AI Aulis

AI NEWS IN ENGLISH · 21 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 21 November 2026Paper

New recurring scheduled tasks in Claude Code can run for up to 30 days

AI-generated contentWritten by AI

github.com, huggingface.co, labmemo.com, automater.ai, claude-news.today, code.claude.com · 7 sources · Sources published: 21 November 2026 · Some source publication dates are unknown

Claude Code 2.1.330 lets recurring scheduled tasks and the /loop command run for up to 30 days while the session keeps running. Dangerous rm commands now ask for approval more often, and in unattended runs they are refused.

Anthropic released version 2.1.330 of its coding agent Claude Code on 20 November. The longer limit applies only to new tasks, the company’s release notes say: tasks created before the update still end at seven days. The news site Automater noted that a task meant to run for more than a week must therefore be recreated after the update. Spending limits and notifications for a month-long unattended run also need to be recalibrated, the site said.

More rm commands aimed at the root directory, the home directory, a top-level folder or the working folder now ask for approval. If nobody can answer, the command is refused. As a result, a CI run that previously deleted a build folder may stop midway, Automater said. The site advises adding an allow rule to the settings rather than switching to a more permissive permission mode. The rule should match exactly the command the run needs and should be tested, because a command that does not match it is still refused.

A change to the MCP server deny list affects administrators. When the deniedMcpServers setting blocks a claude.ai connector by name, the block now applies only to that connector, not to others at the same URL. To block them all, administrators must deny the connector’s exact URL. A blocked server disappears from the /mcp view and the claude mcp list output, Claude Code’s documentation says. Automater noted that after the update a name-based block may cover less than before, and the list will not reveal this, because blocked servers are hidden rather than flagged. The documentation also warns that name-based entries are not a security control.

Previously, if claude -p wrote its output to a full disk, it exited with code 0 and no message. It now prints an error and exits with code 1. A full NFS share may still go unnoticed. Spawn-runner hooks on self-hosted runners now start in parallel, by default up to four per orchestrator. If a hook cannot cope with running in parallel, set --hook-concurrency to 1. Most log lines for 403 errors now begin with the reason given by the server, so log alerts that rely on the old wording need updating.

Users in organisations managed in the Enterprise Admin Console get the organisation’s settings and its cloud provider or gateway after signing in. The most important fixes: an MCP server shared by subagents no longer shuts down when one subagent finishes, a worktree is no longer deleted from under another session, and a stopped background subagent stays stopped.

On 21 November, npm’s latest and next tags pointed to version 2.1.330, but the stable tag was still at 2.1.323, the Labmemo blog reported.

Report an error in this story