Adds a "Response Completeness" section to the Maestro system prompt
instructing agents to make each response self-contained with clear
summaries, relevant context, and key details. This ensures users can
understand responses even when only seeing the most recent message.
Updated all agent prompts to explicitly state that only the LAST
message is captured by the system. This ensures agents produce
complete, self-contained responses with all relevant information.
Prompts updated:
- context-grooming.md, context-summarize.md, context-transfer.md
- autorun-default.md, autorun-synopsis.md
- group-chat-participant.md, group-chat-participant-request.md
- group-chat-moderator-system.md, group-chat-moderator-synthesis.md
- maestro-system-prompt.md (new "Response Completeness" section)
Enable saving markdown content directly to remote filesystems
when connected via SSH. The fs:writeFile IPC handler now accepts
an optional sshRemoteId parameter and uses writeFileRemote for
SSH transport.
The native folder dialog can only browse local filesystems, so hide the
browse button when the session is running over SSH (where the default
path would be a remote path). Users can still manually type a path.
Changed the git info popup's remote URL link to use
window.maestro.shell.openExternal() instead of window.open()
so it opens in the user's default system browser.
- Add await to shell.openExternal call to ensure proper async handling
- Show toast notification when no remote URL is found for the repository
- Add error handling with toast feedback for any failures
- Update tests to verify toast notifications are shown on error
Fixes issue where the command didn't provide feedback when failing.
Custom notification commands (e.g., TTS, fabric pipelines) should only
run for regular AI conversation flow toasts, not for synopsis messages.
Added skipCustomNotification flag to Toast interface and set it to true
for synopsis toasts. This prevents the notification command from being
executed when a synopsis completes, while still showing OS notifications
and toast UI.
Document nodes with long unbroken text (like separator lines with ====)
were bleeding outside their boundaries. Fixed by:
- Setting fixed width (280px) on container instead of just maxWidth
- Using break-all word breaking for description text
- Adding webkit line clamping to limit description to 3 lines
- Using text-overflow: ellipsis with overflow: hidden
- Shows "Drop here to ungroup" zone when dragging a session and no ungrouped sessions exist
- Allows users to drag sessions out of groups even when Ungrouped Agents folder is hidden
- Added tests for drop zone visibility and drop handler
Previously, the Test button would show "Stop" after the notification
command started and remain visible even after the process exited,
requiring an 8-second timeout fallback to reset.
Now the component listens for the onCommandCompleted event and
immediately resets to "Test" (via Success state) when the process
exits naturally. The Stop button is only shown while the process
is actively running.
Remove backward compatibility shim and fully migrate from legacy ttsId
property to notificationId across all notification-related code:
- Update NotificationsPanel.tsx to use result.notificationId
- Remove deprecated ttsId from NotificationCommandResponse interface
- Remove legacy TTS alias exports from handler
- Update all test files and mocks to use notificationId
Changed elapsed time calculation for USER history entries to use the
actual task duration (from thinkingStartTime to completion) instead of
time since the last synopsis. This accurately reflects how long the
agent was actively working on the task being synopsized.
Add elapsed time calculation when creating USER history entries via
the saveToHistory/synopsis feature. Time is calculated from lastSynopsisTime
(if a previous synopsis exists) or tabCreatedAt (for first synopsis).
This complements the existing elapsed time tracking for AUTO entries.
- Rename "Speakable Notifications" to "Custom Notifications" in features
- Update Notifications tab description to remove TTS-specific language
- Revise Custom Notification section with broader command examples
- Remove default TTS bias from configuration guidance
The previous fix added document-level dragover/drop handlers to call
preventDefault(), but handleImageDragOver was calling stopPropagation()
which prevented the document-level handlers from firing during the bubble
phase. Switching to capture phase ensures preventDefault() is called at
the document level before element handlers can stop propagation.
