ComposablePipelines: Skill for Claude Code

Composable Pipelines

A runtime-agnostic stack for AI pipelines in Swift: a declarative DSL lowers to a Codable AST, a compiler emits an execution graph, and an observable walker runs it — while an Executor you supply decides what "run a model" means.

Quick start

import ComposablePipelines

struct Summary: Pipeline {
    typealias Output = String
    let document: String

    @State var keyPoints = ""
    @State var summary = ""

    var body: some Pipeline {
        $keyPoints.set {                                    // step 1 → slot
            Model<String>().systemPrompt("Extract the 5 key points.").message(document)
        }
        $summary.set {                                      // step 2 reads step 1
            Model<String>().systemPrompt("Summarize from these points.").input { $keyPoints.get() }
        }
        $summary.get()                                      // pipeline output
    }
}

// Run it (linear flow): compile → walk.
let graph  = PipelineCompiler().compile(pipeline.loweredGraph())
let result = try await PipelineWalker(executor: MyExecutor()).run(graph: graph) { event in print(event) }
let text   = try JSONDecoder().decode(String.self, from: result)   // ExecutionValue == Data (JSON)

Authoring cheatsheet

• Pipeline: struct X: Pipeline { typealias Output = T; @State var … ; var body: some Pipeline { … } }. The graph is inferred from the @State slots each step reads/writes — never wired by hand.
• State: @State var s = "" → write with $s.set { <pipeline> } or $s.set(value); read with $s.get(). End body with the output slot's .get().
• Model (output-only generic): Model<Output>() then chain .systemPrompt(_), .input { $slot.get() } or .message(staticValue), and optionally .tools([…]), .temperature(_), .maxTokens(_), .requirements(ModelSelectionRequirements(traits: […])).
• Primitives: Guardrail(input, rules:allowed:blocked:), While(condition:) { … }, Group { … }, ForEach(in: xs) { x in … }, Summarize(text: $slot, maxTokens:), ClientTask(input: $slot) { value in … }, From(provider, query: $slot), Self.return(value).
• Control flow: use native if / switch on produced state (e.g. if verdict == "trivial").