J.A.R.V.I.S. - Intelligent Autonomous Orchestration

JikiME-ADK's intelligent autonomous orchestration system. Proactive and adaptive development automation inspired by Iron Man's AI assistant.

Overview

J.A.R.V.I.S. (Just A Rather Very Intelligent System) is JikiME-ADK's development-dedicated intelligent orchestrator. Rather than simple command execution, it is an autonomous system that predicts, adapts, and learns, while migration tasks are handled by the partner orchestrator F.R.I.D.A.Y.

Core Philosophy

"I'm not just following orders, sir. I'm anticipating your needs."

Differences from Existing Orchestrators

Feature	Existing Orchestrators	J.A.R.V.I.S.
Exploration	3 agents in parallel	5 agents + dependency analysis
Planning	Single strategy	Multi-strategy comparison and optimal selection
Execution	Fixed sequential/parallel	Situation-adaptive dynamic switching
Error Handling	Simple retry	Self-diagnosis + alternative strategy pivot
Learning	None	In-session pattern learning
Prediction	None	Proactive suggestions for next steps

Architecture

System Structure

┌─────────────────────────────────────────────────────────────────┐
│                    J.A.R.V.I.S. System                          │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  Phase 0: Proactive Intelligence Gathering                      │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌────────┐│
│  │ Explore  │ │ Research │ │ Quality  │ │ Security │ │  Perf  ││
│  │  Agent   │ │  Agent   │ │  Agent   │ │  Agent   │ │ Agent  ││
│  └────┬─────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘ └───┬────┘│
│       └────────────┴────────────┼────────────┴───────────┘     │
│                                 ▼                               │
│                    ┌────────────────────┐                       │
│                    │ Integration Engine │                       │
│                    │  + Dependency Map  │                       │
│                    │  + Risk Assessment │                       │
│                    └─────────┬──────────┘                       │
│                              ▼                                  │
│  Phase 1: Multi-Strategy Planning                               │
│  ┌──────────────┐ ┌──────────────┐ ┌──────────────┐            │
│  │ Strategy A   │ │ Strategy B   │ │ Strategy C   │            │
│  │ Conservative │ │  Balanced    │ │  Aggressive  │            │
│  └──────┬───────┘ └──────┬───────┘ └──────┬───────┘            │
│         └────────────────┼────────────────┘                    │
│                          ▼                                      │
│                ┌──────────────────┐                             │
│                │ Trade-off Matrix │                             │
│                │ Optimal Selection│                             │
│                └────────┬─────────┘                             │
│                         ▼                                       │
│  Phase 2: Adaptive DDD Implementation (Ralph Loop)              │
│  ┌─────────────────────────────────────────────────────────┐   │
│  │  WHILE (issues_exist AND iteration < max):              │   │
│  │    ├── LSP Quality Gate (baseline capture/compare)      │   │
│  │    ├── Diagnostics (LSP + Tests + Coverage)             │   │
│  │    ├── Self-Assessment: "Is approach working?"          │   │
│  │    │   ├── YES → Continue                               │   │
│  │    │   ├── REGRESSION → Ralph alerts → Pivot            │   │
│  │    │   └── NO  → Pivot Strategy                         │   │
│  │    ├── Expert Agent Delegation                          │   │
│  │    └── Verification (zero errors required)              │   │
│  └─────────────────────────────────────────────────────────┘   │
│                         ▼                                       │
│  Phase 3: Completion & Prediction                               │
│  ┌──────────────┐ ┌──────────────────────┐                     │
│  │  Doc Sync    │ │ Predictive Suggest   │                     │
│  │ manager-docs │ │ "You might also..."  │                     │
│  └──────────────┘ └──────────────────────┘                     │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

File	Description
`templates/.claude/commands/jikime/jarvis.md`	J.A.R.V.I.S. slash command
`templates/.claude/commands/jikime/friday.md`	F.R.I.D.A.Y. slash command
`templates/CLAUDE.md`	Dual orchestrator registration
`templates/.jikime/config/quality.yaml`	LSP Quality Gates configuration
`docs/friday.md`	F.R.I.D.A.Y. migration orchestrator documentation
`docs/migration.md`	Migration system documentation

