Approach: Production-First Iterative Refactoring — This skill is built for production enterprise codebases where stability and reviewability matter more than speed. Architecture changes are delivered through iterative refactoring — small, focused PRs (≤200 lines, single concern) tracked in a refactoring/ directory. Critical safety issues ship first; cosmetic improvements come last.

UIKit MVVM Architecture (iOS 13+)

Enterprise-grade UIKit MVVM architecture skill. Opinionated: prescribes Combine-bound ViewModels, Coordinator navigation, constructor injection via factories, ViewState enum, DiffableDataSource, and programmatic Auto Layout. Adopts a production-first iterative refactoring approach — every pattern is chosen for testability, reviewability, and safe incremental adoption in large teams. UIKit MVVM remains the dominant production architecture for large-scale iOS apps.

Architecture Layers

View/ViewController  → UIViewController + UIView. Binds to ViewModel via Combine/closures/GCD. Renders state.
ViewModel Layer      → Zero UIKit imports. Exposes ViewState<T>. Uses @Published (Combine) or closures (GCD).
Coordinator Layer    → Manages navigation flow. Creates ViewModels + VCs. Owns UINavigationController.
Repository Layer     → Protocol-based data access. Hides data source details.
Service Layer        → URLSession, persistence. Injected via protocol. May use GCD or async/await.

Quick Decision Trees

"Should this ViewController have a ViewModel?"

Is there business logic, networking, or complex state?
├── YES → Create a ViewModel (see "Which binding mechanism" below for Combine vs closures)
└── NO → Is it a container/flow controller (tab bar, navigation)?
    ├── YES → Coordinator manages it, no ViewModel needed
    └── NO → No ViewModel needed unless it simplifies testing

"Which binding mechanism should I use?"

Is this a legacy codebase with heavy GCD / completion handlers?
├── YES → Is the goal to extract ViewModels first (keep GCD)?
│   ├── YES → Closures / Bindable<T> + GCD in service layer
│   │   └── Upgrade to Combine later (see migration-patterns.md)
│   └── NO → Adopt Combine during extraction
│       └── @Published + sink (production standard)
└── NO → What is the minimum iOS target?
    ├── iOS 13+ → Combine: @Published + sink
    │   └── One-shot events? → PassthroughSubject
    └── < iOS 13 → Closures / Bindable<T> wrapper

"Where do dependencies come from?"

ViewModel always receives dependencies via constructor:
  init(service: NetworkServiceProtocol)

Who creates the ViewModel?
├── Coordinator (recommended)
│   └── Coordinator owns factory, creates VM with deps, passes to VC init
├── VC creates it (simpler apps)
│   └── VC receives deps via init, passes to VM init
└── DI Container (large apps, 20+ screens)
    └── Container resolves protocols, coordinator pulls from container

"How should I handle navigation?"

Is there more than one navigation flow (auth + main, tabs)?
├── YES → Coordinator pattern with parent/child hierarchy
│   └── AppCoordinator → AuthCoordinator / MainCoordinator → TabCoordinator
└── NO → Single Coordinator wrapping UINavigationController
    └── ViewModel signals navigation via closures, never UIKit imports

Workflows

Default workflow: Analyze & Refactor (below). New screen creation applies the same patterns but from a clean slate. In production enterprise codebases, most work is iterative modernization — not greenfield.

Workflow: Analyze & Refactor Existing MVC Codebase

When: First encounter with a legacy UIKit MVC codebase — the most common enterprise scenario.

1. Scan for anti-patterns using the detection checklist (references/anti-patterns.md)
2. Create refactoring/ directory with per-feature plan files (references/refactoring-workflow.md)
3. Write each issue with full description (Location, Severity, Problem, Fix) — titles alone get forgotten
4. Categorize issues by severity: 🔴 Critical → 🟡 High → 🟢 Medium
5. Plan Phase 1 PR: fix critical safety issues only (≤200 lines per PR)
6. Execute one PR at a time. New findings go to refactoring/discovered.md with full descriptions, NOT into current PR
7. After completing each fix: mark the task - [x] and update the Progress table
8. Proceed through phases: Critical → ViewModel extraction → Coordinator → Combine bindings → DI

Workflow: Create a New Screen

When: Building a new feature screen from scratch. Apply enterprise patterns from the start.

1. Define the data model and repository protocol (references/testing.md for mock pattern)
2. Create ViewModel: plain class with private(set) var state: ViewState<T> — zero UIKit imports (use @Published + Combine or closures per the binding mechanism decision tree)
3. Add // MARK: - sections: Properties, Init, Actions, Computed Properties
4. Create the ViewController with constructor injection: init(viewModel: MyViewModel)
5. Wire Combine bindings in setupBindings() called from viewDidLoad
6. Build UI programmatically with Auto Layout (references/layout-approaches.md)
7. Add Coordinator route and factory method (references/coordinator-navigation.md)
8. Create test file with mock repository (references/testing.md)

Workflow: Extract ViewModel from Massive ViewController

When: Refactoring an existing MVC screen to MVVM incrementally.

1. Identify all state properties in the VC (data, loading flags, error state)
2. Create ViewModel class — move state properties to @Published private(set) var
3. Move business logic methods from VC to VM (networking, validation, formatting)
4. Add Combine @Published to VM state, Set<AnyCancellable> to VC
5. Replace direct state access in VC with viewModel.$property.sink bindings
6. Remove all import UIKit from ViewModel — compiler will flag violations
7. Write tests for the ViewModel (now possible since VM has no UIKit dependency)
8. Verify VC only does: bind, render, forward user actions to VM

Workflow: Introduce Coordinators to Existing MVVM

When: Navigation is scattered across ViewControllers. Adding Coordinators incrementally.

