Kill the entire D-Bus rendering pipeline. The GUI now uses plan9port's
9p command as a subprocess for all server communication:
- Chat streaming: persistent 'read chat' process, readyReadStandardOutput
delivers tokens directly to QML (no polling, no signals, no offsets)
- State: 'read statewait' process, restarts on each state change
- Session list: 'read session/idx' one-shot
- Submit/stop/kill: 'write prompt/ctl' one-shot
The QML is stripped to fundamentals: session list + monospace streaming
text (last 8KB tail) + input field. No ChatBlockModel, no ThemeManager,
no OllieDBusClient, no syntax highlighting, no markdown rendering.
This eliminates the multi-second GUI freezes caused by:
- QTextDocument relayout on 200KB+ accumulated text
- loadEarlier cascade creating hundreds of delegates
- KSyntaxHighlighting::Repository blocking init
- Offset mismatch triggering synchronous D-Bus calls
Features will be re-added incrementally on this proven foundation.
D-Bus retained only for plasmoid/tray/KRunner (desktop integration).
- Detect mermaid/plantuml/drawio/mxgraph code fences as 'diagram' sub-blocks
- Render via local servers (PlantUML :8080, Mermaid :8081, draw.io :8082)
- Fall back to CLI tools (plantuml, mmdc, drawio) if servers unavailable
- Cache rendered SVGs in ~/.cache/*/diagrams/ keyed by content hash
- Show rendered image with white background (dark theme friendly)
- Collapsible source toggle (▶/▼) below rendered diagram
- Middle-click to invalidate cache and re-render
- Left-click opens fullscreen pannable/zoomable viewer
- Fullscreen: scroll-wheel zoom, drag to pan, Esc/right-click/dbl-click to close
- Both kf5 and kf6 builds
Adds a filterText property to ChatBlockModel that filters all
parsed blocks by case-insensitive substring match on content
or tool/call name. Toggle the filter bar with Ctrl+F; Escape
or the close button clears the filter and hides the bar.
Shows match count while filtering.
Incremental parsing caused phantom blocks and disappearing content.
The 150ms coalescing on the daemon side is sufficient for perf.
updateVisibleWindow() extraction kept for clean separation.
reparseIncremental() only re-parses from the last block boundary
instead of the entire chat text. Fixes UI lag and typing blockage
during long streaming responses.
Also extracted updateVisibleWindow() for reuse.
- Replace Qt.callLater scroll debounce with Timer (16ms) to coalesce
rapid contentHeight changes and eliminate scroll flickering
- Add execute button (play icon) on code blocks for scripted languages
(bash, python, perl, ruby, etc.) — launches in terminal
- Fix assistant bubble width: always use 95% of available width
- Add minimum width (280px) to code fence blocks
- Move emoji popup into prompt RowLayout so it appears near input
- Disable wheel scroll inertia (direct contentY) while keeping flick
inertia intact
- Apply all fixes to both KF5 and KF6 QML files
QML scope issues made JS-based formatting unreliable. New approach:
- FormattedContentRole in the model computes pretty-printed HTML in C++
- Call blocks: always prettyJson()
- Tool blocks: detect JSON/diff and format, otherwise escaped HTML
- QML just binds text: block.formattedContent || ""
The onVisibleChanged approach failed because block.blockContent could
not resolve through the Loader scope. Now using direct binding:
- call blocks: text bound to chatModel.prettyJson(block.blockContent)
- tool blocks: detect JSON/diff, format accordingly
- prettyJson() uses QJsonDocument for proper indentation + KSyntaxHighlighting
- Only evaluated when visible (lazy)
- Model shows last 40 blocks initially, positioned at end
- Scrolling near top triggers loadEarlier() which prepends 30 blocks
- Scroll position maintained after prepend via contentY adjustment
- All blocks still parsed (for searching later) but only tail rendered
- Reduced initial render cost significantly
- Cache parseSubBlocks per block (only reparse when content changes)
- Cache highlightedHtml results (code+language key)
- Prevents expensive KSyntaxHighlighting and regex on every render
- Also fixes scroll loop with _scrolling guard
Replace flat TextArea chat view with structured ListView:
- ChatBlockModel parses raw chat text into typed blocks (user,
assistant, reasoning, call, tool, error, info)
- Sub-blocks parsed from markdown: prose, fenced code, diffs
- Code blocks rendered with KSyntaxHighlighting (auto light/dark theme)
- Diff blocks with red/green line coloring
- Copy button on code and diff blocks
- File paths linkified and clickable (opens with xdg-open)
- Reasoning/call/tool blocks collapsible (click to expand)
- User blocks highlighted with subtle accent background
- Error blocks in red
- Scroll pinning preserved (auto-scroll when at bottom)
Levi Neely
Levi Neely