Skip to content

Claude Code

v2.1.236

4 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Claude Code versions and changes →

Claude Code 2.1.236 adds ANTHROPIC_DEFAULT_MODEL, an environment variable that sets the model new sessions start on. A /model pick still overrides it and, unlike with ANTHROPIC_MODEL, persists across restarts, say the August 19 release notes on GitHub. If your team sets ANTHROPIC_MODEL only to give everyone a starting model, the new variable does that job without undoing people’s /model picks at the next restart.

Read-deny rules on macOS

On macOS, wildcard read-deny rules in the sandbox, such as **/.env, now take precedence inside allowed read regions, cover the contents of directories they match, and can’t be bypassed by renaming the denied file. Mac users who rely on such a rule to keep secrets files out of reach of sandboxed commands get that coverage by updating. Read-deny rules are getting a bypass fix again: version 2.1.224 closed one in which entries written with a trailing slash, such as denyRead: "~/.aws/", could be silently bypassed on Linux and macOS.

When a session goes idle

Cross-session SendMessage gets a notify_when_idle option: one Claude Code session can ask another on the same machine to send a single notice when it next goes idle. The option is opt-in and one-shot, needs no polling, and works on macOS and Linux.

A rapid burst of messages to one session is now refused up front once it would exceed what that session’s inbox accepts; previously, the extra messages were reported as sent while they were dropped. Two other ways messages went missing were closed earlier: 2.1.224 stopped “Message sent” from appearing when the write to a teammate’s inbox had failed, and 2.1.235 made SendMessage refuse messages too large for cross-session delivery up front rather than silently drop them. Because typing @ with another session’s name has sent through SendMessage since 2.1.232, a burst of such mentions meets the same refusal.

An idle session whose /goal is parked behind long-running background work now checks in on its own after 30 minutes, then after an hour and after two, instead of waiting for you to return. Version 2.1.234 added a check-in for goals kept waiting 30 minutes or more by background tasks, with CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0 to opt out.

Auto mode, typos, SIGTERM

Auto mode gets three changes. Monitor allow rules are set aside while it is active, so Monitor commands are reviewed the same way Bash commands are. On Bedrock, Vertex AI and Foundry, and when telemetry is disabled, the classifier now uses the same defaults as on the Claude API, including severity-scored classification. And a repo’s status.showUntrackedFiles=no setting can no longer fool the git status check into reporting a clean tree.

Pressing Enter on a mistyped slash command, or on one unavailable in the session, now reports it instead of running the closest fuzzy match; prefixes and aliases still run. In print and SDK mode, SIGTERM no longer records an interrupted turn or synthetic tool denials before exiting. Running commands are still terminated and the process still exits with code 143, so a script that checks for that code sees no change.

Regressions and startup fixes

Since 2.1.229, removing the directory a session had switched into broke clipboard copy, background housekeeping, background sessions and local MCP logs; 2.1.236 fixes that, along with skills hot-reload in SDK and VS Code sessions raising an error on every skills change after the session’s working directory was deleted. A subprocess that fails to start, such as powershell.exe on WSL with Windows interop disabled, no longer causes unhandled promise rejections, a regression from 2.1.234.

A fullscreen renderer that fails to start once now falls back to the classic renderer, where it used to exit on every later launch. VS Code gets screen reader support for the transcript, with live announcements and per-turn heading navigation; /usage shows the usage-credits spend row to Team and Enterprise members; Remote Control marks a session offline within seconds of the CLI exiting; and recaps are capped at 400 characters.

Until this release, the managed-settings approval prompt sometimes failed to appear at startup while still taking the first keypress as approval, so a key pressed for any other reason could approve the settings.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.