exists for Linux: tarball, one-line installer, desktop entry, icon install, appcast, and notify-only updater. Native package formats (.deb, AppImage, Flatpak, etc.) are still undecided.
• The long-term Linux performance direction remains a GPUI-owned glyph-atlas grid renderer so Linux stops paying layout-time text shaping cost for terminal frames.

Done

Phase 0 — groundwork

• Portable crates build on Linux.
• con UI binary builds on Linux.
• Unix-domain control socket at /tmp/con.sock.
• Linux CI smoke coverage in .github/workflows/ci-portable.yml.
• GPUI Linux runtime dependencies wired in CI.

Phase 1/2 — architecture decision

• Design plan in docs/impl/linux-port.md.
• Feasibility and interop studies in docs/study/.
• Architecture frozen to a local backend rather than waiting for upstream Ghostty/GPUI embedding hooks.
• Shared libghostty-vt layer reused with Windows.

Phase 3/4 — preview backend and renderer

• con-ghostty/src/linux/ facade, Unix PTY spawn, resize, exit, transcript/search plumbing, VT feed path.
• con-app/src/linux_view.rs owns the Linux terminal view path.
• Live local shell through Unix PTY.
• PTY output feeds libghostty-vt and snapshots drive visible rows.
• Per-row styled-cell renderer with SGR colors and common text attributes.
• Cursor rendering and inverse/cursor interaction correctness.
• Theme palette sync from settings.
• IoskeleyMono font-family normalization.
• Client-side decorations and Linux caption controls.
• Transparent rounded window; KWin Wayland blur where supported.
• Alt-screen placeholder flash fixed.

Phase 5 — near-term performance and correctness follow-ups

• Row-level StyledText cache landed in linux: cache styled rows in preview renderer #65.
• Direct PTY output wake landed in perf: reduce non-macOS terminal command latency #68.
• Shared CON_GHOSTTY_PROFILE VT snapshot timing for Linux/Windows comparison.
• Stale visible rows after htop / vim exit fixed before perf: reduce non-macOS terminal command latency #68 merge.
• CJK IME commit/preedit input path landed in PR [codex] Add Windows and Linux CJK IME input #99, including cursor-relative candidate bounds and width-aware CJK/wide preedit positioning.

Phase 6 v0 — release path

• Linux tarball release path.
• install.sh Unix dispatcher and Linux install flow.
• Desktop entry and icon install.
• Sparkle-shaped appcast for notify-only updater.
• Settings update card can surface Linux update availability.
• Generic non-macOS libghostty-vt Zig target selection landed in PR Fix:Linux WSL illegal-instruction startup && inline LaTex #100 for WSL / older-CPU release portability.

Remaining work

Renderer and performance

• Replace the preview per-row StyledText paint path with a GPUI-owned glyph-atlas grid renderer.
• Avoid layout-time shaping for steady terminal frames.
• Run native hardware performance smoke on real Linux desktops for ls, htop, nvim, alt-screen enter/exit, prompt redraws, and bursty command output.

Input, selection, and terminal UX

• Mouse reporting: button, drag, wheel, and xterm/SGR semantics.
• Selection: drag, copy, scrollback interaction, clipboard integration.
• Bracketed paste / DECCKM regression matrix across shells and TUIs.
• CJK IME validation matrix across native Wayland/X11 desktops and common IME engines.
• Scrollback UX and search polish.
• Focus/click behavior across X11, Wayland, GNOME, KDE, and non-KWin compositors.

Desktop integration and packaging

• Decide native artifact strategy: .deb, AppImage, Flatpak, or another distribution-native path.
• Validate installer/update behavior on real GNOME/KDE desktops, not only cloud VM / Crostini / llvmpipe environments.
• Clarify blur/transparency expectations per compositor: KWin Wayland supports app-requested blur; X11/mutter/sway generally do not expose the same app-driven blur protocol.

Verification notes

• Crostini remains useful for build/startup smoke checks, but it is not the primary runtime signal for Linux UI quality.
• The latest confirmed runtime fix from the user was Windows CJK IME input in PR [codex] Add Windows and Linux CJK IME input #99; Linux IME still needs native desktop validation.
• The next useful validation pass is native Linux performance and input testing after the perf: reduce non-macOS terminal command latency #68 / [codex] Add Windows and Linux CJK IME input #99 / Fix:Linux WSL illegal-instruction startup && inline LaTex #100 merge builds.

