Back to Blog
claude-code weekly-update managed-agents vim-mode plugins community

What's Up Claude Code: Week of April 20th

Olivier Legris ·

TL;DR

This was a consolidation week, not a headline-launch week. Claude Code v2.1.116–v2.1.119 tightened the day-to-day workflow: faster /resume, better plugin dependency handling, more serious Vim support, persistent /config settings, and cleaner MCP + OAuth behavior across the board. Anthropic’s April 23 post-mortem also mattered: three widely-felt Claude Code quality regressions turned out to be harness bugs, not underlying model decay, and all were fixed in v2.1.116+ with subscriber usage resets. On the platform side, memory for Claude Managed Agents entered public beta on April 23, 2026, and Haiku 3 was fully retired on April 20, 2026. The other big signal: Anthropic’s Opus 4.6 hackathon winners showed, again, that Claude Code’s most interesting use cases are increasingly coming from people who are not full-time software engineers.

1. The Release Train: v2.1.116 → v2.1.119

Four releases landed this week, and the pattern was clear: less about new surface area, more about making Claude Code feel sharper and more dependable in long-running use.

  • v2.1.116 (Apr 20) — much faster /resume on large sessions, deferred resources/templates/list for faster MCP startup, inline thinking progress text, better /config search, /doctor available mid-response, missing plugin dependencies auto-installed on reload/update, and a dangerous-path safeguard so sandbox auto-allow no longer lets rm/rmdir bypass protection on critical directories.
  • v2.1.117 (Apr 22) — concurrent MCP startup becomes the default, plugin dependency recovery improves again, Opus 4.6 / Sonnet 4.6 default effort for Pro/Max moves to high, native macOS/Linux builds swap Glob and Grep for embedded bfs and ugrep, and the context meter bug for Opus 4.7’s 1M window gets fixed.
  • v2.1.118 (Apr 23) — Vim visual mode (v) and visual-line mode (V) land, /cost and /stats fold into /usage, themes become properly customizable from /theme, hooks can directly invoke MCP tools with type: "mcp_tool", and DISABLE_UPDATES arrives for locked-down environments.
  • v2.1.119 (Apr 23) — /config settings now persist to ~/.claude/settings.json, --from-pr grows beyond GitHub.com to GitLab / Bitbucket / GHE URLs, PowerShell auto-approval reaches parity with Bash, MCP reconfiguration and subagent MCP setup run in parallel, and a long list of polish fixes lands around fullscreen scrolling, paste behavior, /plan, /doctor, and plugin/MCP reliability.

The practical takeaway: the product is getting more opinionated about staying usable in larger, longer, messier sessions. Faster resume, better config persistence, and fewer MCP/OAuth edge-case failures matter more than a flashy experimental flag.

Anthropic’s April 23 engineering post-mortem adds important context here. Three specific bugs explain a lot of the “Claude Code got worse” discourse from the last month:

  • March 4 — effort silently dropped from high to medium
  • March 26 — a caching bug wiped thinking history on every turn
  • April 16 — a verbosity prompt change capped responses to roughly 25 words

All three were fixed by the v2.1.116+ line, and Anthropic reset subscriber usage limits after publishing the write-up. That does not mean every complaint about Opus 4.7 was imaginary, but it does mean a lot of the perceived regression was tooling-layer damage rather than model collapse.

One more release-level item worth knowing: CLAUDE_CODE_FORK_SUBAGENT=1 became publicly relevant this week. It lets parallel subagents share the parent cache prefix, which materially cuts duplicated cache cost in multi-agent runs. This is exactly the kind of feature that signals Claude Code is getting more serious about orchestration economics, not just orchestration syntax.

2. Config, Themes, and the Shift Toward a More Personal Claude Code

The most important quiet change this week is probably config persistence.

Starting in v2.1.119, settings adjusted through /config now persist into ~/.claude/settings.json and respect the normal project/local/policy precedence chain. Combined with named custom themes in v2.1.118, the direction is obvious: Claude Code is becoming less like a disposable CLI process and more like a configurable working environment.

Three details worth calling out:

  • /theme is no longer cosmetic trivia. You can create and switch named themes, edit them as JSON, and plugins can ship themes too.
  • /usage is becoming the center of gravity for operational awareness, absorbing /cost and /stats.
  • prUrlTemplate in v2.1.119 is a small but meaningful enterprise feature: it lets teams point the footer PR badge at an internal review surface instead of assuming GitHub.com.

None of this is dramatic individually. Together, it looks like Claude Code maturing from “smart terminal tool” into “developer workspace with defaults, preferences, and org policy.”

