Checkpoint what is already done

A checkpoint is a small Codable record. It names the workflow, the current step, the inputs you still need, and the outputs you already have. Write it after every successful step, before you start the next model call.

import Foundation
import FoundationModels

enum WorkflowStep: String, Codable {
    case plan
    case gather
    case decide
    case act
    case done
}

struct WorkflowCheckpoint: Codable, Sendable {
    var id: UUID
    var step: WorkflowStep
    var userPrompt: String
    var plan: String?
    var gathered: String?
    var decision: String?
    var transcriptData: Data?
    var updatedAt: Date
}

actor CheckpointStore {
    private let url: URL

    init(url: URL) {
        self.url = url
    }

    func save(_ checkpoint: WorkflowCheckpoint) throws {
        let data = try JSONEncoder().encode(checkpoint)
        try data.write(to: url, options: .atomic)
    }

    func load() throws -> WorkflowCheckpoint? {
        guard FileManager.default.fileExists(atPath: url.path) else { return nil }
        let data = try Data(contentsOf: url)
        return try JSONDecoder().decode(WorkflowCheckpoint.self, from: data)
    }
}

Save the file in the app container. Use an atomic write so a kill in the middle of a save does not leave half a JSON. Do not put secrets in the checkpoint if the file is later backed up. Redact first, as in Stop prompt attacks and leaks.

Store a serialized transcript next to the step fields when the next step still needs that conversation. The exact encode API for Transcript can vary by SDK. Confirm it in the Xcode 27 docs. If you cannot encode the transcript yet, store your own message list and rebuild instructions on resume.

Read Make several agents agree first, Talk to tools with MCP, and Stop prompt attacks and leaks if a saved step is a vote tally or a tool result.

Next, open that file and continue instead of starting over.