Dark Factory Experiment

Why it's ranked here

The project is compelling because its autonomy has explicit boundaries. Independent validation ignores the implementation plan, protected governance and security areas reject automated edits, and deterministic dispatch replaces model judgment for scheduling. It also exposes the experiment through a substantial working product. However, a critical orchestration component lives outside the repository, limiting reproducibility from this source alone.

What's good

The factory design turns useful safety ideas into enforceable mechanics. Agents cannot rewrite their governing rules. Review uses fresh context and checks results against the original issue. Per-node budgets and intake limits constrain cost and abuse. The application also combines keyword and vector retrieval, preserves transcript timestamps, streams responses, and protects user-scoped data with authentication.

Tradeoffs

This is not full specification-to-production autonomy. Humans still create issues and promote releases. The external cron dispatcher is required but absent from the repository. Running the application needs Python, Bun, Postgres with vector support, and paid external services for models, embeddings, and transcript retrieval. Channel synchronization is synchronous and sequential, so callers may need long request timeouts.

How to use it well

Study it when designing an agent-managed engineering pipeline, especially if you want concrete patterns for holdout review, immutable governance, label-based state, bounded retries, and deterministic scheduling. The included application provides a realistic test bed for those ideas. Do not treat it as a self-contained factory distribution, an offline transcript tool, or a database-free starter.

Observed

Primary languages

Python backend and TypeScript React frontend

Interfaces

Browser application and session-authenticated FastAPI HTTP API with Server-Sent Events

Install surface

Requires Python 3.11+, uv, Bun, Postgres 16+ with pgvector, and an OpenRouter API key

Application stack

FastAPI, React 18, Vite, Tailwind CSS, asyncpg, Postgres full-text search, and pgvector

Factory structure

Archon workflows handle triage, implementation, validation, regression testing, and merging through GitHub state

Repository boundary

The cron-driven bash orchestrator is explicitly not included in the repository

Testing surface

Documented validation includes ruff, mypy, pytest, TypeScript checking, Biome, Vitest, and browser scenarios