Backend improvements:
- Add cache auto-clear on settings save (class-settings-v2.php)
- Hooks updated_option action
- Clears connection test transients when local backend settings change
- Add reasoning_content streaming support (class-local-backend-provider.php)
- Handles thinking models like Claude extended thinking
- Captures chunk['choices'][0]['delta']['reasoning_content']
Documentation:
- Add FRONTEND_AND_CHAT_FIX_SUMMARY.md with all fixes
- Add FRONTEND-REFACTOR-PHASE2.md with modularization plan
Note: Splitting effort deferred - will continue iterating on monolith
Next: Fix session history not appearing bug
Phase 1: Core Client
- New class-memanto-client.php: Singleton PHP client for MEMANTO API v2
- Health check with 5-min transient caching
- Agent CRUD (ensure, activate, deactivate sessions)
- Memory operations (remember, batch_remember, recall, recall_recent)
- Auto re-activation on expired session tokens (401 retry)
Phase 2: Write-Through Memory Hooks
- New class-memanto-context-enhancer.php: Orchestrates remember/recall
- Fires on: user message, plan generated/approved/rejected,
section written, block refined, config saved, session start/end
- All hooks via do_action() — zero coupling to MEMANTO when disabled
Phase 3: Context Enrichment
- Context builder injects recalled memories into AI prompts
via build_memanto_context() in build_working_context()
- 3-recall strategy: recent post memories, semantic search, user preferences
- Deduplication by content hash
Phase 4: Cross-Session Restore
- New REST endpoints: /memanto/restore, /memanto/preferences
- restore_session() recalls 15 recent memories + user preferences on editor load
- build_session_restore_message() creates AI-ready system message
- get_user_preferences_for_new_post() extracts tone/audience/length/language
- Frontend: 🧠 Restored badge in status bar with memory count tooltip
- Preference carry-over: auto-fills post config from stored user preferences
- deactivate_session() called on session end (triggers MEMANTO summary)
- Badge clears on new conversation start
Settings UI:
- MEMANTO Context Keeper section with enable toggle, URL, API key, test connection
- Settings registered via class-settings-v2.php + tab-memanto.php view
Graceful degradation: all MEMANTO calls guarded by is_active(),
frontend catches silently, plugin works identically when disabled.
Root causes of writing getting stuck:
1. Model returns empty response for a section → now detected early with
actionable error message including model name
2. Model responds but without ~~~ARTICLE~~~ divider (happens with fallback
models like Gemini) → now treats entire response as markdown content
3. Stream ends without 'complete' event (error/exit in PHP) → JS timeline
entries lingered as 'active' forever. Now deactivated on stream close.
4. Error messages in execution flow now use structured formatAiErrorMessage
with retry button instead of raw text
Also: deactivateActiveTimelineEntries called in catch block so errors
properly clear the 'Writing section X' status indicator.
Bug 1 - Session opens with empty messages:
- loadChatHistory effect was re-running on every currentSessionId change,
racing with openSessionById and overwriting loaded messages
- Removed currentSessionId from effect dependencies (only runs on mount/postId)
- Added recovery: if session has 0 messages but has post_id, try fetching
from the post-based conversation endpoint as fallback
Bug 2 - Start New Session shows old messages:
- startNewConversation now sets isHydratingSessionRef=true before changing
session state, preventing the persistence effect from saving stale data
- Fully resets: messages, plan, agentMode, keyword suggestions, providerInfo
- loadPostSessions called AFTER state reset to avoid stale renders
Also fixed:
- Legacy fallback now only fires when no session was resolved at all
(prevents loading old post_meta data over session data)
Session issues fixed:
- Removed auto-draft-only gate for showing unassigned sessions on new posts
(now shows on any post that has no linked sessions)
- Auto-link unassigned session to current post when user opens it
- Added beforeunload/pagehide flush to persist messages before page close
(prevents data loss from 700ms debounce not firing)
- Added warning log when session loads with 0 messages for debugging
UI fix:
- Override WP editor h3 shrinkage (11px/uppercase → 15px/normal/white)
- Fix h2/h4-h6 headings in response content for dark theme readability
Outline error messages:
- Separate empty response from parse failure with distinct actionable messages
- Show model name + token count on empty response for debugging
- Reassure user that parse failures are usually one-time issues
Phase 1 - UI Theme Consistency:
- Chat messages now use consistent dark theme (removed jarring white bg)
- Plan cards restyled with rounded borders, fills, colored status badges
- Timeline entries use humanist sans-serif instead of monospace
- Error messages now structured (icon + title + detail + action link)
- Input area unified with dark theme cohesion
Phase 2 - UX Flow:
- Added contextual placeholder text per agent mode in textarea
- Added visual mode indicator badge (Chat/Planning/Writing)
- Simplified welcome screen (single 'Continue' + collapsible history)
- Added slash command/mention discovery hint in empty input
- Added write confirmation when editor has existing content
- Added 30s streaming heartbeat (reassurance when model is slow)
Phase 3 - Error Handling:
- Added DB table health check on sidebar init
- Improved 'no API key' error with settings link
- Shows in-chat warning when provider fallback triggers
- Auto-fallback to registry fallback model on unavailability
- isLoading always resets via try/finally pattern
- Implement local backend AI provider with Ollama integration
- Add Brave Search API integration for real-time search suggestions
- Add image generation manager with multiple AI providers
- Create hybrid provider system with local/cloud fallback
- Add comprehensive settings UI with provider management
- Implement Gutenberg sidebar with writing assistance controls
- Add SEO schema generation for AI-generated content
- Multiple provider support: OpenRouter, local backend, Codex
