Implements AI-powered multi-intent classification for Cortex V2.
When Reflex Arc (Layer 1) detects complexity, it escalates to
the Orchestrator for intelligent classification using Claude 3.5 Haiku.
E2.S1 - System prompt for Orchestrator
- ORCHESTRATOR_SYSTEM_PROMPT constant with Dutch instructions
- Multi-intent detection (en, daarna, ook, eerst, vervolgens)
- Pronoun resolution rules (hij/zij → active patient)
- JSON-only output format specification
E2.S2 - Context formatting for AI
- formatContextForPrompt() with emoji structure
- Dutch labels for shifts and views
- Graceful null handling, token limiting (max 5 agenda, 3 recent)
E2.S3 - AI classification endpoint
- classifyWithOrchestrator() calling Claude 3.5 Haiku
- Builds IntentChain from AI response
- Tracks tokens used and processing time
E2.S4 - IntentChain parsing with Zod
- parseAIResponse() with Zod schema validation
- Strips markdown code blocks
- Partial extraction for malformed responses
- Graceful fallback to intent: 'unknown'
E2.S5 - POST /api/cortex/classify hybrid endpoint
- Reflex first, escalate to Orchestrator if needed
- Response includes handledBy: 'reflex' | 'orchestrator'
- Auth check, feature flag check, request validation
E2.S6 - Graceful fallback
- buildFallbackChain() with confidence cap (0.6)
- buildChainFromReflex() for simple chains
- classifyWithTimeout() wrapper (5s default)
New files:
- lib/cortex/orchestrator.ts (~780 lines)
- app/api/cortex/classify/route.ts (~180 lines)
Key deliverable: "Zeg Jan af en maak notitie" correctly
parsed into 2 actions in an IntentChain.
Progress: 27 SP / 48 SP (56%) - E0, E1, E2 complete
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Epic 0 status: ⏳ To Do → ✅ Done
- All 5 stories (E0.S1-S5) marked as complete
- All done criteria checkboxes marked as checked
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Adjusted confidence threshold for executing simple commands from >0.9 to >=0.7 in architecture documentation and code.
- Updated API endpoint paths from `/api/intent/*` to `/api/cortex/*` for intent classification and execution.
- Modified related tests to reflect the new confidence threshold and endpoint changes.
This update enhances the system's responsiveness to user commands while maintaining clarity in API structure.
- Changed default interface preference in LoginForm component from 'klassiek' to 'cortex'.
- Updated redirect path logic to reflect the new 'cortex' preference.
- Modified button label and state management to align with the new interface preference terminology.
- Adjusted type definition for InterfacePreference to include 'cortex' instead of 'swift'.
