The tab layout of the open project lives in <data_dir>/layout/<project-id>.json
(written atomically, 256 KiB cap both ways, a broken file reads as null),
never in webview storage. New bool setting editor.tabs.preview (default on).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
deploy.browse, readRemote, writeRemote, renameRemote, mkdirRemote,
deleteRemote (confirm: true required), compare, sync (confirmDelete
required for deletions). deploy.upload saves dirty buffers under its
paths first (saveFirst, default true) without queueing uploads on save.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Plum Night, River, Paper and Reed (designed by Amy, approved by Daniel)
join the brand pair. theme.list/current/select/import, a picker with
live preview (view.selectTheme) and "Import Theme…" (view.importTheme);
imported themes are validated and stored in <data_dir>/themes. Editor
font family and line height settings apply live. Docs: docs/themes.md.
Gate: fmt, clippy and the rest pass up to cargo test, where only the
unrelated lutra-terminal PTY test a_full_buffer_stops_the_shell timed
out under load; to be re-run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Servers per project in the config dir (no secrets), secrets in the
keychain via a swappable SecretStore, test connection with the host-key
and missing-secret questions as values, trust host key, upload,
download, cancel, log. One job at a time; uploads on save queue.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- crates/vcs chunk.rs: chunks of a file (buffer text vs index, HEAD vs
index), a minimal patch applied with git apply --cached, revert as a
byte edit; token refusal when buffer, index or HEAD changed; CRLF
(autocrlf and stored), no final newline, BOM, staged renames, first
commit; binary and untracked refused. Real-git tests + property test.
- engine: git.chunks, git.stageChunk, git.unstageChunk, git.revertChunk
(revert edits the buffer under the lock, undoable in the editor);
cmd::start_then; git.lineChanges shares the vcs line diff.
- UI: chunk popup on a gutter marker click (Stage / Unstage / Revert /
Previous / Next), hollow staged markers, vcs.showChunk, diff viewer
hint, "partly staged" badge in the Changes window.
- Docs: G8 row, Git parity row evidence, ui-parity, architecture.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
git.ignore appends one anchored, escaped literal pattern per path to the
work tree root's .gitignore, skipping patterns already there. The Changes
window gets a row context menu (right click, Shift+F10, context-menu key)
with Open and Add to .gitignore.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each command module keeps its own COMMAND_IDS next to its commands; the
engine listing test checks their union against the registry, so a missing
or extra command still fails without one shared list. The menu and palette
tests derive their expectation from the listing and check grouping and
ordering rules; the catalog's runtime keys are scanned from the sources.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- crates/vcs: blame.rs (git blame --porcelain, --contents - from a
buffer, revisions, untracked files and empty repositories give no
lines; property-tested parser, ISO dates without a date crate)
- engine: git.blame {path, rev?} -> {lines, version} through start_with;
an open buffer is blamed as it would be saved (line ending, BOM)
- ui: blame.ts gutter toggled per file (vcs.toggleBlame, Git menu, no
shortcut), date + author per run, bar coloured by age from tokens,
hover card with the full commit message, click opens the commit in
the Log (vcs.showCommit); refreshed on edits and gitChanged
- docs: architecture "Blame (M4.7)"
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- crates/vcs: stash.rs (append) with Repo::stash, stash_list, stash_apply
and stash_drop; refusals decided structurally (no commit, nothing to
stash, unmerged files, conflicts after apply, missing or changed entry)
plus hints for overwritten and untracked files. Works around git leaving
untracked files behind under --literal-pathspecs. Repo::lock is
pub(crate), as for M4.8.
- engine: git.stash, git.stashList, git.stashApply {index, pop, hash?},
git.stashDrop through the M4.6 hook; gitChanged after changes.
- ui: stash dialog (vcs.stash, vcs.showStashes; Git menu, no shortcuts),
two buttons in the Changes window, dirty buffers saved first, drop asks.
