Journey Builder

Build one user journey at a time. Each journey is a realistic path through the app tested like a real user — with screenshots at every step. Each journey MUST cover specific spec requirements and verify ALL of their acceptance criteria with real implementations and real outcomes. Map each journey to its spec requirements in journey.md. A journey is complete when every mapped acceptance criterion has: (1) a real implementation (no placeholders/simulations), (2) a test step that exercises it, and (3) a screenshot proving it works.

Depth-chain principle: A journey is a chain of actions where each step produces an outcome that the next step consumes or verifies. Example: setup → create a recording → verify it appears in library → play it back → check transcript syncs → search for it → delete it → verify it's gone. Every step must exercise something NEW. If you've already clicked a button and verified it works, do not click it again.

Prerequisites

spec.md in project root
journeys/ folder (created if missing)
Testable app (XCUITest for macOS, Playwright for web)

Step 0: Load Pitfalls (MANDATORY — do this FIRST)

Before ANY other work, fetch and read ALL pitfall files from the shared pitfalls gist:

# List all files in the pitfalls gist
gh gist view 84a5c108d5742c850704a5088a3f4cbf --files

Then read EVERY file:

gh gist view 84a5c108d5742c850704a5088a3f4cbf -f <filename>

Read each file completely. These contain hard-won solutions to blockers that WILL recur. Apply every relevant pitfall to your work in this session. Do NOT skip this step.

Adding New Pitfalls

When you encounter a blocker and find the solution, create a new pitfall file in the gist:

gh gist edit 84a5c108d5742c850704a5088a3f4cbf -a <category>-<short-name>.md <<'EOF'
# <Title>

## Problem
<What went wrong — exact error message if possible>

## Root Cause
<Why it happened>

## Solution
<Exact fix — code, config, or command>

## Prevention
<How to avoid this in future journeys>
EOF

Name files by category: xcodegen-*.md, xcuitest-*.md, codesign-*.md, swiftui-*.md, etc.

Step 0.5: Copy Template Files (macOS only)

Check if the UI test target has the shared helper files. If missing, copy them from the skill templates:

The {skill-base-dir} placeholder refers to the plugin root (plugins/autocraft).

TEMPLATES="{skill-base-dir}/templates"
UI_TEST_DIR=$(find . -name "*UITests" -type d -maxdepth 1 | head -1)

# Copy JourneyTestCase base class (snap helper with dedup + window launch fix)
if [ -n "$UI_TEST_DIR" ] && [ ! -f "$UI_TEST_DIR/JourneyTestCase.swift" ]; then
  cp "$TEMPLATES/JourneyTestCase.swift" "$UI_TEST_DIR/"
fi

After copying, run xcodegen generate so Xcode picks up the new files.

Also ensure project.yml has these settings on the UI test target (prevents sandbox blocking file writes and missing windows):

  MyAppUITests:
    type: bundle.ui-testing
    settings:
      base:
        BUNDLE_LOADER: ""
        TEST_HOST: ""
        ENABLE_APP_SANDBOX: "NO"
    entitlements:
      path: MyAppUITests/MyAppUITests.entitlements
      properties:
        com.apple.security.app-sandbox: false

Step 1: Check Journey State

Read journey-state.md in the project root (create if missing). This file tracks which journeys are complete:

# Journey State

| Journey | Status | Test Duration | Last Updated |
|---------|--------|---------------|--------------|
| 001-first-launch-setup | polished | 12m30s | 2026-03-28 |
| 002-settings-model | in-progress | 3m15s | 2026-03-28 |

Decision logic:

Find the first journey where status is in-progress or needs-extension — work on that one
Only if ALL existing journeys are polished with all acceptance criteria covered, pick the next uncovered path from the spec
A journey is polished ONLY when: all tests pass, all 3 polish rounds done, AND every acceptance criterion from every requirement listed in the journey's ## Spec Coverage section is covered — meaning a real implementation exists (no placeholders, no simulations), a test step exercises it, and a screenshot captures the outcome. The criterion count in the journey's Spec Coverage must match the count in spec.md. A journey is NOT polished if any criterion from any mapped requirement lacks a screenshot.

Step 2: Read spec + existing journeys

Read spec.md. For every requirement, list ALL its acceptance criteria — do not skim. Read every journeys/*/journey.md. For each journey, note which acceptance criteria it has mapped AND whether each criterion has screenshot evidence (a screenshot whose step matches the criterion). You now have two sets:

Fully implemented criteria: appearing in a journey's ## Spec Coverage section AND having a corresponding screenshot
Uncovered criteria: not in any journey's Spec Coverage, OR in a journey but lacking screenshot evidence

This two-set distinction is your working ground truth for the rest of this run.

