Skip to main content

ios

Develop and verify iOS, macOS and Watch apps, audit web parity, or publish Apple changes to TestFlight.

Source facts

Repository
glowingkitty/OpenMates
Last source activity
October 1, 2026 at 14:43
Detected SKILL.md language
English
Stars
50
Forks
3

Install options

The review-first prompt is selected by default. You can switch to a direct command or download a local copy.

Review the source files

Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.

Showing SKILL.md

SKILL.md
Source instructions · Read-only preview
name
ios
description
Develop and verify iOS, macOS and Watch apps, audit web parity, or publish Apple changes to TestFlight.
user-invocable
true
argument-hint
<task description>
## Instructions This skill sets up context for working on the Apple app (`apple/OpenMates/`). Use it for native Swift implementation, Apple/web parity audits, testability work, and Linux-side planning before Mac/Xcode verification is available. For release-only requests, reuse the assigned workspace and existing verification, then go directly to [TestFlight release](#testflight-release). The development and parity loops below apply when implementation or further verification is needed; they do not require restarting completed work before publishing it. For new shared features, app skills, focus modes, embeds, memory types, and provider-backed behavior, Apple parity is the last phase. Before changing Apple implementation, confirm the spec or session contract has completed evidence for CLI implementation/testing against the dev server, npm SDK and pip SDK parity/testing locally against the dev server when applicable, GitHub Actions CI/daily-test reproduction only after local CLI and SDK success, web implementation/testing, and user confirmation that deployed dev web behavior works and looks correct. If any earlier phase is missing, stop Apple implementation and record the missing gate unless the spec contains an explicit waiver or accepted external blocker. Mocked OpenMates API calls, mocked SDK clients, stubbed servers, direct function calls, and fixture replay do not count as completed CLI evidence. ### Mac repository-scoped deletion safety (TASK-752) Prefer `scripts/apple_remote.py` for supported remote operations. On a local Mac, use XcodeBuildMCP when available or local `xcodebuild`/`xcrun` commands; XcodeBuildMCP availability is not a prerequisite for native verification. Set `DEVELOPER_DIR` for the installed Xcode when the active developer directory points to CommandLineTools. Keep DerivedData, package checkouts and temporary build files inside the repository. Preserve installed apps and account data. Use Simulator for isolated Debug component previews. When delivery outside Simulator is TestFlight-only, do not install or launch a locally built macOS app or replace the installed app with a Debug build. Respect the user's chosen delivery route (for example TestFlight) for real-device testing. Keep simulator builds signed for authentication tests: disabling signing can make Keychain writes fail with `-34018` even without an explicit access group. Use Xcode's normal simulator signing or `CODE_SIGN_IDENTITY=-`; reserve `CODE_SIGNING_ALLOWED=NO` for compilation-only checks, not login evidence. An `UNSUPPORTED_REMOTE_OPERATION` means the helper lacks the capability; it is not a deletion stop. Continue with authorized local tools, or report a remote capability gap. Do not repeatedly invoke helpers the policy does not admit. Policy v2 permits cleanup only inside its verified OpenMates and openmates-marketing checkouts, including descendants. Checkout roots, their parents, outside paths and original media remain protected. Symlink targets and native subprocess cleanup must obey the same resolved scope. Keep temporary files, browser profiles and caches inside repository subfolders. Alternate tools do not authorize outside-scope deletion or bypass an active stop. For login recovery, inspect the installed binary's version, signature and actual entitlements before attributing a failure to current source. Native passkeys require the chosen relying-party domain in the signed associated domains and a valid provisioning profile. An ad-hoc Personal Team build that strips these capabilities cannot prove passkey login. Verify password login and passkey login separately, including encrypted account unlocking; one does not prove the other. On `MAC_NO_DELETE_STOP`, immediately end the affected task with a final response. Give the exact outside-scope manual deletion command and reason when known; never invent a target from a signal alone. Only the user may resolve outside deletion. No coordinator relay, transcript role, timeout or confirmation flag clears a stop. The named historical blanket-policy stop is retained with an explicit v2 policy transition; that is not proof of manual