- Tests: vcs integration and property tests, engine test, 13 Vitest tests.
- Docs: architecture section, G8, tooling checklist, roadmap, Git row
evidence (score left to the PM).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Backend (crates/vcs/src/merge.rs, cmd/git_merge.rs via start_with):
git.merge {branch} -> {result: merged|upToDate|conflicts, conflicts},
git.mergeAbort, git.conflictVersions {path} -> {base?, ours, theirs,
binary, tooLarge} from the index stages :1:/:2:/:3:, git.markResolved
{paths}. Repo::lock becomes pub(crate) so merge and abort hold the
mutation lock.
UI (ui/src/components/vcs/merge/): merge dialog with branch picker,
conflicts list from the shared status store, a three-pane resolve view
(ours / editable result / theirs) with accept left/right per chunk,
accept all, previous/next conflict, Apply through the file's buffer then
mark resolved, keep-a-side for modify/delete, abort after confirmation.
Own three-way line merge on @codemirror/merge's diff. UI commands
vcs.merge, vcs.resolveConflicts, vcs.abortMerge (no default shortcuts);
the Changes window opens conflicted rows in the resolve view.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Owner decision D23: Lutra runs fetch, pull (merge or rebase per the new
git.pull.mode setting), push (upstream set on the first push) and clone
through the user's git. Any HTTPS or SSH server works; sign-in stays with
git's credential helper (its own window is allowed for network calls only)
and the ssh agent. Terminal prompts stay off, ext:: is refused, and URL
userinfo is redacted in every message, progress line and remote listing.
- vcs: Call::network with a 30 min timeout, a stderr progress sink and a
Cancel flag that kills git's process group; remote.rs with clone_repo,
fetch, pull, push (--porcelain rejections), remotes and add/set-url/
remove; explained refusals (rejected push, sign-in, host key, not found,
unreachable, no upstream, local changes, conflicts).
- engine: git.fetch/pull/push/clone/cancel/remotes/remoteAdd/remoteSetUrl/
remoteRemove, one network call at a time, gitProgress events,
gitChanged after changes.
- ui: status-bar progress with Cancel and ahead/behind, vcs.fetch/pull/
push/clone/manageRemotes (Git menu, no default shortcuts), Clone
Repository and Manage Remotes dialogs; desktop pick_dir IPC.
- Tests against local bare repositories only (vcs 8, engine 3, Vitest 11)
plus unit and property tests; docs updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- crates/vcs log.rs: Repo::log (topological, all refs, paging, literal
message/author filters, a path's history with --follow), show_commit
(message and files against the first parent, renames) and show_at
(a file at any commit); parsers property-tested.
- Engine git.log, git.showCommit, git.showAt in cmd/git_log.rs through
the step-1 hook.
- Log tab in the bottom panel: graph lanes carried across pages, ref
badges, filters, lazy paging, keyboard; commit details whose files open
a read-only diff at that commit in M4.3's viewer.
- vcs.showLog and vcs.showHistory (menu, tree context menu, active
editor tab), no default shortcuts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- git.lineChanges {id} -> {ranges: [{kind, startLine, endLine}], version}:
the buffer against its index blob, line diff with similar (Myers +
Replace, no Compact) off the lock, CRLF-safe; index texts cached per
path and dropped on gitChanged (State::emit -> git_lines::observe)
- editor gutter markers for added, modified and deleted lines, queried
150 ms after edits, on show and on gitChanged, mapped through edits
- vcsStatus store (git.status on projectOpened and gitChanged, no
polling) colouring file names in the project tree and editor tabs with
the --lt-vcs-* tokens; contrast also checked on the active tab
- tests: unit and property tests, real-repo engine tests, Vitest
- docs: architecture, parity, gaps, roadmap, ui-parity, ROADMAP
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
New crate lutra-vcs (L3): a safe git runner (argument arrays, no prompts,
GIT_OPTIONAL_LOCKS=0 for reads, timeouts 30 s / 120 s for hooks, output
caps, one spawn function for H1's kill-on-close helper), porcelain v2 / -z
parsers (property-tested), Repo operations (status, stage, unstage before
the first commit, commit and amend, show at HEAD/index, branches, checkout
with remote tracking, branch creation checked by check-ref-format) and a
.git poll (RepoWatch).
