Feb 2026 Update – Clarifications on Setup and Features

Status note (updated 2026-07). This was built in February 2026, before Claude Code shipped native subagents, skills, and first-class plugins. Much of the orchestration and spawning here is now covered by those native features, so a project starting today would lean on them first. What still holds up is the confidence-based routing, the file-based memory / handoff patterns, and the lifecycle-hook integration. Kept as a reference architecture and a record of the multi-agent patterns, several of which the platform has since absorbed.

Here's a straightforward update addressing the Run This First script, the database, and how to use the main script.

1. Run This First Script

The Run This First script is no longer needed. It was a temporary wrapper used during early development when we had separate patch files for monitoring and troubleshooting. The current main deployment script (LegendaryTeamDeploy.sh or .ps1) now includes all those features directly. You can delete Run This First — it's redundant.

2. Database Requirement

No database is required. The optional SQLite artifact index from earlier versions has been removed from the core script because the system works perfectly with simple file-based storage:

Session memory: session-state.json

Codebase tracking: codebase-map.json

Specs/tasks: OpenSpec YAML files

This keeps the setup lightweight and dependency-free. If you have a very large project and want advanced search, you can add SQLite manually later, but it's not necessary for normal use.

3. How to Use the Script

In your project root folder:

Create the file:

Linux/macOS/WSL2: nano LegendaryTeamDeploy.sh
Windows: Create LegendaryTeamDeploy.ps1

Paste the full script from the latest version.

Run it:

Linux/macOS/WSL2:textchmod +x LegendaryTeamDeploy.sh
./LegendaryTeamDeploy.sh

Windows:textpowershell -ExecutionPolicy Bypass -File LegendaryTeamDeploy.ps1

Start Claude Code:textclaude
Run bootstrap:text/bootstrap
Reply to prompts:
"discovery complete — proceed"
"specs approved"

The team is now active.

For returning sessions: @chief resume session.

This process works the same for empty folders, existing codebases, or projects with a pre-existing .claude folder — the script preserves everything and upgrades safely.

What About the SOP? (Orchestration SOP.md)

The Orchestration SOP.md is the unbreakable rulebook for your Legendary Team — it defines the exact, mandatory sequence @chief and all agents must follow every time.
What It Does

Enforces strict hierarchy (@chief only orchestrates)

Defines the 13-step bootstrap process

Sets task execution rules (parallel teams, human approval gates)

Bans chat TODOs, requires pr-agent review, blocks code without "specs approved"

Includes emergency triggers, golden rules, and quality gates

It's project-agnostic — works for any project.

How to Use It

After running the deploy script and /bootstrap

Paste the full content of Orchestration SOP.md into the Claude session

@chief will acknowledge:

"S.O.P. LOCKED IN — ALL AGENTS NOW FOLLOW THIS EXACT FLOW FOREVER"

From then on, every interaction follows the SOP exactly.

Why It's Essential

Without it, agents fall back to default Claude behavior (drift, rebellion possible)
With it, your team is disciplined, safe, and predictable

You only paste it once per project (or when updating the SOP).

__
** Most Recent Updates
✅ @Planner - Dependency-aware task decomposition
✅ @Verifier - Quality assurance and scoring
✅ @ReflectionAgent - Self-critique and improvement
✅ /swarm-planner - Structured execution plans
✅ /parallel-task - Wave-based parallel execution
✅ /spawn-subagent - Dynamic agent spawning
✅ Enhanced hooks with reflection triggers
✅ Updated SOP with planning/iteration sections

What Is This?

The Legendary Team is a complete orchestration system for Claude Code that provides:

25 Specialized Agents - From @chief (orchestrator) to @DatabaseAgent, @SecurityAgent, etc.
Confidence-Based Routing - Auto-proceed (≥70%), queue for review (40-69%), or block (<40%)
File-Based Memory - Ledgers, handoffs, and plans that survive context clears
Token Optimization - 96.7% reduction with lite agents + self-escalation
Lifecycle Hooks - SessionStart, PreToolUse, PostToolUse, PreCompact, SessionEnd

Quick Start

See INSTALLATION.md for complete installation instructions.

# One-liner install into existing project
cd ~/your-project && \
git clone https://github.com/RegardV/LegendaryTeam_For_Claude .legendary-tmp && \
cp -r .legendary-tmp/.claude . && \
cp .legendary-tmp/CLAUDE.md .legendary-tmp/"Orchestration SOP.md" . && \
rm -rf .legendary-tmp && \
bash LegendaryTeamDeploy.sh

Then in Claude Code:

/bootstrap

How It Works

The Deploy Script (`LegendaryTeamDeploy.sh`)

The deploy script initializes your project with:

Created	Purpose
`.claude/settings.json`	Registers hooks with Claude Code
`.claude/session-state.json`	Session tracking
`.claude/codebase-map.json`	File change tracking
`thoughts/` directories	Memory system (ledgers, handoffs, plans)
`CLAUDE.md`	Entry point (if not exists)

Important: The script preserves existing files. It only creates stubs if files don't exist.