deletion or a reusable resume bypass. See `docs/architecture/apple-no-delete-safety.md` for scope and host limitations. Legacy build/sync/cleanup instructions below require an admitted typed operation; they do not authorize arbitrary remote execution. ### Step 1: Load iOS rules and docs 1. Read `.claude/rules/apple-ui.md` (design tokens, forbidden controls, file mappings) 2. Read `apple/AGENTS.md` (architecture, primitives, cross-references) 3. Read `apple/CLAUDE.md` (implementation notes) ### Step 2: Choose Work Mode Use one of three modes depending on the environment and task: - **Mac implementation mode:** Xcode is installed locally. Use XcodeBuildMCP or local Xcode commands for build/test verification and the available computer-use tool for app interaction and screenshots. - **Remote Mac verification mode:** Codex runs on Linux/dev server, but a trusted Mac is reachable through operator-provided SSH configuration. Use SSH to run `git`, `xcodebuild`, and `xcrun simctl` on the Mac. Never commit hostnames, IPs, usernames, SSH aliases, tailnet names, auth keys, device names, or personal local paths to repo files. - **Linux parity audit mode:** XcodeBuildMCP is unavailable. Do static source audits, generate parity specs, update mappings, compare web test IDs with Apple accessibility identifiers, and prepare Mac verification checklists. Do not claim runtime parity until a Mac build/simulator pass verifies it. ### Step 3: Verify the selected Xcode tooling (Mac only) When using XcodeBuildMCP, call `session_show_defaults`. If project/scheme/simulator are not set: ``` session_set_defaults: projectPath: apple/OpenMates.xcodeproj scheme: OpenMates_iOS simulatorName: <pick latest available iPhone> ``` Skip this step on Linux and record `Mac verification required` in the final summary. For local CLI verification, select the installed Xcode with `DEVELOPER_DIR`, inspect project schemes and simulator destinations, and use repository-local DerivedData and result bundles. Missing MCP tools do not block this route. For remote Mac verification mode, use only local runtime configuration supplied by the operator, such as `~/.ssh/config`, environment variables, or the current chat. Before building, verify key-based SSH access, check the Mac checkout with `git status --short`, and avoid overwriting local user changes. Use generic placeholders in notes and committed docs. ### Step 4: Identify affected Swift files or parity surface From the user's task description, identify which Swift files will be touched. For each Swift file: 1. Read the `// ─── Web source` header block to find the mapped Svelte/CSS files 2. Read those Svelte components and CSS files BEFORE writing any Swift code 3. Note exact pixel values, class names, and responsive breakpoints from the CSS If the Swift file has no web source header, check the mapping table in `.claude/rules/apple-ui.md`. For focused component debugging, read [isolated Apple component previews](../../../docs/architecture/apple/component-previews.md). Build/install the Debug app in Simulator, select the component and state through URL, launch arguments, or environment, and test the production controls with local fixture actions. Compare to the user-confirmed rendered web preview at its exact `chrome=0` URL and matching viewport; registry metadata is not visual approval. Run the focused parser/UI tests and then verify the change in its real parent flow. Keep TestFlight delivery for testing outside Simulator. Register new fixture variants in the launch configuration before native tests. Use the rendered accessibility role when querying controls, wait for the expected state rather than existence alone, and measure the component canvas separately from the Simulator window. Clipping must also bound interaction and accessibility to the visible viewport; inspect screenshots alongside accessibility bounds. Streaming parity includes paragraph-by-paragraph updates, live embed references, the rendered web animation timing, and reduced motion. Exercise partial citation syntax, later embed hydration, opening a citation before the response completes, and stream finalization. Cross a responsive breakpoint while a message is mounted: assert that existing paragraphs retain their identity and interaction state rather than remounting the entire renderer. Run iPhone and iPad where the larger viewport exposes a different layout branch. See the `message` streaming fixtures and `ProgressiveMessagePreviewUITests` in the preview guide. For animated fullscreen overlays, wait for the real presentation-completion signal before capturing button coordinates; visibility or hittability alone may be true while a control is moving. Preserve production animation and reduced motion behavior. Auth fixtures must exercise actual form/state transitions with local transports, and remain separate from real