Integrate blitz-dom headless renderer and refactor document model by kevincarlson · Pull Request #9 · AppThere/loki

kevincarlson · 2026-05-04T12:46:27Z

Summary

This PR integrates the blitz-dom headless DOM renderer into the Loki document suite and refactors the document model layer to support bidirectional Loro CRDT serialization. The changes enable rendering documents through a layout engine while maintaining CRDT-based collaborative editing capabilities.

Key Changes

Blitz-DOM Integration

Added complete blitz-dom crate (patches/blitz-dom/) with headless DOM implementation including:
- Core document model (document.rs, node/) with style and layout support
- Layout engine (layout/) with taffy-based box model, inline layout, tables, and lists
- Event handling system (events/) for mouse, keyboard, and IME input
- Form handling and stylesheet loading (form.rs, net.rs)
- Font metrics provider for text measurement
- Default HTML/CSS stylesheet (assets/default.css)

Document Model Refactoring

Restructured loki-doc-model::loro_bridge from single file to modular architecture:
- loro_bridge/mod.rs — public API (document_to_loro, loro_to_document)
- loro_bridge/write.rs — serialization (Loki → Loro)
- loro_bridge/read.rs — deserialization (Loro → Loki)
- loro_bridge/inlines.rs — inline content handling
Moved loro_mutation.rs to loro_mutation/mod.rs with submodules:
- loro_mutation/text.rs — character-level mutations
- loro_mutation/block.rs — block-level structural mutations (split/merge)
Added bidirectional document conversion support with loro_to_document function

Editor Navigation

Added loki-text/src/editing/navigation.rs with cursor navigation functions:
- navigate_left, navigate_right, navigate_up, navigate_down
- navigate_home, navigate_end
- Cross-page navigation support with clamping at page boundaries

Render Cache Refactoring

Simplified PageCache to track tier and dirty metadata instead of GPU textures
Removed RenderQueue and mock_texture.rs (GPU rendering now handled externally)
Updated retier.rs to focus on tier reassignment logic
Refactored DocPageSource to bridge layout cache between document model and page source

Component Updates

Updated DocumentState in loki-text/components/document_source.rs to use Arc<PaginatedLayout>
Modified DocumentView to integrate with PageCache and GPU rendering
Added navigation imports to editor.rs route

Test Updates

Enhanced loro_bridge_tests.rs with comprehensive serialization/deserialization tests
Updated loro_mutation_tests.rs to cover block-level mutations
Simplified GPU integration tests in loki-render-cache

Minor Fixes

Fixed type conversions in OOXML/ODF mappers (e.g., twips_to_pt, f64::from)
Added #[allow(dead_code)] annotations for internal model types
Updated documentation comments across multiple modules
Added rustc-hash dependency to loki-doc-model

Implementation Details

Blitz-DOM uses Servo's style engine for CSS parsing and computed values, with taffy for layout
Document mutations are tracked through damage flags for incremental layout
Loro CRDT containers are mapped to document structure with schema keys for metadata, styles, and content
Navigation functions work in layout points (shared coordinate space with PaginatedLayout)
Form submission and stylesheet loading are deferred through an operation queue in DocumentMutator

