chore(deps): bump workspace deps to latest, @agentclientprotocol/sdk to 2.2.0

Merged
OK
17d ago
+1027-1065

Every workspace dependency on latest, except the Expo SDK 58 upgrade of apps/mobile-benchmark. agent-device 0.21.35 and playwright-core 1.64.1 were already latest.

  • @agentclientprotocol/sdk 1.7.0 -> 2.2.0 (major). 2.0 removed ClientSideConnection / AgentSideConnection, so packages/acp/src/session.ts and tests/fixtures/scripted-agent.ts now use the client() / agent() app API with request(methods.agent.session.*). Imports come from @agentclientprotocol/sdk/unstable, because prompt usage and usage_update are unstable in 2.x.
  • Live ACP adapter pins: claude-agent-acp 0.99.1, codex-acp 2.5.0.
  • ai 7.0.159 and all @ai-sdk/*, oxc 0.156, oxlint 1.90, fallow 3.38, @onkernel/sdk 0.139, @limrun/api 0.68.1, smolmachines 1.32.4, @types/node 26.6.6.
  • Mobile benchmark: latest Expo SDK 57 patches (expo 57.0.31, react-native 0.86.3), matching expo install --check. The native fingerprint changes, so CI rebuilds the app.
  • @ai-sdk/openai 4.0.106 sends JSON decision state as input_json (data) instead of input_text. Updated packages/decision/tests/openai.test.ts to the new wire shape.
  • fallow 3.34+ treats the StandardSchemaV1 interface and its merged namespace as a duplicate export. .fallowrc.json ignores that one export.
const connection = client({ name: 'e2e' })
  .onRequest(methods.client.session.requestPermission, ({ params }) => answerPermission(params, ours, calls, turn.own))
  .onNotification(methods.client.session.update, ({ params }) => observe(params, ours, calls, turn, onOwnTool))
  .connect(ndJsonStream(stdin, stdout));
await connection.agent.request(methods.agent.session.prompt, { sessionId, prompt });

Left out: Expo SDK 58 (react-native 0.88, react 19.3, Stripe RN 0.84). It's a native upgrade with re-recording, so it gets its own PR.

Verified

Ran it locally: yes

  • node packages/acp/tests/live/acp.ts against real Claude Code (claude-agent-acp 0.99.1, sonnet) and Codex (codex-acp 2.5.0). Both record, then replay. The todo flow, secret login and tool inventory pass. The wrong assertion fails with ASSERTION_FAILED as expected. Usage and cost come through, no secret in artifacts, and none of the agent's own tools ran.
  • pnpm check green. pnpm test green apart from acp > skips actions batched behind the verdict, a flake that is already on main (2/15 full acp runs fail on origin/main).
  • Not run: OpenAI Decisions API live with input_json (no key locally), and a native benchmark build (CI builds it per fingerprint).