S234: From Vitto to Jarvis — Phase 1 Progress Report

Audience: Non-developers, stakeholders, product team Date: 2026-03-18 Session: S234.2 Status: Phase 1 Complete

What Is This Project?

Vitto is our AI assistant. Today, when Vitto receives a complex task, it handles everything one step at a time — like a single employee working alone on a big project. The "Jarvis Evolution" makes Vitto smarter by giving it the ability to deploy multiple sub-agents simultaneously to work on different parts of a task in parallel, like a manager delegating work to a team.

This is Phase 1 of an 8-week plan that will ultimately give Vitto:

• Parallel sub-agent execution (Phase 1 — done)
• Virtual machine environments for running code (Phase 2)
• The ability to use your Google, GitHub, Slack accounts on your behalf via OAuth (Phase 3)
• Expert sub-agents you can train for specific tasks (Phase 4)
• A "soul" — personality that adapts to how you work (Phase 5)
• Microsoft Teams integration (Phase 4)

What Was Built in Phase 1

The Sub-Vitto Engine

Before: Vitto could only do one thing at a time. If it needed to search for security patterns AND analyze code structure AND look up past decisions, it did each sequentially — slow.

After: Vitto can now deploy up to 5 "sub-Vittos" that work simultaneously. Each sub-Vitto is a focused specialist that:

• Gets a specific task ("search for security patterns in auth module")
• Works independently with its own tools
• Reports back results
• Doesn't interfere with other sub-Vittos running in parallel

Think of it like a meeting where the manager says: "Alice, research X. Bob, analyze Y. Carol, look up Z. Report back in 5 minutes." — instead of the manager doing all three tasks themselves.

What Users Will See

In the chat panel, when Vitto deploys sub-agents, users see a collapsible card showing:

• How many agents were deployed
• What each one is doing
• Real-time status (running, completed, failed)
• How long each took
• A synthesis of all results

┌─────────────────────────────────────┐
│ Deployed 3 agents            2.3s   │
│                                     │
│ ✓ Security Review (in-process)      │
│ ✓ Pattern Search (in-process)       │
│ ✗ Code Analysis (failed)            │
│                                     │
│ Synthesis: 2 tasks completed...     │
└─────────────────────────────────────┘

Safety Built In

• No infinite loops: Sub-Vittos cannot deploy their own sub-Vittos (recursion blocked)
• Budget controls: Each sub-Vitto has its own token/iteration limits
• Concurrency cap: Maximum 5 parallel sub-Vittos (configurable)
• Graceful failures: If one sub-Vitto fails, others continue working

Timeline of Changes

Step 1: Event Scoping System (SubVittoSSEProxy)

What: Created a system to keep sub-Vitto activity separate from the main Vitto's activity stream. Why: Without this, when 3 sub-Vittos run simultaneously, their status updates would mix with the main Vitto's updates — creating confusion in the user interface. This proxy labels each event so the frontend knows which sub-Vitto sent it.

Step 2: Type Definitions

What: Defined the data structures for sub-Vitto tasks, tiers (in-process, VM, browser), and results. Why: These are the "contracts" that ensure all parts of the system speak the same language. Every component references these shared definitions.

Step 3: Orchestrator Service (Core Engine)

What: Built the central service that manages parallel execution — receives a list of tasks, assigns them to sub-Vittos, monitors progress, handles failures, and synthesizes results. Why: This is the brain of the operation. It decides how many sub-Vittos to run simultaneously, ensures budget limits are respected, and produces a unified summary when all sub-tasks complete.

Step 4: Deploy Agents Tool

What: Created the deploy_agents tool that Vitto's AI can call when it decides a task would benefit from parallel execution. Why: Vitto needs a way to actually trigger parallel execution. This tool follows the same plugin pattern as all other Vitto tools (memory recall, web search, etc.) — it's registered in the tool system and available during conversations.

Step 5: System Wiring

What: Connected all the new pieces together — registered the orchestrator service, plugged the tool into the registry, and made the deployment pipeline aware of the new capabilities. Why: The individual components work in isolation, but they need to be wired into the existing system to function as a whole.

Step 6: Engine Refactoring (Most Critical)

What: Replaced the old, duplicated execution loop in the sub-agent system with a call to the shared engine that powers all of Vitto. Why: Previously, sub-agents had their own 160-line execution loop that missed features like automatic stuck detection, crash recovery, and real-time streaming. By using the shared engine, sub-agents now get all of these for free — reducing code, eliminating bugs, and ensuring consistency. This is the "one engine, many contexts" architecture principle.

Step 7: Chat Panel Renderer

What: Built the visual component that shows sub-Vitto activity in the chat interface. Why: Users need to see what's happening when Vitto deploys parallel agents. This card group shows real-time status with completion indicators, timing, and error details.

Step 8: Integration Verification

What: Ran comprehensive tests to verify all pieces work together. Why: With 22 files changed across backend and frontend, we need confidence that nothing is broken.

Numbers

What's Next

Phase 2 Preview

• VM Agents: Sub-Vittos that can actually write code, run tests, and install packages in sandboxed virtual machines (not just search/analyze)
• MongoDB Cleanup: Removing 2 orphaned database collections and their ghost references

Known Issues (Pre-Existing)

These test failures existed BEFORE Phase 1 work began and are unrelated to the sub-Vitto engine:

1. phase-executor.spec.ts (20 failures) — Missing UserDossierService provider in test setup
2. vitto-core.controller.spec.ts (10 failures) — Same root cause: missing UserDossierService
3. daily-digest-rich.spec.ts (suite failure) — Constructor expects 6 arguments, test provides 5 (missing VittoObservabilityService)

A separate handoff has been created for another agent to fix these.

Key Decisions Made

Owner: fguijarro Session: S234.2 Branch: session-234-235 Spec: docs/superpowers/specs/2026-03-18-vitto-jarvis-comprehensive-evolution-design.md