模型决定下限,Harness 决定上限。
ArchIToken is an open-source AEC AI-Native CDE Workflow OS for the Architecture · Engineering · Construction (AEC) industry. It combines Harness Engineering, OpenBIM CDE workflows, Speckle interoperability, backend-native file runtimes, AI gates, approvals, audit trails and private deployment into one engineering system.
ArchIToken is the active product, repository, codebase and package/API compatibility name.
ArchIToken = AEC AI-Native + Harness Engineering + OpenBIM CDE Workflow OS
Product positioning:
ArchIToken = AEC AI-Native + Harness Engineering + OpenBIM CDE Workflow OS
ArchIToken does not fake proprietary kernels or use frontend derivatives as substitutes for real format support. Open formats are handled through native/open runtime paths; proprietary authoring formats such as RVT, DWG, DGN, Tekla, Navisworks, Office, PKPM and Glodon data must enter through backend workers, licensed adapters or enterprise services with auditable source binding.
Its compliance boundary is wider than AEC: production, logistics, customs, tax, finance, accounting, HR, organization governance, AI, data, cybersecurity and software workflows must also bind to the relevant regulator, standard, evidence and review state.
ArchIToken uses GitHub documents as the engineering source of truth. Some source files and document paths still contain the historical ArchIToken name during migration.
Read these first:
AGENTS.md02-architecture/ARCHITOKEN-SOURCE-OF-TRUTH.md02-architecture/CONSTITUTION.md02-architecture/POSITIONING_AND_COMPETITIVE_STRATEGY.md02-architecture/PROFESSIONAL_STANDARDS_COMPLIANCE.md02-architecture/MODULES.md02-architecture/MODULE-REGISTRY.md02-architecture/ARCHITECTURE.md
| Area | Decision | Principle |
|---|---|---|
| Project identity | ArchIToken | Single active product identity |
| Repository/codebase identity | ArchIToken | Active repository and compatibility identity |
| Engineering philosophy | Technology serves goals | No language/framework religion |
| Product position | AEC AI-Native + Harness Engineering + OpenBIM CDE Workflow OS | Open engineering runtime with backend-native file support |
| Cross-domain compliance | Professional roles + regulators + standards/codes | AEC, production, logistics, customs, tax, finance, accounting, HR, organization, AI, data, cybersecurity and software all need role, source, evidence and review state |
| Core backend | Rust / Cxx first | Python / Go / C++ / Perl / Shell / CUDA / WASM allowed when useful |
| Format runtime | Backend native first | Frontend viewers may render supported open formats, but Office/CAD/BIM proprietary support must come from backend workers or licensed adapters |
| Frontend | Next.js 16.2.6 + React 19.2.5 + TypeScript 6.0.3 + WASM | Application engineering base |
| Rendering | WebGPU first, Three.js r184 compatible | Three.js is ecosystem/compat layer, not the only rendering path |
| Registry | Registry over Enum | Modules, agents, tools, models, routers, renderers, geometry kernels, rules |
| Router | Internal unified Router | OpenRouter is only one external adapter, not the internal architecture |
| Schema | OpenAPI + AsyncAPI + JSON Schema + IFC Schema + Module Schema | Multi-schema truth system |
| Data | Capability composition | Structured, unstructured, vector, time-series, object, cache/state, audit |
| Cache/state | Zedis first | Not a single database religion |
| Deployment | k8s + Docker + local private deployment | Docker Compose for local dev, Kubernetes for production |
| AI gate | Generator != Evaluator | Planner → Generator → Evaluator → RuleChecker → SchemaValidator → Approver |
architoken/
├── .github/ Issue, PR and workflow templates
├── 01-product/ PRD, personas, success metrics
├── 02-architecture/ Constitution, source of truth, modules, registry, architecture
├── 03-frontend/ Next.js primary production frontend · React · TypeScript · WASM
│ └── tests/e2e/ Playwright E2E tests
├── 03-frontend-vite/ Experimental Phase 7 Open AEC sidecar workbench
├── 04-backend/
│ ├── Cargo.toml Rust workspace
│ ├── harness-core/ Router, RollbackGuard, RAG, RBAC, schema gates
│ ├── file-parsers/ DWG / DXF / IFC / STEP / PDF / XML
│ ├── shared/ Domain types, registries, module schema
│ ├── agent-orchestrator/ Agent orchestration adapters and prompt trees
│ ├── migrations/ PostgreSQL schema + RLS policies + module registry tables
│ ├── openapi.yaml REST API contract
