---

🧠 What It Does

When you give Claude Code a large engineering goal, this skill activates and:

1. 🔍 Decomposes the goal into atomic, testable tasks with task-type hints (bug fix, feature, refactor, perf, security)
2. ⚡ Auto-detects your stack — language, framework, test runner, build system, linter
3. 💰 Routes each task to the cheapest model tier capable of handling it (saves tokens)
4. 🧠 Advisor pattern — cheap executor consults Opus at critical decision points for higher accuracy at lower cost
5. 🔁 Executes with smart parallel/sequential decisions — explicit rules prevent wasted multi-agent token burn
6. ♻️ Reuse-first — Reader scans for existing components before Builder writes; QA rejects duplicates
7. ✅ Verifies every task through a 9-point Done Gate (build, tests, lint, regression, security, QA…)
8. 🛠️ Self-heals — failures trigger repair loops with convergence detection and smart termination
9. 🧭 Anti-drift checks — Leader verifies goal alignment every 3 tasks, catches scope creep before it wastes tokens
10. 📊 Periodic self-eval — fresh QA subagent evaluates cumulative quality every 4-5 tasks (DRY, architecture, plan alignment)
11. 🧹 Context engineering — just-in-time retrieval, structured notes, artifact-based communication, compaction strategy

No task is marked DONE until build passes, tests pass, and the leader agent explicitly approves.

---

🚀 Quick Install

npx sean-agentic-orchestrator

Then open Claude Code and type /agentic-orchestrator. That’s it.

---

📦 All Install Options

Option A — npx (Recommended)

npx sean-agentic-orchestrator

or

npx sean-agentic-orchestrator@latest

Automatically copies skill files to ~/.claude/skills/agentic-orchestrator.

Option B — Global Install

npm install -g sean-agentic-orchestrator

Option C — Git Clone

git clone https://github.com/seancheick/Sean-Agentic-Orchestrator.git /tmp/ao
mkdir -p ~/.claude/skills
cp -r /tmp/ao ~/.claude/skills/agentic-orchestrator
rm -rf /tmp/ao

Option D — Manual

1. Download this repo as a ZIP
2. Extract and copy the folder to ~/.claude/skills/agentic-orchestrator

✅ Verify Installation

/agentic-orchestrator

Claude will activate the orchestrator and begin the intake phase.

---

🛠️ Supported Stacks

Works with any tech stack. Includes built-in commands, pitfalls, and conventions for:

Stack	Detection
⚛️ Next.js / React / TypeScript	package.json, tsconfig.json, next.config.*
🐦 Flutter / Dart	pubspec.yaml, .dart files
🐍 Python / Django / FastAPI	pyproject.toml, manage.py, requirements.txt
🦀 Rust	Cargo.toml
🐹 Go	go.mod
🍎 Swift / iOS	Package.swift, .xcodeproj
☕ Spring Boot / Java / Kotlin	pom.xml, build.gradle
🐘 Supabase / PostgreSQL	supabase/ directory, .sql files
📦 Monorepos	turbo.json, nx.json, pnpm-workspace.yaml

---

💡 Usage

Invoke Directly

/agentic-orchestrator

Add user authentication with JWT tokens, refresh token rotation,
and role-based access control to my Express API.

Natural Trigger Phrases

The skill also activates on:

• “Break this down and execute…”
• “Orchestrate the implementation of…”
• “Plan and execute this feature…”
• Any multi-step engineering goal

What You’ll See

Planning phase — structured task list with model tiers, risk levels, dependencies:

T1: Set up JWT signing/verification utilities    [Mid] [Low Risk]
T2: Create auth middleware                       [Mid] [High Risk]
T3: Add refresh token rotation endpoint          [Mid] [High Risk]
T4: Implement role-based access control          [Mid] [Medium Risk]
T5: Add auth tests                               [Mid] [Low Risk]

Execution phase — verified output after each task:

TASK EXECUTION
Task ID: T1  |  Status: QA_REVIEW
Files: src/utils/jwt.ts — signing/verification helpers
Build:  PASS
Tests:  PASS (12 passed, 0 failed)
Lint:   PASS

Done Gate — explicit leader sign-off before the next task begins.