https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

…a GPU readback Path B (CPU readback → data URI) was chosen because register_texture takes ownership of wgpu::Texture, which conflicts with PageCache's ownership model. - loki-render-cache: add data_uri: Option<Arc<String>> to CachedPage and set_data_uri() to PageCache; add readback.rs (gpu feature) with texture_to_data_uri() that copies each rendered texture to CPU memory and encodes it as a PNG data URI; wire readback into render_queue worker after every Rerender and Downsample job; add image/base64 deps behind gpu feature - loki-renderer: rewrite DocumentView to snapshot cache state into a Vec before rsx! (dropping the MutexGuard before the Dioxus diffing pass), then render img { src } with the data URI when available, or a blank placeholder while the first settle render is still in progress https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Replace the RenderQueue + CPU-readback path (Session 7) with Blitz's native CustomPaintSource compositing model. Removed: - loki-render-cache/src/render_queue.rs — background render thread is wrong; CustomPaintCtx is frame-scoped and cannot escape render() - loki-render-cache/src/readback.rs — CPU copy_texture_to_buffer path no longer needed - loki-render-cache/src/mock_texture.rs — only existed to support the removed readback path - Byte-budget eviction (evict_cold_overflow) from retier.rs — Blitz owns texture lifetime via TextureHandle; we no longer track bytes - texture/data_uri fields from CachedPage; image/base64 deps from Cargo Added: - loki-renderer/src/page_paint_source.rs — LokiPageSource implements CustomPaintSource, one instance per page. render() reads the current CacheTier, allocates a wgpu texture, paints via loki-vello, registers with Blitz, and reuses the handle when tier/generation/size unchanged. - PageCache::page_count_by_tier() for tracing - impl Default for PageCache Updated: - DocumentView: new PageTile sub-component calls use_wgpu exactly once per page (satisfies Dioxus hook-count invariant); canvas element carries the registered handle ID - RendererState::new() drops device/wgpu_queue params; on_settle() logs via page_count_by_tier() - PageCache::insert() now takes (index, tier) only; no texture arg Quality: cargo clippy -p loki-render-cache clean; 25/25 tests passing. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Replace the hardcoded `current_generation = 0u64` in LokiPageSource with a real generation signal so Loro mutations trigger page re-renders. Option B chosen: DocPageSource holds Arc<AtomicU64> (starts at 1). DocumentState is in loki-text which loki-renderer does not depend on, so Arc<Mutex<DocumentState>> would require an undesirable cross-crate coupling. The atomic gives callers a simple advance_generation() handle with no new crate dependencies. Changes: - doc_page_source.rs: replace OnceLock<PaginatedLayout> with Mutex<Option<(u64, PaginatedLayout)>>; add generation: Arc<AtomicU64>; add current_generation(), advance_generation(), layout_for_generation(). Generation starts at 1 so LokiPageSource (texture_generation = 0) always renders on first frame. Check + recompute are atomic under one lock. - page_paint_source.rs: replace `current_generation = 0u64` with `self.source.current_generation()`; replace `self.source.layout()` with `layout_for_generation(current_generation)`. Guard dropped before render_to_texture to minimise lock hold time. - renderer_state.rs: replace layout() with layout_for_generation; drop guard before acquiring cache lock to avoid two simultaneous locks. - document_view.rs: same pattern; guard dropped before rsx! macro. Quality: loki-render-cache clippy clean; 25/25 tests passing. loki-renderer clippy unverifiable (pre-existing fontique build failure). https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Root cause: yeslogic-fontconfig-sys 6.0.0 removed direct symbol exports from the static-linkage path — FcCharSetAddChar and friends are now only defined inside dlib::external_library!() which is gated on the dlopen feature. The fontconfig-dlopen feature was not included in the system feature bundle so the #[cfg(not(feature = "fontconfig-dlopen"))] block in patches/fontique/src/backend/fontconfig.rs compiled and failed with 28 unresolved import errors. Fix: add "fontconfig-dlopen" to the system feature in both patches/fontique/Cargo.toml and Cargo.toml.orig. fontconfig 2.15.0 is present on this system so dlopen resolution succeeds at runtime. Also rename `gen` variable (reserved keyword in Rust 2024 edition) to `doc_gen` in renderer_state.rs, document_view.rs, and doc_page_source.rs — these were masked by the fontique failure until now. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

One warning in loki-renderer itself: - clippy::question_mark in page_paint_source.rs:191 — let...else returning None rewritten with ? operator Third-party dep warnings (loki-layout: select_header/select_footer dead_code) are out of scope for this crate's clippy pass; addressed in Task D workspace sweep. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