Dual Orchestrator Architecture

Dual Orchestration System

JikiME-ADK separates roles with two specialized orchestrators:

Orchestrator	Role	Command	Scope
J.A.R.V.I.S.	Development	`/jikime:jarvis`	New feature implementation, refactoring, bug fixes
F.R.I.D.A.Y.	Migration	`/jikime:friday`	Legacy → modern framework conversion

For migration-related detailed documentation, refer to docs/friday.md and docs/migration.md.

Routing Logic

IF migration keywords detected (migrate, convert, legacy, transform):
    → F.R.I.D.A.Y. activation
ELIF development keywords detected (implement, build, fix, refactor):
    → J.A.R.V.I.S. activation
ELSE:
    → Default: J.A.R.V.I.S.

Usage

Basic Usage

bash

# Intelligent autonomous execution
/jikime:jarvis "Add JWT authentication"

# Safe strategy (conservative approach)
/jikime:jarvis "Refactor payment module" --strategy safe

# Fast strategy (aggressive approach)
/jikime:jarvis "Fix typo in README" --strategy fast

# Enable automatic loop
/jikime:jarvis "Implement user dashboard" --loop --max 20

# Resume previous work
/jikime:jarvis resume SPEC-AUTH-001

Command Options

Option	Description	Default
`--strategy`	Execution strategy: auto, safe, fast	auto
`--loop`	Enable automatic error fix loop	config
`--max N`	Maximum iteration count	50
`--branch`	Auto-create feature branch	config
`--pr`	Auto-create PR on completion	config
`--resume SPEC`	Resume previous work	-

Intelligence Features

1. Proactive Intelligence Gathering (Phase 0)

5 specialized agents perform analysis simultaneously:

Agent	Role	Output
Explore Agent	Codebase structure, architecture pattern analysis	Related file list, dependency map
Research Agent	External documentation, library best practices	Implementation patterns, API references
Quality Agent	Test coverage, code quality baseline	Quality metrics, technical debt assessment
Security Agent	Pre-scan potential security impacts	Security considerations, OWASP checklist
Performance Agent	Performance impact prediction analysis	Bottleneck risks, optimization opportunities

2. Multi-Strategy Planning (Phase 1)

For all tasks, 2-3 approach strategies are generated and compared:

Strategy Types

Strategy	Risk	Speed	Rollback	Test Coverage
Conservative	Low	Slow	Easy	100%
Balanced	Medium	Medium	Medium	85%
Aggressive	High	Fast	Difficult	70%

Automatic Strategy Selection Algorithm

IF risk_score > 70:
    SELECT Conservative (safety first)
ELIF risk_score > 40:
    SELECT Balanced (balance)
ELSE:
    SELECT Aggressive (speed first)

OVERRIDE: Can be manually specified with --strategy flag

3. Adaptive Execution (Phase 2)

LSP Quality Gates

During Phase 2 execution, LSP-based quality gates are automatically applied:

Phase	Condition	Description
plan	`require_baseline: true`	Capture LSP baseline at phase start
run	`max_errors: 0`	Requires zero errors/type errors/lint errors
sync	`require_clean_lsp: true`	Clean LSP state required before PR/Sync

Configuration location: .jikime/config/quality.yaml → constitution.lsp_quality_gates

Ralph Loop Integration

J.A.R.V.I.S.'s self-diagnosis loop integrates with LSP Quality Gates:

Ralph Loop Cycle:
  1. Code Transformation (agent task execution)
  2. LSP Diagnostic Capture (post-transformation diagnostics)
  3. Regression Check (comparison against baseline)
  4. Decision: Continue or Pivot

When LSP regression is detected, J.A.R.V.I.S. automatically considers pivoting.

