Design principles

1. The app owns introspection; the driver owns input

On native platforms, the in-app SDK knows the state: view hierarchy, coverage, logs, exceptions, custom extractors. The driver causes the state to change through taps, swipes, typed text, and deep links. The Go runner decides what to do.

Splitting these responsibilities is what makes the system work across iOS and Android with one spec surface. Neither the driver nor the SDK alone is sufficient.

• The driver can read a coarse accessibility tree, but not the real UIView or View hierarchy, not coverage, not in-process logs.
• The SDK can see everything inside the app, but cannot dispatch UI events the way the OS would. Touch injection goes through the OS UI-test pipeline (XCTest on iOS, UIAutomator on Android), which the OS treats as real input.

On web, the driver speaks the Chrome DevTools Protocol and handles both input and introspection. No in-app SDK is needed.

2. One TypeScript surface across platforms

Spec authors write against state.ax, state.logs, state.snapshots, and so on, regardless of iOS, Android, or web. Platform differences (back button semantics, system alerts, coverage format) are absorbed in the Go runner and the drivers.

Corollary: if a concept only exists on one platform, it does not belong in the spec API. It belongs behind a feature flag or an extractor.

3. The driver is an interface

DeviceDriver has two production implementations: sidecar (gRPC to the native sidecar) and chrome (CDP, for web), plus a mock for tests. The runner never knows which is wired in. Adding a new platform means adding a new implementation; nothing else changes.

4. Hot loops bypass the sidecar (native)

On native, per-step introspection (hierarchy dump, coverage read, pause and resume) goes over a local Unix socket directly to the SDK. Only physical UI events go through the sidecar's gRPC surface.

A 30-minute run is about 10,000 steps. Every step has at least one hierarchy dump and one coverage read. If those went through the JVM sidecar, the JVM hop would be the bottleneck. Instead the hot path is a 2 ms round-trip to an in-process Swift or Kotlin SDK.

On web, CDP is fast enough that a separate introspection channel is not needed.

5. Deterministic where it can be

A seeded PRNG drives action selection. Spec evaluation is pure given state and snapshots. The bundle hash and seed are recorded in meta.json.

sanderling does not attempt byte-exact replay. Animation timing, keyboard popup timing, and system daemons are non-deterministic on mobile, and the cost of trying to suppress that is not worth the payoff. Same seed produces a similar trajectory, which is usually enough to reproduce the bug.

6. Fail honest

If coverage is not available (release build, instrumentation off), tell the user. Do not pretend exploration is guided when it is random. If the SDK is not linked, say so. If a property is unparseable, fail the run at startup, not step 1000.

The alternative, graceful degradation that silently weakens guarantees, is how testing tools lose trust.

7. Specs are authoritative; no hidden setup

There is exactly one authoring surface: the TypeScript spec. There is no separate YAML for login, no fixtures directory, no setup.sh. Login, onboarding, permission prompts, and teardown are all expressed as action generators or extractors, evaluated in the same loop as the rest of the spec.

This is intentional. A test harness with two authoring languages (YAML plus code, JSON plus code) splits concerns in a way that always drifts. Something works in one surface and not the other, and debugging requires holding both in your head.

8. A property that cannot fail is worse than one that fails

A spec fails loudly when it is wrong about the app. It fails silently when it is wrong about itself, and that is the failure this project has to design against.

Two shapes recur. A property is vacuously true when a fact it reads is missing, so it never fires and every run is green: state.lastAction was hardcoded null on web, ax.findAll on a selector path returned nothing there, and no element on iOS carried text because the mapping only ever read AXValue. In each case a correct property proved nothing. A property is degenerately false when a missing fact is read as a value instead: folio's balances parsed as 0 on web, so the check became |0 - 0| === typedAmount and fired on every healthy submit. Both shapes look exactly like a working property from the outside.

So an unreadable fact is unknown, never a default. Extractors return null rather than zero or empty string, and a property given an unknown declines to judge instead of convicting. The same rule covers arithmetic the representation cannot hold: integer cents in a float64 stop being exact past 2^53, and a comparison that cannot pass is not a comparison that failed.

Which means a green run is evidence only if you can point at a step where the property actually fired, and a red one only if the witness holds real values. Read the witness, not the exit code. Every bug listed above was found that way, and each had already survived a run that looked fine.