[CahidArda]

Daily upstream-compat run (2026-10-21). This branch is #72 (eve-extension rebuilt on eve 0.81.1, peer floor >=0.81.0; @tanstack/ai 0.70.0 etc.) plus one commit: ai 7.0.146 -> 7.0.148 in devDeps + examples. devDep/example-only, so no new changeset (the one from #72 stands). If #72 is merged first, this reduces to that single commit.

Held on purpose (not fixed here)

• @tanstack/ai-persistence stays at 0.7.2 (latest 0.10.0). Since 0.8.0 runPersistenceConformance has an activities case; upstashPersistence has no ActivityStore, and the test does not pass skip: ['activities'], so the suite fails. Published package is unaffected (ActivityStore is optional upstream).
• packages/eve + examples/eve-demo stay on eve ^0.65.0: src/sandbox.ts imports MutableNetworkSandboxSession (:63) and uses onSessionStop/onSessionDelete (:475, :487), renamed onSandboxStop/onSandboxDelete in eve 0.75.

Verified on this branch: lint, typecheck, sync-version --check, build, examples build; tanstack-ai-demo pnpm e2e 6/6 and tanstack memory contract 13/13 (each on a fresh Redis); eve eval with AGENTKIT_MOCK_MODEL=1 passes in eve-demo and eve-extension-demo. The full pnpm test on a throwaway Redis had 19 failures, 18 of which are Exceeded max index count of 1 (inconclusive; the other is a timing assertion 59997 <= 59996). UPSTASH_BLOB_TOKEN suites were skipped.

Note for review: published eve-extension 0.16.0 (built on eve 0.74.0) still builds clean on eve 0.81.1, so the breaking floor bump in #72 is not forced by a break; it only aligns the floor with the new stamps.

🤖 Generated with Claude Code

https://claude.ai/code/session_018nEKvzcXcgq9gsHixbtLdQ

[canblmz1]

I maintain Prefix-Safe JSON. This SDK upgrade is a concrete place to evaluate an optional provider-boundary regression: memory.test.ts calls execute directly, while save_memory persists through AgentMemory.add.

I can prepare one scripted-provider test using published prefix-safe-json@0.6.3, your candidate's exact AI/OpenAI versions and existing isolated Redis setup. It would compare native, small raw-native and PSJ admission for healthy and duplicate-original-ID responses, checking stored memories, result IDs and continuation. Your Responses path calls for core's existing adapter, not the Chat Completions companion; no production route change. Small native checks may give identical results; this would not address the quota/timing failures reported here.

Would one scoped save_memory test earn a place in your SDK-upgrade checks?

[CahidArda]

Closing as a duplicate: these daily upstream-bump PRs all bump the same pins to newer versions. #74 (eve 0.82.0, ai 7.0.148, @tanstack/ai 0.71.0) supersedes it. The branch is left in place.