Claude docs changes for May 4th, 2026 [diff]

Executive Summary

LLM hooks now have clarified ok: false behavior: for Stop/SubagentStop hooks, Claude keeps working using the reason as its next instruction; for all other events, the turn ends and the reason appears as a chat warning that Claude does not see
autoMemoryDirectory security restriction tightened — now rejected from both project and local settings (not just project settings), since either file in a cloned repo could redirect memory writes to sensitive paths
npm update command for Claude Code corrected from npm update -g to npm install -g @anthropic-ai/claude-code@latest

None

The npm command to update Claude Code was corrected to npm install -g @anthropic-ai/claude-code@latest (previously npm update -g @anthropic-ai/claude-code). [line 389] [Source]

Clarified that ok: false behavior differs by hook type: for Stop and SubagentStop hooks the reason is fed back to Claude and the turn continues; for all other events the turn ends and the reason appears in chat as a warning line that Claude does not see. [line 701] [Source]

Added explanation of the LLM hook ok: false semantics: for Stop and SubagentStop the reason is returned to Claude as its next instruction and the turn continues; for all other supported events, the turn ends and the reason appears as a chat warning line. Also notes this is equivalent to "continue": false from a command hook, and points to command hook fields for alternative blocking semantics. [lines 2318-2320] [Source]

Clarified that autoMemoryDirectory must be set in user settings at ~/.claude/settings.json specifically. [line 298] [Source]
The value must now be documented as requiring an absolute path or a ~/-prefixed path. Security restriction expanded: the setting is now rejected from both project and local settings (previously only project settings), since a cloned repository could supply either file to redirect auto memory writes to sensitive locations. The --settings flag is explicitly listed as an accepted source. [lines 301-303] [Source]

Updated autoMemoryDirectory description to reflect the expanded security restriction: now accepted from policy and user settings, and from the --settings flag; rejected from project or local settings to prevent cloned repos from redirecting memory writes to sensitive locations. [line 159] [Source]

No significant changes. (Only obfuscated contact email link updates in adaptive-thinking.md, extended-thinking.md, and files.md.)