- Android voice assistant: RecognitionService, VoiceInteractionSession with startAssistantActivity, es-HN speech recognition
- Voice transcript sent to first alive terminal via WebSocket, opens FloatingTranscriptDebug on correct session
- PiP window: fix close button using getCurrentWebviewWindow(), add mini/restore toggle, remove alwaysOnTop
- Add webview-close and window-destroy permissions to capabilities
- Pixel art ocean scrollbar on /transcript-debug respecting scroll nav mode settings
- Widget improvements: terminal list service, input widget provider, updated layouts
Server now checks that transcript .jsonl files exist before creating
terminals, preventing dead sessions from --resume errors. Frontend
shows error banner in modal when resume fails. Fixed race condition
where init() would overwrite FAB terminal selection after page refresh
by guarding with pendingSwitchTarget flag.
- Add hooks/notify.ps1 as single hook handler for all events
- Refactor settings.local.json to use notify.ps1 instead of inline PS
- Add Notification hook, auto-detect agent from session_id/transcript
- Rename agent 'main' to 'claude' across server routes and terminal
- Add loading overlay and error state for terminal switching transitions
- Add transitionError ref to useTranscriptDebug composable
Server persists hookHistory[] per agent (cap 500, resets on SessionStart),
synced realtime via session-state-patch. Frontend computes event counts
by macro type and renders color-coded badges at the ribbon start.
Mock mode also accumulates badges during demo sequence.
- Add POST /create-terminal endpoint with MAX_TERMINALS=5 limit
- Server creates PTY, runs command, registers and broadcasts atomically
- Frontend startTerminal() calls server first, connects in reconnect mode
- Remove registerTerminalOnServer() — server handles registration
- Separate broadcast-only WS clients from PTY clients (no phantom "main" PTY)
- All broadcast functions use broadcastToAll() helper
- Fix resume existing flow to create terminal with --resume flag
- Remove auto-creation of terminal sessions from init/selectSession/switchAgent
- Clients only connect to existing alive terminals from server registry
- Remove localStorage persistence (agent/sessionId) — state derived from server
- Refine session-state types: new AgentStatus values, LastError interface
- UI improvements: AgentBadge, ChatContainer, UserInput, BashCard updates
- Simplify claude-hook routes, update session-state service
- Add SessionStateManager (server/services/session-state.ts) as source
of truth for agent status, tools, approvals, and notifications
- Integrate into terminal server with state patches broadcast via WS
- Add /add-approval and /resolve-approval endpoints so approval
lifecycle is tracked centrally and broadcast to all clients
- Add permissionMode field to AgentSessionState
- Frontend store (session-state.ts) + WS service (session-state-ws.ts)
consume snapshots and patches from terminal server (4103)
- Rewrite useGlobalApproval to derive pending approvals from
centralized state — resolving on one client now clears all others
- Migrate useTranscriptDebug: processing, hookMeta, serverRegistry
now derived from session state store; remove 5s registry polling
- hooks-approval.ts notifies terminal server on add/resolve
Remove FloatingTerminal (#1), AgentBar/FloatBubble (#2), and all related
components, composables, types, handlers, routes, and CSS. Clean up orphaned
references in ToolsDropdown, whisperSocket, and claude-hook comments.
Transcript-debug is now the sole chat/agent system.
Deleted: 15 files (~3500 lines)
Edited: 12 files (-717 lines net)
- Add ResumeTerminalButton component with floating terminal modal
(drag, resize, glass morphism, TerminalNavButtons bar)
- Add useEphemeralTerminal composable for temporary PTY sessions
that auto-run `<agent> --resume <sessionId>` and cleanup on close
- Add /kill-session POST endpoint to terminal server for ephemeral
session cleanup
- Integrate button in ChatContainer header (ID row) and status bar
- Pass selectedAgent to ChatContainer from TranscriptDebugPage
- Add PermissionRequest and Stop approval hooks to local Claude config
- Unify PermissionApproval into multi-mode card (permission, plan, question)
- Support allowAlways, deny-with-reason, and AskUserQuestion answering
- Add cross-process broadcast fallback (HTTP to sync server)
- Fix approval scripts to default to .claude/debug/ for local agent
Components are now .vue files in user-components/<folder>/ parsed at runtime.
Replaces 6 DB MCP tools with 2 (list_fs_components, load_fs_component).
Adds vue-parser, fs-components API, and file watcher for live reload.