│ └── deny.toml cargo-deny config
├── 05-infra/
│ ├── docker/ Dockerfile per service + local compose
│ ├── k8s/ Kubernetes manifests
│ ├── phase8/ Phase 8 runtime, observability and scale baselines
│ ├── rainbond/ Optional China private deployment template
│ └── ci/ GitHub Actions
├── 06-workers/ File conversion and domain worker adapters
├── 07-deployment/ Production runbook
├── 08-sdk/ Auto-generated clients
├── config/ Shared seed configuration
├── docs/ Technical notes, ADRs, audits and historical records
└── tools/ Local tooling and load-test utilities
git clone https://github.com/ActiveInAI/ArchIToken.git
cd ArchIToken
docker compose -f 05-infra/docker/docker-compose.yml up -d
# → frontend at http://localhost:3000
# → gateway at http://localhost:8080
# → agent at http://localhost:7001 1 · personal_center · 个人中心
2 · marketing_service · 市场客服
3 · planning_management · 计划管理
4 · concept_design · 方案设计
5 · standard_library · 标准族库
6 · detailed_design · 深化设计
7 · quantity_costing · 计量造价
8 · material_logistics · 材料物流
9 · production_manufacturing · 生产制造
10 · construction_management · 施工管理
11 · digital_twin · 数字孪生
12 · digital_archive · 数字档案
13 · finance_management · 财务管理
14 · human_resources · 人力资源
15 · ai_center · AI中心
16 · settings_center · 设置中心 (side-car)
All 16 modules are peers — no rigid “business vs. horizontal” split. Future modules can be added or retired through registry entries rather than global enum rewrites.
Registry mechanism:
Rust: trait Module + ModuleRegistry
Python: ModuleSpec adapter when useful
Database: modules table + module_id TEXT foreign keys
Frontend: Module Schema driven UI
Implementation maturity: the gateway exposes runtime capabilities, module files, transactions, generation, artifacts, openBIM, assets, conversion jobs, viewer commands, skills, MCP tools, and knowledge source routes. Production profile now requires PostgreSQL, S3-compatible object storage, NATS/Temporal queue endpoints, telemetry export, non-development auth secrets, and configured external generation provider routes; development fallback remains isolated for local testing.
Includes standards, family components, templates, materials, drawings, models, construction methods, rule libraries, and versioned engineering knowledge.
Includes inventory, suppliers, prices, purchase planning, cutting lists, BOM, packaging, loading, delivery, site stacking, receiving, loss control, and batch tracking.
Includes production planning, process routing, cutting optimization, CNC files, welding, coating, QC, MES/ERP integration, component coding, packaging, shipping, and progress feedback.
Includes construction plans, schedule, quality, safety, logs, AR, 360 capture, 3D scanning, oblique photography, drones, construction robots, IoT, visual progress comparison, defect detection, rectification loop, and completion documents.
Not soft guidelines — violations should fail CI.
Highlights:
- 第 3 条 · 技术服务目标,不做语言或框架信仰
- 第 6 条 · Registry 替代 Enum
- 第 7 条 · 多 Schema 协同真源
- 第 9 条 · 内部统一 Router,OpenRouter 只是外部适配器
- 第 12 条 · AI 不自评: Generator 与 Evaluator 分离
- 第 15 条 · 前端工程基座与 WebGPU 优先不冲突
- 第 16 条 · 数据库是能力组合,不是单一产品信仰
- 第 18 条 · 部署基线: k8s + Docker + 本地私有化
- 第 19 条 · 文档即环境,仓库文档是唯一真源
- 第 20 条 · 模块必须全面定义输入、输出、规则、Schema 与审计
Read the full text: 02-architecture/CONSTITUTION.md
ArchIToken's first production target is a 520 ㎡ three-storey heavy-steel villa in 贵州黔东南, Q355B structure, 300 mm grid, 45-day delivery, ¥680k budget. Real project; real forcing function. When the system can close the project's 16 modules end-to-end, the platform baseline has succeeded.
The full repository gate is:
04-backend/scripts/smoke-production-readiness-all.shIt runs the deterministic repository contract, frontend type/lint/test/build,
Rust tests and clippy, Python worker and agent tests, Phase 8 production
readiness smoke, and git diff --check. A release is not production-ready
until this passes and the environment-specific production smoke/load evidence
also passes against real deployed dependencies.
See CONTRIBUTING.md. Read the Constitution first and open an RFC before touching core architecture. During active development, fast-moving ecosystems such as LangChain, LangGraph, OpenAI and Anthropic may use bounded compatible ranges; release, CI, deployment and production artifacts must remain reproducible through lockfiles, constraints files, image digests or explicit release tags.
Apache-2.0 OR MIT · maintainer: ActiveInAI