No binary #[ignore] tests exist in loki-renderer. The 3 ignored entries reported by cargo test are doc-tests in /// ```ignore blocks: - renderer_state.rs:11 — use_hook / provide_context snippet; requires Dioxus component context; correctly ignored - doc_page_source.rs:91 — layout_for_generation usage example; can't run standalone; correctly ignored - scroll_driver.rs:45 — use_drop cancellation snippet; requires Dioxus async executor; correctly ignored All three are intentional and require no change. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

…loki-opc loki-opc (1 fix): - part_names.rs: collapse nested if into if && let (clippy::collapsible_if) loki-layout (36 fixes across flow.rs, flow_para.rs, para.rs, resolve.rs, lib.rs): - dead_code: remove select_header/select_footer - collapsible_if: collapse if/if-let chains with && - useless_conversion: remove .into() on String/&str - extend→append: use Vec::append for full drain - loop_variable: use iter_mut().skip().take() instead of index - is_multiple_of: replace x % n == 0 with .is_multiple_of(n) - ptr_arg: &mut Vec<T> → &mut [T] where only iteration is needed - redundant_closure: |x| f(x) → f - clone_on_copy: remove .clone() on Copy types (LayoutColor) - map_or_is_some_and: .map_or(false, f) → .is_some_and(f) - doc_list_item_without_indentation: add 2-space indent in resolve.rs - assign_op_pattern: x = x + y → x += y loki-odf (partial — 8 of 76 fixes; remaining in follow-up commit): - unused imports in lib.rs, export.rs, mod.rs - missing backticks in constants.rs - package.rs: partial cleanup https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

…_integration test loki-odf — 67 warnings fixed across 8 files: - unnested_or_patterns (3): flatten nested Some("a") | Some("b") → Some("a"|"b") - match_same_arms (15): merge End(_)|Eof identical break arms with | - manual_let_else / single_match_else (6): rewrite match-None-return as let…else - map_unwrap_or (3): .map().unwrap_or(false) → .is_some_and(); .map_or(d, f) - collapsible_if (5): collapse nested if/if-let chains with && - field_reassign_with_default (2): use struct initializer with ..Default::default() - cast_lossless (2): u8 as u16 → u16::from(u8) - cast_possible_wrap (1): as i32 → .cast_signed() - redundant_closure (2): |v| v.into_owned() → Cow::into_owned - ref_option (1): &Option<String> params → Option<&String> - useless_conversion (4): remove identity quick_xml::Error::from(e) - too_many_lines (3): #[allow] with comment on XML event-loop functions - single_match (1): match → if local == ... - missing_errors_doc (2): add # Errors sections - len_zero (1): archive.len() == 0 → archive.is_empty() - case_sensitive_file_extension_comparisons (6): use .eq_ignore_ascii_case() - items_after_statements (1): move use to top of function - consider_adding_semicolon (1): add trailing semicolons loki-render-cache/tests/gpu_integration.rs: - Rewrote stale test that referenced removed RenderQueue, PageCache::new(budget), and RetierResult.evicted from pre-Session-8 architecture - New tests: retier_marks_pages_for_rerender (metadata path, no GPU needed) + allocate_and_downsample_texture_succeeds (GPU path, skips if no adapter) https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

… complete loki-ooxml (147 → 0 warnings across 25 files): - cast_lossless: i32 as f64 → f64::from(i32) - unnested_or_patterns: flatten nested Some("a")|Some("b") → Some("a"|"b") - collapsible_if (14): collapse if/if-let chains with && - missing_backticks_in_doc (11): wrap identifiers in backticks - match_same_arms (10): merge identical arms with | - map_or / is_some_and (7): simplify map().unwrap_or patterns - redundant_closure (7): |x| f(x) → f - manual_let_else (5): match-None-return → let…else - needless_pass_by_value (2): T param → &T where value not consumed - large_enum_variant (2): #[allow] with comment (boxing adds alloc overhead) - field_reassign_with_default (2): struct initializer with ..Default::default() - cast_precision_loss (2): #[allow] with comment (precision acceptable for measurements) - wildcard_imports (1): expand import - unnecessary_wrapping (1): Option<T> return → T - too_many_lines (1): #[allow] with comment (XML event-loop function) - assigning_clones (2): use clone_from or direct assignment - binding_too_similar (4): rename shadowing variables - dead_code: #[allow] on model structs not yet wired to output pipeline loki-text (10 warnings across 6 files): - question_mark (2): if is_none / match → ? - collapsible_if (3): collapse nested if/if-let with && - deref_addrof (1): remove explicit * deref - too_many_arguments (1): #[allow] with comment on hit_test_document - clone_on_copy (2): remove .clone() on Signal/Navigator (both Copy) - iter_last (1): filter().next_back() → rfind() Final state: cargo clippy --workspace --no-deps -- -D warnings → 0 errors cargo test --workspace → 0 failures https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

