AI Engineering Orchestrator

An orchestration and context-management layer above AI coding agents (Claude Code, Codex CLI, GitHub Copilot, or any CLI-based agent). It turns an AI coding agent into a structured engineering workflow with roles, skills, project memory, token-efficient context selection, quality gates, and optional multi-agent orchestration.

Why

Sending an AI agent your entire project on every request wastes tokens and dilutes focus. This extension builds a minimal context package — current task, relevant role, relevant skills, relevant files (scored and budgeted), relevant decisions and conventions — instead of the whole repository.

Core features

Provider-agnostic: adapters for Claude Code and Codex CLIs, a best-effort GitHub Copilot handoff, and a prompt-package fallback that always works.
Context engine: local file indexing, relevance scoring (current file / dependencies / tests), a token budget, and multi-level compression (strip comments → extract relevant symbol → summarize) before anything is dropped.
Secret redaction: API keys, tokens, JWTs, and private key blocks are redacted before any content leaves your machine.
Project memory: .ai-orchestrator/ stores requirements, decisions, tasks, and plans as small, incrementally-updated documents — never a single giant file.
Workflow engine: an 11-phase pipeline (DISCOVER → ... → RELEASE) that dynamically skips phases based on task size.
Quality gates: Implementation, Testing, Security, and Review gates that run your project's own npm scripts.
Multi-agent mode: Planner → Architect → Backend/Frontend → QA → Security → Reviewer, each step reading only the one prior artifact it needs — not the full conversation history.
Role permissions: roles carry explicit allowedActions / restrictedActions; a multi-agent step is skipped, not forced, if its role lacks permission.
Approval levels: READ_ONLY / SUGGEST / EDIT / EXECUTE / AUTONOMOUS. Destructive actions always require confirmation regardless of level.

Getting started

Open a workspace folder.
Run AI Orchestrator: Initialize Project to scaffold .ai-orchestrator/ and detect your stack.
Run AI Orchestrator: Analyze Requirement or AI Orchestrator: Create Plan to turn a plain-language ask into a structured artifact.
Run AI Orchestrator: Show Context on any file to see exactly what would be sent to an AI provider, and why.
Run AI Orchestrator: Run Multi-Agent Workflow for a full planner-to-reviewer pipeline on an objective.

See the AI Orchestrator view in the activity bar for Tasks, Decisions, Roles, and Skills.

Commands

Command	Purpose
Initialize Project	Detects stack, scaffolds `.ai-orchestrator/`
Analyze Requirement	Natural language → structured requirement
Create Plan	Objective → structured implementation plan
Run Multi-Agent Workflow	Planner → ... → Reviewer pipeline
Show Context	Shows the scored, budgeted context for the active file
Export Prompt Package	Builds a redacted, budgeted prompt for manual use
Select Role / Select Skills	Sets the active role/skills for context selection
Run Quality Gates	Runs Implementation/Testing/Security/Review gates
Review Changes	Shows current branch and working-tree changes
Show Decisions	Opens the decision log
Switch Provider	Chooses the active AI provider
Show Dashboard	Status: provider, role, skills, token savings, recent decisions

Configuration

See aiOrchestrator.* settings, including context.maxTokens, context.maxFiles, security.redactSecrets, and workflow.approvalLevel.

Privacy

Indexing, scoring, memory, and token estimation are all local. Nothing is sent externally except through an explicitly selected provider, and secret redaction runs first.