a
adaminsley@gmail.com
I tested Apple's Siri AI Developer Preview (iOS 27.3 beta) with expo-app-intents 0.5.16 on Expo SDK 58, on a real iPhone.
The package works unchanged.
Siri called tools written in
app-intents/
, including with the app closed and with the phone locked. A few preview-specific pieces needed workarounds, and I'd like to know whether you'd welcome contributions for them.
Reproduction and write-up: https://github.com/airowe/siri-ai-expo
Setup:
Xcode 27.3 beta (27C5020n), iPhone 16 Pro on iOS 27.3 beta (24C5037i) with Siri AI on, Release build signed for development.
### What worked with no package changes
  • @AppIntent(.tool)
    intents in
    app-intents/
    . Siri called them and spoke their dialog, with the app open, with it closed, and with the phone locked.
  • @AppEntity(.generable)
    parameters filled in by Siri.
  • Entities published from JavaScript with
    setEntityCatalogAsync
    and
    reindexEntitiesAsync
    , found by Siri through Spotlight.
  • A tool that calls an authenticated API and refreshes an expired token inside the Siri request.
### Where I needed workarounds
  1. Compiling the tool code.
    The tool macros only exist in the 27.3 SDK. I gated them behind a Swift compilation condition set by my own config plugin:
```js
entry.buildSettings.SWIFT_ACTIVE_COMPILATION_CONDITIONS =
"$(inherited) SIRI_AI_PREVIEW"
;
```
  1. Deployment target.
    @Property(title:description:)
    is iOS 27.3+ only, and a property wrapper can't be limited to newer versions with an availability check. Any entity that uses it raises the whole app to iOS 27.3.
  2. Prebuild.
    expo prebuild
    regenerates
    ios/
    , so I need a script to re-add the Siri Evaluations test target after every prebuild. A Skills Bundle added in Xcode is presumably lost the same way (not yet verified).
  3. Siri answers from indexed data instead of calling the tool.
    This is the one I'd most like to see in the docs:
- With order status in the entity and its Spotlight description, Siri called the status tool in
3 of 8
runs. In the rest it gave the same facts in its own words, from the index.
- With only the order's identity indexed (id, name, synonyms) and the tool reading status when it runs:
4 of 4
completed runs.
- A tool with nothing indexed (a cart fetched from an API):
5 of 5
.
- The rule: index what identifies a record, never the answer the tool exists to give.
  1. Locked phone.
    Tools that read a token need Keychain items readable after first unlock (
    kSecAttrAccessibleAfterFirstUnlockThisDeviceOnly
    ). With that, Siri read an authenticated result with the phone locked.
  2. Testing.
    Xcode 27.3 ships
    SiriEvaluations.framework
    for devices only, not for the Simulator.
### Question
Would you welcome contributions for any of these, or would you rather wait until the APIs leave beta?
  • A config plugin option that sets the preview compilation condition and keeps the Siri Evaluations target and Skills Bundle through prebuild.
  • A tool example for
    npx expo-app-intents init
    .
  • Guidance in the docs and the bundled agent skill, covering points 4 and 5.
I'm also following
runBackgroundTask
(#51595), which covers tools answering from JavaScript, so I've left that out.