Self-Diagnosis Loop

With each iteration, J.A.R.V.I.S. asks itself:

"Is the current approach showing progress?"
- Is the error count decreasing?
- Is the test pass rate improving?
- Are LSP diagnostic results improving?
"Should I switch to a different strategy?"
- Trigger: 3 consecutive iterations without improvement
- Trigger: LSP regression detected
- Action: Pivot to alternative strategy
"Is this a pattern I've seen before?"
- Check for similar error patterns within session
- Immediately apply learned solutions

Pivot Decision Tree

IF no_progress_count >= 3:
    IF current_strategy == "aggressive":
        PIVOT → "balanced"
    ELIF current_strategy == "balanced":
        PIVOT → "conservative"
    ELSE:
        REQUEST → user_intervention

4. Predictive Suggestions (Phase 3)

Predicts and suggests next steps based on completed work:

markdown

## Completed: JWT Authentication

### Predictive Suggestions

Based on this implementation, you might also want to:

1. **Add refresh token mechanism** - Extend session when JWT token expires
2. **Implement rate limiting** - Prevent brute force attacks on auth endpoints
3. **Add password reset flow** - Common feature paired with authentication
4. **Set up audit logging** - Track authentication events for security

Would you like me to start any of these?

Strategy Details

auto (Default)

J.A.R.V.I.S. analyzes task complexity and automatically selects the optimal strategy:

Task Type	Analysis Result	Selected Strategy
Simple (single domain)	Risk < 40	Direct expert delegation
Medium (2-3 domains)	Risk 40-70	Sequential workflow
Complex (4+ domains)	Risk > 70	Full parallel orchestration

safe (Conservative)

Applies maximum validation and safeguards:

More user checkpoints
Smaller incremental changes
Comprehensive testing at each step
Rollback points at every Phase

fast (Aggressive)

Fast execution for simple or urgent tasks:

Minimal checkpoints
Parallelize everything possible
Skip optional validation
Prioritize quick completion

Agent Delegation

Delegation Rules

[HARD] All implementation tasks must be delegated to expert agents:

Task Type	Assigned Agent
Backend logic	backend
Frontend components	frontend
Full-stack features (DB→API→UI)	fullstack
Test generation	test-guide
Bug fixes	debugger
Refactoring	refactorer
Security fixes	security-auditor
Performance optimization	optimizer
API design	specialist-api
GraphQL schema	specialist-graphql
Microservices architecture	specialist-microservices
Mobile apps (RN/Flutter)	specialist-mobile
Desktop apps (Electron)	specialist-electron
Real-time communication (WebSocket)	specialist-websocket
UI design system	designer-ui
Angular 15+ development	specialist-angular
JavaScript/Node.js development	specialist-javascript
PHP/Laravel/Symfony development	specialist-php
Python/FastAPI/Django development	specialist-python
Rust system programming	specialist-rust
SQL query optimization	specialist-sql
TypeScript advanced types	specialist-typescript
Vue/Nuxt development	specialist-vue

Output Format

During Execution

markdown

## J.A.R.V.I.S.: Phase 2 (Iteration 3/50)

### Strategy: Balanced (auto-selected)
### Risk Score: 45/100

### Current Status
- [x] User model created
- [x] JWT token generation
- [ ] Login endpoint ← in progress
- [ ] Token validation middleware

### Self-Assessment
- Progress: YES (2 errors → 1 error)
- Pivot needed: NO
- Confidence: 85%

### Active Issues
- ERROR: src/auth/login.ts:45 - undefined 'hashPassword'

Fixing...