Engine: git.status, git.stage, git.unstage, git.commit, git.show,
git.branches, git.checkout, git.createBranch on pending outcomes, off the
engine lock; gitChanged after changes and debounced 300 ms on .git or
project changes; setting git.executable (section Git).
Tests against real temp repositories, engine tests through run_waiting,
docs (architecture, gaps G8, Git parity row, roadmap M4 row).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Engine (`cmd/lsp_nav.rs`, on the pending outcomes of 20d1d28):
- `lsp.definition` (F12) and `lsp.references` (Shift+F12, sorted, with
line previews) return project-relative locations; locations outside
the project, such as PHAR stubs and symlink escapes, are dropped.
- `lsp.prepareRename` uses the word at the cursor when the server has
no prepareProvider.
- `lsp.rename` (F2) applies the workspace edit (`lsp_nav/edit.rs`):
1. Check everything first: inside the project, ranges fit (UTF-16,
CRLF), open buffers unchanged since the request, no file
operations, no incomplete edit.
2. Write closed files atomically through the VFS, restoring earlier
writes when a later one fails.
3. Change open buffers through `buffer.edit` (dirty, undoable).
4. Send each written file to Phpactor as didOpen/didSave/didClose
plus didChangeWatchedFiles; without this its index kept the old
names. Then the summary notification `notification.lsp.renamed`.
UI (`components/editor/nav/`):
- Mod+click goes to the definition, with a pointer cursor while Mod is
held.
- A location picker for several definitions (exclusive modal).
- A References tab in the bottom panel, grouped by file.
- An inline rename field with a use/file count (exclusive modal).
- `BufferSync.beginUndoableRemote` puts the engine's rename edit into
the editor history as one undo step per buffer.
- Catalog strings and tokens only.
Fake LSP server: `--script FILE` with scripted capabilities and
answers (additive next to M3.4's arms).
Tests:
- `crates/engine/tests/lsp_nav.rs` (fake): definition; bad positions;
references; prepareRename; a multi-file rename incl. a closed CRLF
file; refusals for outside the project, a bad range, a stale buffer,
bad names and no server.
- The ignored real-Phpactor test covers definition, references and a
rename and back across 3 files.
- Vitest `ui/tests/nav/`.
Live (Phpactor 2026.06.25.0, Windows, port 7935, tick on):
- definition 93 ms cold / 7-18 ms warm
- references ~140 ms
- prepareRename ~11 ms
- rename 28-59 ms
- index follows written files in 140 ms
- Phpactor 57 MB working set, ~1.6% of a core while ticking idle
Gate: 13 steps, 197 s, Vitest 484 + 1 skipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Engine (crates/engine/src/cmd/lsp_assist.rs), on M3.5's pending outcomes:
- lsp.completion {id, position, triggerCharacter?}: items with label,
kind, detail, Markdown documentation, insertText, textEdit,
additionalTextEdits, sortText; sorted, capped at 500 (cut lists are
incomplete), snippets turned into plain text.
- lsp.hover {id, position}: Markdown contents and range.
- lsp.signatureHelp {id, position}: signatures with UTF-16 parameter spans.
- The request is queued after every didChange; the answer is converted
in the reply callback, off the engine lock. Positions are checked
against the buffer. Server strings and lists are capped.
- lutra-lsp-fake answers completion and signature help.
Editor (ui/src/components/editor/assist/):
- Completion source with Lutra's triggers (identifiers, $, ->, ?->, ::,
namespace \; not in strings or comments unless asked), the server's
ranking as boost, text edits and imports applied in UTF-16 (CRLF too),
Lucide kind icons.
