Swift GUI verifiable closed-loop (agent skill)

When to use

Use this skill when an agent (Codex CLI, Claude Code, Xcode agent) is implementing or refactoring SwiftUI GUI code on:

• macOS (SwiftUI/AppKit)
• iOS (SwiftUI/UIKit, typically via Simulator)

and you need a machine-verifiable iteration loop:

1) change code 2) run deterministic checks 3) capture immutable evidence 4) decide next step strictly from evidence

This skill prioritizes deterministic CLI artifacts over “eyeballing” GUI outcomes.

Versioning note:

• Apple moved to year-based OS version numbering for the “26” generation. This skill intentionally calls out both “pre-26” versions (macOS 15, iOS 18) and “26” versions (macOS 26, iOS 26) to avoid ambiguity in mixed environments.

Core idea (high-level)

A reliable GUI loop is typically hybrid:

• Deterministic core (fast inner loop): pure logic/state tests (Swift Testing / XCTest), reducer/view-model tests, dependency-injected integration tests.
• Deterministic UI evidence (mid loop): snapshot tests (especially text/hierarchy strategies).
• Small GUI smoke (outer loop): minimal XCUITest flows + accessibility audits + rich attachments.
• Immutable evidence store: always keep the .xcresult bundle and derive summaries/attachments from it.

Inputs you must collect (one-time per project)

• Workspace or project: App.xcworkspace or App.xcodeproj (optional when scripts/ui/ui_loop.sh can auto-discover one at the package root)
• Scheme: App
• Test plan (recommended): Smoke (a .xctestplan attached to the scheme)
• Destination:
• macOS: platform=macOS (optionally include arch=arm64 or arch=x86_64)
• iOS Simulator: platform=iOS Simulator,name=<device>,OS=<version> (prefer a simulator UDID when you need strict repeatability)
• Package root (optional): pass --package-root <dir> when the script should search somewhere other than the repo root for a package/Xcode container
• Optional: derived data directory for repeatable runs

Keep these as constants in your project docs (e.g., AGENTS.md) so agents never guess.

If the target project already has an older copied scripts/ui/ui_loop.sh, refresh the managed scripts before using newer options:

/path/to/swift-gui-verifiable-loop/scripts/project/update_ui_loop_tools.sh \
  --apply \
  --platform macos \
  /path/to/target-repo

Use --platform ios for iOS-only apps and --platform both for shared macOS/iOS repos. This updater is intentionally limited to the skill-owned scripts; update project docs and test launch harnesses separately.

Step-by-step closed-loop workflow

Step 0 — Make the UI verifiable by construction

Do this once, then keep enforcing it.

1. Push behavior out of views

• Views render state; they do not own business logic.
• Use MVVM (@Observable view-models) or a reducer architecture (e.g., TCA).

1. Add deterministic entry points

• Launch args / env vars / custom URL schemes should let tests jump into a state directly.

1. Treat Accessibility as an automation contract

• Every actionable control gets a stable identifier.

1. Keep GUI smoke tests small and semantic

• Prefer “prove one contract” tests over long pseudo-human scripts.

(Details + templates: see references/REFERENCE.md and assets/templates/.)

Step 1 — Run a full deterministic verification pass (baseline)

Use the orchestrator script (recommended).

macOS example:

scripts/ui/ui_loop.sh \
  --scheme App \
  --test-plan Smoke

Agent-safe macOS UI example (preferred when full-screen XCTest screenshots are a concern):

scripts/ui/ui_loop.sh \
  --workspace App.xcworkspace \
  --scheme App \
  --test-plan Smoke \
  --destination 'platform=macOS' \
  --reuse-build \
  --system-attachment-lifetime keepNever \
  --sanitize-screenshots keep \
  --delete-raw-attachments

iOS Simulator example:

scripts/ui/ui_loop.sh \
  --scheme App \
  --test-plan Smoke \
  --destination 'platform=iOS Simulator,name=iPhone 16,OS=18.0'

(For iOS 26 environments, use OS=26.0 in the destination.)

Outputs per run:

• <artifacts-dir>/<run-id>/results.xcresult (immutable evidence)
• <artifacts-dir>/<run-id>/toolchain.txt (environment fingerprint)
• <artifacts-dir>/<run-id>/summary.json (machine-readable test summary)
• <artifacts-dir>/<run-id>/xcodebuild-*.log (captured xcodebuild logs unless VERBOSE=1)
• <artifacts-dir>/<run-id>/attachments/** (exported screenshots/attachments; sanitized when --sanitize-screenshots is used)
• <artifacts-dir>/<run-id>/attachments_raw/** (raw export only when screenshot sanitization is enabled and raw deletion is not requested)
• <artifacts-dir>/<run-id>/attachment_sanitization.json (sanitizer report when a transforming policy is used)
• <artifacts-dir>/<run-id>/xctestrun-attachment-policy.json (.xctestrun patch report when attachment lifetime policy is patched)
• <artifacts-dir>/<run-id>/diagnostics/** (crash logs, diagnostics)

Default artifacts dir: ./.artifacts/ui (add /.artifacts/ to your project’s .gitignore).

If you prefer manual commands, see references/xcresult-bundles.md.

Platform notes (read once)

• macOS UI tests may require Accessibility/Automation permissions for the UI test runner. See references/macos-ui-testing-permissions.md.
• For unattended agent runs, prefer a prepared disposable macOS VM when available

(for example a GhostVM xcode-ui-ready snapshot) over repeatedly running on a real host that still needs interactive approval.

