Claude Code 2.1.218 Flips Fork Skills to Background

Claude Code's July 22 release quietly flipped a default: any skill with context: fork in its frontmatter now runs in the background, per the v2.1.218 release notes.

If you wrote a fork skill expecting it to block while it worked, it now returns control immediately. Whatever step consumed its output reads nothing, and nothing errors: the skill succeeds, and the next step improvises with an empty hand-off.

The opt-out is background: false, set per skill. The change pairs with an earlier one: 2.1.215 stopped /verify and /code-review from firing automatically. Together they make the trade explicit: you invoke review tools on purpose, and they cost zero main-thread context in return.

For /code-review that's a real gain. Its output is verbose, and it used to bury the actual conversation under a wall of findings you'd already read.

The same release closed an escalation path. Hooks declared in an agent file's frontmatter no longer run unless that file's folder has accepted workspace trust, per the CHANGELOG. Before this, cloning a repo with a vendored plugin directory could run hooks on load.

The team laid out its philosophy in a Simon Willison fireside chat. Every tool needs "a distinct function from every other tool," and they're trending toward fewer tools, not more. The file edit tool survives that bar because it's also a permission surface, not because it wraps a capability the model already has.

The take

Grep ~/.claude/skills/ for context: fork and audit every hit. Anywhere a downstream step consumes that skill's output, set background: false. Both fixes, the background default and the hook-trust boundary, landed in the same July 22 release, v2.1.218.