login or account-creation proof. Never report callback counters, parser success, or an isolated fixture as complete parent-flow or cross-device synchronization parity. Use the regular wide embed preview on phones as well as desktops. The user rejected the obsolete tall mobile preview fixture: the actual web chat uses the same 300-by-200 card. A variant name or horizontal size class must not override that confirmed reference. Fit only when the available container is narrower. For workspace parity, measure the actual web chat with an embed open while opening and closing the sidebar and settings. Window breakpoints and remaining content width are different inputs: sidebar mode changes at 600, settings at 1100, and the chat/embed split needs at least 1024 points of content (400-point chat plus 10-point gap). Recheck these values against current rendered source. Keep transcript and embed views mounted across resize and pane transitions; deterministic tests must preserve route, scroll position and existing paragraph identity. Reuse production pane controls, not fixture-only replacement panels. Retained native scroll views may still exist in XCUITest while hidden: assert visible-workspace bounds and actual action enabled/hittable state, then reopen and operate the same control to verify restoration. Keep the pane-state fixture short; use the separate long transcript fixture for render-window stress. Continuation parity requires the same available record set, last-opened profile, nonempty draft state and deletion fences, eligibility and ordering before visual comparison. Cover cold cache, older paginated drafts and another device opening a chat. A synthetic carousel alone cannot establish account-level synchronization. For feature parity audits, identify the product surface first, then read all relevant web and Apple sources. Good vertical slices include: - sync and offline lifecycle - message processing and persistence - AI streaming - notifications, unread counts, and banners - auth and recovery - settings pages - embeds and skill outputs - chat shell, composer, and message UI ### Step 5: Start session ```bash python3 scripts/sessions.py start --mode <feature|bug> --task "<task>" ``` ### Step 6: Linux parity audit loop When working without Xcode/Mac access: 1. Read `docs/architecture/apple/parity-matrix.md` if it exists. 2. For broad UI parity work, read `docs/plans/apple-ui-parity-program/plan.yml` and use its chat-first inventory before adding one-off native tests. 3. Run or update `scripts/apple_parity_audit.py` to compare web `data-testid` values with Apple `.accessibilityIdentifier(...)` values and refresh `test-results/apple-parity-inventory.json`. 4. For chat parity, run `scripts/apple_chat_parity_audit.py`; it verifies the chat guardrails and the `programs.apple_ui_parity_program` inventory section. 5. Read the relevant web Playwright specs and source files. 6. Read the corresponding Apple Swift files. 7. Compare protocol names, data models, state machines, error states, loading states, offline behavior, and testability identifiers. 8. Document findings in `docs/architecture/apple/<surface>-parity.md`. 9. Separate findings into `Confirmed from source`, `Likely gap`, `Intentional native difference`, and `Needs Mac verification`. 10. Produce a Mac verification checklist with exact screens, actions, expected screenshots, and tests to add/run later. The Apple UI parity program uses a hybrid gate: deterministic structure, identifier, fixture, native-test, known-renderer, fallback, behavior, and ordering gaps are blocking; visual/style differences stay warning-only until a specific surface and state are reviewed and promoted with explicit tolerance and artifact rules. Linux audits may update docs, scripts, and safe static identifiers. Do not make broad Swift behavior changes that require compiler validation unless the user explicitly asks and accepts Mac verification later. ### Step 7: Build-run-verify loop (Mac only) After each code change: 1. Build and run: use `build_run_sim` (or manual xcodebuild fallback) 2. Screenshot the simulator: use `screenshot` 3. Compare against the web app (use `firecrawl_scrape` with `screenshot` format on the equivalent page) 4. If mismatch: re-read the CSS, fix the Swift code, repeat 5. When verification is complete, shut down any simulator booted by this session unless the operator asks to keep it running For Apple UI parity work, success requires visual proof. Include either a simulator screenshot artifact, a UI test artifact, or an explicit blocker/skip reason. For composer, header, settings, auth, upload, audio, and navigation controls, tests must prove rendered visibility and clickability, not only that an accessibility element exists. ### Step 8: Remote Mac CLI verification loop When XcodeBuildMCP is unavailable but SSH to a trusted Mac is available: 1. Run `python3 scripts/apple_remote.py status`. 