VS Code Extension · v2.0

Spec-Driven Development
with AI autonomy

One panel. Five tabs. Write a spec, drag a task to In Progress, and watch the AI implement it — locally or in the cloud, following Clean Architecture from the first file.

Download VSIX View on GitHub

VS Code Marketplace release coming soon — install manually from GitHub Releases in the meantime.

16+
Supported stacks
4
AI providers
14
Domain Agent Shells
0
Intermediary servers
Write
spec.md
Kanban
Board
AI
implements
spec.md
updated
git commit
#SPEC-XXX
🏠
Clean Arch
structure

What is Spec-Driven Development?

SDD is the core idea behind Alpaquitay. Your spec.md is the single source of truth for requirements, progress, and AI context.

spec.md — the source of truth

## Epic: Authentication - [x] Design auth flow SPEC-001 ✓ - [ ] Implement JWT provider SPEC-002 - [ ] Add refresh token rotation SPEC-003 ## Epic: Dashboard - [ ] Build metrics chart component SPEC-004 - [ ] Add export to CSV SPEC-005
Step 1

Write epics and tasks in plain Markdown. AI can generate the initial spec from your goal description.

Step 2

The Kanban board derives its state from the checkboxes — no separate database, no sync problems.

Step 3

Drag a task to In Progress. The AI plans files, generates code, and writes them to your workspace.

Step 4

The task is auto-marked done in spec.md. Your git commit references #SPEC-002 for full traceability.

Correction

Drag a completed task back. Type what needs fixing. The AI re-implements with your feedback applied.

For everyone

Non-technical stakeholders read spec.md and understand the full project state. No Jira login needed.

Features

Everything in one panel — no context switching.

Spec Editor

View and edit spec.md in the panel. Regenerate from AI. Convert existing YAML, OpenAPI, or Gherkin files into spec format.

Kanban Board

Four columns: Backlog, Todo, In Progress, Done. State derived from spec.md — drag to implement, drag back to correct.

💬

Chat

Conversational AI with streaming responses. Stack-aware system prompt. Task work streams progress here in real time.

Git Integration

Last 40 commits. Commits referencing #SPEC-XXX are linked to tasks. Full traceability from code to requirement.

Skills

Built-in: Refactor, Generate Tests, Create File, Generate from Spec, Project Builder, Daily Standup. Custom skills via UI or TypeScript.

📜

Hierarchical Memory

Auto-extracts class names, exported functions, and completed features into .alpaquitay/memory.json. Keeps AI context coherent across sessions.

📋

Stack Detection

Scans package.json, requirements.txt, pom.xml, go.mod, .csproj. Generates architecture-specific prompts for 16+ stacks.

🏠

Clean Architecture

Generates Java, React, and Spring+React projects with strict Clean Architecture layers: domain → application → infrastructure. ISO 12207 compliant.

🔒

Privacy First

No Alpaquitay servers. Code and prompts go directly to your provider. API keys in OS keychain, never on disk.

Instant Kanban — Deferred Quality Pipeline

Kanban card moves to Done in < 3 s after file generation. Format, validate, build, and test run asynchronously — results stream to chat. >95% faster than the blocking pipeline.

📊

Auto Architecture Diagram

When you open the Architecture tab on a new or code-less project, Alpaquitay infers components from spec.md epics and workspace folders and generates the diagram automatically — no manual node placement.

Architecture

How the extension is built — and how it builds your projects.

1. Extension Host Overview

All processing happens inside the VS Code Extension Host (Node.js). The WebView SPA communicates via the VS Code postMessage API — no HTTP servers, no ports, no data leaving your machine unless you invoke a cloud AI provider.