Update notification tests to clearly document and verify that:
- Custom notification commands have NO whitelist and NO validation
- ANY executable path, binary name, or arguments are allowed
- Shell pipelines with redirects and pipes are fully supported
Also rename TTS terminology to generic notification terminology:
- resetTtsState → resetNotificationState
- getTtsQueueLength → getNotificationQueueLength
- getActiveTtsCount → getActiveNotificationCount
- clearTtsQueue → clearNotificationQueue
- getTtsMaxQueueSize → getNotificationMaxQueueSize
- Add onCommandCompleted alongside deprecated onTtsCompleted
This ensures the feature remains open-ended for user flexibility.
Without preventing default on both dragover and drop at the document level,
Electron/Chromium can reject subsequent drag-and-drop operations after the
first drop. This adds document-level handlers that call preventDefault() on
both events, ensuring the drop zone remains valid across multiple operations.
- Remove TTS command whitelist validation - users can now configure any
command for custom notifications (e.g., fabric piped to 11s)
- Enable shell mode for spawn to support pipes and command chains
- Add visual feedback to Test button (running/success/error states)
- Remove speech bubble from AI response messages in TerminalOutput
- Rename internal TTS terminology to more generic "notification command"
- Update tests to reflect new behavior
Fixes overlay not covering sidebar and file panel by rendering the
modal via React portal to document.body, ensuring proper z-index
stacking across all UI layers.
Stats DB initialization failures were only logged locally, providing no
visibility into production issues. This change adds Sentry reporting for:
- Database corruption detection (integrity check failures)
- Native module loading failures (architecture mismatches)
- Database creation failures
- General initialization failures
Created reusable Sentry utilities in src/main/utils/sentry.ts that
lazily load @sentry/electron to avoid module initialization issues.
Add comprehensive crash detection for renderer process failures that
Sentry in the renderer cannot capture (because the process is dead,
broken, or failed before Sentry initialized):
- render-process-gone: captures crash, kill, OOM, launch-failed
- unresponsive/responsive: tracks frozen window states
- crashed: handles page-level crashes
- did-fail-load: captures page load failures (network, invalid URLs)
- preload-error: captures preload script failures before app loads
- console-message: forwards critical renderer errors (TypeError, etc.)
Reports to Sentry from main process with detailed context.
Auto-reloads renderer after non-intentional crashes.
When users select "All" for Max Output Lines, Infinity is stored but
JSON.stringify(Infinity) produces null. On app restart, this null value
was being loaded directly, causing:
- null * 1.5 = 0 for maxHeight calculation
- All AI responses appearing fully collapsed (only "Show all X lines")
Now we detect null values for maxOutputLines and restore them as
Infinity, matching the user's original "All" selection.
Root cause: JSON cannot represent Infinity, so it becomes null during
serialization to electron-store.
Fixes the regression where maxOutputLines setting was lost on restart.
Add a save icon to the action buttons at the bottom of AI responses
that opens a modal allowing users to save the markdown content to a
file. The modal includes:
- Folder path input with a browse button for filesystem navigation
- Filename input with automatic .md extension handling
- Error handling and validation
- Remove "Pencil" and "Image" labels to show icon-only for those controls
- Update Thinking description to "toggles display of agent thinking"
- Make label prop optional in TourIcon component
- Convert tourSteps.ts to tourSteps.tsx to support JSX content
- Add descriptionContent/descriptionContentGeneric fields to TourStepConfig
for rendering JSX after the text description
- Create TourIcon component that renders actual Lucide icons inline
- Update input area step to show PenLine, ImageIcon, History, Eye, Brain,
and Keyboard icons matching the actual UI controls
- Widen tooltip to 480px when step has extra JSX content
- Restore original description text, move icon explanations to JSX content
When loading remote content (e.g., README.md from SSH), the loading
indicator now takes over the entire main panel instead of showing
alongside the terminal output. This prevents the awkward half-filled
appearance.
Apply the same fix from the marketplace handlers to Symphony:
- getRegistry: Falls back to expired cache when GitHub fetch fails
- getIssues: Falls back to expired cache when GitHub fetch fails
This prevents showing empty data when the network is unavailable
but valid (though expired) cache data exists.
When the Playbook Exchange modal opens and the manifest cache is expired,
we attempt to fetch fresh data from GitHub. Previously, if this fetch
failed (network error, timeout, etc.), we would return an empty manifest
showing "No playbooks available" - a confusing UX.