On Completion

markdown

## J.A.R.V.I.S.: COMPLETE

### Summary
- SPEC: SPEC-AUTH-001
- Strategy Used: Balanced
- Files Modified: 12
- Tests: 34/34 passing
- Coverage: 92%
- Iterations: 5
- Self-Corrections: 1 (pivoted from aggressive at iteration 3)

### Changes Implemented
- JWT token generation and validation
- Login/logout endpoints
- Token refresh mechanism
- Authentication middleware
- Comprehensive test suite

### Predictive Suggestions
1. Add rate limiting to auth endpoints
2. Implement password reset flow
3. Set up audit logging

<jikime>DONE</jikime>

Workflow Integration

Dual Orchestrator Workflow

JikiME-ADK operates two orchestrators independently according to their roles:

J.A.R.V.I.S. (Development):
  Individual:  /jikime:0-project → /jikime:1-plan → /jikime:2-run → /jikime:3-sync
  Integrated:  /jikime:jarvis "task" → Full automatic execution

F.R.I.D.A.Y. (Migration):
  Individual:  /jikime:migrate-0-discover → 1-analyze → 2-plan → 3-execute → 4-verify
  Integrated:  /jikime:friday "migrate to X" → Full automatic execution

Command System

Type	Commands	Orchestrator	Purpose
Workflow (Type A)	0-project, 1-plan, 2-run, 3-sync	J.A.R.V.I.S.	Fine-grained control per development stage
Migration	migrate-0 ~ migrate-4	F.R.I.D.A.Y.	Stage-by-stage migration control
Utility (Type B)	jarvis, test, loop, fix	J.A.R.V.I.S.	Quick execution and automation
Utility (Type B)	friday	F.R.I.D.A.Y.	Migration automation

Orchestrator Comparison

Aspect	J.A.R.V.I.S. (Development)	F.R.I.D.A.Y. (Migration)
Purpose	New feature implementation, improvements	Legacy → modern framework conversion
Input	Task description, SPEC	Legacy source code
Stages	4 stages (0-project ~ 3-sync)	5 stages (0-discover ~ 4-verify)
Output	Code, documentation	Migrated project, verification report
Methodology	DDD	DDD + behavioral comparison verification
Completion Marker	`<jikime>DONE</jikime>`	`<jikime>MIGRATION_COMPLETE</jikime>`

Limitations & Safety

Limitations

Maximum 3 strategy pivots (requests user intervention after that)
No pivoting during critical operations (migration, deletion)
Only in-session learning supported (no cross-session learning)

Safeguards

[HARD] All implementations delegated to expert agents
[HARD] User confirmation required before SPEC creation
[HARD] Completion marker required: <jikime>DONE</jikime>
[HARD] LSP Quality Gate: Zero errors required in run phase
Rollback points created at each Phase
Automatic alerts when LSP Quality Gates detect regression

Parallel Execution Features

Commands that can be used in conjunction with J.A.R.V.I.S.'s 5-way parallel exploration:

Command	Description	Parallel Execution Pattern
`/jikime:perspective`	Simultaneous analysis from 4 perspectives (Arch, Sec, Perf, Test)	4-way parallel sub-agents
`/jikime:verify pre-pr`	Verification including Adversarial Review	3-way parallel verification

Multi-Perspective Analysis

The /jikime:perspective command uses a parallel analysis pattern similar to J.A.R.V.I.S.'s Intelligence Gathering:

J.A.R.V.I.S. Phase 0:        /jikime:perspective:
├── Explore Agent            ├── Architecture Agent
├── Research Agent           ├── Security Agent
├── Quality Agent      ←→    ├── Performance Agent
├── Security Agent           └── Testing Agent
└── Performance Agent

Adversarial Review

The Adversarial Review executed in Phase 8 of /jikime:verify pre-pr runs 3 sub-agents in parallel for verification:

Subagent	Role
False Positive Filter	False positive filtering
Missing Issues Finder	Missed issue detection
Context Validator	Context verification

Skill Reference

Detailed documentation on parallel execution patterns:

templates/.claude/skills/jikime-workflow-parallel/SKILL.md
templates/.claude/skills/jikime-workflow-parallel/modules/parallel-patterns.md
templates/.claude/skills/jikime-workflow-parallel/modules/synthesis-strategies.md

Best Practices

When Should You Use J.A.R.V.I.S.?