20-page document was allocating 20 × vello::Renderer — one per PageCanvas (LokiDocumentSource) — consuming ~2.83 GB. vello::Renderer is stateless between calls (holds compiled shaders and staging buffers tied to the Device, not per-page data), so all pages for a document can share one instance. Changes: - Add `shared_renderer: Arc<Mutex<Option<vello::Renderer>>>` to `DocumentState` so all `LokiDocumentSource` instances for the same document share one Arc. - `LokiDocumentSource` clones the Arc from `DocumentState` at construction; the per-instance `renderer: Option<vello::Renderer>` field is removed. - `resume()`: first source to run creates the renderer under a lock; the remaining N-1 sources find it already initialised and skip Renderer::new(). - `suspend()`: sets the shared renderer to None (idempotent; first caller does the real drop, subsequent callers are no-ops). - `render()` Phase 5: locks the shared renderer instead of using as_mut(). Sequential render loop in anyrender_vello (values_mut()) means no contention. - Early-return guard simplified: device/queue check retained; renderer initialisation failure is now handled gracefully via the Option::as_mut()? return at the Phase 5 lock site. Memory reduction: 20-page doc 20 × 141 MB → 1 × 141 MB for the renderer. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

…yout via DocumentState Each LokiDocumentSource previously owned a CachedLayout holding a full DocumentLayout (PaginatedLayout + FontDataCache). For a 20-page document this was 20 independent copies of the same data, ~17 MB/copy = ~340 MB total. The fix stores exactly one PaginatedLayout in DocumentState.paginated_layout (Arc<PaginatedLayout>) and shares it across all page sources. Changes: - Remove CachedLayout struct and layout_cache field from LokiDocumentSource - Remove needs_relayout() method — staleness detection moves to DocumentState - Add layout_stamp/layout_generation/layout_canvas_width/layout_preserve_for_editing to DocumentState: bumped whenever the layout is recomputed, enabling per-source texture reuse guards to detect staleness without per-source layout copies - Add shared_font_cache: Arc<Mutex<FontDataCache>> to DocumentState (same Arc pattern as shared_renderer); glyphs loaded while painting page 0 are reused for pages 1-N on the same frame; cache reset on any layout recompute - render() Phase 2: the first source to find the layout stale recomputes and writes to DocumentState; all others on the same frame read the fresh copy. Blitz sequential execution makes this race-free without extra synchronisation. - render() reuse guard: texture_stamp (was texture_generation) now tracks layout_stamp instead of doc generation; added texture_preserve_effective to detect cursor-mode changes that require re-layout without a generation bump - Phase 4: paint_single_page receives shared_font_cache via Arc lock instead of per-source FontDataCache; paint_layout removed (always paginated) - suspend(): drop layout_cache line removed; texture_stamp/texture_preserve_effective cleared alongside handle/size Memory reduction: 20 × CachedLayout (DocumentLayout + FontDataCache, ~17 MB each) → 1 × Arc<PaginatedLayout> + 1 × Arc<Mutex<FontDataCache>>. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Each LokiDocumentSource previously created its own FontResources in resume(), allocating a parley::FontContext (full Fontique font database scan) and a parley::LayoutContext (shaping scratch space) per page canvas — 20 instances for a 20-page document at ≈20 MB each ≈ 400 MB. parley::LayoutContext docs: "designed to be a global resource, one per application." All fields are scratch buffers overwritten at the start of each layout call, so sharing under a Mutex is safe. Blitz sequential rendering means no contention. Changes: - Remove font_resources: Option<FontResources> from LokiDocumentSource - Add