How the factory works

fab is the factory client — the open-source reference orchestrator. It takes an intake brief and drives it through a phased pipeline to shipped, merge-gated software.

The pipeline

Work moves through phase-scoped sessions, each with its own slice of a 78-role roster (curator/engineer naming; flat, no coordinator). The phases run from intake and design through implementation, and every change meets an evidence-bound merge gate before it lands.

The merge gate

Four gate roles vote with transcripts and citations:

Role	Checks
pr-reviewer	correctness, scope, convention adherence
qa-security	security posture, test coverage
build-verifier	the build + gates actually pass
artifact-auditor	the output matches the contract

A verdict without evidence is auto-downgraded to a reject. Claims don’t ship; evidence does.

Transports

fab runs on four transports, selected by FAB_RUNTIME:

• Managed Agents (default)
• the in-process sdk loop
• per-session-isolated sdk-k8s
• claude-cli

Inference backend is separate (FAB_INFERENCE): the Anthropic API, Bedrock, or Claude Platform on AWS.

The standards

The bar every output is graded against lives in the catalog’s standards — the language toolchain, version currency, the platform-tenant contract, the LLM policy, the quality rubric’s ten dimensions, the testing rubric, observability and SLOs, the telemetry pipeline, resource naming, resource tagging, and the SEO baseline. See Templates & catalog for what each one covers. Anyone reading the public surface gets the bar and a working factory.