Claude Code Daily Briefing - 2026-04-24

Release Summary

Version	Date	Key Changes
v2.1.118	4/23	Vim visual mode, `/usage` consolidation, custom themes, MCP tool hooks, `DISABLE_UPDATES` env var
v2.1.117	4/22	Native bfs/ugrep search, `/model` persistence, Pro/Max effort default `high`

Full Release Notes

New Features & Practical Usage

v2.1.118 — Vim Visual Mode, MCP Hook Integration, Custom Themes (4/23)

Several notable features landed in v2.1.118.

Vim Visual Mode: v (visual mode) and V (visual-line mode) now support selection, operators, and visual feedback. j/k in NORMAL mode navigate history and select the footer pill at input boundaries.

Hooks Can Call MCP Tools Directly: A new "mcp_tool" hook type lets hooks invoke MCP server tools natively. This eliminates the need to shell out via Bash for MCP integrations in hook workflows.

/cost + /stats Merged into /usage: The two separate commands are now consolidated into a single /usage command.

Named Custom Themes: Create and switch between named themes directly from /theme.

# Vim visual mode
# v → start selection, apply operators like d/y

# Hook calling MCP tools (in settings.json)
# "type": "mcp_tool" to invoke MCP server tools from hooks

# Consolidated usage command
/usage

# Create and switch custom themes
/theme  # → create named themes and switch between them

GitHub v2.1.118

Developer Workflow Tips

Use Tests as an External Source of Truth

Without tests, Claude verifies its own work using its own judgment — and that judgment degrades as context fills up. Tests create an external source of truth that remains accurate throughout the entire session.

In practice:

Ask Claude to write failing tests before implementation
Each red-to-green cycle gives Claude unambiguous feedback
Even at 90% context usage, test results stay reliable

# Test-first workflow
claude "Write tests for this feature first. Don't implement yet."
# → Confirm tests fail, then:
claude "Now implement to make the tests pass."

Claude Code Best Practices

Turn Repeatable Tasks into Skills — “Once a Week = Make a Skill”

If you do something more than once a week with consistent output requirements, it should be a Skill. Content formats, code review patterns, research synthesis, outreach templates — any repeatable pattern qualifies.

Why Skills over CLAUDE.md: Instructions in CLAUDE.md consume context every session. Skills only load when context matches, preserving your ~100 instruction budget while keeping domain-specific knowledge available on demand.

# Create personal skills (available across all projects)
mkdir -p ~/.claude/skills
# → Save repeatable patterns as code-review.md, pr-template.md, etc.

Builder.io — 50 Claude Code Tips

Security & Limitations

(No new security issues as of 4/24 — recent highlights below)

MCP STDIO design vulnerability (4/20): Disclosed by OX Security; Anthropic declined protocol-level fix. Avoid untrusted MCP server configs
Pro plan A/B test (4/22): Rolled back. Existing subscribers unaffected
Mythos unauthorized access (4/21): Via third-party contractor; Anthropic investigating

Ecosystem & Plugins

Amazon Investment Expands — Up to $25B Additional, $33B Total (4/20)

The previously reported $5B investment is part of a larger deal. Amazon agreed to invest up to $25B more in Anthropic (on top of $8B already invested, totaling up to $33B), valuing the company at $380B. The remaining $20B is tied to commercial milestones.

Notably, Amazon committed $50B to OpenAI just two months earlier — a “both sides” strategy that signals how critical AI infrastructure has become.

CNBC | About Amazon

Infisical Agent Vault — Secret Proxy for AI Agents

An HTTP credential proxy that lets AI agents access API keys, database credentials, and other secrets without holding them directly. The agent requests secrets through the proxy rather than storing them locally — useful for Claude Code automation pipelines and MCP server setups.

GitHub

Community News

“AI Fatigue Is Growing”: Users are expressing exhaustion with AI integration across every platform. Default-on AI features, excessive marketing, and declining UX quality are the main drivers. The contrast between opt-in tools like Claude Code and uninvited AI insertions highlights how user agency shapes adoption. Hacker News
Martin Fowler: “Technical, Cognitive & Intent Debt”: Analyzes three layers of debt from LLM-generated code. As coding costs decrease, verification becomes the expensive bottleneck. Abstraction design, meaningful naming, and acceptance criteria remain the human’s core role. Martin Fowler
Qwen 3.6-27B Released: Alibaba released a 27B-parameter multimodal dense model with thinking modes and image/video support. A strong showing in the open-source model landscape. Qwen

Minor Changes

/fork optimization: Writes a pointer instead of the full parent conversation to disk. Significantly faster forking on large sessions
DISABLE_UPDATES env var: Blocks all update paths. Useful for version pinning in CI/CD or managed deployments
WSL inherits Windows managed settings: Windows-side managed settings automatically apply in WSL environments
Auto mode "$defaults" support: Include "$defaults" in autoMode.allow, autoMode.soft_deny, or autoMode.environment to layer custom rules on top of defaults
“Don’t ask again” option: Auto mode opt-in prompt now offers a “Don’t ask again” choice
Multiple MCP OAuth fixes: Credential save crashes, token refresh, and authentication flow fixes

Interesting Projects & Tools

KakaoTalk AI Auto-Reply — On-Device AI Chat Responder: An Android app that reads KakaoTalk notifications and generates responses using the Gemma 4 model entirely on-device, with no internet required. Supports per-room persona and condition-based response control. A practical example of privacy-preserving on-device AI. 5 points, 9 comments. GitHub
Infisical Agent Vault — Secret Store for AI Agents: An open-source HTTP credential proxy that lets AI agents securely access secrets without holding them directly. Useful for reducing secret exposure risk in Claude Code automation workflows. GitHub