shared_font_resources: Arc<Mutex<FontResources>> to DocumentState; initialised with one FontResources::new() at DocumentState construction - LokiDocumentSource::new clones the Arc (same pattern as shared_renderer) - resume(): remove FontResources::new() call — nothing left to do for fonts - render() Phase 2: lock shared_font_resources for layout_document, release before writing results back to DocumentState - apply_mutation_and_relayout: eliminate local FontResources::new() (the 21st instance, created on every keystroke); clone shared_font_resources Arc from DocumentState instead, drop the doc_state lock before acquiring the font resources lock to avoid lock-ordering deadlock Memory reduction: 20 × FontResources (≈400 MB) → 1 × FontResources. Expected total after all three sessions: ~320 MB for a 20-page document. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Implements Session 3b-1: block-level structural mutations for the Enter key (split) and Backspace-at-offset-zero (merge) editing flows. - Refactor loro_mutation.rs → loro_mutation/{mod,text,block}.rs (file was over 300 lines after additions) - Add MutationError variants: InvalidByteOffset { offset } and NoPreviousBlock - split_block: splits LoroText at a UTF-8 byte boundary, inserts a new LoroMap into the LoroMovableList, copies KEY_TYPE / KEY_PARA_PROPS / KEY_DIRECT_CHAR_PROPS from the source block - merge_block: appends block N's text to block N-1's LoroText, removes block N from the list, returns the byte-offset join point for cursor placement - 15 new unit tests covering split middle/start/end/Unicode/invalid- boundary/OOB, merge basic/offset/at-zero/OOB/removes-block, and a split→merge round-trip - Export split_block and merge_block from loki-doc-model crate root https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Session 3b-2: implements all five navigation/editing stubs in editor.rs. Navigation logic extracted to editing/navigation.rs (pure layout functions, no Dioxus/Loro dependencies) to keep editor.rs as a thin event dispatcher. New editing/navigation.rs: - navigate_left / navigate_right: grapheme boundary movement; cross-paragraph on same page (verified via PageParagraphData::block_index lookup) - navigate_up / navigate_down: cursor_rect → canvas-local coords → hit_test_page; cross-page deferred (TODO(3b-3)) - navigate_home / navigate_end: cursor_rect y-center → hit_test_point at x=0 or x=100_000 pt on the same line - 8 unit tests covering left/right at boundaries, home/end, up/down at edges editor.rs changes: - ArrowLeft/Right: call navigate_left/right with get_block_text closure; Shift+Arrow extends selection (focus only), plain Arrow collapses both - ArrowUp/Down: call navigate_up/down; same Shift semantics - Home/End: call navigate_home/navigate_end - Enter: call split_block + apply_mutation_and_relayout; cursor moves to (paragraph+1, offset 0); TODO(3b-3) for page_index recompute after split - Backspace at offset 0: call merge_block + apply_mutation_and_relayout; cursor moves to (paragraph-1, merged_offset) - All doc_state locks use unwrap_or_else(|e| e.into_inner()) for poison safety https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Keyboard events were not reaching the onkeydown handler because the scroll container's tabindex="0" only receives focus when its background is clicked directly — not when child canvases (which have no tabindex) receive the click. Fix: add on_keydown: EventHandler<Rc<KeyboardData>> prop to WgpuSurface and move tabindex="0"/onkeydown to WgpuSurface's outer wrapper div so that the element that receives pointer clicks also holds keyboard focus. Handler logic stays in editor.rs and is forwarded up via the prop. Rc<KeyboardData> is used because KeyboardData does not implement Clone; Rc<T>: Clone provides cheap shared ownership of the event data. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

