• Status: Accepted (supersedes the hub graph of Architecture before #116)
• Issue: #116 (branches: engine #137, job-scout #139; spikes #113, #114, #115)

Until #116 every workflow ran one fixed graph: input guards → router ⇄ agent → finalize → output guards, with routers: { main: { maxHops } } and budget in @Workflow. Other shapes (a review cycle, several entries, two routers) could not be expressed, the graph was not visible in the code, and runaway loops ended quietly ("max hops reached", "budget exhausted") with a best-effort answer.

• The workflow file is the graph: @Workflow({ flow }) lists transitions between classes — from(A).next(B), from(Router).routes(), chain(A, B, C), node(Class, "name"), Self. Nodes are @WorkflowStart, @Router, @Agent, @WorkflowFinish classes. (Names as of today: at #116 the steps were called to / choose and the nodes @Entry / @Conclusion; a router's targets later moved into its own routes, so the flow says only routes().)
• Checked at assembly, all at once: 15 rules (graph.*, router.*) before any model call; every violation is reported in one GraphRuleError naming the classes. Relations "by class" are a startup check, never compile time (component classes are empty to the compiler).
• Run as a real LangGraph graph: one node per flow node (<kind>.<name>), conditional edges for a router's choice, workflow starts from START, workflow finishes to END. An agent node runs its loop as a compiled subgraph, so the approval interrupt() pauses the whole run and nothing before it repeats.
• Routers are nodes with routes: @Router({ prompt, model, maxVisits?, routes }) — the prompt says how to choose, each route what the choice means; texts normalised, routes sent sorted by name. No built-in finish: returning to the user is a route to a workflow finish. A failed decision fails the run.
• Limits fail, never stop quietly: settings().limits({ perRun: { steps, cost }, perDay: { cost } }) and maxVisits; steps count agent and router visits; LimitExceededError carries the key, path and spend, and the run is a failed Tern.
• Removed: the hub builder (graph.ts, routing.ts, the hub router node), routers and budget in @Workflow, AgentsConfig and profiles; @Workflow({ agents }).
• The existing guard, approval, compaction, knowledge and agent nodes run inside the new graph until #119 / #120 replace them; #100 (after a human decision the turn ends) is kept by the router node.

• A workflow's shape is code review-able and testable; describe lists it as transitions (a drawing comes with #127).
• Hitting a limit or a router failure (e.g. a Jev timeout) now fails the turn instead of answering with what is there — visible in the Tern and the CLI, at the price of an error message where the hub answered.
• Route texts matter: Jev picks by them, so "send the answer" must say when (see Routers).
• Eval keeps its own account (<workflow>:eval), now capped by limits.perDay.cost.
• Profiles cannot change the flow, routers or limits (they are code, versioned with the workflow).