2. Run `python3 scripts/apple_remote.py doctor` to verify Xcode, simulator, checkout, scheme, and Watch-test readiness with sanitized output. 3. Confirm the remote Mac has a clean checkout or only the current session's expected changes. 4. Run `git pull --ff-only` in the remote checkout only when it is clean. 5. Build with `python3 scripts/apple_remote.py build-ios --simulator "iPhone 17"`, `build-macos`, or `build-watch` depending on the affected surface. 6. For targeted tests, use `test-ios --only-testing "OpenMatesTests/<TestName>"`, `test-ios --only-testing "OpenMatesUITests/<TestName>"`, or `test-macos --only-testing "OpenMatesMacUITests/<TestName>"`. 7. For startup/crash/log-status/screenshot-status evidence, use `verify-ios-startup`, `verify-macos-startup`, or `verify-watch-startup` instead of hand-rolled `simctl` steps. Add `verify-ios-startup --fresh-install` only when first-run or clean-container state is required. The startup verifiers clean temporary artifacts by default; use a targeted native UI test artifact for durable visual evidence. 8. Current Watch unit tests live under the iOS unit-test target. Use `test-ios --only-testing "OpenMatesTests/<WatchTestName>"` unless `doctor` reports a dedicated Watch test scheme. 9. After verification, run `python3 scripts/apple_remote.py cleanup` for any simulator booted by this session unless the operator asks to keep it running. 10. Clean up only temporary artifacts created by the current session, such as copied screenshots or throwaway build logs. Do not delete unrelated DerivedData, caches, or local checkout changes. 11. Keep private connection details out of repo files and final summaries; refer to the remote host only generically. 12. For visual or interaction parity, capture a simulator screenshot after the app reaches the changed state and record where the artifact lives. If the screenshot cannot be captured, report the sanitized blocker and do not claim visual parity complete. ### Reminders - **Always read the web CSS before writing Swift.** The CSS is the spec. - **Never hardcode colors, spacing, fonts, or radii.** Use generated tokens. - **Never edit `*.generated.swift` or `Icons.xcassets` manually.** They come from `npm run build:tokens`. - **Icons.xcassets uses template rendering.** Use `.renderingMode(.original)` only for icons with built-in colors (e.g. openmates favicon). - **Responsive sizing matters.** Check `.mate-profile-small-mobile`, container width breakpoints, and `horizontalSizeClass` in Swift. - **No native product UI.** See forbidden controls table in apple-ui.md. - **Parity is surface-level, not file-level.** Do not force one Swift file per Svelte/TS file. Maintain traceable many-to-many mappings instead. - **Use stable identifiers.** Apple `.accessibilityIdentifier(...)` values should match web `data-testid` names when the product concept is the same, unless a native-only control needs a native-specific name. - **Document native-only differences.** Example: native offline storage of the last 100 chats is allowed if it preserves web sync semantics and is explicitly documented in the parity spec. ### TestFlight release Read [the Apple release procedure](../../../docs/architecture/apple/testflight-release.md) for publication, signing preflight, constrained-Mac execution, retries and cleanup. Use `scripts/apple_testflight_release.py` for a local release Mac; use `scripts/apple_remote.py deploy-latest-testflight --branch dev` only for a configured remote release Mac. One owner runs native builds and uploads while other authorized work proceeds without competing Xcode or Simulator processes. Before expensive archives, reconcile the latest dev inputs and run the focused lint, test-annotation and signing-input checks described in the release guide. Use existing installed tools so a lint command does not unexpectedly install packages or run generators while release inputs are frozen. Reuse source-bound receipts and the existing ExportOptions plist. Keep the source frozen during archive/upload; preserve and rebuild stale archives only when the fingerprint changed. Upload acceptance and TestFlight availability are distinct: verify the intended version/build for both iOS and macOS through App Store Connect or the TestFlight UI, and validate the embedded Watch companion in the signed iOS archive. If API credentials are unavailable, a successful UI availability check suffices; report processing as unverified until there is that evidence. The Apple upload has no Vercel token dependency. Publish accompanying web changes through the normal scoped deploy helper on the user's approved dev-host route when the Mac lacks deployment credentials; do not request a new token before checking that existing route. Preserve deployment and signed-entitlement gates.
View on GitHub