WebView — isolated browser context (5 tabs)
Spec Tab
Kanban Tab
Chat Tab
Git Tab
Settings Tab
↕  acquireVsCodeApi().postMessage  /  onDidReceiveMessage

Extension Host — Node.js process
MainPanel
MessageRouter
SkillRegistry
DeepAgentSkill
AIClient
ProjectContextBuilder
ModelCatalog
MCPExecutor
AlpaquitayConfig
HierarchicalMemory
Anthropic APISSE / Messages API
OpenAI APISSE / Chat Completions
Ollamalocal / localhost:11434
LM Studiolocal / localhost:1234
filesystem MCPread, write, list
git MCPlog, commit, diff

2. Clean Architecture Layers (Robert C. Martin)

Dependencies always flow inward. Infrastructure depends on Application; Application depends on Domain. The Domain layer has zero framework imports — it is pure business logic.

Infrastructure Web · Persistence · Configuration
Spring @RestController · JPA @Entity · @Configuration beans · Adapter classes that implement domain ports
infrastructure/web/rest infrastructure/persistence infrastructure/config
depends on ↓  —  this direction only, never reversed
Application Use Cases · Orchestration · DTOs
Service classes that coordinate domain objects · Input/Output DTOs that cross layer boundaries · No framework annotations
application/service application/dto
depends on ↓  —  this direction only, never reversed
Domain Business Rules · Entities · Port Interfaces
Pure Java POJOs (or TypeScript classes) · Repository interfaces (ports) that infrastructure implements · Zero framework imports allowed
domain/model domain/port
— cross-cutting —
Shared Cross-cutting Concerns
Global exception handlers · Utility classes · Constants accessible by all layers without creating circular dependencies
shared/exception shared/util

3. Generated Project Structures

Alpaquitay scaffolds every new project with the correct package hierarchy from the first file. The AI is instructed to respect these boundaries in every subsequent code-generation step.

Java · Spring Boot · Maven

src/main/java/com/example/myapp/

├── domain/
│   ├── model/
│   │   └── User.java          # pure POJO
│   └── port/
│       └── UserRepository.java # interface

├── application/
│   ├── service/
│   │   └── UserService.java
│   └── dto/
│       ├── UserRequest.java
│       └── UserResponse.java

├── infrastructure/
│   ├── persistence/
│   │   ├── UserEntity.java    # @Entity
│   │   └── UserJpaRepository.java
│   ├── web/rest/
│   │   └── UserController.java# @RestController
│   └── config/
│       └── BeanConfig.java

└── shared/
    └── exception/
        └── GlobalExceptionHandler.java

React · Vite · TypeScript (Clean Architecture)

src/

├── domain/
│   └── model/
│       └── user.ts             # TS interface

├── application/
│   ├── use-case/
│   │   └── getUser.ts
│   └── store/
│       └── userStore.ts        # Zustand/Pinia

├── infrastructure/
│   ├── api/
│   │   └── userApi.ts          # fetch/axios
│   └── storage/
│       └── localStorage.ts

├── presentation/
│   ├── components/         # reusable UI
│   ├── pages/              # route-level views
│   └── layout/             # shell / nav

└── shared/
    ├── hooks/
    └── utils/

Spring Boot + React · Full-Stack Monorepo

my-project/

├── backend/                          # Spring Boot (Maven)
│   └── src/main/java/com/example/
│       ├── domain/               # model + port (no Spring deps)
│       ├── application/          # service + dto
│       ├── infrastructure/       # web/rest + persistence + config
│       └── shared/               # exception + util

├── frontend/                         # React + Vite (Clean Architecture)
│   └── src/
│       ├── domain/               # TS interfaces (pure)
│       ├── application/          # use-case + store
│       ├── infrastructure/       # api + storage
│       ├── presentation/         # components + pages + layout
│       └── shared/               # hooks + utils

├── docker-compose.yml
└── spec.md

4. SDD Task Execution Pipeline

When you drag a task card to "In Progress," this is the exact sequence of operations inside the extension.

1
Drag event (WebView)

The SPA fires task:start over postMessage with the task ID and full title from spec.md.

2
Context assembly (Extension Host)

ProjectContextBuilder reads stack files (package.json, pom.xml, go.mod…) and builds the master prompt. HierarchicalMemory injects known classes, exports, and completed features from .alpaquitay/memory.json.

3
File plan — AI round-trip 1

DeepAgentSkill sends the task + context to the AI and requests a JSON file plan: which paths to create or modify, in what order, with what Clean Architecture layer each file belongs to.

