Claude Code v2.1.218: Quieter Reviews, a Guard on Team Memory, and Two Knobs the Notes Skip#
Published on July 23, 2026
Part of the Claude Code Version Tracker series. | Official Env Vars | Official Changelog
Claude Code v2.1.218 (npm 2026-07-22) is mostly cleanup, and much of it lands where you feel it: a full agent session, a shared memory store, an auto-mode router working turn by turn. The changelog runs past forty items. The binary also carries four new environment variables it never mentions. Here is what actually changes the work.
What ships for you#
Code review stops eating your context. /code-review now runs as a background subagent, so a review no longer floods the conversation you were in the middle of.[1] Skills marked context: fork move to the background by default too, and /deep-research starts only when you ask for it. Heavy work gets its own room.
One keystroke stops wiping your session. The left arrow key could discard the whole conversation with no undo. That is fixed. Multi-line paste no longer folds into a single line with stray j characters where the newlines were, and /context stops reporting stale token counts from before a compaction.
Runaway spend gets a brake. After a context-overflow error, older builds would retry the same doomed request in a loop. That loop is closed, which is real money back on metered plans. Failed MCP connections now show the HTTP status and error text in claude mcp list and /mcp, so a broken server tells you why.
Two security fixes are worth naming. Agent frontmatter hooks no longer run when they live in an untrusted folder, and fork-session lineage survives compaction in headless and SDK runs. Auto mode also interrupts you less: dangerous-rm, background-&, and suspicious-Windows-path checks stopped popping dialogs for calls the analyzer can already clear.
What the binary shows that the notes skip#
Two named variables appear in the build for the first time, and neither reaches the changelog.
CLAUDE_CODE_AUTO_MODE_CLASSIFIER_QUEUE. Auto mode picks a model for every turn using a background classifier. This variable puts those calls in a queue, with a depth and a wait bound (queueDepth, queueWaitMs) sitting beside a cumulativeDroppedTokens counter. When the router gets busy, requests line up instead of piling on, and the build now tracks what it drops. For anyone leaning on auto mode across long sessions, that is the seam between smooth routing and a stalled turn.
CLAUDE_CODE_DISABLE_MEMORY_MASS_DELETE_HOLD. Claude Code memory now syncs to account and team stores, so a delete you make on your machine can travel to everyone else. This variable controls a brake. By default the code holds a bulk delete once it crosses a threshold (max(floor, fraction × count)); set the variable and the threshold jumps to infinity, so the hold never fires. Left alone, it is the seatbelt that keeps one bad wipe from clearing a team's shared memory. You would flip it off only to script large, deliberate cleanups.
Two more arrive as pure server-gated flags, CLAUDE_CODE_GORSE_PLOVER and CLAUDE_CODE_JUNIPER_SUNDIAL. Each is a bare env override backed by a tengu_* gate with no wired behavior visible yet. A third, CLAUDE_CODE_ALDER_WICKET, was in the last build and is gone from this one. Anthropic keeps a steady stream of these noun-codename gates, adding and dropping them between releases, which is the clearest outside read on how much ships dark and flag-controlled before it earns a name.
The read#
The useful half of this release is quiet on purpose: fewer ways to lose a session, a review flow that stays out of your context, a spend loop shut, a couple of security holes closed. The telling half is what the notes leave out. Both silent knobs guard something that recently went shared, auto mode's per-turn routing and team memory, which points at where the load is landing. When a build starts bolting queues and brakes onto a subsystem, that subsystem is getting worked hard.
This analysis is conducted for educational and research purposes under fair use principles. All trademarks and software referenced belong to their respective owners. This content is not intended to infringe on any intellectual property rights, circumvent any protections, or encourage unauthorized access to proprietary systems.
Sources#
- Claude Code Official Changelog — v2.1.218 release notes
Related Versions#
- Claude Code v2.1.217 — Subagent Concurrency and Nesting Caps — hard fan-out limits on the multi-agent runner
- Claude Code v2.1.212 — Session Budget Caps and a Silent /bug Command — per-session subagent and web-search ceilings
- Claude Code v2.1.210 — Org Memory Sync and Session-Ref Propagation — account-scoped memory, remote session refs
- Claude Code v2.1.209 — Auto Mode Grows Up — two-model auto mode, Sonnet 5 permission classifier
- CLAUDE_CODE_EXPERIMENTAL_OBSERVER_AGENTS — The Undocumented Observer Agents Flag — a feature shipped silently in v2.1.200