…andler Temporary diagnostic to trace whether keyboard events reach the on_keydown handler after the tabindex focus fix. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

handle_click walked up the DOM from the hit element but called clear_focus() for any element that wasn't input/label/a. This meant clicking a wgpu canvas (or any non-input child) unconditionally cleared keyboard focus, preventing onkeydown from ever reaching a parent div with tabindex="0". Fix: read is_focussable() before the element data borrow each iteration. After the match arm, if the node is focusable, call set_focus_to() and return instead of continuing the walk. The clear_focus() at the bottom is only reached when no focusable ancestor exists at all. Vendors blitz-dom 0.2.4 to patches/blitz-dom for this change. Remove when upstream blitz-dom implements tabindex click-to-focus for non-input elements. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Temporary println! probes at every layer to pinpoint where the click→ focus→keydown chain breaks: - patches/blitz-dom/src/events/mouse.rs: "BLITZ FOCUS: setting focus to node N" when handle_click finds a focusable ancestor - wgpu_surface.rs outer div: OUTER DIV ONCLICK, OUTER DIV ONFOCUS, OUTER DIV ONKEYDOWN — distinguishes whether focus lands but keyboard events don't reach the Dioxus handler - editor.rs on_keydown: "KEY: ..." at the top before any guards Remove all println! once the root cause is confirmed. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Two new print layers to isolate stale-node hypothesis: - patches/blitz-dom/src/events/driver.rs: DRIVER_KEYDOWN prints focussed_node_id and whether the node still exists in the DOM tree at the moment KeyDown is processed. Stale node → re-render replaced the outer div after set_focus_to was called. - patches/dioxus-native-dom/src/dioxus_document.rs: DIOXUS_HANDLER prints the full bubble chain for every KeyDown event, and for each node whether get_dioxus_id resolves to a Dioxus ElementId. A chain of nodes with no dioxus_id confirms the focused node no longer has data-dioxus-id (stale after re-render). https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Adds println! inside clear_focus() itself to prove it's being called and identify whether it fires between set_focus_to and the next KeyDown. Expected output sequence after click + key: BLITZ FOCUS: setting focus to node N OUTER DIV ONCLICK CLEAR_FOCUS: clearing focus from node N ← where does this come from? DRIVER_KEYDOWN: focussed_node_id=None, ... Once the CLEAR_FOCUS line appears we know exactly which code path calls it (handle_click end-of-walk, or somewhere else). https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

When a mouse click triggers an on_mousedown handler that updates a signal, Dioxus re-renders the component tree during the next poll(). This replaced the focused blitz node (e.g. the outer tabindex="0" div) with a new node ID, and the subsequent handle_click default action walked a stale tree and called clear_focus(). Fix by snapshotting the focused node's stable Dioxus ElementId before render_immediate() and re-applying set_focus_to() after render if the blitz node ID changed. Also exposes focus_node_id as pub on BaseDocument (in the blitz-dom patch) so dioxus-native-dom can read it. Remove all diagnostic println! statements added during debugging. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

Add numbered debug printlns to trace the full keyboard event path: - DBG0: poll() focus state before/after render_immediate - DBG1: focus_node_id seen at KeyDown dispatch time in driver - DBG2: whether on_keydown handler fires in editor, and what editor mode - DBG3: cursor_state.focus value at handler entry - DBG4: insert_text / apply_mutation_and_relayout result https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

flush_is_focussable() was only called in ElementData::new(). Dioxus's template renderer creates elements with Vec::new() (no attrs) then applies static attributes like tabindex="0" via set_attribute(), so is_focussable was always left as false for template-created elements. handle_click()'s tabindex-focusable check therefore never fired, and every canvas click fell through to clear_focus(). Fix: call flush_is_focussable() in set_attribute() and clear_attribute() when tabindex or disabled changes. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

loro_to_document() produces Document::new() with an empty StyleCatalog because the catalog is not stored in the CRDT. apply_mutation_and_relayout() was therefore rendering with no styles on the first keystroke, causing the document to lose all formatting (fonts, paragraph spacing, headings, etc.). Fix: before re-computing the layout, copy styles and source from the previously stored document in doc_state into the reconstructed document. The catalog is effectively read-only during an editing session, so copying it from the last good state is safe. Also remove all DBG0-DBG4 diagnostic println!s now that the focus and formatting bugs are resolved. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