4
Code generation — AI round-trip 2+

For each file in the plan, a second AI call generates the full content. SSE streaming sends tokens to the Chat tab in real time so you can watch the implementation happen.

5
Write to workspace (MCP)

MCPExecutor calls the filesystem MCP server to write each file at the planned path. Directories are created as needed, preserving the Clean Architecture layout.

6
spec.md update + git commit

The task checkbox is ticked in spec.md. A git commit is created via MCP with message feat: [task title] #SPEC-XXX for full traceability.

7
Memory extraction

HierarchicalMemory scans the new files for class names, exported functions, and interfaces. These are appended to .alpaquitay/memory.json so the next task has full context.

5. AI Provider Chain

Every provider is abstracted behind the same AIClient interface. The active provider, model, and token limit are resolved from AlpaquitayConfig — zero hardcoded values in the pipeline.

User message or task drag → AIClient.streamChat(messages, { provider, model, maxTokens })
↓  provider routing from AlpaquitayConfig.preferredProvider  +  ModelCatalog
AnthropicSSE · Messages API
OpenAISSE · Chat Completions
Ollamalocal SSE · /api/chat
LM Studiolocal · OpenAI-compatible
↓  async token stream  →  chat:token events  →  WebView renders live
chat:done → task marked complete · spec.md updated · git commit created

The Architecture to Win a Vertical

Stop building better prompts. Build autonomous agents that own entire workflows end-to-end. Every industry will have a winner — the moat is domain knowledge, not the LLM.

Domain Agent Shell — Hexagonal Architecture (Ports & Adapters)

Based on TOGAF ADM, ArchiMate 3.2, and BIAN Service Domain patterns. Dependencies always point inward; adapters implement ports; the core never touches infrastructure.

DOMAIN AGENT SHELL
Process Definition
Replaces spec.md
SOP / BPMN
Domain Tools
Real APIs
ERP / CRM / TMS
ALPAQUITAY AGENT ENGINE
decompose execute validate
🧠
Domain Memory
Persistent context
per learner / case
🔒
Compliance Guardrails
Domain-specific
validation rules
TOGAF ADMB → C → D → E
ArchiMate 3.2Application Layer
BIANService Domain
Hexagonal ArchPorts & Adapters
4+1 ViewsLogical · Process · Physical
ISO Alignedper Vertical

4+1 Architectural Views Applied to Every Domain Shell

1
Logical View
Pure domain model — entities, value objects, aggregates. Zero framework imports. The business language of the vertical encoded in TypeScript types.
domain/model.ts
2
Development View
Hexagonal package structure: domain/ports/application/infrastructure/. Dependency rule enforced at compile time.
src/domains/{vertical}/
3
Process View
Objective → decompose → primary port → use case → secondary port → adapter → API. Guardrails check before every output commit.
IDomainAgentShell.run()
4
Physical View
VS Code Extension Host (in-process). Local storage at .alpaquitay/{vertical}/. Any Alpaquitay AI provider as the generation backend.
infrastructure/adapters/
+1
Scenarios
Each use case is a named scenario: "Practice past perfect at B1", "Assess my level from this text", "Give me 5 business phrases". Scenarios drive port design — not the other way around.
shell.run('use-case-id', params)

Impact Domains — One Agent to Win Each Vertical

The question is not which AI model you use. The question is: which repetitive process do you know better than anyone in your industry?