References

• Plan: docs/impl/linux-port.md
• Milestone notes: docs/impl/linux-port-tracker-updates/2026-04-23-styled-cell-renderer.md, docs/impl/linux-port-tracker-updates/2026-04-23-release-pipeline.md, docs/impl/linux-port-tracker-updates/2026-04-26-render-latency-followup.md
• Postmortem: postmortem/2026-04-23-linux-styled-cell-renderer.md
• Related fixed issue: crash on WSL ubuntu 22.04.5 #97
• Related PRs: Linux preview: real PTY pane, styled cells, CSD + transparent rounded chrome, KWin blur #58, Linux release pipeline: one-liner installer + tarball + notify-only auto-update #59, linux: cache styled rows in preview renderer #65, perf: reduce non-macOS terminal command latency #68, [codex] Add Windows and Linux CJK IME input #99, Fix:Linux WSL illegal-instruction startup && inline LaTex #100

[wey-gu]

./target/debug/con        
[2026-04-22T05:12:27Z INFO  con] con starting (pid 2388)
[2026-04-22T05:12:27Z INFO  con] config loaded
[2026-04-22T05:12:28Z INFO  con] gpui application created
[2026-04-22T05:12:28Z INFO  con] Release channel: con Dev
[2026-04-22T05:12:28Z INFO  zbus::connection::handshake::common] write_command; command=Auth(Some(External), Some([49, 48, 48, 48]))
[2026-04-22T05:12:28Z INFO  zbus::connection::handshake::common] write_commands; commands=[Auth(Some(External), Some([49, 48, 48, 48]))] extra_bytes=None
[2026-04-22T05:12:28Z INFO  tracing::span] read_command;
[2026-04-22T05:12:28Z INFO  zbus::connection::handshake::common] read_commands; n_commands=1
[2026-04-22T05:12:28Z INFO  zbus::connection::handshake::common] write_commands; commands=[NegotiateUnixFD, Begin] extra_bytes=Some([108, 1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 109, 0, 0, 0, 1, 1, 111, 0, 21, 0, 0, 0, 47, 111, 114, 103, 47, 102, 114, 101, 101, 100, 101, 115, 107, 116, 111, 112, 47, 68, 66, 117, 115, 0, 0, 0, 2, 1, 115, 0, 20, 0, 0, 0, 111, 114, 103, 46, 102, 114, 101, 101, 100, 101, 115, 107, 116, 111, 112, 46, 68, 66, 117, 115, 0, 0, 0, 0, 3, 1, 115, 0, 5, 0, 0, 0, 72, 101, 108, 108, 111, 0, 0, 0, 6, 1, 115, 0, 20, 0, 0, 0, 111, 114, 103, 46, 102, 114, 101, 101, 100, 101, 115, 107, 116, 111, 112, 46, 68, 66, 117, 115, 0, 0, 0, 0])
[2026-04-22T05:12:28Z INFO  zbus::connection::handshake::common] read_commands; n_commands=1
[2026-04-22T05:12:28Z INFO  zbus::proxy] {}; task_name="org.freedesktop.portal.Settings proxy caching"
[2026-04-22T05:12:30Z INFO  gpui_wgpu::wgpu_context] Found 2 GPU adapter(s):
[2026-04-22T05:12:30Z INFO  gpui_wgpu::wgpu_context]   - llvmpipe (LLVM 15.0.6, 256 bits) (vendor=0x10005, device=0x0000, backend=Vulkan, type=Cpu)
[2026-04-22T05:12:30Z INFO  gpui_wgpu::wgpu_context]   - llvmpipe (LLVM 15.0.6, 256 bits) (vendor=0x10005, device=0x0000, backend=Gl, type=Cpu)
[2026-04-22T05:12:30Z INFO  gpui_wgpu::wgpu_context] Testing adapter: llvmpipe (LLVM 15.0.6, 256 bits) (Vulkan)...
[2026-04-22T05:12:31Z INFO  gpui_wgpu::wgpu_context] Selected GPU (passed configuration test): llvmpipe (LLVM 15.0.6, 256 bits) (Vulkan)
[2026-04-22T05:12:31Z INFO  gpui_wgpu::wgpu_context] Selected GPU adapter: "llvmpipe (LLVM 15.0.6, 256 bits)" (Vulkan)
[2026-04-22T05:12:31Z INFO  gpui_linux::linux::wayland::window] Server-side decorations requested, but the Wayland server does not support them. Falling back to client-side decorations.
[2026-04-22T05:12:31Z INFO  con::model_registry] Fetching model registry from https://models.dev/api.json
[2026-04-22T05:12:31Z INFO  con_ghostty::vt] VtScreen::new: ghostty_terminal_new(cols=80, rows=24, scrollback=10000)
[2026-04-22T05:12:31Z INFO  con_ghostty::vt] VtScreen::new: terminal=0x7f5887061000
[2026-04-22T05:12:31Z INFO  con_ghostty::vt] VtScreen::new: ghostty_render_state_new(NULL_alloc)
[2026-04-22T05:12:31Z INFO  con_ghostty::vt] VtScreen::new: render_state=0x7f5887040800
[2026-04-22T05:12:31Z INFO  con_ghostty::vt] VtScreen::new: ghostty_render_state_row_iterator_new(NULL_alloc)
[2026-04-22T05:12:31Z INFO  con_ghostty::vt] VtScreen::new: row_iter=0x7f5887030080
[2026-04-22T05:12:31Z INFO  con_ghostty::vt] VtScreen::new: ghostty_render_state_row_cells_new(NULL_alloc)
[2026-04-22T05:12:31Z INFO  con_ghostty::vt] VtScreen::new: row_cells=0x7f5887030100
[2026-04-22T05:12:31Z INFO  con_ghostty::linux::pty] linux pty spawning shell program="/usr/bin/zsh" cwd=None cols=80 rows=24
[2026-04-22T05:12:31Z INFO  con_core::harness] [classify] discovered 219 shell aliases
[2026-04-22T05:12:31Z INFO  con_ghostty::linux::pty] linux pty received first output chunk bytes=16 preview=""
[2026-04-22T05:12:32Z INFO  con::model_registry] Model registry updated from models.dev