1. Start with one flow (e.g., auth flow or a tab's navigation stack)
2. Create Coordinator protocol and AppCoordinator (references/coordinator-navigation.md)
3. Move VC creation from VCs/Storyboard segues to Coordinator's factory methods
4. Replace performSegue / pushViewController calls with ViewModel navigation closures
5. Wire Coordinator as UINavigationControllerDelegate for back-button cleanup
6. Add child coordinator lifecycle management (didFinish delegate pattern)
7. Test Coordinator with mock UINavigationController (references/testing.md)

Code Generation Rules

<critical_rules> Whether generating new code or refactoring existing code, every output must be production-ready and PR-shippable — small, focused, and testable. ALWAYS:

1. ViewModels import only Foundation and Combine — never UIKit
2. Use @Published private(set) var for state properties modified only by the ViewModel
3. Use ViewState<T> enum for async data — never separate boolean flags
4. Inject dependencies via constructor with protocol types
5. Bind in viewDidLoad using setupBindings() — store in Set<AnyCancellable>
6. Always [weak self] in sink closures when stored in cancellables
7. Always .receive(on: DispatchQueue.main) before UI updates in sink
8. Add // MARK: - sections: Properties, Init, Lifecycle, Bindings, Actions
9. Use programmatic Auto Layout — translatesAutoresizingMaskIntoConstraints = false
10. Keep every generated file ≤ 400 lines. Extract child VCs or ViewModel extensions when approaching that limit.
11. Before modifying a ViewController, output a brief <thought> analyzing its current dependencies and retain cycles. </critical_rules>

When generating tests, ALWAYS:

1. Use protocol mocks with var stubbed* and var *CallCount tracking
2. Test through public interface, never test private methods
3. Use XCTestExpectation + sink + dropFirst() for Combine publisher tests
4. Use await fulfillment(of:) for async tests — NEVER wait(for:) in async contexts. When you write async test code in a response, the actual code example must call await fulfillment(of: [expectation], timeout: 2.0), not wait(for: [expectation], timeout: 2.0). Mentioning the modern API only in a note while the code uses wait(for:) still hangs the test — it's the classic cooperative-pool deadlock. In XCTestCase this is a synchronous blocking call that cannot pump the queue, so the expectation never fulfills:

// ❌ WRONG — deadlocks in async test function
func test_load_updates_items() async {
    let exp = expectation(description: "items updated")
    let cancellable = sut.$items.dropFirst().sink { _ in exp.fulfill() }
    await sut.load()
    wait(for: [exp], timeout: 2.0)  // cooperative-pool deadlock
    _ = cancellable
}

// ✅ CORRECT — releases the cooperative thread via await
func test_load_updates_items() async {
    let exp = expectation(description: "items updated")
    let cancellable = sut.$items.dropFirst().sink { _ in exp.fulfill() }
    await sut.load()
    await fulfillment(of: [exp], timeout: 2.0)
    _ = cancellable
}

5. Include memory leak detection with addTeardownBlock { [weak sut] in XCTAssertNil(sut) }

Fallback Strategies & Loop Breakers

<fallback_strategies> When refactoring legacy code, you may encounter stubborn Swift compiler errors. If you fail to fix the same error twice, break the loop:

1. Combine Type Erasure: If you get generic type mismatch errors with Combine AnyPublisher, append .eraseToAnyPublisher() to the pipeline or fall back to closures instead of fighting the type system. Why: Combine's generic types nest deeply (e.g., Publishers.Map<Publishers.Filter<...>, Output>), and type erasure is the standard solution.
2. DiffableDataSource Generics: If the compiler complains about Hashable conformance or type differences in NSDiffableDataSourceSnapshot, verify your CellViewModel uses a unique UUID instead of complex nested generic models. Why: DiffableDataSource compares items by hash — nested models with mutable state produce inconsistent hashes and silent data corruption.
3. Revert and Restart: If a ViewController refactor spirals into 50+ compiler errors, stop. Propose reverting the changes and breaking the problem into two smaller phases (e.g., extract networking first, then migrate state). Why: UIKit's implicit dependencies cascade — changing one property often breaks unrelated outlets, delegates, and data sources. </fallback_strategies>

Confidence Checks

Before finalizing generated or refactored code, verify ALL:

□ No duplicate functionality — searched codebase for existing implementations
□ Architecture adherence — follows patterns already established in the project
□ Naming conventions — matches existing project naming style
□ Import check — ViewModel imports only Foundation + Combine, NOT UIKit
□ ViewState — used for all async data, no separate isLoading/error booleans
□ Combine bindings — [weak self] in every sink, .receive(on: .main) before UI updates
□ DI — dependencies injected via protocol, not accessed via singletons
□ Coordinator — navigation handled by Coordinator, not by VC pushing other VCs
□ Memory management — deinit cancels Tasks, no retain cycles in closures
□ Tests — corresponding test file exists or is created alongside
□ PR scope — changes within defined scope, new findings go to `refactoring/discovered.md`
□ File size — new files ≤ 400 lines; existing oversized files have a split task logged

Companion Skills

Before generating async ViewModel or migrating completion handlers: determine the project's concurrency approach. If unclear from context, ask the user.

If unclear, ask: "Is the team migrating to Swift Concurrency or keeping GCD/OperationQueue?"

References