---

🤖 Agent System

9 specialized agents coordinate to execute your work:

Agent	Role
👑 Leader	Owns the plan, controls progress, runs done gates, anti-drift checks
📋 Planner	Decomposes goals into atomic tasks
📖 Reader	Cheap context gathering and targeted file scanning
🔨 Builder	Implements changes with strict scope control
🐛 Debug	Investigates failures, isolates root causes
🧪 Test	Validates behavior, identifies missing coverage
🔍 QA/QC	Reviews quality, rejects incomplete work
🔒 Security	OWASP scanning, secrets detection, auth validation
⚡ Perf	Profiling, bottleneck detection, optimization validation

---

✅ The 9-Point Done Gate

Every task must pass ALL gates before being marked DONE:

#	Gate	What It Checks
1	📐 Scope	All requested work is implemented
2	🎯 Acceptance	Success criteria are satisfied
3	🏗️ Build	Zero errors, zero new warnings
4	🧪 Tests	Pass — new tests where gaps existed
5	🧹 Lint	Passes (if configured)
6	🔄 Regression	No breakage in adjacent code
7	🔒 Security	No new vulnerabilities (boundary tasks)
8	💎 QA	Maintainable, follows conventions
9	👑 Leader	Explicit approval

Any gate fails → task status = FIX_REQUIRED → automatic repair loop.

---

💰 Model Routing

The orchestrator routes each task to the cheapest capable model tier — no wasted tokens:

Tier	Model	Used For
🟢 Cheap	haiku	Reading, scanning, extraction, summaries
🟡 Mid	sonnet	Implementation, testing, QA, bug fixes
🔴 High	opus	Architecture, ambiguous root-cause, critical reviews
🧠 Advisor	opus (consulted)	Plan review, high-risk decisions, phase sign-off

New in v2.3: The Advisor Pattern — instead of running entire tasks on the most expensive model, a Mid-tier executor consults a High-tier advisor only at critical decision points. Result: +2.7% accuracy, -11.9% cost.

---

🆕 What’s New

Version	Feature	Impact
v2.4.0	Code Reuse Enforcement — Reader scans for existing components; Builder reuses; QA rejects duplicates	Eliminates bloat and duplicate code
v2.4.0	Clean & Simple Code (Principle #9) — 3+ uses to abstract, ~40 line max, no speculative code	Stops over-engineering
v2.4.0	Periodic Self-Eval — fresh QA subagent evaluates every 4-5 tasks (DRY, architecture, alignment)	Catches quality issues mid-project
v2.4.0	Harness Config — reasoning effort HIGH, hooks for enforcement, context discipline	More accurate responses
v2.4.0	Opus 4.7 alignment — model IDs updated, adaptive-thinking xhigh guidance, self-verification awareness	Accurate routing on current model lineup
v2.3.1	Anti-Drift Checks — Leader verifies goal alignment every 3 tasks	Catches scope creep before it burns tokens
v2.3.1	Parallel Invocation Rules — explicit decision table for when to parallelize vs stay sequential	Prevents wasted multi-agent overhead (~15x token cost)
v2.3.1	Task-Type Hints — starting pipeline suggestions for bug fix, feature, refactor, perf, security	Faster planning, but planner adapts based on actual task
v2.3.0	Advisor Pattern — Opus consulted at plan review, critical decisions, and phase sign-off	+2.7% accuracy, -11.9% cost
v2.3.0	Context Engineering — just-in-time retrieval, structured notes, artifact-based agent communication	Prevents context bloat, extends effective session length
v2.3.0	Smart Repair Termination — convergence detection, quality thresholds, self-consistency trap avoidance	Stops wasted repair cycles, saves tokens

All enhancements validated against Anthropic’s published guidance on building effective agents.

---

📁 Skill Structure

agentic-orchestrator/
├── SKILL.md                        # Core orchestrator — principles, loop, done gate
└── references/
    ├── agent-roles.md              # 9 specialized agent definitions
    ├── advisor-pattern.md          # Advisor consultation strategy & integration
    ├── context-engineering.md      # Token efficiency & memory management
    ├── harness-config.md           # Accuracy optimization settings & hooks
    ├── model-routing.md            # Full escalation/de-escalation policy
    ├── output-templates.md         # 8 structured output templates
    ├── stack-patterns.md           # 9 tech stack command & pitfall guides
    └── failure-recovery.md         # Recovery patterns & escalation ladder

---

⚙️ Configuration

Zero configuration needed. The skill auto-reads your project’s existing config:

• CLAUDE.md — project instructions
• package.json / Cargo.toml / go.mod / pyproject.toml — scripts & deps
• tsconfig.json — TypeScript settings
• .eslintrc / biome.json / rustfmt.toml — linter config
• turbo.json / nx.json — monorepo config

---

🤝 Contributing

Contributions welcome!

1. Fork this repository
2. Edit references/ for stack patterns, agent roles, or recovery patterns
3. Edit SKILL.md for core orchestration logic
4. Test locally: copy to ~/.claude/skills/agentic-orchestrator and run /agentic-orchestrator
5. Submit a PR with description of what changed and why

Adding a New Stack

Edit references/stack-patterns.md:

1. Add detection (what config files indicate this stack)
2. Add build/test/lint commands
3. Add common pitfalls
4. Add conventions to check

---

📄 License

MIT License — see LICENSE for details.

---