Quickstart: fork a starter

examples/workflows/ holds forkable starters: standalone, self-contained reference implementations of the canonical workflow shapes from the design corpus. The lifecycle is deliberately fork-and-forget — copy the directory out, and it’s your code. No upgrade path, no compatibility promise; mast’s CI only keeps the in-repo originals building. Starters never import each other, and duplication between them is accepted on purpose.

v0.1 ships two shapes:

LLM-as-router

A cheap SingleTurn classifier reads the input, and the graph dispatches to the matching Task-mode specialist — with a Default fallback for anything the classifier can’t place. Runs offline with deterministic in-process fakes:

go run ./examples/workflows/llm-as-router

Or pipe your own tickets, one per line:

echo "I was double-charged on my last invoice" | go run ./examples/workflows/llm-as-router

Customization points, in the order you’ll hit them: the categories table (the whole routing domain), the two model seams (swap the fakes for e.g. gemini.NewModel(...) — keep the classifier SingleTurn: one call, cheap tier), per-specialist tools, and an optional HITL approval gate.

Fan-out-fan-in

A planner emits N independent work items, workers execute them in parallel under a concurrency cap, a join barrier collects results, and a summarizer folds N results into one output. Every node is deterministic, so it runs with no model at all:

go run ./examples/workflows/fan-out-fan-in

Or pipe a fleet of service names:

echo "web, api, worker" | go run ./examples/workflows/fan-out-fan-in

The README in the starter documents the two constraints this shape lives under (no HITL inside parallel branches — escalation belongs after the join — and which of the three concurrency knobs actually binds).

How to fork one

cp -r examples/workflows/llm-as-router ~/src/my-router
cd ~/src/my-router
go mod init example.com/my-router && go mod tidy
go run .

Each starter’s README lists its customization points and the design docs that govern the shape. Smoke tests come along for free (go test ./...).

When you outgrow a fork

That’s the tell you want the full runtime: wrap your graph in a workload bundle, get budgets, durable HITL, the sessions surface, and metrics from the daemon — or embed via the library. More shapes (supervisor+workers, sequential pipeline, map-reduce, adversarial verifier, autonomous loop) land with the reference-graph library — see the roadmap.