archiet-microcodegen 0.1.0

Archiet's core algorithm in one file: PRD text -> Flask app ZIP, in <700 LOC of pure stdlib.

archiet-microcodegen

Archiet's core algorithm in one file. PRD text → manifest → genome → rendered Flask app → ZIP, in <700 LOC, with zero imports from app.*, agents.*, or templates. Pure stdlib. No LLM calls.

Inspired by Karpathy's microgpt.py philosophy: this file is the complete algorithm; everything else in Archiet is just efficiency on top.

Install

pip install archiet-microcodegen

Use

archiet-microcodegen path/to/prd.md --out ./out/

Or pipe to a ZIP:

archiet-microcodegen path/to/prd.md > app.zip

As a library:

from archiet_microcodegen import microcodegen

prd_text = open("prd.md").read()
zip_bytes = microcodegen(prd_text)

What you get

The generated output is a working Flask + Postgres CRUD app with:

• JWT-cookie auth (register / login / logout / me) — never localStorage
• Full CRUD per extracted entity (list / create / get / update / delete)
• Per-tenant row scoping (no Query.all() data leaks across users)
• JSON error handlers, CORS scoped to FRONTEND_URL, per-ZIP random secrets
• A single-page HTML demo at / that lets you exercise every entity from a browser
• A docker-compose.yml with healthcheck-gated Postgres startup
• A Dockerfile ready for Cloud Run / Fly / Railway

The whole algorithm runs in four stages:

1. parse_prd(text) → manifest — regex extraction of entities, fields, user stories, integrations
2. manifest_to_genome(manifest) → genome — maps to the canonical IR Archiet uses internally
3. render_genome(genome) → {path: content} — string.Template-based Flask rendering
4. pack(files) → bytes — stdlib zipfile

Why this exists

This isn't the production Archiet pipeline — it's the spec. The full platform (https://archiet.com) adds:

• LLM-powered PRD extraction (chunked, overlap+dedup)
• 12+ stack renderers (Flask, FastAPI, Django, NestJS, Laravel, Go, Java, Rails, .NET, Tauri+Rust, Salesforce, SAP CAP, Dynamics)
• Capability emission, frontend (Next.js / Expo), stub-filling, quality scoring, delivery gates
• Cross-stack parity enforcement and verification

But none of that changes the core algorithm. If a bug doesn't reproduce in microcodegen, it's in an efficiency layer, not the spec. If a new top-level genome key isn't in microcodegen, the genome doesn't have that key yet.

What's NOT included

• No LLM calls (the deterministic zone — see CLAUDE.md LLM-BOUNDARY)
• No frontend, no mobile, no payment integration, no rate limiting, no audit logging
• No Alembic migrations (uses db.create_all on first boot)
• No multi-stack output (Flask only; that's the algorithmic baseline)

For the full platform, see https://archiet.com.

License

MIT. See LICENSE.

Links

• Source: https://github.com/aniekanasuquookono-web/archiet/blob/main/scripts/microcodegen.py
• Issues: https://github.com/aniekanasuquookono-web/archiet/issues
• Full platform: https://archiet.com