Foreword

Foreword: Migration Guide (February to May 2026)

Published May 2026

Foreword: Migration Guide (February to May 2026)

Short reference for what changed in the three months between editions. Most changes are non-breaking; a few are worth a CLAUDE.md update.

Models

Then (Feb 2026): Sonnet 4.5 was the default · Now (May 2026): Sonnet 4.6 is the default · Action: No action; updated automatically.
Then (Feb 2026): Opus 4.5 was the flagship · Now (May 2026): Opus 4.7 is the flagship (4.6 also available) · Action: Update CLAUDE.md model references if any.
Then (Feb 2026): Haiku 4.5 was the cheap option · Now (May 2026): Haiku 4.5 still is · Action: No change.
Then (Feb 2026): Cache was new and explicit · Now (May 2026): Cache is default in Claude Code · Action: Remove any manual cache_control if the default is fine.
Then (Feb 2026): “Opus for all teammates” was uncontroversial · Now (May 2026): Sonnet works for most teammates now · Action: Consider downgrading routine teammates to Sonnet.

Claude Code

Then: Desktop app was a CLI-in-window wrapper · Now: Desktop app is a multi-session Mission Control. See Chapter 15.
Then: Only the CLI for CI · Now: CLI for GitHub Actions, Routines for cloud-run automation. See Chapter 16.
Then: No headless cost caps documented · Now: --max-tokens is the standard pattern. Always set it.
Then: Agent teams experimental · Now: Still experimental but more stable. our team uses them on project-alpha.
Then: Subagents had MCP inheritance bugs · Now: Fixed. Subagents now inherit dynamically-injected MCPs.
Then: No Agent SDK credit pool · Now: New monthly Agent SDK credit pool from June 15, 2026.
Then: /learn was the main way to grow · Now: /learn still is, plus auto-memory and CLAUDE.md evolution loops.

Cowork

Mostly new. Cowork was not covered in the Feb edition. The headline:

Three modes: Dispatch (mobile-initiated), scheduled tasks, live artifacts. See Chapter 18.
Plugin marketplace shared with Claude Code, catalog skews toward knowledge work.
CLAUDE.md per engagement is the central pattern. See Chapter 19.
Org controls (group spend limits, OTel, SSO) for Team and Enterprise plans.

Microsoft 365 and Chrome

All new. See Chapter 17.

Excel, PowerPoint, Word: GA on all paid plans.
Outlook: public beta on all paid plans.
Chrome: beta on all paid plans.

Plugins and connectors

Then: ~33 plugins in official marketplace · Now: 101 plugins (33 Anthropic-built, 68 partner-built) as of March 2026.
Then: Plugin ecosystem dominated by code tools · Now: Knowledge-work plugins are now a major category (Slack, Notion, Drive, Zoom, M365).
Then: No Zoom connector · Now: Zoom connector new for 2026. Pulls recordings, transcripts, chat.
Then: M365 connectors via beta extensions · Now: M365 plugins GA. Excel/PowerPoint/Word in production use.

CLAUDE.md changes worth making now

Add the no-AI-slop list

If your CLAUDE.md doesn’t have one, add it. The our standard list is in Chapter 19.

Add the em-dash rule

Same. Models still love em dashes.

Update model references

If your CLAUDE.md mentions specific models in agent definitions, audit them. Sonnet 4.6 is the new default for routine work; reserve Opus 4.7 for high-stakes reasoning.

Add the “no command chaining” rule

If you haven’t already. The rule prevents Bash tool calls that chain commands with && , ; , or || . See Chapter 1.

Add a “Decision Making” section

The our pattern from project-alpha: “Always use a team of agents to make decisions democratically. Always use a team of adversaries, supporters, and neutrals. Always go through three rounds of deliberation.” Worth borrowing.

Add the debugging discipline section

Three rules: never guess, never claim a fix works until verified, three failed fix attempts means the architecture is wrong. See Chapter 23.

Workflow updates worth making

Switch routine review jobs to Sonnet 4.6

If you have CI workflows running Opus on every PR, downgrade to Sonnet 4.6. The 1.2pp accuracy gap on SWE-bench is rarely worth the cost premium for routine review.

Adopt the project-alpha CI pattern

PR reviewer → terraform plan reviewer → build failure triage. Three jobs, three different token caps, three different prompts. See Chapter 16.

Move recurring work to scheduled tasks

Anything you do every morning, every Monday, every week. Schedule it. See Chapter 20.

Add a daily documentation review

Xavier’s scheduled task pattern (see Chapter 20). Catches CLAUDE.md drift before it compounds. Worth ten minutes to set up.

Things that changed under the hood

Subagents inherit dynamically-injected MCP servers.
Isolated-worktree subagents can Read/Edit files inside their own worktree.
PostToolUse hooks can replace tool output for all tools (previously MCP-only).
Subagent and SDK MCP server reconfiguration connects servers in parallel, not serially.
Multiagent orchestration: lead agents can delegate to specialist subagents working in parallel on a shared filesystem.

Things you can safely delete from your old setup

Hardcoded claude-opus-4-5-* model strings. Use opus , sonnet , haiku or the latest dated snapshot.
Manual cache_control markers in Claude Code prompts (the default works).
Subagent workarounds for MCP inheritance bugs (the bugs are fixed).
Settings that forced single-session mode in Desktop (the new app is multi-session by default).

Things that are still the same

CLAUDE.md model still loads at session start.
Memory, auto-memory, and the /memory command unchanged.
Permission model (allow-list in settings.local.json) unchanged.
Hook lifecycle events unchanged.
Plugin marketplace install command unchanged.
Agent and skill definition syntax unchanged.

A quick CLAUDE.md upgrade checklist

Open your current CLAUDE.md.
Add the no-AI-slop list (Chapter 19).
Add the em-dash ban.
Add “no command chaining” (Chapter 1).
Update any model references to the 2026 lineup.
Add a “Decision Making” section if your project benefits from adversarial deliberation.
Add the debugging discipline section.
Run /revise-claude-md to catch anything else.
Commit the changes. Reload your session.

Total time: about 15 minutes. ROI: every subsequent session is better.