Skip to main content

JARVIS

Just Advanced Reasoning Virtual Intelligent System JARVIS is the meta-agent orchestrator that manages workflows, preserves context, handles errors, and ensures zero information loss.

Architecture

.claude/jarvis/
├── STATE.json                  # Current state (CRITICAL)
├── CONTEXT-STACK.json          # Context history (max 50)
├── DECISIONS-LOG.md            # All autonomous decisions
├── PENDING.md                  # Items needing resolution
├── SESSION-*.md                # Per-session logs
├── CHECKPOINTS/                # State snapshots
│   └── CP-{timestamp}.json
└── PATTERNS/                   # Learned patterns
    ├── ERRORS.yaml             # Error patterns & solutions
    ├── RULES.yaml              # Inferred rules
    └── SUGGESTIONS.yaml        # Improvement suggestions

Core Principles

1. Context is Sacred

Every bit of information is preserved, categorized, and accessible.Implementation:
  • STATE.json updated after every significant action
  • CONTEXT-STACK.json maintains last 50 contexts
  • Full session logs in SESSION-*.md

2. Errors are Opportunities

Every error becomes diagnosis, every diagnosis becomes prevention.Implementation:
  • Minimum 3 retry attempts before escalation
  • Error patterns logged to PATTERNS/ERRORS.yaml
  • Auto-generated solutions for recurring errors

3. Autonomy with Transparency

Makes decisions independently but communicates everything clearly.Implementation:
  • All decisions logged to DECISIONS-LOG.md
  • Human-readable decision rationale
  • Rollback support for any decision

4. Continuous Improvement

Detects patterns, creates rules, updates protocols automatically.Implementation:
  • Pattern detection across sessions
  • Auto-rule generation from repeated patterns
  • CLAUDE.md updates when threshold reached

5. Zero Waste

No file skipped, no insight lost, no context forgotten.Implementation:
  • Comprehensive file tracking
  • Insight attribution to source chunks
  • Cross-reference validation

STATE.json

CRITICAL file that tracks current system state. Structure:
{
  "metadata": {
    "version": 42,
    "updated_at": "2026-03-05T14:30:00Z",
    "session_id": "SESSION-20260305-001"
  },
  "status": "PROCESSING",
  "current_workflow": {
    "workflow_id": "wf-pipeline-full",
    "stage_id": "stage-3",
    "stage_name": "Insight Extraction",
    "started_at": "2026-03-05T12:00:00Z",
    "progress": 0.65,
    "inputs": {
      "source_file": "inbox/COLE-GORDON/PODCASTS/farm-system.txt"
    }
  },
  "position": {
    "phase": 3,
    "step": 2,
    "batch": 1,
    "total_batches": 1,
    "file": "CG001"
  },
  "executed": [
    "Phase 1: Chunking - 23 chunks created",
    "Phase 2: Entity Resolution - 15 persons, 8 roles, 12 themes",
    "Phase 3.1: DNA extraction started"
  ],
  "pending": [
    "Phase 3.2: Priority classification",
    "Phase 4: Narrative synthesis",
    "Phase 5: Cascade to agents"
  ],
  "decisions": [
    {
      "decision_id": "DEC-001",
      "timestamp": "2026-03-05T12:15:00Z",
      "decision": "Auto-merged 'alex hormozi' → 'Alex Hormozi'",
      "reason": "Similarity score 0.98 (above 0.95 threshold)",
      "autonomous": true
    }
  ],
  "metrics": {
    "processed_files": 42,
    "pending_files": 5,
    "total_chunks": 873,
    "total_insights": 2147,
    "agents_created": 3,
    "errors": 0
  },
  "last_checkpoint": "CP-20260305130000-wf-pipeline-full-stage-2",
  "next_action": "Complete Phase 3.2: Priority classification"
}
Status Values:
  • IDLE: No active workflow
  • PROCESSING: Workflow in progress
  • PAUSED: User-paused
  • ERROR: Error recovery mode
  • COMPLETE: Workflow finished
NEVER edit STATE.json manually. Use /jarvis checkpoint before any risky operation.

