Claude Code Daily Briefing - 2026-11-14

Release Summary

VersionDateKey Changes
v2.1.32511/13Over 140 changes, including a security fix for decoding zstd-compressed responses, switching sandboxed Bash commands to require permission checks by default, and tightened handling of read-only git commands in untrusted repos
v2.1.32311/12Added the experimental opt-in Code Mode feature, version-range enforcement via engines["claude-code"] in plugin.json, and several sandbox security fixes (covered in the 11/13 briefing)

v2.1.325 arrived on 11/13, just a day after v2.1.323 (11/12). There are no notable new feature announcements, but it’s a substantial release with over 140 changes centered on stability and security - including a fix for a security flaw in decoding zstd-compressed responses and a change that makes sandboxed Bash commands go through the standard permission-check flow by default.

Full release notes


New Features & Practical Usage

No new Anthropic product or model announcements today

A direct check of Anthropic’s official newsroom confirms there have been no new product or model feature announcements since the Claude Fable 5.2 / Mythos 5.2 release on 11/9 and the Nasdaq IPO roadshow kickoff on 11/10 (both covered in earlier briefings), through 11/14. v2.1.325 itself is also a release focused on security and stability fixes rather than new features. Instead, today’s Developer Workflow Tips section covers the sandbox and permission changes in v2.1.325, and the Security & Limitations section covers the state of the 11/13 service incidents.


Developer Workflow Tips

Sandboxed Bash commands now require permission checks by default (v2.1.325)

Previously, Bash commands run inside the sandbox executed immediately without approval; starting with this version, they now go through the standard permission-check flow like any other command. The sandbox still limits the blast radius of a command, but simply running inside it no longer automatically skips approval. If you want commands to keep running without prompts as before, you’ll need to change a setting.

{
  "sandbox": {
    "autoAllowBashIfSandboxed": true
  }
}

If your organization enforces the sandbox via managed settings, this value needs to go into your highest-priority managed settings source for commands to keep running without prompts.

If you run Claude Code in sandbox mode and your workflow assumes commands execute immediately without approval, know that the sudden increase in permission prompts after upgrading to v2.1.325 is caused by this change - and only restore the old behavior with the setting above if you actually need to. Full release notes

Read-only git commands in untrusted repos now also require confirmation (v2.1.325)

For repos that aren’t your own project and haven’t been explicitly trusted, even read-only git commands now prompt for confirmation first. If you want to keep working with such a repo without prompts, you need to pass that folder explicitly via --add-dir rather than the additionalDirectories setting.

# To run git commands in another untrusted repo without prompts
claude --add-dir ../other-repo

If you’ve been keeping another repo outside your monorepo open and regularly running read-only commands like git log or git diff against it, starting with v2.1.325 you’ll need to explicitly pass that repo via --add-dir to avoid the confirmation prompt every time. Full release notes

Scheduled tasks and /loop now ask for confirmation before running unattended skills (v2.1.325)

When a scheduled task or /loop tries to run a skill that declares allowed-tools or hooks, it now prompts for confirmation first. If you want it to keep running fully unattended, you need to explicitly add that skill to your allow list.

Add Skill(skill-name) to permissions.allow → runs unattended without confirmation

If you’ve hooked a skill with hooks into a scheduled task or a self-paced /loop to run overnight, failing to add this setting ahead of the v2.1.325 upgrade could leave the next run stuck waiting on a confirmation prompt. Full release notes


Security & Limitations

Claude service status - two incidents on 11/13 (auth errors, elevated errors), and the Microsoft 365 connector incident is resolved

A direct check of the official status.claude.com confirms two incidents affected claude.ai and Claude Code on 11/13 (both now resolved). The first was an authentication error affecting claude.ai and Claude Code, lasting from 09:58 to 11:58 UTC (about 2 hours). The second was a broader increase in errors across claude.ai (including mobile), the Claude API, and Claude Code, lasting from 13:09 to 15:38 UTC (about 2.5 hours), which also affected the Docs and Sheets connectors, Claude Security, Claude Managed Agents, and Claude Meetings - causing some scheduled tasks and routines to run late or get skipped. Separately, the Microsoft 365 connector incident (which started on 11/12 and was still listed as “ongoing” in the 11/13 briefing) was resolved at 16:40 UTC on 11/13, for a total duration of about 24.5 hours. GeekNews also saw posts from users who experienced these outages, sharing stories of work getting cut off right before a deadline, with comments advising against relying on a single vendor and recommending keeping another model or a local model as a backup.

If you experienced authentication errors or slow responses on claude.ai or Claude Code over the past few days, it was likely due to these now-resolved 11/13 incidents - and if you keep running into repeated outages on critical work, it’s worth considering a backup path with another model or a local model, as this case illustrates. Claude Status · GeekNews

Security flaw in decoding zstd-compressed responses fixed (v2.1.325)

This release only states that a security flaw in the process of decoding zstd-compressed responses has been fixed; no specific attack scenario or scope of impact has been disclosed.

Even for a security fix with no disclosed details, it’s safer not to shrug it off - make sure you’re running v2.1.325 or later, where this flaw has been fixed. Full release notes


Ecosystem & Plugins

BeatBrowser - an MIT-licensed tool that turns your existing Chrome profile into Claude Code’s agent browser (11/13)

Instead of spinning up a separate browser and logging in again, or exporting and importing cookies, this tool lets an agent drive your browser while keeping your existing Chrome profile and login sessions intact. It’s made up of a Chrome extension, a local bridge, and an MCP server, and Claude Code, Codex, and OpenCode connect to it over a local stdio MCP to use tools for reading pages, navigation, interaction, and networking. The agent works in a background window, a human can pause or take over at any time, and CAPTCHAs, QR logins, and OTPs are handed off to the human. Local use is free and requires no account or API key.

# GitHub: BeatAPI/BeatBrowser

If you’ve needed browser automation in Claude Code but found spinning up a dedicated browser or moving session cookies around a hassle, it’s worth trying connecting your agent directly to your already-logged-in Chrome instead. GitHub · GeekNews

D:bo - a service that connects your personal design judgment to Claude Code and Codex via an API (11/13)

Instead of re-explaining design instructions like “keep it minimal” or “highlight just one button” every time you build a UI, this service has you answer 15 visual-preference questions once to build your own design profile, then reuse it across tools. The resulting profile can be downloaded for free as a designPersona.md file for use in Cursor or ChatGPT, or connected via API so that Claude Code and Codex pull in and apply that profile directly. It doesn’t yet learn from every choice you make, and the creator says they plan to refine the profile based on future feedback.

Separate from shared project rules or a design system, if you’re tired of re-explaining your own taste to Claude Code every time, it’s worth testing directly whether this kind of personal design-profile connection holds up against Rules or a plain Markdown file. D:bo · GeekNews


Community News


Minor Changes

All of the following are from v2.1.325 (11/13).



Interesting Projects & Tools