Fabro: a dark factory as version-controlled workflow graphs
Mar 2026 Helmkamp/Fabro: Graphviz workflows with human gates, multi-model routing, cloud sandboxes, git checkpoints, and verification fix-loops — walk away while the line keeps running.
Fabro (Bryan Helmkamp / fabro-sh; MIT; open fabro-sh/fabro, ~1.6k★) is the open-source answer to babysitting a fleet of Claude Code / Codex tabs: encode the process as a Graphviz workflow graph, let agents, shell commands, and human hexagon gates execute it, and intervene only where the graph says so. The pitch is explicit about the continuous-factory problem always-on teams keep hitting — “extend average time between disengagements,” run agents 24/7 via an API server that queues while the laptop sleeps, and treat builds and tests as gates, not suggestions, with failures triggering automatic fix-loops.
The harness pieces are structural, not prompt poetry. Nodes are agent sessions, commands, conditionals, or human approvals; CSS-like stylesheets route cheap breadth to Haiku-class models and coding stages to frontier ones, with cross-provider critique. Every stage git-checkpoints code plus execution metadata onto branches so you can resume, revert, or audit. Cloud sandboxes (Daytona) isolate network and filesystem; fabro sandbox ssh and preview links keep the laptop out of the blast radius. Runs emit retros (cost, duration, files touched, LLM narrative) and a unified event stream you can query — the continuous-improvement loop the docs say broke when every sub-process boundary lost its data. A single Rust binary, no Node/Python/Docker required for the engine itself.
Honest packaging: Fabro is a factory framework, not a measured throughput claim — the receipts are the workflow graphs, the verification-first marketing, and an active nightly release train (still pushing mid-September). It sits beside NearForm Last Light and StrongDM-style attractors as “own the toolchain” rather than rent another REPL.
Why it matters for code review
Always-on capacity only beats bursty Claude Code mastery when “done” is a graph edge through a gate, and when human attention sits on approval hexagons instead of every intermediate diff. Fabro’s review surface is the morning queue of checkpointed, verification-gated stages — FOMO that nothing useful ran overnight is often missing a durable workflow, not missing another peak-session prompt.