S234: Vitto → Jarvis Evolution — Session Complete

20+ commits | 1,036/1,036 tests passing | 14 lessons captured in MCP intelligence

What Was Delivered

Phase 1: Sub-Vitto Engine

• SubAgentService refactored to use AgentLoopService.run() (eliminated 160 lines of duplicated loop)
• SubVittoOrchestratorService for parallel dispatch with Promise.allSettled()
• SubVittoSSEProxy for scoped event streaming
• deploy_agents + create_agent tool plugins with chat renderers
• Recursion depth guard (MAX_DEPTH=1)

Phase 2: OAuth MVP (Google/GitHub)

• OAuthService with PKCE + AES-256-GCM token encryption
• @Public() callback endpoint, token health cron
• oauth_status + oauth_api_call tool plugins with OAuthRenderer
• Connected Services tab in channel-manager UI

Phase 3: Generation Pipeline Fix

• Root cause: SSE events sent file metadata but NOT content → VFS empty → preview blank
• Fix: Include file content in module-gen.module-generated event + forward through tool.progress wrapper

Phase 4: Conversation Memory (CRITICAL)

• Root cause: Every user message started a fresh agent loop with zero history
• Fix: Load last 20 messages from AgentSessionService, inject into agent loop
• Also fixed: memory recall 403 (internal token resolved to wrong role)

Phase 5: Intelligence Improvements

• Context-aware tool filtering (20+ → 5-8 tools per request via INTENT_TOOL_MAP)
• Error recovery with per-tool hints + tool blocking after 3 failures
• Multi-step planning in system prompt for complex tasks
• Self-reflection phase before expensive tools
• Anti-hallucination guardrails (4 rules)

Phase 6: Memory Architecture Redesign

• Token budget priorities inverted: user identity now priority 9 (SACRED)
• Session boundary detection (NEW SESSION vs FIRST INTERACTION)
• Auto-extracted facts → UserDossier sync (learnFact + updateExpertise)
• Mem0-style memory consolidation (ADD/UPDATE/NOOP with Jaccard similarity)

Phase 7: WhatsApp Fixes

• Quick reply deduplication (contextual body instead of extractLastSentence)
• Tool-specific button suggestions
• Mood inference throttling (80+ calls → ~4 per 20-min task)
• Module persistence retry (3x with backoff)

Phase 8: Infrastructure

• SelfHostedProvider sentinel mode (stops ERROR spam after 3 failures)

MCP Intelligence Captured

What's Left for Next Session

P1: Memory Improvements

1. Graph memory enrichment — KnowledgeGraphService exists but is underutilized. Wire entity relationships from dossier facts into graph queries for relational reasoning ("Who works with whom on what?")
2. Embedding-based consolidation — current Jaccard similarity misses semantic duplicates ("likes Angular" vs "prefers Angular framework"). Use VittoMemoryEmbeddingService for the consolidation comparison step
3. Memory importance decay tuning — importance decay exists but hasn't been calibrated against real usage patterns

P1: Proactive Intelligence

4. Anticipatory pattern detection — learn from user behavior to auto-create proactive checks. Needs usage analytics pipeline to detect recurring patterns (e.g., "user checks builds every morning at 9am")
5. Task observer — auto-escalate stalled long-running tasks ("Your deploy has been running for 30 min vs avg 2 min — should I stop it?")

P1: Personality Evolution

6. Dynamic tone adaptation — EvoEmo-style personality evolution during conversations based on engagement signals
7. Communication style learning — track whether user prefers verbose/terse, code-first/explanation-first, and adapt

P2: Advanced Patterns

8. Online correction replay — when user says "no, that's wrong", inject the correction into the current conversation context immediately (not just store for next session)
9. Cross-session context bridging — "We were working on X yesterday" should recall that specific session's context, not just a summary

Files Modified (Key)

Backend (backend-vitto-api)

• engine/agent-loop.service.ts — conversation history injection, error recovery, tool failure tracking
• engine/vitto-engine.service.ts — conversation history loading, intent classification, context-aware tools
• engine/channel-processing.service.ts — channel history, quick reply fix
• engine/context-assembly/web-prompt-builder.service.ts — token budget priorities, session awareness, planning, self-reflection, anti-hallucination
• engine/memory-persistence.service.ts — facts → dossier sync
• engine/channel-response-enricher.service.ts — contextual quick replies
• mind/memory/vitto-memory-store.service.ts — memory consolidation
• tools/tool-registry.service.ts — TOOL_CATEGORIES + INTENT_TOOL_MAP
• tools/plugins/module-gen.plugin.ts — SSE file content forwarding
• tools/plugins/orchestration.plugin.ts — create_agent tool
• module-generation/phases/execute.phase.ts — SSE file content inclusion
• connectivity/backend-ai-client.service.ts — module persistence retry

Backend (backend-ai-api)

• chat/chat.controller.ts — memory recall 403 fix
• infrastructure/services/sandbox/self-hosted.provider.ts — sentinel mode

Frontend

• libs-ai/vitto-chat/services/vitto-emotion.service.ts — mood throttling
• libs-ai/vitto-hub/components/sub-agent-manager/ — marketplace tab

How to Verify

# All tests pass
cd APIs/backend-vitto-api && npx jest --no-cache --forceExit
# Expected: 105 suites, 1036 tests, 0 failures

# Restart backends to activate
yarn serve:backend-vitto  # port 3003
yarn serve:backend-ai     # port 3001

Testing Guide

See: docs/english/sessions/S234/S234-vitto-jarvis-testing-guide.md