- Hover documentation (350 ms) and editor.quickDocumentation (Mod+I).
- Signature help on ( and , following the cursor; editor.parameterInfo.
- editor.complete.
- Lutra's own Markdown renderer without HTML (XSS-tested), tokens only.
- Requests wait for the buffer sync to be idle and use runWaiting.
- New dependency @codemirror/autocomplete (MIT, already transitive).
Tests: 6 engine tests against the fake server, 9 unit tests, an ignored
real-Phpactor latency test, 29 Vitest tests; palette and menu tests list
the new UI commands.
Measured with the real Phpactor (Windows 11, debug build, 25 ms tick):
- Completion 61-109 ms while typing (median 76 ms; ~107 ms in the app).
- Hover and signature help ~16 ms median.
- A useful list 2.3 s after opening a file in a 20-class project, ~65 s
for 2,000 classes (indexing).
- Phpactor 56-71 MB working set, ~6% of one core while typing.
Live-checked in the app on port 7934. Docs: Code completion 0% -> 5%
(est.), feature depth ~4.1% -> ~4.6%, breadth 26/155, architecture, G6,
ui-parity, language-parity, roadmap, README.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- One Phpactor per open project on its own session thread, started when
the toolchain is ready; ordered document sync from buffer.edit and the
pump; lspStatus, diagnosticsChanged, diagnostics.list, lsp.status,
lsp.restart; app.quit stops the server synchronously.
- Phpactor 2026.06.25.0 on Windows: keep-alive $/lutra/tick (its event
loop only runs on stdin input), in-process diagnostics and a polling
watcher through initializationOptions (amphp subprocesses fail), 500 ms
shutdown timeout.
- Editor lint extension (underlines, gutter dots, hover; tokens only) and
a Problems tab in the bottom panel; problems.open Mod+Alt+P.
- Tests: engine tests against the fake server (incl. sync proptest), an
ignored real-Phpactor test, a TCP test, Vitest for lint and Problems.
- Docs: architecture, gaps G6, parity (Inspections 4% est.), ui-parity,
roadmap, ROADMAP.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- crates/terminal: portable-pty sessions (ConPTY on Windows) with reader,
writer and waiter threads, UTF-8 chunking, a 1 MiB output buffer with
backpressure, a non-blocking input queue, resize, kill and exit codes;
default shell pwsh -> powershell -> cmd on Windows, $SHELL or /bin/sh.
- Engine: terminal.create/input/resize/close, terminalOutput (capped per
terminal and per pump, round-robin) and terminalExited through the pump;
terminals close with the project and on app.quit.
- UI: Terminal tab in the bottom panel (tabs next to Find), xterm.js with
the fit addon, sub-tabs, copy/paste, a focus rule that leaves plain Ctrl
keys to the shell, terminal.open (Mod+Shift+T), --lt-terminal-* ANSI
tokens in both themes (text colours AA) and a 4.5:1 minimum contrast.
- Tests: real-shell crate tests (UTF-8, backpressure), engine and TCP
control tests, Vitest with a mocked xterm; live check on port 7936;
docs and parity rows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- crates/mcp (L5): JSON-RPC 2.0 over stdio, dual era: MCP 2026-07-28
(per-request _meta, server/discover, -32022) and legacy initialize
(2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05). Tools list_commands,
execute_command (app.quit only with confirm: true), screenshot (image
content), read_buffer, get_events (ring of 1000 events via
events.subscribe); resources lutra://buffer/<path> for open buffers.
Engine errors become tool errors with the engine code. Lines capped at
16 MiB, guard around every request, parser and tool args fuzzed.
- lutra mcp [--port PORT] [--no-start] in apps/lutra-cli: connects through
the rendezvous file and token, or starts Lutra.
- Engine: read-only buffer.list command.
