The 60s stale threshold was triggering during normal use (reading, thinking)
causing constant reconnects and presence join/leave toast spam. Now force-
reconnect only happens when the tab was hidden for 3+ minutes and becomes
visible again. Regular activity (mouse/keyboard) only resets the inactivity
timer without ever forcing a reconnect.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
On initial page load, the Supabase browser client's auth session may not
be ready yet (still loading from cookies). The Realtime channel subscription
silently fails without a valid access token. Now connect() awaits getSession()
first, which ensures the token is available and also refreshes expired tokens
on reconnect after inactivity.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The WebSocket can go stale during inactivity without triggering any error
status (channel still shows 'joined' but socket is dead). Now when the user
returns after 60+ seconds of inactivity, the connection is torn down and
re-established to guarantee a fresh WebSocket. This ensures CRDT broadcasts
actually reach other clients after returning from idle.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
CRDT broadcasts are fire-and-forget, so any updates missed during a
disconnection were permanently lost. Now when a client connects (or
reconnects), it broadcasts its full Yjs doc state. When an existing client
sees a new user join, it also broadcasts its full state. Yjs merges
handle deduplication automatically, so this converges all clients.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Increase CRDT DB persist debounce from 2s to 30s since it's only needed for
crash recovery (CRDT handles real-time sync). Increase LocalStorage draft
save debounce from 1s to 5s.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The CRDT auto-persist was broadcasting state-refresh after each save, causing
other clients to fetch from DB and overwrite their local variables/characters
with stale values. Since CRDT already syncs nodes/edges via yjs-update
broadcasts, the state-refresh from auto-persist was redundant and destructive.
Manual save still broadcasts state-refresh as intended.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add connection timeout (15s) to handle stale initial subscribes, inactivity
pause (5min) to save resources when idle, and automatic resume on user activity
or tab focus. The heartbeat now detects unhealthy channel states and consecutive
failures to trigger reconnects. Unexpected CLOSED status also triggers reconnect
instead of staying disconnected silently.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Instead of relying on Yjs broadcast serialization (which has delivery
issues), use a lightweight state-refresh broadcast event. When any
client persists (manual save or CRDT auto-persist), it broadcasts
state-refresh. Other clients fetch the latest flowchart_data from the
database and update their local state and CRDT.
Added isSuppressed flag to CRDTManager to prevent broadcast/persist
loops during initialization and refresh operations.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When a user saves, the full Yjs document state is broadcast so all
connected clients converge, even if they missed incremental updates.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The yjs-update broadcast listener was added after the Supabase channel
was already subscribed, which meant it never received messages. Moved
the listener registration to the builder chain before .subscribe()
(matching how cursor/node-lock listeners work), and removed the broken
isRemoteUpdateRef guard that caused ReferenceErrors preventing local
changes from reaching the CRDT.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The profiles table RLS policies only allowed users to view their own
profile, causing the share feature to fail when searching for users
by email. Added a policy allowing any authenticated user to read profiles.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add isSaving as a prop so the save button properly shows disabled/spinner
state. Add a back arrow link to /dashboard for editor navigation.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
ChoiceNode was passing incompatible props (optionId, optionLabel, onSave,
onRemove, onCancel) to OptionConditionEditor which expects (condition,
onChange, onClose). This caused a build failure preventing the component
tree from rendering.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The Yjs map observers were calling notifyNodesChange/notifyEdgesChange
for local changes, creating a feedback loop that reset React Flow state
mid-drag. Remote notifications are already handled by the connectChannel
handler, so the observers only need to schedule persistence.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The merge at bcbad57 lost utility functions, type definitions, and
handlers from the developing branch. This restores: imports (useRef,
useEffect, ConditionalEdge, ContextMenu), constants (AUTOSAVE_DEBOUNCE_MS),
types (ContextMenuState, ConditionEditorState), utility functions
(fromReactFlowNodes, fromReactFlowEdges, saveDraft, loadDraft, clearDraft,
flowchartDataEquals, isValidFlowchartData, convertToRenpyFormat), context
menu handlers, proper FlowchartEditorProps (userId, userDisplayName,
isOwner), ReactFlow container height (h-screen), and fixes stale closure
dependency arrays for characters/variables.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Use SECURITY DEFINER helper functions to break circular dependencies
between projects and project_collaborators table policies.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
GHMiranda
Gustavo Henrique Santos Souza de Miranda