14 Live shells · 5 planned · All extending BaseDomainShell · 0 TypeScript errors

LIVE
🇬🇧
English Mastery
CEFR · ISO 17024 · ISO 21001
CEFR A1→C2 grammar drills, IPA pronunciation, daily business phrases, level assessment from free text.
practice-grammar · assess-level · get-daily-phrases
LIVE
💻
Software Engineer
ISO/IEC 25010 · 12207 · SOLID
Code review, SOLID analysis, tech debt detection, design pattern suggestions, complexity estimation.
review-code · analyze-solid · detect-tech-debt
LIVE
🏛
Software Architect
ISO/IEC 42010 · TOGAF · C4
Architecture assessments, ADR generation, C4 diagrams, tech radar, quality attribute evaluation.
assess-architecture · create-adr · generate-c4
LIVE
Developer
ISO/IEC 12207 · Clean Code
Feature implementation, bug debugging, code refactoring, explanation, and test generation.
implement-feature · debug-issue · generate-tests
LIVE
QA
ISO/IEC 29119 · IEEE 829
Test plans, test case generation, bug triage, coverage evaluation, quality gate definition.
create-test-plan · triage-bug · define-quality-gate
LIVE
🚀
DevOps
DORA Metrics · ISO/IEC 27001
CI/CD pipeline design, DORA assessment, deployment planning, IaC generation, runbooks.
design-pipeline · assess-dora · generate-iac
LIVE
🛡
DevSecOps
OWASP SAMM · ISO/IEC 27001
Secure pipeline design, threat modelling, SAMM assessment, findings triage, SBOM generation.
threat-model · assess-samm · generate-sbom
LIVE
🔒
Security
NIST CSF · ISO/IEC 27001/27005
Compliance audit, penetration test planning, risk register, incident response, CSF assessment.
audit-compliance · plan-pentest · respond-incident
LIVE
🔥
Infrastructure
ISO/IEC 27001 · ITIL v4
Capacity planning, network design, SLA creation, disaster recovery, monitoring configuration.
plan-capacity · design-network · plan-dr
LIVE
Cloud (AWS/Azure/GCP)
AWS WAF · ISO/IEC 27017
Multi-cloud architecture design, Well-Architected reviews, cost optimisation, migration planning, IaC.
well-architected-review · optimize-cost · plan-migration
LIVE
📊
Marketing
ISO 9001 · IAB Standards
Campaign planning, audience segmentation, SEO analysis, content creation, ROI measurement.
plan-campaign · analyze-seo · measure-roi
LIVE
🔄
Process
ISO 9001 · BPM CBOK · Six Sigma
Process mapping, gap analysis, value stream mapping, ISO compliance audit, Six Sigma optimisation.
map-process · value-stream-map · iso-compliance
LIVE
🧠
AI Expert
ISO/IEC 42001 · EU AI Act · NIST AI RMF
LLM evaluation, RAG architecture, prompt engineering, AI system design, EU AI Act governance & MLOps.
evaluate-llm · design-rag · assess-governance
LIVE
📈
Business Expert
ISO 56002 · OKR Framework · BMC
Business Model Canvas, SWOT/PESTLE analysis, OKR definition, financial modeling, market analysis, business cases.
design-business-model · define-okrs · financial-model
ROADMAP
💰
Finance
ISO 20022 · BIAN
Invoice → reconciliation → accounting entry. Credit assessment, regulatory reporting, multi-jurisdiction tax rules.
ERP API · QuickBooks · Tax DB
ROADMAP
Legal
ISO/IEC 27001 · GDPR
Contract → clause extraction → risk flagging. Jurisdiction-specific clause library. Chain of custody for discovery documents.
Legal DB · NLP · eSign API
ROADMAP
🚚
Logistics
ISO 9001 · GS1
Order → routing optimization → carrier assignment. Exception detection, SLA monitoring, customs classification by HS code.
TMS API · Carrier APIs · GS1 DB
ROADMAP
👥
Recruiting
ISO 30405
JD → candidate sourcing → bias-free scoring → interview scheduling. GDPR-compliant candidate data lifecycle.
ATS API · Calendar · LinkedIn
ROADMAP
Healthcare
ISO 13485 · HL7 FHIR
Patient intake → triage → care plan. Clinical protocol adherence, medication safety guardrails, outcome tracking.
FHIR API · Clinical DB · Safety Rules

Multi-Agent Orchestration Stack

Every request flows through a unified pipeline: privacy protection → knowledge augmentation → multi-agent execution → recursive refinement → learning.