- Broadcast PermissionRequest hook as claude-permission WS event
- Fix dedup bug where undefined requestId blocked all permission cards
- Add sendInput() to AgentTerminal for char-by-char PTY responses
- Make AskUserQuestion options clickable (sends option number to PTY)
- Add Approve/Reject buttons for ExitPlanMode plan cards
- Permission Allow/Deny now sends y/n to PTY instead of broken HTTP call
- Add :key to PromptBar to force remount on agent switch, fixing shared terminal session bug
- Raise AgentTerminal z-index above PromptBar backdrop so floating terminal is visible/clickable
- Send prompt text char-by-char (15ms delay) matching FloatingVoice pattern for Claude Code compat
- Guard xterm dispose against unloaded addons to prevent errors on agent switch
- Widen PromptBar panel from 360px to 420px to fit all ChatInput buttons
- Add transcript-engine service that parses Claude Code JSONL transcripts
with session listing, message extraction, token/stats analysis, and caching
- Add transcript REST routes (sessions list, latest, by session ID, section filtering)
- Rewrite ConversationHistory to fetch from /api/transcript/* instead of mock data
- Add session pills for switching between conversation sessions
- Add stats bar footer with model, duration, tokens, and tool count
- Add TranscriptSession/TranscriptMessage types, ChatInput, InputSettings,
PromptBar updates, TranscriptCard, and useVoiceCapture composable
Replace hardcoded PowerShell status hooks with stdin-forwarding hooks
that send full Claude Code hook data (tool_input, tool_response, prompt,
session_id, model, etc.) to /api/claude-hook endpoint.
- PowerShell hooks read stdin JSON and POST to /api/claude-hook
- Server derives status for backward-compat FAB animations
- Server extracts assistant_response from transcript on Stop events
- New /api/claude-permission endpoint with Promise-based allow/deny flow
- HookNotifications.vue: toast system showing session, prompt, tool use,
tool results, notifications, and final assistant response
- WebSocket broadcast for claude-hook and claude-permission message types
- Add StatusTree component with collapsible directory hierarchy for staged/unstaged/untracked files
- Replace flat file lists in GitPage with tree view showing file type icons and git status badges
- Add AgentBar arc dock with per-agent terminal frame and voice modal
- Update ejecutor settings with hooks for claude-status reporting
- Add ui.json configs for Main (purple) and Ejecutor (red) agents
- AgentBar: fused arc-shaped dock at bottom with dynamic glow
- Quick press opens styled terminal frame mockup
- Hold opens voice modal with Web Speech API streaming transcription
- Responsive: full-width mobile, max-width on tablet/desktop/4K
- Agents API: serve uiConfig from ui.json in agent directories
- Agents page: route, store, toolbar integration
- Auto-save rendered Vue components to DB on render_vue_component
- Soft delete (archive) instead of hard delete for components
- Tags support for component categorization
- Gallery limited to 10 most recent items per section
- Upsert with ON CONFLICT for component saves
- PUT endpoint for partial component updates
- Collapsible toolbar with animated toggle button
- Window Controls Overlay support for PWA titlebar
- Compact header mode (32px) with hidden dot toggle
- Dynamic theme-color meta sync for Windows titlebar
Replace the empty dynamic canvas placeholder with a gallery showing
saved canvases, snapshots and Vue components. Users can create new
canvases, restore snapshots, load components, and manage canvas
toolbar/archive settings from the gallery.
- Backend: soft delete (archive) instead of hard delete, status column
- Frontend: CanvasGallery component with grid, search, settings popover
- Show canvas name in global header when viewing a project canvas
- Remove ProjectsPage (replaced by gallery), clean all references
- MCP tools: project category available on canvas page, update handlers
Server is now source of truth for autoRequest. When a client enables it,
all other clients lose it. Broadcast includes autoRequest per client,
frontend syncs from server state on each torch-update.
- Named clients persisted in localStorage, editable from dropdown
- Auto-request: clients can auto-receive torch when no holder exists
- Early torch init in App.vue (fires before WebMCP, awaited later)
- Deferred torch connection in toolRegistry for race condition safety
- TorchButton rewritten as dropdown with name editor, toggle, client list
- Server broadcasts client names, supports update-name message
- MCP tool handlers display client names
Implements save/restore system that captures HTML base, injected CSS,
executed scripts, and floating Vue windows with their full definitions.
Adds 4 MCP tools, backend CRUD API, Pinia store, and script logger.
- Add torchHandlers.ts with 5 tools:
- list_torch_clients: list connected browsers with metadata
- get_torch_status: show current torch holder
- transfer_torch: transfer control to another browser
- request_torch: request MCP control
- release_torch: release MCP control
- Add 'transfer' message type to server torch-handler
- Add torch category to all pages
- Export helper functions from torch.ts
- Consolidate git and torch WebSocket servers on port 4105
- Create separate handlers for git and torch in handlers/ directory
- Combine TorchButton with connection status into single pill button
- Remove StatusBar (now redundant with TorchButton)
- Remove auto-assign torch on register/disconnect
- Remove auto-connect to MCP on page load
- Connection only happens when user explicitly requests torch
- Add torch WebSocket server on port 4106
- Remove HTTP polling, use WebSocket for instant sync
- Torch state changes broadcast immediately to all clients
- Auto-reconnect on disconnect
- Add port 4106 to kill-ports script
- Add /api/torch endpoints for torch state management
- Torch system uses HTTP polling instead of WebSocket
- Only browser with torch connects to MCP
- Other browsers disconnect and poll for torch state
- Auto-assign torch to first registered client
- Auto-reassign torch when holder disconnects
This approach requires no changes to WebMCP library.