The Hooks System

Hooks are registered in .claude/settings.json and fire automatically:

Hook	When	Purpose
`SessionStart.js`	Session begins	Load ledgers and handoffs
`PreToolUse.js`	Before Edit/Write	Validation, budget checks
`PostToolUse.js`	After Edit/Write	Track changes, update map
`PreCompact.js`	Before compaction	Block compaction, require handoff
`SessionEnd.js`	Session ends	Cleanup, extract learnings

The Agents

25 specialized agents in three tiers:

Always Loaded:

@chief - Master orchestrator
@ConfidenceAgent - Routes tasks by confidence score

Loaded On-Demand:

@DatabaseAgent, @UIAgent, @TestAgent, @SecurityAgent, etc.

Token Optimization:

.claude/agents-lite/ - Minimal prompts (~60-100 words)
.claude/agents-full/ - Complete definitions (for self-escalation)

The Memory System

File-based memory that survives context clears:

thoughts/
├── ledgers/              # Current session state
│   └── CONTINUITY_CLAUDE-*.md
├── shared/
│   ├── handoffs/         # Cross-session knowledge
│   │   └── handoff-*.md
│   └── plans/            # Execution plans
│       └── plan-*.md
└── templates/            # Standard formats

File Structure

your-project/
├── .claude/
│   ├── agents/           # 25 agent definitions
│   ├── agents-lite/      # Token-optimized versions
│   ├── agents-full/      # Full definitions for escalation
│   ├── hooks/            # Lifecycle hooks (JS files)
│   ├── rules/            # Behavioral rules
│   ├── skills/           # Best practice patterns
│   ├── commands/         # Slash commands
│   ├── settings.json     # Hook registration
│   ├── session-state.json
│   └── codebase-map.json
├── thoughts/             # Memory system
├── OpenSpec/             # Specifications (if using)
├── CLAUDE.md             # Entry point
└── Orchestration SOP.md  # Full documentation

Key Commands

Command	Description
`/bootstrap`	Initialize team, run on session start
`/review-queue`	Show tasks waiting for human review
`/team-status`	Show active agents and progress
`/approve-task [id]`	Approve a queued task
`/swarm-planner [task]`	Generate dependency-aware plan
`/emergency-stop`	Halt all operations

Confidence-Based Routing

Tasks are scored 0-100 and routed automatically:

Score	Tier	Action
≥70%	Auto-proceed	Executes immediately, no approval needed
40-69%	Queue	Added to review queue, continues other work
<40%	Block	Requires human approval before proceeding

Scoring factors:

+40: Similar task succeeded before
+30: Clear spec/requirements exist
+20: Known patterns available
-20: Security implications
-30: Conflicting requirements
-40: Destructive operations

The 25 Agents

Orchestration

@chief - Master orchestrator, spawns teams, manages queue
@ConfidenceAgent - Scores tasks, routes to tiers
@SessionOrchestrator - Memory and continuity management

Planning & Quality

@Planner - Task decomposition with dependencies
@Verifier - Quality assurance, plan validation
@ReflectionAgent - Self-critique, continuous improvement

Task Execution

@DatabaseAgent - Schemas, migrations, CRUD
@UIAgent - Components, styling, responsive design
@TestAgent - Unit/integration tests (≥80% coverage)
@E2ERunner - Playwright E2E tests
@SecurityAgent - Auth, encryption, audits
@PerformanceOptimizer - Profiling, benchmarking
@RefactorAgent - Code cleanup, dead code removal
@DocAgent - Documentation, README, API docs
@BugResolver - Bug diagnosis, root cause analysis
@ArchitectureAgent - System design decisions
@InfrastructureAgent - Deployments, scaling

Guardians

@CodebaseCartographer - Tracks all file changes
@TechStackFingerprinter - Detects technology stack
@DiscoveryProtector - Drift detection, blocks on mismatch
@OpenSpecPolice - Enforces specs as source of truth
@SpecArchitect - Manages spec backups and rollbacks
@InfraGuardian - Validates infrastructure config
@TeamBuilder - Rebuilds agents from stack detection
@ProjectAnalyzer - Deep scan, technical debt detection

Documentation

Document	Description
INSTALLATION.md	Installation instructions
Orchestration SOP.md	Complete operational guide
CLAUDE.md	Entry point (read by Claude Code)
.claude/hooks/README.md	Hooks documentation
thoughts/README.md	Memory system guide

Troubleshooting

Hooks not firing

Check .claude/settings.json exists and is valid JSON
Verify hook files exist: ls .claude/hooks/*.js
Make hooks executable: chmod +x .claude/hooks/*.js

Agents not found

Run the deploy script: bash LegendaryTeamDeploy.sh
Check agents exist: ls .claude/agents/

Memory not persisting

Check thoughts/ directories exist
Verify ledger files are being created
Run /bootstrap to initialize

Philosophy

Clear, Don't Compact - Never compact context. Clear and restore from ledgers.
Specs First - Document before coding. OpenSpec is truth.
Confidence Routing - Let data decide what needs human review.
Token Efficiency - Lite agents with self-escalation when needed.
Human Control - AI assists, humans decide on critical paths.

Version

Version: 2026-ultimate
Status: Production-ready
Last Updated: 2026-02-10

License

See LICENSE file for details.

Built for Claude Code | GitHub