CentralBrainAgent Unified entry point — session history, RAG augmentation, pipeline orchestration, self-learning
PrivacyGuard 12 PII patterns · GDPR Articles 5 & 17 · ISO 27018 · Masks before AI calls
RAGEngine BM25-lite retrieval · ISO seed chunks · Learns from high-score outputs (≥80)
OrchestratorAgent Decomposes objectives → assigns tasks to best-scoring shells → parallel execution respecting dependency graph
MetaheuristicEngine — auto-selects algorithm by problem size
n ≤ 3 → Greedy n ≤ 10 → Genetic Algorithm
pop=20, gen=50, elitism 10%
n > 10 → Simulated Annealing
T₀=1.0, α=0.95, 200 iter
RecursiveRefinement Rubric-based quality scoring → if score < threshold and depth < max → regenerate and recurse

Reference Implementation: English Mastery Shell

The reference implementation — every architectural layer demonstrated with production-ready TypeScript. CEFR A1→C2.

CEFR Levels (ISO 17024 Aligned)

A1Beginner~500 words
A2Elementary~1,500 words
B1Intermediate~3,500 words
B2Upper-Intermediate~7,000 words
C1Advanced~12,000 words
C2Proficient~20,000 words

Use Cases (Primary Ports)

practice-grammar CEFR-level grammar drill session with explanations
get-daily-phrases 5 phrases with IPA transcription & practice script
assess-level Detect CEFR level from free text or exercise history
submit-exercise Evaluate answer, give feedback, update progress
get-progress Weekly insight, skill scores, streak, next milestone

Hexagonal Package Structure

domain/ model.ts — CEFRLevel, Exercise, Lesson, DailyPhrase, LearnerProgress — zero framework imports
↓ ports expose, ports require
ports/ input.ts (primary) — IPracticeGrammar, IGetDailyPhrases, IAssessLevel  |  output.ts (secondary) — IEnglishAIPort, ILessonRepository, ISpeechPort
↓ use cases implement input ports, depend on output ports
application/ PracticeGrammarUseCase · GetDailyPhrasesUseCase · AssessLevelUseCase · SubmitExerciseUseCase
↓ adapters implement output ports
infrastructure/ AIProviderAdapter (wraps Alpaquitay AI) · LessonStorageAdapter (.alpaquitay/english/data.json)

AI Providers & Models

Choose your provider in Settings. The model list, context window, and max-tokens field update dynamically from the built-in ModelCatalog — no hardcoded values.

Anthropic Claude
Cloud
claude-opus-4-7200k ctx · 32k out
claude-sonnet-4-6200k ctx · 64k out
claude-haiku-4-5200k ctx · 8k out
OpenAI
Cloud
gpt-4o128k ctx · 16k out
gpt-4o-mini128k ctx · 16k out
o1 / o1-mini200k ctx · 32k out
gpt-3.5-turbo16k ctx · 4k out
Ollama
Local · Free

Runs any GGUF/GGML model on-device. Auto-detected at localhost:11434. No API key required. Small-model mode activates automatically for models under 4B params.

LM Studio
Local · Free

GUI model manager with OpenAI-compatible API. Auto-detected at localhost:1234. Supports any model loaded in LM Studio.

Hierarchical Memory

A living index of your project that grows as the AI works — stored locally in .alpaquitay/memory.json. Never sent to any server.

LevelWhat is storedExample key
projectName, description, architecture decisionsgoal, stack
componentMajor subsystemsauth, dashboard
moduleSpecific modules within a componentjwt-provider
sectionLogical sections within a moduletoken-refresh
featureCompleted spec tasks + output filesSPEC-001
packageExternal packages and why chosenjsonwebtoken
classClass names and their source filesUserService
methodExported functions and their filescalculateTotal
annotationInterfaces, types, enumsAuthPayload
keyvalueAd-hoc key-value factsdb_url
configConfiguration entriesmax_upload_mb

Get started in 2 minutes

Install the extension, open your project folder, and press Ctrl+Shift+A to open the hub or Ctrl+Alt+A for the command menu.

code --install-extension alpaquitay-ai.alpaquitay-ai

Or search Alpaquitay AI in the Extensions panel inside VS Code.