Skip to content
AI Primer
release

Claude Code 2.1.116 adds 67% faster /resume and safer sandbox rm checks

Claude Code 2.1.116 shipped 24 CLI changes, including faster resume on large sessions, stricter guardrails around rm and rmdir, and automatic plugin dependency installs. It also updates terminal input behavior and model surface area for agent workflows, so teams should upgrade if they rely on the CLI.

3 min read
Claude Code 2.1.116 adds 67% faster /resume and safer sandbox rm checks
Claude Code 2.1.116 adds 67% faster /resume and safer sandbox rm checks

TL;DR

  • Claude Code 2.1.116 shipped 24 CLI changes, and the release thread leads with a /resume speedup of up to 67% on 40MB+ sessions.
  • According to the changelog thread, sandbox auto-allow no longer bypasses danger checks for rm or rmdir targeting /, $HOME, or other critical directories.
  • The changelog thread also says plugin reloads and background updates now auto-install missing dependencies from marketplaces you already added, while MCP startup defers resources/templates/list until the first @ mention.
  • The CLI surface thread shows a small interface shakeup too, with new env vars such as ANTHROPIC_CONFIG_DIR and CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT, plus removal of model entries including claude-haiku and haiku-4.

You can inspect the release page, browse the CLI surface diff, check the metadata diff, and compare the unchanged prompt stats. The odd little detail is that prompt files and prompt tokens stayed flat in the CLI surface thread, even though the bundle still grew by 256 kB.

/resume

The headline change is about session recovery, not model behavior. The release thread says /resume is up to 67% faster on sessions above 40MB and recovers large conversations without getting stuck on dead-fork entries.

The changelog thread pairs that speedup with two fixes that matter on oversized transcripts: /branch no longer rejects conversations above 50MB, and /resume no longer fails by silently opening an empty conversation when a large session file cannot be loaded.

Sandbox and plugin installs

The safety fix is blunt and useful. According to the changelog thread, sandbox auto-allow used to bypass the dangerous-path check for rm and rmdir, and 2.1.116 closes that gap for /, $HOME, and other critical directories.

The same release also patches a few reliability edges around automation:

  • Plugin reloads and background plugin auto-update now install missing marketplace dependencies automatically, per the changelog thread.
  • The Bash tool now shows a hint when gh hits GitHub API rate limits, again in the changelog thread.
  • An intermittent API 400 tied to cache-control TTL ordering during parallel request setup is listed as fixed in the changelog thread.

Terminal and MCP behavior

A lot of the patch notes are terminal paper cuts, but there are enough of them to read as a cleanup pass across the whole CLI surface. The changelog thread lists fixes for Indic-script alignment, Kitty protocol keybindings like Ctrl+- and Cmd+Left/Right, Ctrl+Z hangs when launched through wrappers such as npx or bun run, fullscreen scrolling glitches, modal overflow on short terminals, and duplicate scrollback in inline mode.

The same post adds a few workflow tweaks:

  • MCP startup is faster with multiple stdio servers because resources/templates/list is deferred until the first @ mention, per the changelog thread.
  • /doctor can open while Claude is still responding, according to the changelog thread.
  • The thinking spinner now shows inline progress states such as "still thinking" and "almost done thinking," again in the changelog thread.

CLI surface changes

The CLI surface thread adds the release's least flashy but most reference-worthy inventory. It reports eight new env vars, a new claude-empty- model entry, removal of claude-empty-7, claude-haiku, claude-sonnet-4-, and haiku-4, plus a 256 kB bundle increase only 2 days and 15 hours after 2.1.114.

The same inventory says prompt files and prompt tokens did not change at all in 2.1.116, which points the delta at code and CLI wiring rather than prompt-layer edits in the CLI surface thread.