chore(deps): bump workspace deps to latest, agent-device to 0.21.36

Merged
OK
14d ago
+480-470

Every workspace dependency on latest, except the Expo SDK 58 upgrade of apps/mobile-benchmark (react-native 0.88, react 19.3, Stripe RN 0.84, @expo/fingerprint 0.21). That one is a native upgrade with re-recording, so it stays out like in #1491. Expo SDK 57 packages are already on their latest patches. playwright-core 1.64.1 was already latest.

  • agent-device and @agent-device/testmu 0.21.35 -> 0.21.36. The two mobile unit tests that hardcode agentDeviceVersion move with it.
  • @agentclientprotocol/sdk 2.2.0 -> 2.4.0. Adds types only, and PROTOCOL_VERSION is still 1.
  • @limrun/api 0.68.1 -> 0.69.0 (devDep). The breaking change removes iOS keychain archive encryption, which packages/limrun doesn't use. The peer range stays >=0.59.0 <1.
  • @sindresorhus/slugify 3 -> 4 (root devDep). The only breaking change is requiring Node 22. docs:check anchors still resolve: 999 links across 70 pages.
  • vitest and @vitest/expect 5.1.0, oxc 0.157.0, ai 7.0.162 and all @ai-sdk/*, next 16.4.4, fallow 3.38.1.
  • Changeset: patch for e2e, acp, mobile, limrun and testmu, the published packages whose runtime deps changed.

Verified

Ran it locally: yes

  • testbed suite: 82 passed, 1 skipped.
  • Web benchmark deterministic suite on built packages: 41 files, 91 tests passed.
  • iOS mobile benchmark on 2 sims with agent-device 0.21.36: 55 passed, 2 failed.
    • setOrientation rotates the window is a known failure on this machine's sim and also fails on main.
    • infinite scroll reaches the target item was a moving-target flake: 3/3 passes on rerun.
  • Golden device trees (AGENTS.md asks for a re-capture after an agent-device bump): captured iOS and Android on main (0.21.35) and on this branch (0.21.36) with the same devices.
    • The projected trees are byte-identical between main and the branch on both platforms, so the bump doesn't change the snapshot shape.
    • Android trees match the committed goldens.
    • iOS captures here come from agent-device's private-AX backend, not the XCTest backend the committed goldens were recorded with. That holds on main too, so it's local backend selection, not this bump. Committing those trees would break stripe-paymentsheet takes its title from the navigation bar, so the goldens are left as they are.