Anthropic shipped four Claude Code releases between November 9 and 13, from 2.1.321 through 2.1.325. Coverage went to Code Mode and to Fable 5.2 pricing. The lines that will actually page someone sit in 2.1.325, dated November 13: a set of permission defaults moved toward ask-first, and every one of them lands hardest on runs nobody is watching. I read the changelog end to end and cross-checked it against how the trackers summarized the release. If you run scheduled tasks, /loop, or headless claude -p jobs in CI, four changes concern you directly.
The four defaults that moved
Quoting the changelog, compressed but not paraphrased where it matters:
- Sandboxed Bash commands now "follow the normal permission flow by default". The old behavior, auto-allow inside the sandbox, lives behind a flag now.
- "Scheduled tasks and
/loopto ask before running a skill that declaresallowed-toolsor hooks". An explicit allow rule keeps them unattended. - "Always allow" in Desktop, IDE and SDK hosts offers a rule "only when saving it would allow that command again", so a scheduled task that relied on one can now wait for you.
- Read-only git commands "ask first in a repository that isn't the project's or one you trusted", and the fix is a flag, not a settings key.
None of these come with migration notes. They read like routine changelog lines. Each one converts a silent approval into a prompt, and a prompt in an unattended window is a hang.
Sandboxed bash is no longer a silent approval
{
"sandbox": {
"autoAllowBashIfSandboxed": true
}
}
Before 2.1.325, the deal was simple: the command runs inside the sandbox, so the prompt was skipped. That made the sandbox the fastest way to run CI-style jobs without babysitting. Now the default is the normal flow, ask first, and the flag above is the documented way back.
Two details from the changelog matter for teams. If managed settings turn the sandbox on, the flag has to sit in your highest-ranking managed source, not in a user settings file, or the prompts come back. And read the setting name for what it is: auto-allow bash if sandboxed. Flipping it back on means trusting the container, so check what your sandbox configuration actually blocks before you do.
Scheduled tasks ask before skills that carry hooks
A skill that declares allowed-tools or hooks can run tools without a per-call prompt; that is the point of the declaration. From 2.1.325, scheduled tasks and /loop ask before entering such a skill. The explicit opt-out:
{
"permissions": {
"allow": [
"Skill(deploy-checks)"
]
}
}
Naming the skill in permissions.allow restores the unattended run for that skill only, which is the right shape: you whitelist the skill you reviewed, not all skills forever. If your task runs several skills, each needs its own line, and each line is a small standing decision that someone should be able to explain six months from now.
"Always allow" quietly narrowed
In Desktop, IDE and SDK hosts, a rule is offered only when saving it would allow that command again. The changelog states the consequence itself: "a scheduled task that relied on one can now wait for you." The failure mode is annoying because nothing errors. The task pauses on a prompt nobody answers, and you find out later from a stale branch or a missed report. If you maintain unattended hosts, inventory the rules that were saved from interactive sessions and move the ones that matter into settings files under version control, where a scheduled task can actually rely on them.
Git in untrusted repos
Read-only git commands now ask first when the repository is not the project's own and not one you trusted. The changelog is specific about the mechanism: name it with --add-dir, not the additionalDirectories setting. That distinction does real work. --add-dir records a trust decision for the session, while a settings key would silently widen what never prompts. For agent workflows that hop between checkouts, this is the change most likely to produce a first-run surprise.
Why these shipped together
2.1.325 also fixes "a security issue in how zstd-compressed responses are decoded", and the weekly trackers counted permission and sandbox fixes across the whole November 9 to 13 run, including read-only commands that ran outside the sandbox without a prompt. The direction is consistent with the rest of the month: fewer silent approvals. We documented 1,379 secrets that reached public GitHub files through casually saved don't-ask-again rules; an allow rule written once, in the wrong context, outlives the reason you wrote it.
One more hardening line deserves attention: the sandbox now refuses sandboxed writes to .claude settings, skills, agents, commands and hooks in added directories, as it already did in the launch folder. Your configuration is now protected from the sandboxed process itself. That matters when an agent works in repos that carry MCP servers or plugins you do not fully control, the same exposure we walked through for MCP's git server fixes in v2.4.0.
Audit before you update
- Search every settings file in play for
sandboxandpermissionskeys, and list which of them exist to keep unattended jobs prompt-free. - List your scheduled tasks and
/loopuses. For each skill they run, check its frontmatter forallowed-toolsor hooks, and add aSkill(name)allow rule only for the ones you have actually read. - Inventory "Always allow" rules saved from interactive sessions on hosts that also run unattended work.
- For headless jobs, look at
permissions.denyInsteadOfPromptingfrom 2.1.319 as well. It is the opposite trade: a call that would prompt gets denied and the turn continues. A stalled run and a denied run both fail the task, but the denied one fails fast and visibly, which is easier to catch in CI. It can also produce a quietly incomplete result, so it belongs on jobs where failure is loud either way. - Update on a day when someone can watch the first unattended run.
None of this makes Claude Code harder to use interactively. It makes unattended execution a policy you declare instead of an accident of defaults, and the declarations live in files you can review, diff, and pin. For the rest of the November releases, including the plugin version ranges that now gate installs, see our 2.1.323 breakdown.
Top comments (0)