Step 3: Pick or extend a journey

If extending an existing journey (status is in-progress or needs-extension):

Read the existing journey.md and test file
Run the test and measure duration
Check which acceptance criteria from the mapped spec requirements are NOT yet covered. For each uncovered criterion: implement the real feature if missing, add a test step that exercises it, and take a screenshot proving it works. A journey is not done until ALL mapped acceptance criteria are covered.
Update journey.md with the new steps

If creating a new journey:

Find the longest uncovered user path
Create numbered folder: journeys/{NNN}-{name}/
Write journey.md as a depth-chain: each step produces output the next step uses
Spec mapping (MANDATORY — no cherry-picking): At the top of journey.md, list which spec requirements this journey covers. For each mapped requirement, you MUST list ALL of its acceptance criteria — not a subset. Count the criteria in spec.md for that requirement and list every one by number.

CORRECT (all criteria listed for each requirement):
```
## Spec Coverage
- P0-0: First Launch Setup — criteria 1, 2, 3, 4, 5, 6 (all 6)
- P0-2: Window Picker — criteria 1, 2, 3, 4, 5 (all 5)
- P0-3: Screen + Audio Recording — criteria 1, 2, 3, 4, 5, 6, 7 (all 7)
```
WRONG — DO NOT DO THIS (omits criteria):
```
- P0-2: Window Picker (criteria 1, 2, 5)   ← FORBIDDEN: criteria 3 and 4 silently dropped
```
If a criterion requires data only available from a prior journey, defer it explicitly:
```
- P0-2: Window Picker — criteria 1, 2, 3 (this journey); criteria 4, 5 → journey 005 (requires recording created in journey 003)
```
Each deferred criterion MUST appear in exactly one future journey's Spec Coverage. Every criterion from every mapped requirement must be owned by exactly one journey.

Every criterion listed MUST be implemented and tested by the end of the journey.
Include: complete workflow (create → use → modify → verify → clean up), edge cases, error recovery, data persistence checks

Anti-repetition rule (HARD): Before finalizing, scan the test for repeated interactions. If the same element is clicked more than twice, or the same navigation path is traversed more than once, it is padding — remove it. Coverage must be achieved through feature depth (more acceptance criteria verified), NEVER through repeating interactions already performed. Clicking through 5 model cards once is testing; clicking through them 3 times is waste. Downloading multiple models exercises the same code path — one download verifies the download flow.

Step 4: Write the test

One test file. Act like a real user. Screenshot after every meaningful step via XCTAttachment (macOS) or Playwright locator screenshot (web). Name: {journey}-{NNN}-{step}.png. The extract script adds elapsed-time prefixes (T00m05s_) automatically — you do NOT add timestamps in code.

Snap helper with built-in timing measurement (MANDATORY)

Every journey test MUST use a snap() helper that measures the gap since the last screenshot and writes it to a timing file in real-time. This is the enforcer — no gap > 5s goes unnoticed.

macOS — use JourneyTestCase base class (preferred):

If JourneyTestCase.swift was copied in Step 0.5, subclass it:

final class MyJourneyTests: JourneyTestCase {
    override var journeyName: String { "001-first-launch-setup" }

    override func setUpWithError() throws {
        app.launchArguments = ["-hasCompletedSetup", "NO"]
        try super.setUpWithError()  // clears timing, creates dirs, launches app, ensures window
    }

    func test_MyJourney() throws {
        let icon = app.images["myIcon"]
        XCTAssertTrue(icon.waitForExistence(timeout: 10))
        snap("001-initial", slowOK: "app launch")
        // ...
    }
}

JourneyTestCase provides:

snap(_:slowOK:) — screenshot + timing + disk write + dedup (skips if identical to previous)
setUpWithError() — clears timing, creates dirs, launches app, opens window if needed
tearDownWithError() — terminates app

Use .exists instead of waitForExistence (CRITICAL for speed)

waitForExistence(timeout: N) polls the accessibility tree every ~1 second. When an element doesn't exist, the full timeout is burned. This is the #1 cause of slow journey tests.

Rule: one waitForExistence per view transition, .exists for everything else.