[wey-gu]

Landed d1091be (linux: add first live terminal pane).

What changed:

• Linux GhosttyView now starts a local PTY-backed shell instead of rendering the placeholder card.
• The pane paints a sanitized transcript in GPUI and forwards keyboard input for a first interactive bring-up.
• PTY resize now uses an estimated grid from pane bounds/font size so shells are not stuck at 80x24.
• PTY transcript sanitization strips ANSI/OSC control sequences and handles carriage-return rewrites so prompts render legibly before the real grid renderer lands.

Current scope / limits:

• This is not the final Linux renderer yet. It is a live transcript pane, not a full cell-accurate terminal grid.
• No mouse reporting, selection, ANSI styling, alternate-screen fidelity, or scrollback UX yet.

Validation:

• cargo check -p con
• cargo check -p con-ghostty --target x86_64-unknown-linux-gnu

[wey-gu]

Linux update:

• retired the experimental PTY -> sanitized-transcript pane as the primary bring-up path; that abstraction was fundamentally wrong for real shells
• moved libghostty-vt into a shared con-ghostty::vt module so Linux and Windows now consume the same terminal-state contract
• taught the Linux backend to build/link libghostty-vt, feed PTY bytes into VtScreen, and render the Linux pane from VT snapshots instead of transcript text
• restored Linux shell startup to normal interactive-shell behavior now that VT parsing is present
• bracketed paste / DECCKM queries now come from VT state on Linux too
• simplified Linux window bootstrap away from the custom client-decoration path while we validate focus/chrome on real desktops

Validation on this branch:

• cargo check -p con
• cargo check -p con-ghostty --target x86_64-unknown-linux-gnu

Current status: Phase 4 is now genuinely in progress. The remaining Linux work is no longer “make a shell prompt appear from raw transcript text”; it is “finish the GPUI-owned renderer/input/chrome path on top of real VT state.”

