[turnipdabeets]

💡 Motivation and Context

Part of the 4.0 stack, on top of #873. It doesn't depend on the macOS/tvOS/watchOS bump in #906. This replaces the vendored Reachability (Ashley Mills, about 400 lines built on SCNetworkReachability) with a small wrapper around NWPathMonitor. Apple deprecated the SCNetworkReachability APIs in the iOS 17.4 / macOS 14.4 SDKs. NWPathMonitor is available on every platform we support.

The wrapper keeps the same internal interface (connection, onReachable/onUnreachable, startNotifier/stopNotifier), so the queues and PostHogContext barely change.

Behavior

	Before	After
dataMode = .any	pauses when unreachable	pauses when path.status == .unsatisfied
dataMode = .wifi	sends on any reachable path that isn't cellular (Wi-Fi, ethernet)	same: !path.usesInterfaceType(.cellular)
VPN on demand / connection on traffic	reachable	reachable (.requiresConnection)
Path needs a connection with no on-demand flag (e.g. VPN waiting for a password)	unavailable	reachable (.requiresConnection)
App's cellular data turned off in Settings	reachable, cellular (requests fail, retry backoff)	unavailable (.unsatisfied, reason .cellularDenied), queues pause
$network_wifi / $network_cellular	from the reachability flags	from the path, same meaning
watchOS	not monitored	not monitored (unchanged)

• First path: NWPathMonitor delivers its first path asynchronously. The old API read flags synchronously, so reads of connection wait for that first path, up to 100 ms after init in total; locally it arrives in about 1 ms. If no path has arrived, connection is nil. In that state .any doesn't pause, .wifi stays paused until a path arrives, and events leave out $network_* rather than reporting false.
• startNotifier() no longer throws. NWPathMonitor has no failure path, so the queue's "notifier failed to start" fallback is gone.
• Notifications still go to the main queue by default. They fire on every path update, as the old callback fired on every flags change. NWPathMonitor also reports changes that keep the same connection (e.g. DNS), which now trigger a flush that does nothing. The first startNotifier() reports the current connection, as the old initial check did, so every queue flushes leftover events at launch. Each notification reads the connection when it's delivered, as the old code did.
• Edge differences: the two rows marked in the table. NWPath doesn't separate a connect-on-demand path from one that needs user action, so both count as reachable; when a request then fails, the normal retry backoff applies. With cellular data turned off for the app, the queues now pause instead of sending requests that fail.

Public API: none. #874 already made ReachabilityError internal. The only public symbol removed is the ReachabilityChangedNotification global, which was already @available(*, unavailable) and isn't in the API snapshot.

Objective-C: none of this was ever exposed to Objective-C. Network autolinks like SystemConfiguration did, so there are no podspec or SPM linker changes.

💚 How did you test it?

• New Reachability notifier tests cover the start/update/stop rules: the first start reports the current connection once, paths after start are reported, paths after stop aren't, and with no path the wait is bounded. A notification queued before a newer path reports the connection at delivery, so a stale "unavailable" can't pause the queues while online.
• make test (macOS): all 169 XCTest tests pass. Swift Testing ran 964 tests; 13 issues in PostHogStorage app-group tests fail the same way on the base branch (macOS blocks access to the app-group container on my machine) and are unrelated.
• A side-by-side harness ran the old SCNetworkReachability code and the new code in one process on macOS (Wi-Fi). Both reported WiFi with no mismatches, and both fired one initial onReachable. I didn't test Wi-Fi/cellular switching or going offline.
• xcodebuild build succeeds for iOS, tvOS, watchOS, visionOS and Mac Catalyst. make buildSdk wasn't run as-is because xcpretty isn't installed locally.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed.
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran pnpm changeset to generate a changeset file

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

• Made with Claude Code, from an audit for leftover code in the 4.0 release work. A separate agent reviewed the PR; its findings are fixed here or listed above.
• I mapped .wifi to "not cellular" to keep today's behavior. I rejected !path.isExpensive, which would also pause on Wi-Fi tethered to an iPhone hotspot. That's a real behavior change and would need its own changelog entry.