// SLOW — 238s test (original)
XCTAssertTrue(title.waitForExistence(timeout: 5))     // 5s timeout, polls
XCTAssertTrue(button.waitForExistence(timeout: 5))     // 5s timeout, polls
if optional.waitForExistence(timeout: 3) { ... }       // 3s burned if missing

// FAST — 61s test (3.9x faster)
XCTAssertTrue(title.waitForExistence(timeout: 10))     // wait ONCE for view to load
XCTAssertTrue(button.exists)                           // instant (~50ms)
if optional.exists { ... }                             // instant, no timeout burn

Pattern per phase:

After a view transition (navigation, button click that changes screens), use waitForExistence() on ONE element to confirm the new view loaded
For all other elements in that same view, use .exists (synchronous, ~50ms, no polling)
Use live element references for clicks — they need current coordinates
Repeat after the next navigation

Example:

// Phase 1 — Consent Screen
let consentIcon = app.images["consentIcon"]
XCTAssertTrue(consentIcon.waitForExistence(timeout: 10))  // wait for view
snap("001-consent-initial")

// Everything else is already rendered — .exists is instant
XCTAssertTrue(app.staticTexts["Recording Consent"].exists)
snap("002-consent-title")
XCTAssertTrue(app.buttons["acceptConsentButton"].exists)
snap("003-accept-button")

// Click transitions to next view
app.buttons["acceptConsentButton"].click()
snap("004-accepted")

// Phase 2 — new view, wait once again
let downloadButton = app.buttons["downloadButton"]
XCTAssertTrue(downloadButton.waitForExistence(timeout: 8))  // wait for new view
snap("005-model-selection")
if app.staticTexts["Choose Whisper Model"].exists { snap("006-title") }  // instant

Web (Playwright) — equivalent pattern:

let lastSnapTime = 0;
let snapIndex = 0;

async function snap(page: Page, name: string, journeyDir: string) {
  snapIndex++;
  const now = Date.now();
  const gap = lastSnapTime ? (now - lastSnapTime) / 1000 : 0;
  lastSnapTime = now;
  const status = gap > 5 ? 'SLOW' : 'ok';

  // 1. Write screenshot to disk
  const dir = `${journeyDir}/screenshots`;
  fs.mkdirSync(dir, { recursive: true });
  await page.locator('#app').screenshot({ path: `${dir}/${name}.png` });

  // 2. Append timing to JSONL
  const line = JSON.stringify({ index: snapIndex, name, gap_seconds: +gap.toFixed(1), status });
  fs.appendFileSync(`${journeyDir}/screenshot-timing.jsonl`, line + '\n');
}

5-second gap rule

Every gap between consecutive screenshots MUST be <= 5 seconds. The snap() helper writes each gap to screenshot-timing.jsonl in real-time. The journey-loop's timing watcher monitors this file and will kill the test if a SLOW entry is detected.

If the watcher kills your test, you will be restarted after the orchestrator investigates and fixes the slow gap. To avoid being killed:

Keep all waitForExistence timeouts <= 5s unless the operation genuinely requires longer
For unavoidable long waits (async downloads, app launch), pass slowOK: to the snap call: snap("042-download-done", slowOK: "model download requires async completion"). The watcher ignores SLOW-OK entries.

Add intermediate screenshots inside long wait loops so no single gap exceeds 3s:

// Break a 30s download wait into 3s chunks with progress screenshots
for i in 0..<10 {
    if doneButton.waitForExistence(timeout: 3) { break }
    snap("042-download-progress-\(i)")
}
snap("043-download-done")

Step 5: Run the test and enforce timing

Run only this test. Fix failures. Measure wall-clock time. Extract screenshots:

rm -rf /tmp/test-results.xcresult
time xcodebuild test \
  -project {Project}.xcodeproj \
  -scheme {UITestScheme} \
  -destination 'platform=macOS' \
  -derivedDataPath build \
  -only-testing:{UITestTarget}/{TestClassName} \
  -resultBundlePath /tmp/test-results.xcresult \
  -quiet 2>&1

Acceptance criteria check: After the test run, check which mapped acceptance criteria are NOT yet covered. If any are missing, go back to Step 3 and implement them. Do NOT proceed to polish until all mapped criteria have real implementations.

Timing: The journey-loop's watcher enforces the 5-second gap rule by monitoring screenshot-timing.jsonl in real-time. If your test is killed by the watcher, you'll be restarted after the gap is fixed. See the 5-second gap rule in Step 4.

Step 5.5: Critical Evidence Review (MANDATORY)

A passing test is not proof of correctness. After every test run, verify that the test actually exercised what it claims.