…lization Replace Debug-string stubs in map_para_props/map_loro_block with correct encode/decode helpers so direct_para_props and direct_char_props on StyledParagraph survive a CRDT round-trip. Serialization changes: - ParagraphAlignment: variant name string ("Left", "Center", …) - Spacing: "Exact:<pts_f64>" / "Percent:<pct_f32>" - LineHeight: "Exact:<pts_f64>" / "AtLeast:<pts_f64>" / "Multiple:<mul_f32>" - ListId: bare inner string (previously format!("{:?}", …) which wrapped it in 'ListId("…")') - direct_char_props: flat LoroMap using bool/f64 for scalar fields and Debug strings for UnderlineStyle/StrikethroughStyle/VerticalAlign Deserialization additions: - reconstruct_para_props(): reads KEY_PARA_PROPS sub-map back into ParaProps - reconstruct_char_props_from_map(): reads KEY_DIRECT_CHAR_PROPS sub-map back into CharProps - Helper accessors (get_str/f64/bool/i64_from_map) used by both functions tab_stops and color fields are intentionally deferred (complex types with no clean round-trip format yet). https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

… round-trips - Split loro_bridge.rs (769 lines) into mod.rs, write.rs, read.rs, inlines.rs - Add PageLayout serialization: page_size, margins, orientation, columns - Add six HeaderFooter slots (header/footer/first/even) with recursive block content - Add full inline mark reconstruction: underline, font_size, font_family, letter_spacing, word_spacing, vertical_align, link_url, small_caps, etc. - Add clean tagged-string encode/decode for Spacing and LineHeight enums - Add schema constants for layout, margins, columns, and header/footer keys - Add 8 new round-trip tests covering all newly supported features https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

…coding Replace Debug-format string encoding for color fields with proper encode/decode: - DocumentColor::Rgb → to_hex() / from_hex() (#RRGGBB); non-Rgb variants (Theme, Cmyk, Transparent) are gracefully dropped with a TODO comment - HighlightColor → Debug-format names ("Yellow" etc.) with a new decode_highlight_color() match table for the full palette - Add background_color to the map path (write.rs / read.rs) — was missing - Wire color and highlight_color decode in read_char_props_from_marks - 3 new tests: Rgb color survives, Transparent drops gracefully, Yellow highlight survives https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

claude added 27 commits May 3, 2026 14:04

chore(loki-text): add temporary KEY diagnostic println to onkeydown h…

ea7e7f1

…andler Temporary diagnostic to trace whether keyboard events reach the on_keydown handler after the tabindex focus fix. https://claude.ai/code/session_0152a892tWygTUHBHRHJTE6d

kevincarlson merged commit 999e109 into claude/improve-loki-documentation-rfjR6 May 4, 2026

kevincarlson deleted the claude/virtual-scroll-positioning-cXMOT branch May 4, 2026 12:47

AppThere locked as resolved and limited conversation to collaborators May 4, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

Integrate blitz-dom headless renderer and refactor document model#9

Integrate blitz-dom headless renderer and refactor document model#9
kevincarlson merged 27 commits into
claude/improve-loki-documentation-rfjR6from
claude/virtual-scroll-positioning-cXMOT

kevincarlson commented May 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

kevincarlson commented May 4, 2026

Summary

Key Changes

Blitz-DOM Integration

Document Model Refactoring

Editor Navigation

Render Cache Refactoring

Component Updates

Test Updates

Minor Fixes

Implementation Details

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants