Claude Code Daily Briefing - 2026-05-05
Release Summary
| Version | Date | Key Changes |
|---|---|---|
| v2.1.128 | 5/4 | ZIP plugin support, /mcp tool count, EnterWorktree HEAD-based branching, vim mode fix |
| v2.1.126 | 5/1 | Custom gateway model picker, project purge, WSL2/SSH OAuth |
New Features & Practical Usage
v2.1.128 — ZIP Plugins, MCP Tool Count, EnterWorktree Fix (5/4)
Released the day before Code with Claude SF, v2.1.128 packs feature additions and 20+ bug fixes.
ZIP plugin archives for --plugin-dir: You can now pass .zip files instead of directories, making plugin distribution across teams much simpler.
/mcp tool count display: The /mcp command now shows per-server tool counts and flags servers with 0 tools. This makes it easy to diagnose MCP connection issues at a glance.
EnterWorktree local HEAD branching: Previously, creating a worktree would branch from the remote, silently dropping unpushed commits. Now it branches from local HEAD, preserving your work.
/model picker cleanup: Duplicate Opus 4.7 entries are collapsed and the current model shows as “Opus” instead of “Opus 4.7”.
# Use ZIP plugins
claude --plugin-dir ./my-plugin.zip
# Check MCP server status
# Type /mcp in session → see per-server tool count and 0-tool warnings
Other notable changes:
--channelsnow works with console (API key) authentication- Subprocesses (Bash, hooks, MCP, LSP) no longer inherit
OTEL_*environment variables - Auto mode classifier errors now include hints (retry,
/compact, or--debug) - MCP reconnection tool summaries grouped by server prefix (prevents conversation flooding)
- Vim mode:
Spacein Normal mode now moves cursor right (standard vi behavior)
Developer Workflow Tips
Claude Code Skills — Define Once, Reuse Everywhere
One of the most underutilized features in 2026 is the Skill system. A Skill is a saved, reusable workflow that Claude triggers automatically when the context matches. You teach Claude how to perform a task well once, refine it, then save it — Claude applies it consistently going forward.
Practical examples:
- Code review Skill: “When reviewing PRs, check security, performance, then test coverage in order. Rate severity.”
- Commit Skill: “Before committing, run lint, type check, and related tests. Use conventional commit format.”
- Deploy Skill: “Run smoke tests in staging before proceeding with production deployment.”
# List available skills
/skill
# Save current conversation patterns as a skill
/learner
Unlike CLAUDE.md instructions, Skills activate automatically when context matches, so you don’t have to repeat instructions every session.
Debug MCP Connections with /mcp Tool Count
With v2.1.128’s improved /mcp command, you can quickly identify MCP server issues. Servers showing 0 tools likely have configuration errors or expired authentication. This is especially useful when running multiple MCP servers — a silent failure in one can cause confusing behavior.
Also note that workspace is now a reserved MCP server name. If you have an existing server named workspace, rename it to avoid it being skipped.
// settings.json — rename if using "workspace"
{
"mcpServers": {
"my-workspace": {
"command": "node",
"args": ["./mcp-server.js"]
}
}
}
Security & Limitations
Microsoft Edge Stores All Passwords in Plaintext in Memory (5/4)
Microsoft Edge has been found to decrypt all saved passwords at startup and keep them unencrypted in memory, regardless of whether they’re being used. This makes credentials vulnerable to memory dump attacks — any malware that can read process memory gets access to every stored password.
If you’re using Edge to store API keys or service passwords for your development workflow, consider switching to a dedicated password manager (1Password, Bitwarden, etc.). This is particularly relevant when running Claude Code with --dangerously-skip-permissions, as the Bash tool could theoretically access process memory.
Ecosystem & Plugins
Anthropic Launches $1.5B Joint Venture with Blackstone & Goldman Sachs (5/4)
Anthropic announced a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to accelerate enterprise AI adoption. Anthropic, Blackstone, and Hellman & Friedman are each contributing ~$300M, Goldman Sachs ~$150M, with additional backing from Apollo, General Atlantic, GIC, Leonard Green, and Sequoia Capital.
Instead of traditional consulting, the venture will embed engineers directly inside companies to redesign workflows and integrate AI into core processes. Target sectors include healthcare, financial services, manufacturing, retail, real estate, and infrastructure. Investors get preferred sales access to their portfolio companies.
Hours before, OpenAI announced a similar $10B venture “The Development Company” with TPG, Bain Capital, and 19 other investors — signaling that the enterprise AI market race is now fully underway as both companies eye IPOs.
CNBC | TechCrunch | Fortune
Community News
-
Code with Claude SF D-1 — Livestream tomorrow: Anthropic’s first developer conference is tomorrow (5/6). Keynotes, hands-on workshops, and 1:1 office hours will be livestreamed globally. The 5/7 Extended session (for indie devs and early-stage founders) will not be livestreamed. New model or feature announcements are expected — check for updates immediately after. Code with Claude SF
-
Kimi K2.6 beats Claude, GPT-5.5, and Gemini in coding challenge: Moonshot AI’s open-weights model Kimi K2.6 won an AI Coding Contest programming challenge, beating major frontier models. Open-weights models are increasingly matching or exceeding proprietary models in coding benchmarks. GeekNews
-
OpenAI launches parallel $10B enterprise venture: Just hours before Anthropic’s announcement, OpenAI revealed “The Development Company” at a $10B valuation with 19 investors including TPG and Bain Capital. Both companies are racing to lock in enterprise revenue ahead of potential IPOs this fall. Semafor
Minor Changes Worth Knowing
workspaceis now a reserved MCP server name: Existing servers with this name will be skipped with a warning. Rename required- OTEL_ env var inheritance blocked*: Subprocesses no longer inherit OpenTelemetry variables, preventing unintended telemetry from third-party tools
- Kitty terminal
/exitnotification fix: Stray4;0;desktop notifications on every exit resolved - Large stdin crash loop fix: Piping >10MB input to
claude -pno longer causes crash loops - Parallel shell call stabilization: Read-only command failures no longer cancel sibling calls
- 1M context autocompact fix: Sessions on 1M-context models with smaller autocompact windows are no longer falsely blocked
Recommended Reads
-
“Long-Running Agents”: An analysis of agent architectures that operate autonomously for days or weeks. Key insight: “the 1M token window also eventually depletes, and context rot occurs before that” — new sessions are “like shift workers with no memory of previous shifts.” Covers the convergence pattern across Anthropic, Google, and Cursor (model loop + execution sandbox + session log separation) and practical patterns like checkpoint-resume, delegated approval, and memory-layered context. Substack
-
“Lessons from ~1,000 Interviews at Amazon”: A 17-year Amazon veteran and 600-time Bar Raiser shares that technically qualified candidates fail primarily due to poor communication, not technical gaps. Candidates spend 95% of prep time on technical skills and only 5% on behavioral — but 10 hours of behavioral practice outperforms additional coding study. Pragmatic Engineer
-
“How Linux 7.0 Broke PostgreSQL”: Linux 7.0’s removal of PREEMPT_NONE mode caused PostgreSQL throughput to drop from 98,565 to 50,751 TPS. The root cause is spinlock interaction with 4KB page faults under PREEMPT_LAZY. Enabling Huge Pages (2MB/1GB) resolves the issue, but the tension between kernel engineers proposing rseq and PostgreSQL’s resistance highlights the “don’t break userspace” debate. The Coder Cafe
Interesting Projects & Tools
-
gc-tree — Global context management for AI coding agents: Stores work style preferences, team terminology, and shared knowledge outside individual repositories. In multi-repo environments, context is selectively retrieved rather than fully loaded, reducing token consumption. Supports branch-like workflows and team onboarding data sharing. GeekNews | GitHub
-
Winpodx — Run Windows apps as native Linux windows: Instead of full-screen RDP or Wine, it runs a Windows container (dockur/windows) via Podman in the background and displays individual apps as native Linux windows using FreeRDP RemoteApp. Taskbar grouping, file associations, and Alt+Tab work naturally. Includes an RDPWrap reimplementation in Rust (rdprrap) supporting up to 10 simultaneous RDP sessions. 7 points. GeekNews | GitHub