5.5a. Read all screenshots in order. Look at every screenshot in journeys/{NNN}-{name}/screenshots/ sequentially. Narrate the story they tell: "Consent screen → model selection → downloading → permissions → library → ..."

5.5b. Compare the visual story against journey.md. For each phase in the journey, there must be screenshot evidence that it happened. If a phase has no screenshots (e.g., no recording screenshots, no playback-with-video screenshots), that phase was silently skipped.

5.5c. For every missing phase, investigate WHY. Read the test code for that section. Find which condition caused it to skip:

An if element.exists guard that evaluated to false? → Read the app's SwiftUI code to find why the element wasn't found. Common causes: wrong element type (View with onTapGesture instead of Button), async content not loaded yet (need waitForExistence on child, not just container), missing accessibility identifier.
A waitForExistence that timed out? → Read the app code to find why the element never appeared. Is the view not being presented? Is the data not loading?
An assertion that passed vacuously? → The test asserts something exists, but the surrounding code never runs because an earlier guard skipped it.

5.5d. Fix the root cause. Don't add more if guards or workarounds. Fix the app code or test code so the phase actually executes. Then re-run the test and repeat this review.

Do NOT proceed to Step 6 (Polish) until every phase in the journey has screenshot evidence.

Step 6: Polish loop (3 rounds)

Run this loop 3 times (round 1, 2, 3). Each round does all three phases in order. Each phase produces a NEW timestamped file — never overwrite previous rounds.

Phase A: Testability Review

Review app code touched by this journey. Check:

ViewModels accept protocol dependencies via initializer injection
Use Cases are framework-free (no UIKit/SwiftUI/AppKit imports)
Side effects (network, disk, permissions) behind protocols
DependencyContainer uses real implementations; unit test factory uses protocol-based test doubles (NOT used in the running app)
No shared mutable singletons
State transitions are testable (given/when/then)

Fix issues: extract protocols, add DI, move logic out of Views. Write fast unit tests (< 1s total). Run them.

Write journeys/{NNN}-{name}/testability_review_round{N}_{YYYY-MM-DD}_{HHMMSS}.md.

Phase B: Refactor UI test

Clean up the journey test code:

Readable steps with helpers/page objects where needed
NEVER use sleep() or fixed-time waits to simulate user reading/thinking time. Tests MUST complete as fast as possible. Always wait for elements (waitForExistence(timeout:)) or conditions instead. The only acceptable fixed wait is when no element/condition can be checked (e.g., animation with no completion signal), and even then keep it under 1 second.
Accessibility identifiers on every interactive element

Phase C: UI Review — Design-Driven

Read every screenshot in journeys/{NNN}-{name}/screenshots/. Review with both platform convention AND design quality lenses.

Platform Conventions:

macOS: HIG compliance — toolbar, sidebar, window chrome, system colors
Web: Responsive layout, standard navigation patterns, accessibility

Design Quality: Apply the frontend-design skill principles when reviewing. Evaluate typography, color, spatial composition, visual details, motion, and overall polish level. Flag anti-patterns like generic fonts, flat backgrounds, missing hover states, and unstyled empty states.

Fix all issues in the view code. Re-run the test. Extract fresh screenshots to screenshots/.

Write journeys/{NNN}-{name}/ui_review_round{N}_{YYYY-MM-DD}_{HHMMSS}.md.

Round 3 is the final gate — every screenshot must Pass or have issues explicitly deferred with justification.

Step 7: Final verification + acceptance criteria audit

Run unit tests + this journey's UI test one last time. Both must pass.

Run the acceptance criteria audit: for each criterion mapped to this journey in journey.md, verify (1) the production code implements it for real (grep for placeholder/simulated/fake), (2) the test exercises it, (3) a screenshot captures the result. List any gaps and fix them before proceeding.

Step 8: Update journey state

Update journey-state.md:

Set status to polished ONLY if: (1) all tests pass, (2) all 3 polish rounds are done, AND (3) for every requirement in the journey's ## Spec Coverage section, EVERY one of that requirement's acceptance criteria (as they appear in spec.md) has: a real implementation (grep confirms no placeholder/simulated/fake), a test step that exercises it, and a screenshot that captures the outcome. Count the criteria in spec.md for each mapped requirement — the count must match what is listed in the journey.
Set status to needs-extension if ANY criterion from ANY mapped requirement is missing an implementation, test step, or screenshot — including criteria listed in spec.md but absent from the journey's ## Spec Coverage section.
Record the ACTUAL measured wall-clock time from the xcodebuild test run (for reference, not as a gate)
Record the current date

Step 9: Commit

New commit (never amend). Include: journey.md, all review files, all screenshots, updated journey-state.md. Message summarizes journey, fixes, features covered.

