Who contributed to Vercel on May 17, 2026?

3 developers shipped this update, including dancer, github-actions[bot], and Melkeydev.

What were the notable Vercel updates?

fix(provider/xai): dedupe reasoning-start on multi-summary-part responses to prevent xai 400s on continuations, fix(provider/xai): dedupe reasoning-start on multi-summary-part responses to prevent xai 400s on continuations (#15334), and fix(provider/xai): surface full xai error detail in APICallError.message.

@vercel

Next.js + the Vercel platform - frontend infrastructure for the web

github ↗

Pick a date

Topics: JavaScript Full archive →

The Wire · Showcase

VERCEL AI FIXES XAI PROVIDER BUGS BREAKING MULTI-STEP TOOL CALLING

By RepoJournal · Filed 06:00 UTC on May 17, 2026 · About Vercel

3 people shipped this

dancer @dancer 2 cited

github-actions[bot] @github-actions[bot] 2 cited

Melkeydev @Melkeydev 1 cited

Grok reasoning models are failing 4-12% of the time on continuation steps, and Vercel AI just shipped two fixes that surface the actual errors and stop duplicate reasoning events from tanking requests.

The XAI provider in Vercel AI was silently breaking multi-step tool calls with grok-4.3 and grok-4.20-reasoning [1]. Every reasoning summary part event fired a new reasoning-start marker, which meant when XAI split a reasoning block across multiple parts, Vercel AI sent duplicate markers that violated XAI's API contract. That's been fixed [2]. In parallel, error reporting was completely broken [3]. XAI 4xx failures surfaced as generic "Bad Request" messages while the real error detail lived buried in responseBody, making debugging impossible [4]. Now errors properly parse XAI's actual response schema and bubble up the real message. On the Turborepo side, v2.9.15-canary.3 hit a doc aliasing snag but shipped critical cache fixes [5], while the standalone Vercel CLI fixed its own update check routing [6]. The AI package is holding in pre-release mode and waiting on the next changeset merge [7].

Action items

→ Upgrade @vercel/ai if using XAI providers with reasoning models before next deploy vercel/ai [immediate]
→ Update standalone Vercel CLI to fix self-update checks vercel/vercel [plan]
→ Monitor Turborepo canary for stable 2.9.15 release vercel/turborepo [monitor]

References

[1] fix(provider/xai): dedupe reasoning-start on multi-summary-part responses to prevent xai 400s on continuations ↗ vercel/ai
[2] fix(provider/xai): dedupe reasoning-start on multi-summary-part responses to prevent xai 400s on continuations (#15334) vercel/ai
[3] fix(provider/xai): surface full xai error detail in APICallError.message ↗ vercel/ai
[4] fix(provider/xai): surface full xai error detail in APICallError.message (#15338) vercel/ai
[5] release(turborepo): 2.9.15-canary.3 ↗ vercel/turborepo
[6] [FIX] Fixing a single deployment error ↗ vercel/vercel
[7] Version Packages (canary) ↗ vercel/ai

Quick answers

What shipped in Vercel on May 17, 2026?: Grok reasoning models are failing 4-12% of the time on continuation steps, and Vercel AI just shipped two fixes that surface the actual errors and stop duplicate reasoning events from tanking requests. In total, 3 commits, 5 pull requests, and 11 releases landed.
Who contributed to Vercel on May 17, 2026?: 3 developers shipped this update, including dancer, github-actions[bot], and Melkeydev.
What were the notable Vercel updates?: fix(provider/xai): dedupe reasoning-start on multi-summary-part responses to prevent xai 400s on continuations, fix(provider/xai): dedupe reasoning-start on multi-summary-part responses to prevent xai 400s on continuations (#15334), and fix(provider/xai): surface full xai error detail in APICallError.message.

NEXT.JS OPTIMIZES APP ROUTING ENTRY POINTS WHILE WORKFLOW CUTS STREAM LATENCY 20 PERCENT

Next.js extracted invariant app-page handlers into shared runtime modules while Workflow's stream write batching moves into core, eliminating per-token round trips on idle producers.

+11

javascript 126 shipped 2-min read

@vercel 1 day ago

TWO CRITICAL SECURITY PATCHES LAND IN WORKFLOW AND EVE OVERNIGHT

Next 16.2.11 and PostCSS 8.5.12 patches ship across workflow to seal CVE vulnerabilities, while Eve locks down workflow failure visibility.

javascript 153 shipped 1-min read

@vercel 2 days ago

WORKFLOW 5.0 BETA ARRIVES WITH HOOK TOKEN RETENTION, AI SDK FIXES GEMINI AND BEDROCK HALLUCINATIONS

Workflow shipped its most ambitious beta yet with experimental token retention, while the AI SDK shipped five critical fixes across Gemini, Bedrock, and dynamic tool validation in a single push.

javascript 132 shipped 2-min read

@vercel 3 days ago

NEXT.JS PATCHES FIVE SECURITY HOLES OVERNIGHT WHILE VERCEL CLI GAINS AI GATEWAY BUDGET CONTROLS

Five security advisories dropped across Next.js in the last 24 hours, ranging from origin validation on internal redirects to fetch cache key collisions that could leak cached responses between requests.

+12

javascript 145 shipped 2-min read

Elsewhere on the wire

AI Agents about 9 hours ago

CLAUDE OPUS 5 LANDS ACROSS THE STACK

The newest Anthropic model is now live in langchain, Cline, and llama-index, with native support for extended reasoning and 1M context windows.

ai-agents 28 shipped 1-min read

Local LLMs about 9 hours ago

OLLAMA LANDS LAGUNA SUPPORT AND CRUSHES MEMORY LEAKS WHILE SGLANG HITS V0.5.16 WITH CONFIDENCE-DRIVEN SPECULATIVE DECODING

Ollama shipped three critical performance and reliability fixes for Metal residency and concurrent access patterns, while SGL-Lang released 0.5.16 with a new speculative algorithm hitting 383.7 tok/s on DeepSeek-V4.