🤖 Generated with Claude Code

[github-actions]

posthog-ios Compliance Report

Date: 2026-10-07 20:51:22 UTC
Duration: 230207ms

⚠️ Some Tests Failed

44/45 tests passed, 1 failed

---

Capture Tests

⚠️ 28/29 tests passed, 1 failed

View Details

Test	Status	Duration
Format Validation.Event Has Required Fields	✅	3018ms
Format Validation.Event Has Uuid	✅	3282ms
Format Validation.Event Has Lib Properties	✅	2784ms
Format Validation.Distinct Id Is String	✅	3078ms
Format Validation.Token Is Present	✅	659ms
Format Validation.Custom Properties Preserved	✅	449ms
Format Validation.Event Has Timestamp	✅	2829ms
Retry Behavior.Retries On 503	✅	11899ms
Retry Behavior.Does Not Retry On 400	✅	4977ms
Retry Behavior.Does Not Retry On 401	✅	2758ms
Retry Behavior.Respects Retry After Header	✅	7887ms
Retry Behavior.Implements Backoff	✅	21322ms
Retry Behavior.Retries On 500	✅	5640ms
Retry Behavior.Retries On 502	✅	9357ms
Retry Behavior.Retries On 504	✅	9377ms
Retry Behavior.Max Retries Respected	❌	22055ms
Deduplication.Generates Unique Uuids	✅	3261ms
Deduplication.Preserves Uuid On Retry	✅	9220ms
Deduplication.Preserves Uuid And Timestamp On Retry	✅	16317ms
Deduplication.Preserves Uuid And Timestamp On Batch Retry	✅	7849ms
Deduplication.No Duplicate Events In Batch	✅	2904ms
Deduplication.Different Events Have Different Uuids	✅	3097ms
Compression.Sends Gzip When Enabled	✅	2902ms
Batch Format.Uses Proper Batch Structure	✅	479ms
Batch Format.Flush With No Events Sends Nothing	✅	291ms
Batch Format.Multiple Events Batched Together	✅	3100ms
Error Handling.Does Not Retry On 403	✅	4911ms
Error Handling.Does Not Retry On 413	✅	4883ms
Error Handling.Retries On 408	✅	9212ms

Failures

retry_behavior.max_retries_respected

Expected 4 requests, got 6

Feature_Flags Tests

✅ 16/16 tests passed

View Details

Test	Status	Duration
Request Payload.Request With Person Properties Device Id	✅	3060ms
Request Payload.Flags Request Uses V2 Query Param	✅	3028ms
Request Payload.Flags Request Hits Flags Path Not Decide	✅	2837ms
Request Payload.Flags Request Omits Authorization Header	✅	3092ms
Request Payload.Token In Flags Body Matches Init	✅	2974ms
Request Payload.Groups Round Trip	✅	2938ms
Request Payload.Groups Default To Empty Object	✅	3088ms
Request Payload.Person Properties Distinct Id Auto Populated When Caller Omits It	✅	2960ms
Request Payload.Disable Geoip False Propagates As Geoip Disable False	✅	3013ms
Request Payload.Disable Geoip Omitted Defaults To False	✅	2984ms
Request Payload.Flag Keys To Evaluate Contains Only Requested Key	✅	3339ms
Request Lifecycle.No Flags Request On Init Alone	✅	168ms
Request Lifecycle.No Flags Request On Normal Capture	✅	3005ms
Request Lifecycle.Two Flag Calls Produce Two Remote Requests	✅	5610ms
Request Lifecycle.Mock Response Value Is Returned To Caller	✅	2963ms
Side Effect Events.Get Feature Flag Captures Feature Flag Called Event	✅	3141ms

[dustinbyrne]

AI-assisted static review at 144941560cb468c1eb9992a0092b28db8a2245ab. Regression scenarios were not executed.

[dustinbyrne]

Both earlier review comments are addressed. No demonstrated implementation regression found at 8fd604deca89cb595f659badf9a6fceedc06b28b; the inline test-coverage note is nonblocking, and current tests pass.

Reviewed statically with existing CI, without executing new scenarios. Compliance still reports 43/45 despite its green check: the retry-count failure also appears in base-associated CI, and the additional zero-request case has unresolved harness/PR causality.

AI-assisted review.