S216: Vitto Jarvis Wiring — Close Every Open Loop

Summary

Vitto has 21+ implementation plans and massive infrastructure, but the connections between systems are loose. The same issues recur across sessions because features get designed but not wired end-to-end. This session closes every verified gap with surgical, small changes (~400 lines across 9 existing files + 1 new helper).

Goal: Transform Vitto from "impressive parts, loose connections" into a fully wired Jarvis-like super-agent where every subsystem feeds into every other subsystem.

Non-goal: No new features, no new architecture, no new services. Only wiring what already exists.

Gap Inventory (Code-Verified)

Each gap was verified by reading the exact source line. No assumptions.

Gap 1: Episodic Memory Gate Excludes Channel Messages

• File: outcome-feedback.service.ts:84-87
• Code: Line 85: if (task.type === 'channel-message') return; AND Line 87: if (!task.phases.find(p => p.phase === 'ANALYZE')) return;
• Impact: Two guards block channel messages: (1) blanket type check, (2) ANALYZE phase requirement (most channel messages skip the 4-phase pipeline). Every WhatsApp/Slack/Telegram conversation is invisible to episodic memory.
• Fix: Replace blanket skip with a quality gate: only skip if task had 0 tool calls AND output < 100 chars. Move the ANALYZE phase check to only apply to non-channel tasks. Substantive channel conversations create lightweight episodes.

Gap 2: Skill Composition is Stubbed

• File: skill-composition.service.ts:138
• Code: const output = { executed: true, skillName: nodeName, ...input };
• Impact: Composite skills (multi-step workflows) never actually execute. The skill system looks complete from the outside but does nothing.
• Fix: Wire executeComposite() to call ToolRegistryService.executeTool() for each node, passing output of one step as input to the next.

Gap 3: Skill Extraction Never Notifies Training

• File: skill-extraction.service.ts:71
• Code: Creates skill via skillManager.createSkill() but never calls trainingRegistry.notifySkillExtracted(skill)
• Impact: Training plugins (4 registered, all working) never learn about newly extracted skills. The skills closed loop is broken at the last mile.
• Fix: Add this.trainingRegistry.notifySkillExtracted(skill) after successful creation. ~1 line.

Gap 4: Duplicate Outcome Feedback

• File: vitto-engine.service.ts:440 AND mind.service.ts:129
• Code: this.aiClient.postExecutionOutcome({...}) (engine) + this.outcomeFeedback.sendOutcome(task, lessonsLearned) (mind)
• Impact: Every task completion sends TWO outcome reports to the learning pipeline. Inflates metrics, wastes bandwidth, may create duplicate learned patterns.
• Fix: Remove the direct aiClient.postExecutionOutcome() call from vitto-engine.service.ts:440. The mind.recordExecution() path at line 129 is the canonical one.

Gap 5: Channel Session History Ignores Channel Type

• File: channel-prompt-builder.service.ts:253 AND memory-persistence.service.ts:155
• Code: _channel param unused in both files. Messages stored without channel tag, history loaded without channel filter.
• Impact: WhatsApp, Slack, and Web all load the same unified session history. Vitto can't distinguish "what did we discuss on WhatsApp" from "what happened on Slack." Channel-specific context is lost.
• Fix: (1) Tag messages with channel field when persisting in memory-persistence.service.ts. (2) Filter by channel when loading in channel-prompt-builder.service.ts. Untagged legacy messages included as fallback.

Gap 6: Space Brief Thin for New Spaces

• File: space-intelligence.service.ts:191
• Code: getSpaceBriefForPrompt() returns empty string when profile?.brief is falsy
• Impact: New spaces get zero context in Vitto's system prompt. Vitto doesn't know what the space is about until updateAfterGeneration() runs (which requires a module generation).
• Fix: Add a bootstrap brief that pulls from the space's name, description, and module list when no full brief exists. Falls back gracefully.

