Claude Code Week 17: /ultrareview Goes Public, Session Recap, and Custom Themes
Claude Code versions 2.1.114 through 2.1.119 shipped during the week of April 20–24, 2026, with the headlining change being the public research preview of /ultrareview. Where the initial roll-out required opt-in access, the command is now available to all users: it dispatches a fleet of cloud-hosted bug-hunting agents that analyse your current branch in parallel, then deliver findings back directly to the CLI or Desktop app. The fleet approach means reviewers can be specialised—security agents, performance agents, test-coverage agents—running simultaneously rather than sequentially.
Session recap is a quality-of-life addition that fills a real gap for developers who run long Claude Code sessions in the background or in split-terminal setups. When you return to a terminal that Claude was operating in, a concise recap of what happened while it was unfocused is shown before the next prompt. This makes it practical to switch context, answer a Slack message, or grab coffee without losing track of what Claude did in the interim.
Custom themes via /theme or a plugin let teams standardise the visual style of Claude Code across their organisation, which is especially useful when multiple developers are streaming or recording demos. The Claude Code web interface also received a redesign this week with a new sessions sidebar and drag-and-drop panel layout, making it easier to juggle multiple concurrent tasks.
Read more — Claude Code Docs
IntelliJ IDEA 2026.1.1 Squashes Five-Year-Old Bugs and Spring Performance Issues
JetBrains released IntelliJ IDEA 2026.1.1 in late April 2026, delivering a focused patch release that targets bugs that have been frustrating developers since the 2026.1 major release. The most notable fix restores WSL Python SDK setup, which had regressed for developers using the Windows Subsystem for Linux as their primary Python runtime. Gradle synchronisation failures caused by class casting errors are also resolved, addressing a widespread complaint from projects using non-trivial Gradle plugin configurations.
WildFly deployment support is now reliable again: IDEA can properly connect to the WildFly admin process, so the deployment tool window and run configurations behave as expected. Emmet in remote development environments was similarly broken and is now fixed—a fix that has apparently been pending for nearly five years based on the linked issue history.
For Spring developers the most impactful change is a performance improvement in context actions and code completion for large Spring projects. Teams working on monolithic Spring applications with thousands of beans reported that context-action popups took several seconds to appear; the 2026.1.1 fix brings this back to near-instant response times. Additional fixes cover WSL 2 JDK detection, Ant target execution, WebLogic configuration support, and Find-and-Replace corner cases.
Read more — JetBrains Blog
Karpathy Skills: A Single CLAUDE.md That Makes AI Coding Assistants Dramatically More Useful
A GitHub repository called andrej-karpathy-skills, created by developer Forrest Chang and published in January 2026, has accumulated over 30,000 stars and become one of the most-discussed developer tools of early 2026. The core artifact is a single CLAUDE.md file that encodes four principles distilled from Andrej Karpathy's observations about how LLMs fail at software engineering: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution.
The problem being solved is a recognisable one. LLMs make unchecked assumptions, add speculative features, make changes far outside the scope of what was asked, and fail to surface ambiguities before diving into implementation. Karpathy's original tweet threads on this topic resonated widely because they named failure modes every developer had experienced—bloated pull requests, orthogonal changes that break unrelated tests, and the need to write extensive review comments to undo things that were never requested.
The Karpathy Skills file addresses this by providing explicit behavioral constraints that Claude Code (and other AI coding assistants that support CLAUDE.md or equivalent configuration files) respect at the session level. The four principles map directly to the failure modes: Think Before Coding forces explicit assumption-stating before any code is written; Simplicity First bans speculative features; Surgical Changes restricts edits to the explicitly requested scope; Goal-Driven Execution converts vague instructions into verifiable acceptance criteria before execution begins.
Teams adopting the file report dramatically shorter code review cycles and a substantial reduction in unwanted scope creep. The repository can be installed globally across all Claude Code projects with two commands via the plugin marketplace, making it one of the lowest-friction improvements available to any team already using AI-assisted development.
Read more — BrightCoding
Links & Sources
You May Also Enjoy