3. MCP and Plugin Reliability Keep Getting Serious

This week’s changelog is saturated with MCP and plugin work, which is usually a sign that real usage has finally found the cracks.

The highlights:

  • MCP startup is faster because more of it is now concurrent by default
  • OAuth refresh and re-auth flows got another round of hardening in v2.1.117–v2.1.119
  • Hooks can now call MCP tools directly in v2.1.118
  • Missing plugin dependencies now get auto-installed in more places instead of failing halfway through
  • Managed settings around blocked marketplaces are enforced more consistently

This matters because Claude Code’s extensibility layer is no longer experimental garnish. For a growing share of users, plugins + MCP are the product. When those layers are flaky, Claude Code feels flaky. When those layers become predictable, the tool starts behaving more like infrastructure.

The interesting architectural move is hooks invoking MCP tools directly. That closes a loop between policy/automation and tool invocation that used to require shell indirection. If Anthropic keeps pushing here, the next phase is obvious: fewer “glue bash scripts,” more first-class event-driven Claude workflows.

4. Managed Agents Get Memory, and Haiku 3 Finally Exits

Two platform-level updates from Anthropic matter this week.

On April 23, 2026, the Claude Platform release notes added memory for Claude Managed Agents in public beta under the standard managed-agents-2026-04-01 header. That is a meaningful upgrade: managed agents stop being purely stateless task runners and start accumulating continuity in a first-class way.

The implication is straightforward. Managed agents were already useful for infrastructure-light automation; with memory, they become more compelling for recurring tasks where remembering prior context actually matters: triage loops, long-lived repo maintenance, customer-specific workflows, or any pattern where the agent should improve across runs instead of starting cold every time.

Also on April 20, 2026, Anthropic fully retired claude-3-haiku-20240307. If anyone still had legacy code pinned there, the migration window is over.

This retirement matters less as product news than as a reminder: the Claude stack is now moving fast enough that “we’ll update it later” is starting to become real operational debt.

There was also a quieter but important product signal on April 21–22: Anthropic briefly tested removing Claude Code from the $20/month Pro plan for a small slice of new signups, then rolled the public pricing page back after backlash. Existing subscribers were unaffected, but the episode revealed something real: current usage patterns are stressing the original plan boundaries, and Anthropic is still feeling for the right packaging. If you depend heavily on Claude Code, it is worth assuming pricing and quota structure are not settled yet.

5. Community Signal: The Hackathon Winners Matter More Than the Prize Pool

Anthropic’s April 20 post on the Built with Opus 4.6 Claude Code hackathon is the clearest community signal of the week.

The most interesting detail wasn’t the $100,000 in credits. It was the winner profile:

  • CrossBeam — AI for California permitting workflows, built by a personal injury lawyer
  • Elisa — a block-based, kid-friendly programming IDE, built by a software engineer for his daughter
  • PostVisit.ai — post-appointment healthcare guidance, built by a cardiologist
  • TARA — road-investment appraisal from dashcam footage, built by a transport specialist in Uganda
  • Conductr — a live AI musical bandmate, built by an electronic musician

Anthropic explicitly notes that four out of the five winners were not professional developers.

That is the real headline. Claude Code is still getting better for senior engineers, yes. But the sharper adoption story is that it is increasingly functioning as a capability amplifier for domain experts. People with strong problem taste and weak traditional implementation leverage are now shipping software that would have been out of reach a year ago.

That changes what “community” means. It’s no longer just plugin authors and terminal power users. It’s lawyers, doctors, infrastructure people, educators, and creators using Claude Code as the implementation layer for domain-specific products.

6. Community Events

The calendar keeps widening. Based on the latest live Luma sync, the community is now at 212 events and 237 leaders.

Coming next week (Apr 27 – May 3):

New cities showing up in the current wave include Lille, Milano, Odense, and Zhongzheng District.

Host your own on Luma →

What We’re Watching

  1. Whether config persistence changes behavior — /config now matters more because changes survive restart. That should make people actually tune Claude Code instead of treating settings as ephemeral.
  2. Whether hooks → MCP becomes the default automation pattern — If this sticks, Claude Code workflows get cleaner and less shell-dependent.
  3. Managed Agents with memory — The question is whether memory makes managed agents meaningfully better on recurring operational work, or just more stateful in theory.
  4. The community’s next layer of builders — The hackathon winners suggest the strongest demand may come from domain experts, not just software teams. That’s a larger market and a different product-design problem.
  5. How far Anthropic pushes the “workspace” idea — Themes, persistent config, usage views, org policy, better resume, better MCP stability. The center of gravity is moving away from “single command” and toward “daily operating environment.”