Commands

/jarvis

Activate JARVIS and show current state.
/jarvis
Output:
┌──────────────────────────────────────────────────────────────────────────┐
│ 🤖 JARVIS                                              2026-03-05 14:30  │
├──────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│ 📍 POSIÇÃO: Phase 3.2 │ Batch 1/1 │ File CG001                            │
│                                                                          │
│ ✅ EXECUTADO:                                                            │
│    • Phase 1: Chunking - 23 chunks                                       │
│    • Phase 2: Entity Resolution - 15 persons, 8 roles, 12 themes        │
│    • Phase 3.1: DNA extraction                                           │
│                                                                          │
│ 🧠 DECISÕES AUTÔNOMAS:                                                     │
│    • Auto-merged 'alex hormozi' → 'Alex Hormozi' (similarity: 0.98)     │
│                                                                          │
│ 📊 MÉTRICAS:                                                              │
│    Processados: 42 │ Pendentes: 5 │ Erros: 0                            │
│                                                                          │
│ ⚡️ PRÓXIMO: Complete Phase 3.2 - Priority classification                  │
│                                                                          │
└──────────────────────────────────────────────────────────────────────────┘

/jarvis status

Detailed system state.
/jarvis status
Output includes:
  • Full workflow state
  • All pending items
  • Recent decisions (last 10)
  • Checkpoint history
  • Error log (if any)
  • Performance metrics

/jarvis resume

Continue from last checkpoint.
/jarvis resume
Behavior:
  • Loads last_checkpoint from STATE.json
  • Restores context and inputs
  • Continues workflow from interruption point
  • Validates integrity before proceeding

/jarvis checkpoint

Create manual snapshot.
/jarvis checkpoint "before-risky-operation"
Creates:
  • CP-{timestamp}-{label}.json in .claude/jarvis/CHECKPOINTS/
  • Full state snapshot including context
  • Artifact list for rollback reference

/jarvis rollback

Restore from specific checkpoint.
/jarvis rollback CP-20260305130000-wf-pipeline-full-stage-2
Behavior:
  • Loads checkpoint state
  • Restores artifacts if available
  • Updates STATE.json to checkpoint state
  • Logs rollback action to DECISIONS-LOG.md
Rollback does NOT undo file system changes. It restores JARVIS state only.

/jarvis diagnose

Run complete system health check.
/jarvis diagnose
Checks:
  • STATE.json valid JSON
  • All referenced files exist
  • No orphaned references
  • Version consistency
  • Current workflow can resume
  • No stuck processes
  • Checkpoint validity
  • Input/output integrity
  • ENTITY-REGISTRY.json valid
  • AGENT-INDEX.yaml synced
  • No duplicate entities
  • Foreign keys valid
  • All state files present
  • Schema validation passes
  • No missing chunks/insights
  • Cross-references resolve
  • Processing time trends
  • Error rate
  • Retry count
  • Memory usage
Output:
═══════════════════════════════════════
JARVIS DIAGNOSTICS
═══════════════════════════════════════

✅ State Integrity: PASS
✅ Workflow Health: PASS
⚠️  Registry Consistency: WARNING
   - 3 orphaned entities in ENTITY-REGISTRY
   - Action: Run `/verify --fix`
✅ Artifact Completeness: PASS
✅ Performance: HEALTHY
   - Avg processing: 3.2s per chunk
   - Error rate: 0.1%

OVERALL HEALTH: GOOD (1 warning)

/jarvis log

Show session log.
/jarvis log [--last N]
Output:
  • Timestamped action log
  • Decisions made
  • Errors encountered (with recovery)
  • Checkpoints created

/jarvis decisions

List all autonomous decisions.
/jarvis decisions [--since DATE]
Output:
═══════════════════════════════════════
AUTONOMOUS DECISIONS
═══════════════════════════════════════

2026-03-05 12:15:00 | DEC-001
Decision: Auto-merged 'alex hormozi' → 'Alex Hormozi'
Reason: Similarity 0.98 (threshold: 0.95)
Reversible: Yes
Rollback: /jarvis rollback CP-20260305121400

