Tweak AI test skills

.buildkite/commands/run-ai-e2e-tests.sh

@@ -60,7 +60,7 @@ export SIMULATOR_LLM_PILOT_SITE_URL="$(normalize_site_url "$SIMULATOR_LLM_PILOT_
 
 # ── Defaults ─────────────────────────────────────────────────────────
 APP="${APP:-jetpack}"
-export SIMULATOR_NAME="${SIMULATOR_NAME:-iPhone 16}"
+export SIMULATOR_NAME="${SIMULATOR_NAME:-iPhone 17}"
 TEST_DIR="${TEST_DIR:-Tests/AgentTests/ui-tests}"
 SIMULATOR_LLM_PILOT_REPO_URL="${SIMULATOR_LLM_PILOT_REPO_URL:-https://github.com/Automattic/simulator-llm-pilot.git}"
 SIMULATOR_LLM_PILOT_SOURCE_PATH="${SIMULATOR_LLM_PILOT_SOURCE_PATH:-}"

.claude/skills/ios-sim-navigation/references/json-tree.md

@@ -0,0 +1,48 @@
+# JSON Tree Format
+
+The default tree format (`format=description`) is plaintext and grep-friendly
+and covers ~90% of needs. Reach for `format=json` when you need to walk the
+tree programmatically — for example reading a specific element's `value`
+attribute after typing into it, or finding all elements matching a filter.
+
+```bash
+curl -s 'http://localhost:8100/source?format=json' > /tmp/wda-tree.json
+```
+
+The JSON tree is ~375 KB (vs ~25 KB for description), so save to a file
+and `jq` against it rather than piping it through your conversation
+context.
+
+## Node shape
+
+Each node has these fields:
+
+| Field | Description |
+|-------|-------------|
+| `type` | Element type (`Button`, `StaticText`, …) |
+| `label` | Accessibility label (user-visible text) |
+| `name` | Accessibility identifier (developer-assigned id) |
+| `value` | Current value (text field contents, switch state, …) |
+| `rect` | `{"x": N, "y": N, "width": N, "height": N}` |
+| `isEnabled` | Whether interactive |
+| `children` | Array of child nodes |
+
+## Common jq patterns
+
+```bash
+# Find a node by accessibility identifier.
+jq '.. | objects | select(.name == "post-title")' /tmp/wda-tree.json
+
+# Find a node by visible label.
+jq '.. | objects | select(.label == "Settings")' /tmp/wda-tree.json
+
+# Partial match on label.
+jq '.. | objects | select(.label? // "" | contains("Posts"))' /tmp/wda-tree.json
+
+# Read a text field's current value.
+jq '.. | objects | select(.name == "post-title") | .value' /tmp/wda-tree.json
+```
+
+For reading a single attribute, the targeted `/element/<id>/attribute/<name>`
+endpoint is cheaper than dumping the whole JSON tree. Use the full tree
+only when you genuinely need to enumerate or filter across many nodes.

.claude/skills/ios-sim-navigation/references/raw-actions.md

@@ -0,0 +1,70 @@
+# Raw W3C Actions and Less-Common Gestures
+
+Use these only when `tap.rb` doesn't cover the gesture (long press,
+multi-touch) or when you need to clear a text field. All examples assume
+a bound `SESSION_ID` — see `references/sessions.md`.
+
+```bash
+SID=$(jq -r .session_id /tmp/wda-8100.session)
+```
+
+## Long Press
+
+A `pause` between `pointerDown` and `pointerUp`. Duration is milliseconds.
+
+```bash
+curl -s -X POST http://localhost:8100/session/$SID/actions \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "actions": [{
+      "type": "pointer", "id": "finger1",
+      "parameters": {"pointerType": "touch"},
+      "actions": [
+        {"type": "pointerMove", "duration": 0, "x": X, "y": Y},
+        {"type": "pointerDown"},
+        {"type": "pause", "duration": 1000},
+        {"type": "pointerUp"}
+      ]
+    }]
+  }'
+```
+
+## Clear Text Field
+
+**Both methods below are unreliable on iOS 26** — verified to silently
+no-op against `SearchField` and similar controls. Prefer the field's own
+clear control (the small X icon present on most search/text fields) or
+tap-and-hold to bring up the iOS edit menu.
+
+If you must try the programmatic path, **always read the field's
+`value` attribute afterwards to confirm it actually cleared.** Don't
+trust the HTTP 200.
+
+```bash
+# Select all (Ctrl+A) then delete
+curl -s -X POST http://localhost:8100/session/$SID/wda/keys \
+  -H 'Content-Type: application/json' \
+  -d '{"value": ["\u0001"]}'
+curl -s -X POST http://localhost:8100/session/$SID/wda/keys \
+  -H 'Content-Type: application/json' \
+  -d '{"value": ["\u007F"]}'
+```
+
+Alternatively, if you have an element id:
+
+```bash
+curl -s -X POST http://localhost:8100/session/$SID/element/ELEMENT_ID/clear
+```
+
+## Back Navigation Deep-Dive
+
+To return to the previous screen:
+
+- **Primary**: find a Button inside `NavigationBar`. Its label is
+  typically the previous screen's title. Tap it via
+  `tap.rb text "<Prev Title>"` (with `--wait-aid` for the destination's
+  marker).
+- **Fallback**: edge swipe from `(5, H/2)` to `(W*2/3, H/2)`.
+
+The button approach is more reliable because edge swipes can be finicky
+depending on gesture recognizers.