- Add requestToken() and autoConnect() functions to request tokens directly from WebMCP server
- Add /api/webmcp-request-token endpoint to proxy token requests (for HTTPS/Traefik)
- Add webmcpHttp endpoint configuration for direct WebMCP HTTP API access
- Update App.vue to auto-connect on load with polling fallback
- Add kill-ports script to clear ports 4101, 4102, 4103, 4105 before starting
- Create dedicated git-watcher.ts with its own WebSocket server (port 4105)
- Remove git watcher code from terminal.ts (no more PTY dependency)
- Add /ws/git endpoint for Traefik routing
- GitPage now connects to dedicated git WebSocket instead of terminal
- Create useTerminalRenderer.ts with all xterm.js logic
- Support custom theme, fontSize, fontFamily options
- Add handleReplay() for proper visibility handling
- Add getBufferContent() for copying terminal content
- Refactor FloatingTerminal.vue to use composable
- Refactor TerminalPage.vue to use composable
- Server: Add request-replay message type for on-demand replay
- Server: Remove auto-replay on connect (client requests when ready)
- Fix xterm.js rendering issues with hidden containers (v-show)
- Add file watcher on .git directory in terminal server
- Broadcast git-change events to connected clients
- Frontend auto-refreshes when changes detected
- Visual indicator shows realtime connection status
Add Files tab with browsable project structure and file content viewer.
New components: ProjectTree for navigation, FileViewer for content display.
Backend endpoints: /api/git/tree and /api/git/file.
- Hold FAB to open panel and start recording immediately
- Release to stop recording and send after 1s buffer
- Orange pulsing animation when PTT active
- PTT also works on record button inside modal
- Added stopRecordingAndSend exposed method
- Add Nucleo atom logo with animated orbiting electrons
- Redesign FAB with glassmorphism effect and purple gradient
- Add connection indicator (green dot) when terminal is open
- Update FloatingTerminal header with Nucleo branding
- Add PermissionRequest hook support with red alert animation
- Document Nucleo in README with visual states table
- Create official Nucleo logo SVG in docs/
- Add multiple hook-driven animations for FAB (processing, reading, writing, subagent, sessionStart, notification)
- Create claude-status.ts route to handle status updates from Claude Code hooks
- Broadcast status via WebSocket to all connected clients
- Processing (UserPromptSubmit→Stop): orange pulsing dots
- Reading (Read/Glob/Grep): cyan eye icon with scan animation
- Writing (Edit/Write): green pencil icon with pulse
- Subagent: purple orbital ring animation
- SessionStart: green wake-up ripple effect (3s)
- Notification: yellow bounce animation (2s)
- Tool flash: quick white flash on any tool use
- Add /api/recordings endpoint with full CRUD operations
- Create voice_recordings SQLite table for metadata
- Save audio files to server/recordings/ as .webm
- Store transcription, duration, microphone name, file size
- Auto-save on each Whisper recording completion
- Make server startup async to avoid Bun's 10s timeout
- Add frontend polling to detect when server is ready
- Use PowerShell Get-NetTCPConnection for reliable port detection
- Add starting state to prevent multiple simultaneous starts
- Reduce verbose logging, keep only essential info
- Add dev-dist and nul to gitignore
- Add ffmpeg conversion from WebM/Opus to WAV (16kHz mono PCM)
- Optimize transcription parameters (VAD, temperature, beam_size)
- Add Honduras Spanish context prompt with local expressions
- Fix chunk accumulation display in voice panel
- Add 1.5s recording buffer after releasing Ctrl+Space
- Skip small audio chunks (<5KB) that cause ffmpeg errors
- Use large-v3 model for better accuracy
- Add faster-whisper Python server for GPU-accelerated transcription
- Support dual mode: Web Speech API or Whisper GPU (toggleable)
- Progressive transcription every 3 seconds while recording
- Separate terminal server process (stable during hot-reload)
- Add Ctrl+V paste and Ctrl+C copy support in FloatingTerminal
- Add MCP tools: whisper_start, whisper_stop, whisper_toggle, whisper_status
- Update package.json with separate api/terminal/frontend processes
- Slim, transparent Aero Glass design with backdrop blur
- Draggable window from titlebar
- Resizable from corner handle
- Ctrl+E toggle to open at cursor position
- Remove minimize functionality (close returns to sidebar button)
- Add PWA dev-dist files and lock files