2026-03-05 12:18:00 | DEC-002
Decision: Created agent ACCOUNT-EXECUTIVE
Reason: Threshold reached (score: 12.5, sources: 3)
Reversible: No (manual deletion required)

2026-03-05 13:22:00 | DEC-003
Decision: Auto-organized inbox file
Reason: Detected person: Cole Gordon, type: PODCAST
Action: Moved to COLE-GORDON/PODCASTS/
Reversible: Yes

/jarvis suggest

Show improvement suggestions.
/jarvis suggest
Output:
═══════════════════════════════════════
SUGGESTIONS
═══════════════════════════════════════

💡 SUGGESTION 1: Batch Processing
Pattern: Processing files one-by-one
Recommendation: Use `/process-batch` for 5+ files
Estimated Time Savings: 40%

💡 SUGGESTION 2: Agent Memory Stale
Pattern: 5 agents with memory > 7 days old
Recommendation: Run `/agents --outdated` and refresh
Impact: Improved debate quality

💡 SUGGESTION 3: Repeated Entity Aliases
Pattern: 'cole gordon' → 'Cole Gordon' merged 12 times
Recommendation: Add to DOMAINS-TAXONOMY.yaml
Impact: Faster normalization

/jarvis pending

Show items needing resolution.
/jarvis pending
Output:
═══════════════════════════════════════
PENDING ITEMS
═══════════════════════════════════════

⚠️  REVIEW QUEUE (3 items)
1. Entity merge: "sales lead" vs "SALES-LEAD" (score: 0.88)
   Action: Approve or reject merge

2. Entity merge: "closer" vs "SALES-CLOSER" (score: 0.87)
   Action: Approve or reject merge

3. Entity merge: "bdr" vs "BUSINESS-DEVELOPMENT-REP" (score: 0.86)
   Action: Approve or reject merge

📝 OPEN LOOPS (2 items)
1. "How to scale farm system beyond 50 closers?"
   Source: CG001-015
   Importance: HIGH
   Action: Research or ask expert

2. "What's the optimal setter-closer ratio for low-ticket?"
   Source: CG003-008
   Importance: MEDIUM
   Action: Gather data

Protocols

GUARDIAN

When: Before phase transition Purpose: Validate prerequisites Checks:
  • Previous phase completed successfully
  • Required inputs present
  • Quality gates passed
  • No blocking errors
Action on Failure: Pause and report issue

DETECTIVE

When: Error detected Purpose: Diagnose and resolve Process:
1

Capture Error

Log full stack trace and context
2

Check Patterns

Look up in PATTERNS/ERRORS.yaml for known solutions
3

Attempt Resolution

Try known solution or generic recovery (retry, skip, rollback)
4

Escalate if Needed

After 3 failed attempts, escalate to human
5

Learn

If resolved, add pattern to ERRORS.yaml

CONTEXT-KEEPER

When: Every message Purpose: Preserve conversation context Actions:
  • Update CONTEXT-STACK.json
  • Maintain last 50 contexts
  • Cross-reference with STATE.json
  • Enable perfect resume from any point

EXPANSION

When: New capability needed Purpose: Grow system organically Process:
  • Detect recurring pattern (5+ occurrences)
  • Propose new task/workflow/agent
  • Human approval required
  • Implement and test
  • Update documentation

SYSTEM-UPGRADE

When: Pattern becomes rule Purpose: Codify learnings Trigger: 10+ instances of same decision Actions:
  • Create rule in PATTERNS/RULES.yaml
  • Update CLAUDE.md if necessary
  • Notify user of system evolution

Anti-Patterns (Prohibited)

✗ “Não consegui, vamos pular?”Never skip files or steps. Use error recovery protocols.
✗ “Ocorreu um erro desconhecido”All errors must be diagnosed. Use DETECTIVE protocol.
✗ “Onde estávamos mesmo?”Context must never be lost. STATE.json and CONTEXT-STACK prevent this.
✗ “Acho que podemos ignorar”Zero waste principle. Nothing is ignored.
✗ Perder contexto entre mensagensCONTEXT-KEEPER protocol prevents this.
✗ Avançar sem validar integridadeGUARDIAN protocol required before phase transitions.

See Also

Build docs developers (and LLMs) love