Now we fall back to the expired cache data if available, which is better
than showing nothing. This ensures users can still browse playbooks even
when GitHub is temporarily unreachable.
Changes:
- getManifest: Fall back to expired cache when fresh fetch fails
- refreshManifest: Fall back to existing cache when refresh fails
- Added tests for new fallback behavior
Add hints across implementation-focused prompts to encourage searching
for existing code patterns before creating new implementations. This
addresses the issue of duplicate code being generated by guiding agents
to reuse and extend existing utilities, helpers, and patterns.
Updated prompts:
- autorun-default.md: Task implementation now searches for existing code first
- maestro-system-prompt.md: New "Code Reuse and Refactoring" section
- wizard-document-generation.md: Tasks should be reuse-aware, phases include guidance
- speckit.tasks.md: New "Code Reuse Principle" section before task generation
- speckit.implement.md: Task guidelines include reuse-awareness
- openspec.implement.md: Task guidelines include reuse-awareness
Update input area tour step to describe each icon with visual hints that
match the actual Lucide icons used in the input area. Also update
TourStep component to properly render newlines as line breaks so the
bulleted list of icons displays correctly.
Icons now described:
- Pencil (✎) for expanded prompt editor
- Image (🖼) for file attachments
- History (⏱) for session history toggle
- Eye (👁) for read-only mode
- Brain (🧠) for extended thinking
- Keyboard (⌨) for submit hotkey toggle
Add comprehensive description of input area features during introductory
tour including expanded prompt editor (pencil icon), image attachments,
and toggle buttons for History, Read-Only, and Thinking modes. Also
mention the submission hotkey option and link to Settings for defaults.
Use flexbox layout with flex-shrink-0 on header and sash elements to
prevent them from being clipped. The scrollable entries section now
uses flex-1 min-h-0 to take remaining space while allowing proper
overflow scrolling.
When the computer goes to sleep and wakes up, some settings like
MAX OUTPUT LINES PER RESPONSE were resetting to default values.
This adds Electron's powerMonitor to detect system resume events and
notify the renderer to reload settings from persistent storage:
- Add powerMonitor.on('resume') listener in main process
- Add onSystemResume IPC handler in preload API
- Refactor loadSettings to useCallback for reuse
- Register system resume listener in useSettings hook
Closes#269
Claude Code reliably emits tool execution events in stream-json format,
but Codex and OpenCode may not emit parseable tool events or may answer
from context without using tools.
Updated the 'should emit tool execution events' and 'should track tool
execution state transitions' tests to only assert tool event requirements
for Claude Code, while keeping informational logging for other providers.
The tests still verify the core functionality (correct response) for all
providers, but tool event parsing is now provider-specific.
Fixes integration test failure for Codex tool execution events.
Some git clients produce malformed URLs that mix HTTPS and SSH formats
(e.g., `https://git@github.com:user/repo`). These URLs would cause
window.open() to fail with an invalid URL error.
Added handling to convert these hybrid URLs to proper HTTPS format.
Fixes MAESTRO-43
Move React DevTools profiling instructions from CLAUDE-PERFORMANCE.md
to CONTRIBUTING.md under the Profiling section:
- Full installation and launch commands
- Components and Profiler tab descriptions
- Step-by-step profiler workflow
- Chrome DevTools Performance tab guidance
CLAUDE-PERFORMANCE.md now references CONTRIBUTING.md for profiling
workflow (keeping it focused on code patterns for AI).
Claude ID: 286ae250-379b-4b74-a24e-b23e907dba0b
Maestro ID: b9bc0d08-5be2-4fdf-93cd-5618a8d53b35
Document the standalone React DevTools app for render profiling:
- Installation and launch commands
- Auto-connection via script in src/renderer/index.html
- Components and Profiler tab descriptions
- Basic profiler workflow for identifying render issues
Claude ID: 286ae250-379b-4b74-a24e-b23e907dba0b
Maestro ID: b9bc0d08-5be2-4fdf-93cd-5618a8d53b35