[wey-gu]

Pushed two structural Linux follow-ups on branch linux:

• ee320d2 linux: fix vt wake and host callbacks
• 6c999c7 linux: simplify app window chrome

What changed:

• Linux PTY output now advances the app wake generation, and the Linux view drains surface updates from the normal deferred-work loop instead of depending on the macOS-style wake path alone. That fixes a real bug where the pane could stay stuck on the first empty snapshot even though the PTY and VT parser were alive.
• libghostty-vt on Linux now registers more of the terminal host contract: size, color-scheme, device-attributes, and XTVERSION callbacks in addition to WRITE_PTY. Resize updates now feed those callbacks with current geometry.
• Linux window bootstrap was simplified to reduce Crostini/Wayland hit-testing risk: no platform titlebar at window creation, request server decorations by default, remove Linux-specific custom drag/control plumbing from the in-app top bar, and keep custom caption buttons Windows-only.

Validation here:

• cargo check -p con
• cargo check -p con-ghostty --target x86_64-unknown-linux-gnu

I have not yet verified runtime behavior on real Linux after these commits. The next external smoke test should specifically check two things:

1. whether the terminal pane now reaches a visible prompt instead of staying at Waiting for shell prompt…
2. whether Crostini input/click handling is restored under the simplified Linux chrome path

[wey-gu]

Pushed linux: log vt startup diagnostics on branch linux.

This adds one more real host callback (ENQUIRY) and detailed Linux bring-up diagnostics:

• first PTY output chunks now log hex + ASCII previews, not just the sanitized preview
• WRITE_PTY responses from libghostty-vt now log too
• early VT snapshots now log whether Ghostty is actually producing non-empty cells

The point of this cut is to distinguish:

1. shell blocked waiting on a terminal query response
2. shell emitting printable state but our snapshot/render path losing it
3. shell never emitting a prompt at all

[wey-gu]

Pushed linux: poll screen state in view on branch linux.

Latest Chromebook logs proved the shell and VT parser are both healthy:

• zsh emitted a normal prompt sequence (~, .venv, prompt glyph, DECCKM, bracketed paste)
• libghostty-vt produced non-empty cells in the snapshot (non_empty_cells=7)

So the remaining terminal blocker is in con's Linux view refresh path, not shell startup. This follow-up makes the Linux view poll and adopt screen text directly from deferred work even when the current wake semantics miss an update edge.

[wey-gu]

Pushed linux: simplify terminal text rendering on branch linux.

Latest Linux logs showed the prompt is definitely reaching the view layer:

• PTY output includes a normal zsh prompt sequence
• libghostty-vt snapshot contains non-empty cells
• the Linux view cache contains prompt lines (last=Some(".venv ❯"))

So this follow-up strips the Linux text-pane rendering down to a much simpler path: render the cached lines as one preformatted text block instead of the earlier bottom-stacked flex layout. This is intended to eliminate presentation/layout mistakes while Linux is still on the temporary GPUI text renderer.

[wey-gu]

Pushed linux: fix terminal text paint build on branch linux.

The previous render simplification used a nonexistent GPUI API (whitespace_pre), so the Chromebook build failed and the runtime test was still using the older binary. This follow-up keeps the important rendering simplification but switches back to supported GPUI text APIs, so the build is valid again.

[wey-gu]

Merged the current Linux backend work to \ at 2dda418.

Current state:

• local Linux backend scaffold + shared \ lane is in tree
• Linux build/startup path is working
• PTY/VT reaches a live shell prompt
• remaining work is renderer/input polish and native-desktop verification

Verification note:

• ChromeOS/Crostini is still useful for build/startup smoke checks
• it is no longer the primary environment we should trust for Linux UI behavior
• next Linux runtime verification should move to a native Linux desktop

[wey-gu]

Merged the current Linux backend work to main at 2dda418.

Current state:

• local Linux backend scaffold + shared libghostty-vt lane is in tree
• Linux build/startup path is working
• PTY/VT reaches a live shell prompt
• remaining work is renderer/input polish and native-desktop verification

Verification note:

• ChromeOS/Crostini is still useful for build/startup smoke checks
• it is no longer the primary environment we should trust for Linux UI behavior
• next Linux runtime verification should move to a native Linux desktop