Router Flow Workflows
Status: Implemented · Created: 2026-06-30
Problem
Multi-step model workflows often require a separate orchestration service and a client-specific API. That makes model pools, routing policy, traces, and failure behavior harder to manage alongside ordinary inference.
Router Flow exposes bounded orchestration through the same OpenAI-compatible gateway and recipe policy used for other routes.
Implemented design
Router Flow has three public concepts:
| Concept | Surface | Purpose |
|---|---|---|
| Flow model | vllm-sr/flow | An ordinary entrypoint mapped to a recipe containing workflow decisions. |
| Workflow algorithm | algorithm.type: workflows | The decision-local orchestration policy. |
| Worker pool | modelRefs | The only models workflow steps may invoke. |
The public name does not implicitly filter decisions by algorithm. Declare its entrypoint-to-recipe mapping and the recipe's rules explicitly; a different public name can expose the same workflow recipe.
The planner creates a plan; worker models execute task steps. A planner cannot add a
model that is absent from modelRefs.
Static workflows
Static mode uses an operator-authored role plan. It is suitable when steps, models, dependencies, and synthesis behavior should be deterministic and reviewable.
Independent steps may run concurrently when the plan declares no dependency between them. Dependent steps receive only the bounded outputs declared by the workflow.
Dynamic workflows
Dynamic mode asks a configured planner model for a structured plan. The router parses and validates that plan before any worker call:
- step identifiers must be unique;
- dependencies must form an acyclic graph;
- roles and models must be allowed by the decision;
- tool loops, steps, calls, and output sizes are bounded; and
- invalid plans fail through the configured error policy.
Planner output is untrusted data, not executable configuration.
Function-calling boundary
Each workflow step has its own message and tool-call state. Tool results return only to the step that requested them. The router enforces per-step and workflow-wide limits so a planner cannot create an unbounded agent loop.
Tools still require authorization outside semantic relevance. A workflow may select a tool only when both the route and caller are permitted to use it.
Failure and observability
The workflow trace should distinguish planning, validation, worker execution, tool calls, and final synthesis. Partial worker failure follows the decision's policy; it must not silently disappear from the trace.
Timeouts apply to individual calls and to the complete workflow. Cancellation propagates to outstanding worker calls.
Scope and non-goals
The implemented boundary supports static and dynamic workflows with bounded workers inside the gateway. It does not:
- train a coordinator;
- provide a general-purpose agent framework;
- allow arbitrary code execution;
- widen the decision's worker pool;
- promise that a workflow outperforms its best worker; or
- hide workflow cost behind a single-model response.
Evaluation
Compare Flow with the best single worker and other looper algorithms on the same task set. Report task success, judge criteria where necessary, latency, tokens, upstream calls, plan-validation failures, tool-loop failures, and trace completeness.
Open questions
- Which workloads justify dynamic planning over static plans?
- When should final synthesis use the planner versus a worker?
- Which trace fields are safe for user-facing responses?
- What evidence would justify training or distilling a coordinator?