WHY this matters
v2.1.216 ships one useful new knob and two silent breaking changes worth knowing about.
🚀 New: sandbox.filesystem.disabled
Set in user/managed/CLI settings (NOT project settings — can't self-widen) to sandbox network without filesystem restrictions. Lets subprocesses touch the host filesystem while network egress stays locked to allowedDomains. Docs include a new warning: with auto-allow on, sandboxed commands can now write shell startup files → escalation risk.
⚠️ Breaking: hooks continueOnBlock behavior flip
PreToolUse deny used to return the reason to Claude as a tool error and continue the turn. Now by default the turn ends and reason appears as a chat warning. Add continueOnBlock: true to restore old behavior. PostToolUse also gains this flag.
Who's affected: anyone with deny hooks who expected Claude to see the reason and adapt.
⚠️ Breaking: /rewind skips symlinks & hard-links
Now silently skips symlinked/hard-linked files with a "Restored the code, but skipped N files" warning instead of writing through links. Affects dotfile-manager setups and pnpm hard-links. Use /debug before restore to see skipped paths.
Other
- Chrome integration now requires
/login auth — API keys / setup-token sessions no longer connect (previously tried and got 403)
- Headless JSON-stream now waits 30s for slow consumers (was ~2s) — fixes truncated large responses
CLAUDE_CODE_OTEL_CONTENT_MAX_LENGTH env var for OTEL inline truncation limit (was hardcoded 60KB, requires v2.1.214+)
PR: https://github.com/anthropics/claude-code-docs/pull/1079 (internal fetch PR)
WHY this matters
v2.1.216 ships one useful new knob and two silent breaking changes worth knowing about.
🚀 New:
sandbox.filesystem.disabledSet in user/managed/CLI settings (NOT project settings — can't self-widen) to sandbox network without filesystem restrictions. Lets subprocesses touch the host filesystem while network egress stays locked to
allowedDomains. Docs include a new warning: with auto-allow on, sandboxed commands can now write shell startup files → escalation risk.continueOnBlockbehavior flipPreToolUsedeny used to return the reason to Claude as a tool error and continue the turn. Now by default the turn ends and reason appears as a chat warning. AddcontinueOnBlock: trueto restore old behavior.PostToolUsealso gains this flag.Who's affected: anyone with deny hooks who expected Claude to see the reason and adapt.
/rewindskips symlinks & hard-linksNow silently skips symlinked/hard-linked files with a "Restored the code, but skipped N files" warning instead of writing through links. Affects dotfile-manager setups and pnpm hard-links. Use
/debugbefore restore to see skipped paths.Other
/loginauth — API keys /setup-tokensessions no longer connect (previously tried and got 403)CLAUDE_CODE_OTEL_CONTENT_MAX_LENGTHenv var for OTEL inline truncation limit (was hardcoded 60KB, requires v2.1.214+)PR: https://github.com/anthropics/claude-code-docs/pull/1079 (internal fetch PR)