• If macOS shows an "XCTest is trying to Enable UI Automation" password prompt, preserve the .xcresult, capture TCC attribution with scripts/macos/tcc_attribution_tail.sh, try the documented mitigations once, then ask the human or MDM policy owner to grant the OS permission instead of repeatedly rerunning.
• For agent-facing macOS visual evidence, prefer --reuse-build --system-attachment-lifetime keepNever --sanitize-screenshots keep --delete-raw-attachments after UI tests attach only app-window/root-element screenshots or cropped status-surface screenshots. Use redact-suspect only when privacy is more important than readable PNG evidence. See references/artifact-privacy.md.
• iOS simulator runs benefit from simulator-state and permission control via simctl. See references/ios-simulator-determinism.md.

Step 2 — Add snapshot tests for stable UI surfaces

Recommended: Point-Free SnapshotTesting (see references/snapshot-testing.md).

• macOS: prefer .fixed / .sizeThatFits layouts.
• iOS: device presets (.device(config: ...)) are fine, but only deterministic when you pin the simulator runtime + device model.

Policy:

• Only snapshot isolated view states (empty/loading/error/selected/disabled).
• Prefer text/hierarchy snapshots for stability; use image snapshots selectively.
• Snapshot updates must be explicit (“record mode”), never automatic in CI.

Templates:

• macOS: assets/templates/SnapshotTestTemplate.swift
• iOS: assets/templates/iOSSnapshotTestTemplate.swift

Step 3 — Add accessibility audits to the smoke suite

Add at least one audit per major screen family:

try app.performAccessibilityAudit()

This produces a high-signal, machine-actionable gate.

On macOS, be prepared to ignore narrowly-scoped host/framework noise with an explicit closure when the audit traverses synthetic SwiftUI container nodes or system-owned controls (for example, Touch Bar items) that are outside your app's actionable surface.

Template: assets/templates/AccessibilityAuditUITestTemplate.swift Reference: references/accessibility-audit.md

Step 4 — Keep XCUITests minimal (smoke only)

UI tests are valuable, but operationally fragile. Use them as proof-of-life flows:

• launch → first interactive screen → one key action
• document/window creation
• settings toggle persists after relaunch
• menu bar extra popover/context menu via a deterministic launch harness when direct NSStatusItem clicks are not hittable under XCUITest

Templates:

• macOS: assets/templates/XCUITestLaunchHarnessTemplate.swift (uses click() and shows menu patterns)
• macOS menu bar extras: assets/templates/MacOSMenuBarExtraUITestTemplate.swift
• iOS: assets/templates/iOSXCUITestLaunchHarnessTemplate.swift (uses tap() and includes a basic interruption monitor)

Step 5 — Always enrich failures with artifacts

In UI tests, attach:

• window/root-element screenshots on failure; use XCUIApplication.screenshot() only after confirming it is app-scoped on that platform/runner, and avoid XCUIScreen.main.screenshot() for macOS agent runs
• any relevant exported files
• optional debug JSON state dumps or accessibility-tree text dumps (debug builds only)

.xcresult already stores these; export them after each run.

Templates: assets/templates/XCUITestLaunchHarnessTemplate.swift, assets/templates/AgentSafeUITestArtifactsTemplate.swift Extraction scripts (canonical): scripts/ui/xcresult_export.sh, scripts/ui/xcresult_summary.sh

Decision rules for the agent (strict)

After each code change:

1. Run Step 1 (scripts/ui/ui_loop.sh).
2. If compilation/tests fail:

• fix failures first; do not proceed.

1. If snapshot diffs fail:

• decide whether change is intended.
• if intended: update snapshots via “record mode” (see references/snapshot-testing.md).
• if unintended: fix UI.

1. If XCUITest smoke fails:

• inspect exported attachments/diagnostics.
• reduce flakiness by improving launch harnesses and accessibility identifiers.

1. Only if all deterministic gates pass:

• optionally use Preview screenshots as a qualitative spot-check (not a proof oracle).

Common pitfalls (and how this skill mitigates them)

• GUI nondeterminism → push logic into unit tests + launch harnesses.
• Flaky UI queries → stable accessibility identifiers, concise queries, small smoke flows.
• Hard-to-interpret failures → .xcresult as evidence + exported attachments/diagnostics.
• Sensitive/full-screen screenshots → patch automatic attachment lifetimes, attach app-scoped screenshots, and sanitize exports.
• Tooling churn → record toolchain fingerprint each run; prefer structured xcresulttool subcommands.

See references/REFERENCE.md for deeper troubleshooting and patterns.

Quick file map

• Orchestrator (canonical): scripts/ui/ui_loop.sh
• Evidence extraction (canonical): scripts/ui/xcresult_summary.sh, scripts/ui/xcresult_export.sh
• Attachment privacy helpers: scripts/ui/patch_xctestrun_attachment_policy.py, scripts/ui/patch_xcscheme_attachment_policy.py, scripts/ui/xcresult_sanitize_attachments.py
• macOS permission triage: scripts/macos/tcc_attribution_tail.sh, scripts/macos/collect_tcc_identities.sh
• Existing-project updater: scripts/project/update_ui_loop_tools.sh
• Toolchain fingerprint (canonical): scripts/ui/toolchain_fingerprint.sh
• Templates: assets/templates/*.swift
• Deeper reference: references/REFERENCE.md

iOS simulator helpers:

• scripts/ios/simctl_prepare.sh
• scripts/ios/simctl_privacy.sh