The Wire · Showcase
AI SDK fixes Gemini 3.7 Flash as chat and Eve harden core flows
By RepoJournal · Filed · About Vercel
Gemini 3.7 Flash just broke for minimal reasoning, and the fix that landed overnight says more about the AI SDK's trajectory than any feature release.
The AI SDK fixed a production blocker where Gemini 3.7 Flash rejected 'minimal' and 'none' reasoning values with HTTP 400, forcing the provider to use 'low' instead [1]. That landed in a flurry of patches, including WorkflowAgent surfacing model stream errors without retries [2] and finally making maxRetries control model-call attempts [3]. Meanwhile, Eve reworked onboarding reliability with unified project discovery [5] and safe agent directory initialization [6], while making glob and grep tools opt-in to slim the default model-facing toolkit [4]. Next.js shipped canary 16.3.1-canary.22 with a 32-bit usize fix and test fixture restorations, but the real story is a CI fix ensuring canary releases no longer lag stable after backports [7]. Chat tightened attachment downloads to Google's media API, dropping the unsupported downloadUri fallback [8], and isolated queue dispatch by thread to prevent cross-thread state bleed [9]. The twilio adapter also patched a credentials leak to untrusted media origins [10].
One email a day. Unsubscribe in one click.
Keep up with Vercel in about 3 minutes a day: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
One email a day. Unsubscribe in one click. Read a past issue →
Action items
- → Upgrade @ai-sdk/google to the release containing the Gemini 3.7 Flash reasoning fix vercel/ai [immediate]
- → Review WorkflowAgent maxRetries and error handling for model call retries vercel/ai [plan]
- → Update chat adapter to 4.38.1 to patch the credential leak and media download path vercel/chat [immediate]
- → Monitor Eve onboarding changes before adopting new init and discovery behavior vercel/eve [monitor]
References
- [1] fix: prevent Gemini 3.7 Flash errors for minimal reasoning (#18989) vercel/ai
- [2] fix: surface WorkflowAgent model stream error values without durable-step retries or duplicate callbacks ↗ vercel/ai
- [3] fix: make WorkflowAgent maxRetries control model-call attempts ↗ vercel/ai
- [4] Make glob and grep tools opt-in ↗ vercel/eve
- [5] fix(eve): unify project discovery behavior across CLI ↗ vercel/eve
- [6] fix(eve): safely initialize agent directories ↗ vercel/eve
- [7] [ci] Fix backport canary release dispatch (#97383) vercel/next.js
- [8] fix(gchat): use media api for attachment downloads (#830) vercel/chat
- [9] fix(chat): isolate channel-scoped queue dispatch by thread ↗ vercel/chat
- [10] @chat-adapter/twilio@4.38.1 ↗ vercel/chat