Step 10: Report

Tell the user: which journey, how many steps, test duration, features covered, issues fixed across rounds, unit tests added, and what journey to work on next.

If any blockers were solved during this run, confirm that new pitfall files were added to the gist.

Rules

Load pitfalls first — Step 0 is not optional. Every session starts by reading the gist.
Add pitfalls for every blocker — When you find a solution to a non-obvious problem, add it to the gist immediately via gh gist edit.
No sleep waits — NEVER use sleep(), Thread.sleep(), or fixed-time waits. Tests must complete as fast as possible.
Use .exists, not waitForExistence — Use waitForExistence() ONLY once per view transition. For all other element checks in the same loaded view, use .exists (synchronous, ~50ms). Never use waitForExistence on elements that are already rendered. This is the difference between a 238s test and a 61s test.
3-second gap enforcement — Every gap between consecutive screenshots must be <= 5s. The snap() helper writes screenshot-timing.jsonl in real-time. The journey-loop watcher monitors this file and kills the test on violations. Mark unavoidable long gaps with // SLOW-OK: reason before the snap call.
Acceptance criteria coverage — A journey is not done until every acceptance criterion from its mapped spec requirements has a real implementation, a test step, and a screenshot. Duration is not a target. Extend by covering uncovered criteria, not by repeating the same code path (e.g., downloading multiple models exercises the same download→progress→complete flow — one download is sufficient to verify that flow).
No repetitive padding — NEVER repeat an interaction already performed to pad time. No cycling through the same cards multiple times. No navigating between the same tabs repeatedly. No typing multiple search queries that all produce the same result. Each interaction must test something the previous interactions didn't. If you catch yourself writing "round 2" or "again" in a comment, you are padding.
Actual durations only — Never write estimated durations (e.g., ~5m) to journey-state.md. Always measure from the real xcodebuild test run.
Work on existing journeys first — Check journey-state.md before creating new ones.
NEVER simulate, fake, or stub app features — Do NOT create SimulatedXxxRepository, FakeXxx, MockXxx, or placeholder implementations that bypass real functionality. Every repository, service, and feature MUST use the real framework APIs (ScreenCaptureKit, whisper.cpp, AVPlayer, AVAssetWriter, etc.). If a feature is specified in spec.md, implement it for real — not with Thread.sleep() + fake data. Simulated implementations waste time: they pass tests but deliver zero user value, and every journey built on top of them must be rewritten when real implementations arrive. If a real API requires permissions or hardware that blocks testing, document the blocker and use /attack-blocker to resolve it — do not work around it with a simulation. The only acceptable "fake" is a test double used exclusively in unit tests (never in the running app).
NEVER mock test data in /tmp or anywhere — Do NOT create fake fixture data programmatically in test setUp() methods (e.g., writing JSON files to /tmp/ or NSTemporaryDirectory()). Instead:
1. Use earlier journeys to generate data. Journey tests run in sequence. Earlier journeys (e.g., first-launch-setup, recording) should create real data through UI operations that later journeys can use.
2. Generate data like a real user would. If a journey needs a recording to exist, a prior journey must have created it through the app's actual recording flow via the UI.
3. If UI-generated data is truly impossible (e.g., the feature isn't built yet), you MAY add data programmatically BUT you MUST: (a) document it clearly in journey.md under a ## Programmatic Test Data section explaining what was added and why UI generation wasn't possible, and (b) add a TODO to replace it with UI-generated data once the feature is available.
4. Journey ordering matters. Design journey sequences so that data-producing journeys come before data-consuming journeys. The numbering (001, 002, 003...) defines execution order.
One journey at a time
Real user behavior only — no internal APIs
Every step gets a screenshot (app window only)
Screenshots always go in journeys/{NNN}-{name}/screenshots/
Journey folders are numbered sequentially: 001-, 002-, 003-, etc.
Fix before moving on
Each round produces NEW timestamped files — never overwrite
Unit tests must run in < 1 second total
Only run this journey's tests, not the full suite
Use the extract script for screenshots from xcresult
NEVER edit .xcodeproj manually — always update project.yml and run xcodegen generate

journey-builder