- Tests: protocol units + proptest fuzz, integration over the token
listener, stdio end to end. Docs: architecture (MCP), G7, ROADMAP.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- crates/lsp toolchain: PHP from php.executable or PATH, `php -v` with a
5 s timeout and capped output, version parsing (proptest), PHP >= 8.2.
- Pinned Phpactor 2026.06.25.0 (SHA-256 832510fe...59e9); newer releases
need ext-posix and do not start on Windows.
- Download with ureq + rustls (ring, OS roots via rustls-native-certs):
HTTPS only, 32 MiB cap, SHA-256 verified before an atomic write, retry
with backoff. Tests use a local HTTP server only.
- Engine: settings php.executable, phpactor.path, phpactor.autoDownload
(bool and string setting kinds), toolchain.status,
toolchain.downloadPhpactor, toolchainChanged, notifications; worker
thread off the engine lock, results via the pump. Downloads only after
the host calls Engine::allow_downloads (the desktop does).
- Docs: architecture Toolchain section (network access, D15), G6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Engine: every open PHP buffer keeps a tree-sitter tree, reparsed
incrementally on a worker thread off the engine lock (200 ms debounce,
edit batches recorded by buffer.edit, full parse when the chain breaks).
syntax.outline {id} returns {items, truncated, version}; outlineChanged
{id} at most every 200 ms per buffer. Replaced trees and outlines are
freed on the worker.
UI: left stripe with Project | Structure in MainLayout, the Structure
tool window (WAI-ARIA tree, Lucide kind icons, signatures, jumps via
editor.reveal), view.toggleStructure on Mod+2.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
crates/index: walks the project with the search/VFS rules (vendor as
library), parses .php files from tree-sitter outlines on a worker pool
off the engine lock, updates incrementally from disk and buffers through
an updater thread, caps at 200,000 files, and answers fuzzy file, class
and symbol queries (nucleo-matcher, MPL-2.0 as a separate crate).
Engine: index.status, goto.file, goto.class, goto.symbol; events
indexProgress and indexReady drained by the pump; the index starts on
project.open and is cancelled on close. Tests: unit, incremental,
property (random queries and sources), engine and TCP control tests,
release timing on a generated 20k-file project (3.9-4.2 s).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
crates/search: parallel ignore walk + grep-searcher with the project's
ignore rules (vendor searched as library, node_modules and .git skipped),
literal or regex, case, whole word, include/exclude globs, binary and
> 8 MiB skips, match cap, cancel token, buffer overlay.
Engine: search.find returns {searchId} at once and runs off the lock;
pump_vfs drains results into searchResults (one per file, at most 500
per pump) and searchDone; search.cancel. Testkit: generated search tree.
Tests: library unit and property tests, engine tests, TCP control test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Engine: ui.registerCommands (validated, replaces the previous set),
UI commands in listCommands with ui: true and labelKey, running one
through the engine emits uiCommand {id, params}.
- Frontend: ui/src/commands/ (keymap by key code, fuzzy ranking, Pinia
commands store with handlers, recents and error toasts), the command
palette (Mod+Shift+P), UI commands palette.open, settings.open (Mod+,,
disabled until M1.9b) and view.toggleProjectTree (Mod+1). The editor's
shortcuts.ts moved into the global keymap.
- Tests: engine and control-channel tests for registration and
uiCommand, Vitest for keymap, ranking, store, palette and App wiring.
- Docs: architecture, ui-parity, target-app-parity, gaps, roadmap,
ROADMAP, README.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Dirty buffers are written atomically to <data_dir>/recovery/<project-id>/
about 1 s after the last edit (at most 2 s while typing), from pump_vfs,
never per keystroke. Entries are removed on save, close (also force),
forced project close/open and recovery.discard. project.open emits
recoveryAvailable; recovery.list/restore/discard; restored buffers are
dirty and keep line ending and BOM. Corrupt entries are moved aside and
reported as warning notifications.
