Daily AI Builder Brief - April 26, 2026

Executive Summary

Claude Code 2.1.120 (April 24) ships BrowserBatch, new harness and memory instructions, a read-before Write flow, an automated rule reviewer, and claude ultrareview for headless CI runs.
Same release breaks --resume and --continue with a TypeError; auto-update is reverting affected clients to 2.1.119.
A community survival checklist (April 24) lists eight 2.1.120 regressions including silent Opus 4.7 200k to 1M routing, broken auto-updates, and CLAUDE.md loaded but ignored.
Recommended pin until a cleanup release lands: 2.1.119 (safer) or 2.1.117 (safest).
Anthropic and NEC announced a partnership (April 24): Claude rolls to roughly 30,000 NEC employees and a Japan Center of Excellence built on Claude Code and Cowork.
MCP spec is still 2025-11-25. Working groups stay focused on streamable HTTP scaling, audit trails, and .well-known discovery.
Nothing in broader AI moves the builder calculus in the last 48 hours.

Claude Code

2.1.120 ships, then breaks resume (April 24)

Highlights: BrowserBatch fires multiple browser tool calls in one round trip, harness and memory instructions get cleaner cleanup and cache behavior, the Write tool now reads before writing, an automated rule reviewer flags ambiguous classifier rules, and claude ultrareview runs /ultrareview headless and prints findings to stdout for CI capture. Then --resume and --continue crash with g9H is not a function because a feature-flagged session-restore callback returns undefined when disabled. Auto-update is rolling clients back to 2.1.119 on next check. The official changelog hasn't published the 2.1.120 entry yet. Issue #53086

Eight-regression survival checklist (April 24)

Community list of 2.1.119 and 2.1.120 issues with workarounds: resume crash, silent Opus 4.7 200k to 1M routing (cost and behavior drift), terminal resize duplicates UI, auto-updater runs but doesn't install, /mcp menu freezes on resumed WSL2 sessions, CLAUDE.md loads but isn't consulted, sandbox excludedCommands ignored, macOS worktree git merges hang on Apple Silicon. Pin to 2.1.117 if anything bites. Checklist

Anthropic

NEC partnership and Japan AI Center of Excellence (April 24)

Anthropic's first Japan-based global partner. NEC rolls Claude to roughly 30,000 employees worldwide and stands up a Center of Excellence to build what it calls one of Japan's largest AI-native engineering teams. Joint work targets finance, manufacturing, and cybersecurity, with Claude Code and Cowork landing inside NEC BluStellar Scenario customer offerings. Announcement

MCP Ecosystem

No new spec or major release

Spec stays at 2025-11-25. Working groups continue on streamable HTTP horizontal scaling, audit trails, SSO auth, and .well-known capability metadata for registries and crawlers. Nothing cut in the last 48 hours. MCP Blog

Broader AI

Nothing today.

What This Means For Builders

Don't run latest on Claude Code right now. Pin to 2.1.119 or 2.1.117 in any pipeline until a fix release lands. CI runners with auto-update on will drift through the regression window.
BrowserBatch and claude ultrareview are worth wiring into staging once stable. One round trip for multi-step browser flows and stdout findings for CI capture both close real gaps in agent pipelines.
If you depend on Opus 4.7 200k specifically, watch your bills. The silent 1M-variant routing bug changes pricing and context behavior without warning.