iOS Simulator

Use Apple's xcrun simctl for lifecycle and app operations, and current idb for accessibility/input. Keep native command syntax rather than a second simulator API. Resolve SKILL_DIR to the directory containing this file; it is not the app repository's scripts/ directory.

Establish the target

xcode-select -p
xcodebuild -version
xcrun simctl list --json
idb --help
idb list-targets

Record the Xcode/runtime, exact simulator UDID, app bundle ID, build/configuration, and reproduction steps. Choose from the actual inventory; never choose the first fuzzy name match or silently switch to a different booted device. Use the same explicit UDID for every operation. Inspect effective IDB_COMPANION/IDB_UDID configuration without exposing credentials before assuming a local target.

Only install missing tools or runtimes when authorised. The current upstream route is brew install facebook/fb/idb, which includes client and companion. Full Xcode is required, not standalone Command Line Tools. Check host requirements against installation; do not force an unsupported binary onto a different architecture. The client may connect to an explicitly authorised remote Mac; do not invent a remote-execution tool or expose a companion to the public network.

Run, observe, act, verify

1. Inspect xcrun simctl help COMMAND or idb ui COMMAND --help for the installed version. Unsupported syntax is a setup issue, not a reason to guess fallbacks.
2. Boot only when needed, then wait with xcrun simctl bootstatus "$UDID" -b. A boot request is not proof that the app is ready; an already-booted error is acceptable only after confirming the intended target's actual state.
3. Install the simulator-compatible .app, launch the exact bundle, and capture the initial screenshot/accessibility state.
4. Read the current UI; resolve a unique intended element; take one authorised action; check its expected postcondition. Re-read after layout/navigation changes.
5. Save the smallest useful evidence and distinguish process completion from app behaviour. A screenshot or successful tap alone does not pass a test.

xcrun simctl install "$UDID" /absolute/path/MyApp.app
xcrun simctl launch "$UDID" com.example.MyApp
idb ui describe-all --udid "$UDID"
idb ui text 'test text' --udid "$UDID"
xcrun simctl io "$UDID" screenshot /absolute/path/evidence.png

idb ui text, not idb text, targets the currently focused field. Use synthetic values, not real credentials in arguments. Prefer a verified accessibility identifier when available. Current marker matching is first substring match, not unique equality: inspect candidates before tapping; refuse ambiguity. For installed versions supporting it, guard with --expected-value. A coordinate fallback needs a fresh image, correct point scaling, and a checked target; it is not equivalent to accessibility activation.

See operations and diagnosis for remaining lifecycle, permissions, push, clipboard, recording, logs, and test workflows.

Reliable command outcomes

The optional helper runs a bounded, non-interactive command without a shell:

node "$SKILL_DIR/scripts/run.mjs" --timeout-ms 120000 -- \
  xcrun simctl bootstatus "$UDID" -b

It reports a JSON envelope and nonzero exit for failed launch, nonzero child exit, signal, timeout, or output overflow. Captured output defaults to 1 MiB total; --max-bytes adjusts that bound up to 16 MiB. It does not parse the command's own result, approve actions, redact output, or guarantee termination of remote work. Inspect nested results and verify state. A timed-out write may already have run; do not blindly repeat a tap, purchase, push, or destructive operation.

Run recordings, interactive tools, and streaming logs directly with explicit lifecycle control, not through this bounded helper. Some execution interfaces (such as idb-repl) report execution errors in their output despite exit zero; process status alone cannot validate them.

Safety and acceptance

A simulator app can reach real services. UI input, URLs, push, permissions, uninstall, erase, and delete are mutations, not a blanket safe tier. Use test accounts/backends and the user's authorised scope. Before destructive operations, identify the target and data loss; never default to all or erase to fix an unrelated failure. Keep screenshots/logs private and review them before sharing.

Report the build/UDID, actions, expected versus observed results, evidence paths, and unresolved failures. Simulator success does not establish physical-device performance, store billing, hardware behaviour, or submission readiness.

Maintainer check: node --test "$SKILL_DIR/tests/run.test.mjs". These runner tests exercise local subprocess semantics, not Xcode, idb, or a real application.