UI: restore dialog in ui/src/components/recovery (Restore all, Restore
selected, Discard, time, changed-on-disk hint), catalog keys and runtime
allow-list. Tests: crash/restore with CRLF + BOM, corrupt entries, removal
paths, property test of random edits, headless control-channel test,
Vitest for the dialog. Docs: architecture, parity, gaps, ROADMAP.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- theme::LIGHT next to DARK with the 10 new tokens (editor selection,
active line, cursor, gutter, match, font size, tab active/modified,
notice); light fg-muted, accent and error darkened for WCAG AA
- theme.list, theme.systemPreference, themeChanged event when the
resolved theme changes (settings.set/reset appearance.theme, or the
OS preference in system mode)
- desktop init script and window background from the stored setting;
system mode uses a prefers-color-scheme media query for the first frame
- ui/src/theme.ts applies tokens to :root and color-scheme live; editor,
tabs and notice drop their var() fallbacks
- Vitest WCAG AA contrast check against a Rust-maintained snapshot,
DOM tests, headless control-channel test
- docs: architecture Themes, parity rows, ui-parity, gaps, roadmaps
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
EngineConfig and Engine::with_config; the desktop resolves Tauri's app
config and data dirs. Settings model (appearance.theme, editor.fontSize,
ui.language) with typed validation, persistence in settings.json via
write_atomic (1 MiB cap, unknown keys kept, corrupt file backed up with a
warning notification), settings.list/get/set/reset, settingsChanged and
notification events. Property test over random settings.json content,
control-channel test for persistence and corrupt-file recovery.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The desktop host publishes every engine event, from commands and from a
pump thread that calls Engine::pump_vfs every 100 ms, to control channel
subscribers and to the window as the Tauri event engine://event, in engine
order. ui/src/ipc.ts gets a typed onEngineEvent; its first subscription
calls engine_ready, and the host replays the current project and buffers,
so a project opened from the command line (lutra-desktop --control PORT
PATH) reaches the UI. project.openDialog returns Effect::PickFolder; the
host shows the native folder picker (tauri-plugin-dialog, licence checked)
and opens the choice, one picker at a time, cancel is not an error.
Capabilities shrink to event listen/unlisten.
Contract fix: the M1.5 project UI assumed shapes the engine does not use.
engineEvents.ts now translates engine events ({kind: projectOpened, ...})
and project.children results ({dir, entries, truncated}) into the UI types,
keeping role and truncated; the store maps listings through
toTreeEntries() and the project tests use the real engine shapes.
Verified in the running app on Windows 11: tree of a project given on the
command line, expanding src, a file created on disk appearing live, the
folder picker opening one dialog and opening the chosen folder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
State holds the open project (VFS project, watcher, watched directories)
and versioned buffers. New commands: project.open/close/children,
file.open/save/saveAll, buffer.get/edit/close. buffer.edit takes UTF-16
change sets based on a version and fails with the new stale_version code;
closing dirty buffers or projects without force fails with
unsaved_changes. Commands queue events (projectOpened, treeChanged,
bufferChanged, dirtyChanged, fileChangedOnDisk, ...) that the engine
returns in Outcome.events; Engine::pump_vfs maps changes on disk to
events and reloads clean buffers. project.children watches listed
directories, capped at 256 native / 32 polling, least recently listed
dropped first.
Control channel: events.subscribe streams {"event": ...} lines, with
bounded per-subscriber queues (lutra_control::Subscribers) and method docs
in methods.rs. testkit's EngineBackend and the desktop Host implement the
new Backend::subscribe. Verified with 14 engine integration tests, a TCP
events test and the running app over the control channel.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Command trait with static spec, enabled check and run; Registry rejects
duplicate ids; Engine::execute runs every command inside the guard, which
turns panics into EngineError::Internal and keeps the engine usable.
Effects (Quit) are applied by the host. Theme tokens for Lutra Dark and
Lutra Light render as --lt-* CSS variables and a webview init script.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>