Suitable cases:

New feature implementation (spanning multiple domains)
Large-scale refactoring
Complex bug fixes
When full workflow automation is needed

When individual commands are better:

Single file modifications
When you want to execute only specific stages
When fine-grained control is needed

Recommended Usage Patterns

bash

# Complex new feature
/jikime:jarvis "Implement payment processing system"

# Safety-critical refactoring
/jikime:jarvis "Refactor database layer" --strategy safe

# Simple fix
/jikime:jarvis "Add validation to login form" --strategy fast

# Resume support for long tasks
/jikime:jarvis "Complex feature" --loop --max 30
# ... after interruption ...
/jikime:jarvis resume SPEC-XXX

# Use F.R.I.D.A.Y. for migrations
/jikime:friday "Migrate Vue app to Next.js 16"

Version: 3.2.0 Last Updated: 2026-01-25 Codename: J.A.R.V.I.S. (Just A Rather Very Intelligent System) Inspiration: Iron Man's AI Assistant Changelog:

v3.2.0: Added Related Commands section (perspective, verify), linked parallel execution pattern documentation
v3.1.0: Updated LSP Quality Gates configuration path (quality.yaml), documentation improvements
v3.0.0: Dual Orchestrator (J.A.R.V.I.S. + F.R.I.D.A.Y.), LSP Quality Gates, Ralph Loop integration
v2.0.0: Added Migration Mode (--mode migrate), unified workflow orchestration
v1.0.0: Initial release with Development Mode

J.A.R.V.I.S. - Intelligent Autonomous Orchestration ​

Overview ​

Core Philosophy ​

Differences from Existing Orchestrators ​

Architecture ​

System Structure ​

Related Files ​

Dual Orchestrator Architecture ​

Dual Orchestration System ​

Routing Logic ​

Usage ​

Basic Usage ​

Command Options ​

Intelligence Features ​

1. Proactive Intelligence Gathering (Phase 0) ​

2. Multi-Strategy Planning (Phase 1) ​

Strategy Types ​

Automatic Strategy Selection Algorithm ​

3. Adaptive Execution (Phase 2) ​

LSP Quality Gates ​

Ralph Loop Integration ​

Self-Diagnosis Loop ​

Pivot Decision Tree ​

4. Predictive Suggestions (Phase 3) ​

Strategy Details ​

auto (Default) ​

safe (Conservative) ​

fast (Aggressive) ​

Agent Delegation ​

Delegation Rules ​

Output Format ​

During Execution ​

On Completion ​

Workflow Integration ​

Dual Orchestrator Workflow ​

Command System ​

Orchestrator Comparison ​

Limitations & Safety ​

Limitations ​

Safeguards ​

Related Commands & Features ​

Parallel Execution Features ​

Multi-Perspective Analysis ​

Adversarial Review ​

Skill Reference ​

Best Practices ​

When Should You Use J.A.R.V.I.S.? ​

Recommended Usage Patterns ​

J.A.R.V.I.S. - Intelligent Autonomous Orchestration

Overview

Core Philosophy

Differences from Existing Orchestrators

Architecture

System Structure

Related Files

Dual Orchestrator Architecture

Dual Orchestration System

Routing Logic

Usage

Basic Usage

Command Options

Intelligence Features

1. Proactive Intelligence Gathering (Phase 0)

2. Multi-Strategy Planning (Phase 1)

Strategy Types

Automatic Strategy Selection Algorithm

3. Adaptive Execution (Phase 2)

LSP Quality Gates

Ralph Loop Integration

Self-Diagnosis Loop

Pivot Decision Tree

4. Predictive Suggestions (Phase 3)

Strategy Details

auto (Default)

safe (Conservative)

fast (Aggressive)

Agent Delegation

Delegation Rules

Output Format

During Execution

On Completion

Workflow Integration

Dual Orchestrator Workflow

Command System

Orchestrator Comparison

Limitations & Safety

Limitations

Safeguards

Related Commands & Features

Parallel Execution Features

Multi-Perspective Analysis

Adversarial Review

Skill Reference

Best Practices

When Should You Use J.A.R.V.I.S.?

Recommended Usage Patterns