Gap 7: Daily Digest is Plain Text

• File: daily-digest.service.ts:299 (formatDigestWhatsApp())
• Code: Returns lines.join('\n') — flat markdown-style text
• Impact: The daily digest on WhatsApp is a wall of text instead of an interactive card with action buttons. Users can't tap to drill into alerts or tasks.
• Fix: Use RichMessageComposerService to create a CardMessage with "View Alerts", "View Tasks" quick reply buttons. Falls back to current text for non-WhatsApp channels.

Gap 8: Alert Acknowledgment Handler Missing

• File: channel-processing.service.ts — no handler for "Acknowledge" button payloads
• Impact: When Vitto sends an alert card with an "Acknowledge" button, tapping it does nothing. The inbound button payload has no routing to ProactiveMonitorService.resolveAlert().
• Fix: Add a button payload handler in the channel message processing flow that routes action:acknowledge_alert:{alertId} to proactiveMonitor.resolveAlert().

Gap 9: Voice Note Delivery Stubbed

• File: channel-processing.service.ts:763
• Code: this.logger.log('TTS voice note generated — needs public URL hosting to send');
• Impact: Voice notes are generated (OpenAI TTS) but never delivered. The generated audio buffer is discarded.
• Fix: Upload to Cloudflare R2 (or S3), get public URL, send via channelGateway.sendWhatsAppMedia(). ~20 lines.

Gap 10: Whisper Voice-In (Not Started — DEFERRED)

• Status: No code exists. Requires speech-to-text integration for incoming voice messages.
• Decision: DEFER to S217. Depends on audio ingestion pipeline which doesn't exist yet.

Architecture

Principle: Minimal Wiring, Maximum Impact

Every fix targets an existing file. No new services except one small helper for voice note upload. The total diff is estimated at ~400 lines across 9 files.

Fix Map

Data Flow After Wiring

User Message (WhatsApp/Slack/Web)
  → RequestClassifier (routes to fast/balanced/powerful)
  → ContextAssembly (now includes: space brief [Gap 6], channel-filtered history [Gap 5])
  → AgentLoop (tools + LLM iterations)
  → Task Completion
    → MindService.recordExecution() [single path, Gap 4 fixed]
      → OutcomeFeedback (now includes channel messages, Gap 1 fixed)
        → Creates episodic memory
      → TrainingPluginRegistry.notifyTaskCompleted()
    → SkillExtraction (if pattern detected)
      → TrainingPluginRegistry.notifySkillExtracted() [Gap 3 fixed]
      → SkillComposition (now actually executes, Gap 2 fixed)
  → Response
    → RichMessageComposer (picks card/quick_reply/text)
    → ChannelAdapter (WhatsApp/Slack/Web)
    → Voice note delivery (if short + TTS enabled, Gap 9 fixed)
  → Proactive Alerts
    → Daily Digest (now rich cards, Gap 7 fixed)
    → Alert buttons → Acknowledge handler (Gap 8 fixed)

What This Does NOT Cover

• Gap 10: Whisper voice-in — Deferred to S217. No audio ingestion pipeline exists.
• WhatsApp Flows — Phase 3 of rich messaging. Blocked on WhatsApp Cloud API credentials.
• Cost tracking ($0.00 bug) — Separate investigation needed for CostTrackerService data source.
• New LLM model routing — Covered by S210.5 plan, not this session.

Testing Strategy

Each gap fix gets a focused unit test:

No E2E tests — these are all internal wiring fixes testable with unit mocks.

Risk Assessment

Success Criteria

After S216:

1. Channel conversations appear in episodic memory (verifiable via /vitto/introspection)
2. Extracted skills trigger training plugins (verifiable via training logs)
3. Only ONE outcome per task in learning pipeline (verifiable via ai-agents DB query)
4. WhatsApp users see interactive digest cards with buttons
5. Alert "Acknowledge" buttons work end-to-end
6. Voice notes are delivered (when TTS is enabled)
7. New spaces get a bootstrap brief in Vitto's context
8. Channel history is filtered by channel type