Back to Blog
claude-code weekly-update release-notes worktrees mcp cache community

What's Up Claude Code: Week of May 4th

Olivier Legris ·

TL;DR

The biggest signal this week is not the changelog. It’s that Anthropic is clearly pushing Claude Code toward longer sessions, more headroom, and more continuity: the five-hour limits were doubled, peak-hour throttling got removed for Pro and Max, and Managed Agents Dreaming showed up as a real “we want Claude to remember and improve across sessions” story. That is the week.

The rest matters too — worktrees got clearer, MCP got less flaky, prompt caching got fixed, and there’s still one nasty IPC bug to watch — but those are supporting details. If you’re scanning for the main story, it’s more capacity + more memory + more statefulness.

1. The main story: more headroom, more memory

The Code with Claude 2026 conference in San Francisco was the clearest product signal of the week.

Two things stood out:

  • Five-hour limits doubled for Pro, Max, Team, and Enterprise users.
  • Peak-hour throttling was removed for Pro and Max.

That’s the practical story. Anthropic is trying to make Claude Code feel less fragile for people who use it all day, every day.

The second headline was Managed Agents Dreaming in research preview. The idea is simple but important: a scheduled background process reviews past sessions, extracts patterns, and curates memories so agents can improve over time instead of starting cold every run.

My read: this is the strongest sign yet that Claude Code is moving from “useful CLI” to “persistent working environment.”

2. The release notes that matter

The changelog still had useful stuff in it, but I’d treat it as the supporting cast this week.

  • v2.1.128 cleaned up defaults: random session colors, /mcp tool counts, zip plugin dirs, and a better EnterWorktree branch behavior.
  • v2.1.129 fixed the 1-hour prompt cache TTL bug and added --plugin-url.
  • v2.1.132 made terminals less annoying: session IDs now reach Bash subprocesses and alternate-screen mode can be disabled.
  • v2.1.133 added worktree.baseRef, sandbox path overrides for Linux/WSL, and better skill discovery for subagents.
  • v2.1.136 was the biggest cleanup: MCP survivability, OAuth refresh races, resume bugs, and a wide terminal/IDE sweep.
  • v2.1.137 / v2.1.138 were small follow-ups.

The nice thing here is that the release train mostly supports the week’s real message: Claude Code is getting more predictable, more explicit, and a bit less magical.

3. Community projects worth exploring this week

The ecosystem is not just reacting to Claude Code — it’s shipping useful stuff around it, and these are the projects I’d point people to from this week’s KB:

  • Plannotator — visual plan annotation before execution. Great if you want to review and tweak an agent plan before it runs.
  • pdf-mcp — chunked PDF reading with BM25 + semantic search and OCR. Useful when the source material is too big to dump into context.
  • ccswarm — Rust multi-agent orchestration with git worktree isolation, PTY sessions, and human-in-the-loop approvals.
  • claude-code-best-practices — 11 stack-specific CLAUDE.md templates plus 30+ guides. Good for teams trying to standardize how they use Claude Code.

This is the part I’d keep an eye on next: the community is building the missing layers around review, orchestration, and reusable project context.

4. Fresh signals from this week

Three more things landed that are worth a quick callout.

  • /insights is a good self-coaching loop. Run /insights, drop the report into a fresh session, and ask the model to turn the trace into concrete improvements.
  • Anthropic’s finance templates are a packaging signal. The May 5 drop shipped 10 ready-to-run agent templates — pitch builder, KYC screener, earnings reviewer, model builder, and more — as plugins, skills, and Managed Agents recipes.
  • Issue #57550 is still the bug to watch. A May 9 regression in v2.1.121–v2.1.123 can stall IPC after a tool_use block and leave sessions hanging without a tool_result.

5. What I’m watching

  1. Whether the doubled limits actually change habits — this is the clearest user-visible win of the week.
  2. Whether Managed Agents becomes a real product direction — if it sticks, Claude Code gets a much stronger continuity story.
  3. Whether worktree and session-state changes reduce weirdness — fewer surprises, fewer hidden assumptions.
  4. Whether the community keeps building around memory and session visibility — that feels like the next layer of the ecosystem.
  5. Whether #57550 gets patched quickly enough to avoid ugly workarounds.

6. Community events

The community graph keeps growing. The latest sync shows 223 events and 247 leaders.

Coming next week (May 11–17):

Coming later:

Host your own on Luma →