.claude/skills/ios-sim-navigation/references/sessions.md

@@ -0,0 +1,69 @@
+# WDA Session Management
+
+Most action endpoints (`/element/...`, `/elements`, `/wda/keys`, `/actions`)
+require a session id. `tap.rb` manages this for you automatically — it
+creates a session bound to the foreground app's bundle id, persists it at
+`/tmp/wda-<port>.session`, and reuses it across calls. **Read this only
+when interacting with `/session/...` endpoints directly.**
+
+## Why the bundleId binding matters
+
+Without `bundleId` in the session's capabilities, `/actions` returns
+HTTP 200 but the taps never reach the UI: a silent failure that's easy
+to mistake for "WDA is broken." Read the active bundle from
+`/wda/activeAppInfo` and include it in `alwaysMatch`:
+
+```bash
+BUNDLE=$(curl -s http://localhost:8100/wda/activeAppInfo | jq -r .value.bundleId)
+SID=$(curl -s -X POST http://localhost:8100/session \
+  -H 'Content-Type: application/json' \
+  -d "{\"capabilities\":{\"alwaysMatch\":{\"bundleId\":\"$BUNDLE\"}}}" \
+  | jq -r .value.sessionId)
+echo "$SID"
+```
+
+## Reusing the session `tap.rb` persists
+
+For non-tap curl calls (e.g. `/actions` swipes, `/wda/keys`), reuse the
+session id that `tap.rb` already established:
+
+```bash
+SID=$(jq -r .session_id /tmp/wda-8100.session)
+```
+
+Don't mint a fresh session with `alwaysMatch:{}` for these calls — that
+produces an unbound session whose `/actions` requests return HTTP 200
+but never reach the UI.
+
+## Launching with arguments: use `wda-session.rb`, not `simctl`
+
+Creating a session relaunches the target app even when it's already running
+(`forceAppLaunch` defaults to `YES`), so any arguments passed via
+`simctl launch -key value` are lost — they belong to the original process, and
+the relaunch starts a new one.
+
+So when the app needs launch arguments, don't pass them with `simctl`. Let WDA
+launch the app with them in the call that creates the session, using
+`scripts/wda-session.rb`. It persists the session so `tap.rb` reuses it (no
+further relaunch):
+
+```bash
+ruby scripts/wda-session.rb --bundle com.example.app \
+  --arg -some-flag --arg value
+```
+
+Once that session exists, avoid forcing another relaunch before the arguments
+are consumed (don't `simctl launch` again, don't delete the session file).
+`--wait-quiescence` is off by default because a screen with a spinner can keep
+the app from going quiescent and stall the call.
+
+## When sessions break
+
+If the foreground app changes (you launch a different bundle, or iOS
+pushes you to Springboard), the existing session may stop dispatching
+events. Recreate it with the new bundle id, or just call any `tap.rb`
+command (it auto-rebinds).
+
+Symptoms of a dead session: action requests return HTTP 4xx, or (more
+confusingly) return HTTP 200 with no visible UI effect. `tap.rb`
+recreates the session automatically on the next call.

.claude/skills/ios-sim-navigation/references/troubleshooting.md

@@ -0,0 +1,62 @@
+# Troubleshooting and Tips
+
+## Common Failures
+
+### WDA session expiry
+
+WDA sessions can expire after inactivity, or stop dispatching events
+when the foreground app changes. Symptoms: action requests return HTTP
+4xx, or (more confusingly) return HTTP 200 with no visible UI effect
+(see `references/sessions.md` for the bundleId gotcha). `tap.rb`
+recreates the session automatically on the next call. For direct curl
+work, recreate it bound to the current foreground app using the snippet
+in `references/sessions.md`.
+
+### Stale element coordinates
+
+After animations or screen transitions, previously fetched coordinates
+may be wrong. `tap.rb` always re-resolves the element by aid/text before
+tapping, so prefer it over caching coordinates yourself.
+
+### System alert interception
+
+System alerts (location permissions, notification permissions, tracking
+prompts) can block interactions with the app. If a tap silently does
+nothing:
+
+1. Fetch the tree and look for elements of type `Alert` or `Sheet`.
+2. If found, look for a dismiss button ("Allow", "Don't Allow", "OK",
+   "Cancel") and tap it with `tap.rb text "<button>"`.
+3. Retry the original action.
+
+### App crash detection
+
+If actions consistently fail or the tree looks unexpected, the app may
+have crashed. Check and re-launch:
+
+```bash
+xcrun simctl list devices booted
+xcrun simctl launch <UDID> <APP_BUNDLE_ID>
+```
+
+After re-launching, the next `tap.rb` call will create a fresh session
+automatically.
+
+## Tips
+
+- **Tree coordinates, not screenshot pixels** — screenshots may be at a
+  different resolution than the tree's point-based coordinates.
+- **Vertical swipes**: right-edge x (`screen_width - 30`) avoids
+  accidentally tapping interactive elements in the center.
+- **Slow swipes on tappable items**: gestures may register as a tap.
+  Use `duration: 1000` (1 s) for reliability.
+- **WDA startup time**: minutes on a cold checkout (the build phase
+  runs first); ~60 s once DerivedData is warm.
+- **Reconnecting**: if WDA disconnects, re-run `wda-start.rb`.
+- **Tab bar**: look for elements with type containing `TabBar`. Its
+  children are the individual tabs.
+- **Deep links for navigation**: when the target app supports URL
+  schemes, `xcrun simctl openurl <UDID> <url>` (e.g.
+  `wordpress://post/new`) jumps straight to a screen and skips
+  multi-tap navigation chains. Both faster and less flaky than driving
+  the UI to get there.