Source: github.com/anthropics/claude-code/releases/tag/v2.1.325
The anthropics/claude-code release fixes a security issue in how zstd-compressed responses are decoded. Most of the other entries cover sessions, MCP servers, permission rules and sign-in flows, including a fix for some deny and ask rules that did not apply to certain Bash commands.
The most significant behavior change concerns sandboxed Bash. Those commands now follow the normal permission flow unless sandbox.autoAllowBashIfSandboxed is set to true. Scheduled tasks and /loop also now ask before running a skill that declares allowed-tools or hooks, unless Skill(name) is added to permissions.allow.
Remote Control runs now exit with status 4 when they stop on their own and 69 when they cannot register, so a service manager can restart them. Subagents cut off at the output token limit are asked again for a complete answer.
Why it matters if you build agents
Unattended scheduled tasks and /loop jobs that run skills with hooks or allowed-tools will pause for approval until Skill(name) is added to permissions.allow. Teams that expect sandboxed Bash to run without prompts should set sandbox.autoAllowBashIfSandboxed to true. Anyone supervising claude remote-control with a service manager should handle exit codes 4 and 69.
Get stories like this every weekday
One short email with the day's agent tooling news.