Claude Code 2.1.325: sandboxed isn't silent now
Two Claude Code releases landed on consecutive days this week, and neither made much noise. Reading the official changelog for 2.1.323 (November 12) and 2.1.325 (November 13), though, they share a theme: places where "allow" quietly meant more than I thought it did are being closed. I run scheduled jobs and /loop with this tool, so I read these entries line by line.
A hook's "allow" no longer overrides a failed check
The change in 2.1.323 I care about most: if a tool's permission check fails after a PreToolUse hook returned "allow", Claude Code now decides the call as if no hook had answered. Before, a hook's yes could be the last word even when the normal check then errored. If you have a hook that auto-approves "safe" commands, the failure case now falls back to the regular flow instead of trusting your hook.
The same release fixed several rule-matching gaps. Write and Edit were able to write to paths covered by a Read deny rule. Bash deny and ask rules missed commands run through sudo and env. An rm was judged from the wrong folder, which skipped the dangerous-removal prompt. Two smaller items: an allow rule that looks like it holds a secret now lasts for the session only and isn't saved, and /security-review plus skills' inline git commands now ask before running outside auto and bypass modes.
Sandboxed no longer means silent
2.1.325 changes a default. Sandboxed Bash commands now follow the normal permission flow. To get the old no-prompt behaviour you set sandbox.autoAllowBashIfSandboxed: true. The release notes also show a permissions.denyInsteadOfPrompting setting, added in 2.1.319, where a call that would wait on a prompt is denied and the turn carries on. Together they look like this:
{
"permissions": { "denyInsteadOfPrompting": true },
"sandbox": { "autoAllowBashIfSandboxed": true }
}
That pairing is the one I'd consider for unattended work: the sandbox is the boundary, and anything outside it is refused rather than left hanging on a prompt nobody will answer. I'm describing what the changelog says the settings do; I haven't benchmarked them.
Also in 2.1.325: read-only git commands in a repository that isn't the project's or a trusted one now ask first, and a Bash allow rule that could run a different program than the one it names is fixed. There is also a security fix for how zstd-compressed responses are decoded. The changelog gives no further detail on that one, so I won't guess at the mechanism.
Scheduled tasks and /loop now ask about skills
This is the entry that touches my own setup. Scheduled tasks and /loop now ask before running a skill that declares allowed-tools or hooks. To run it unattended you add Skill(name) to permissions.allow. The logic holds up: a skill with its own tool grants or hooks is code you effectively approved once, and a schedule is exactly where nobody is watching when it fires.
The practical consequence is that any loop that worked yesterday and invokes such a skill may now stall at a prompt. The fix is deliberate and small, one explicit allow rule per skill you trust.
What I'd do this week
- List every skill your scheduled jobs call, and check which declare
allowed-toolsor hooks. - Add
Skill(name)allow rules only for the ones you've actually read. - If you relied on sandboxed Bash running silently, decide whether to set
autoAllowBashIfSandboxedor accept the prompts. - Review any
PreToolUsehooks that return "allow" and confirm you're happy with the new fallback.
None of this is glamorous, but a permission model earns its keep in these edge cases. I'd rather have a loop stop and ask than one that quietly inherits trust it was never given.