Claude Code 2.1.325: sandboxed Bash now asks first
Claude Code 2.1.325 makes sandboxed Bash follow the permission flow. Codex 0.169.0 and Copilot CLI 1.0.112 also tightened defaults. Recheck your unattended agent jobs.
Claude Code 2.1.325 changed a default that many unattended jobs depend on. Before this release, a Bash command that ran in the sandbox ran without a prompt. Now it follows the normal permission flow. On the same day, Codex and Copilot CLI also shipped tighter defaults. If you run coding agents in CI or on a schedule, a job that passed on Thursday can now stop and wait for an answer that nobody gives.
What actually happened
Three releases are dated November 13.
- Sandboxed Bash commands follow the normal permission flow by default. To keep the old behavior, set
sandbox.autoAllowBashIfSandboxed: true. For managed settings, set it in your highest-ranking managed source. - Scheduled tasks and
/loopask before they run a skill that declaresallowed-toolsor hooks. To run it unattended, addSkill(name)topermissions.allow. - Read-only git commands ask first in a repository that is not the project's or one you trusted. Name it with
--add-dirto stop the prompt. - The sandbox refuses writes to
.claudesettings, skills, agents, commands and hooks in added directories. - The release also fixes a gap where some compound Bash commands ran without a prompt, and a case where a Bash allow rule could run a different program than the one it names.
Codex 0.169.0: codex exec keeps project configuration and MCP servers disabled in untrusted directories, also with --sandbox workspace-write. The --trust flag on codex exec or codex review saves project trust explicitly.
Copilot CLI 1.0.112: subagents now respect the session-wide credit limit set with --max-ai-credits. A resumed session also counts the credits it used before.
Why Claude Code 2.1.325 matters for your business
All three changes go in one direction. The sandbox is no longer a reason to skip the question. A cloned repository does not get to load its own config. A subagent does not get its own budget. These are correct defaults. We have covered what happens without them: config files that hijack an agent from a downloads folder.
The cost is that your automation can break without an error. A job that waits for approval looks the same as a job that hangs.
What to do:
- Find each place that runs an agent with no person present. CI jobs, cron jobs, scheduled tasks and
/loopsessions. - Run each one after the upgrade and read the log. Look for a prompt, a skipped MCP server or a skipped skill.
- Add narrow rules, not the old default. Allow
Skill(name)for the one skill the job needs. Do not setautoAllowBashIfSandboxedeverywhere to make the prompts stop. - Trust directories explicitly. Use
--trustin Codex and--add-dirin Claude Code for the repositories you own. - Pin the agent version in CI. Upgrade on your schedule, after you read the release notes.
Key takeaways
- Claude Code 2.1.325: sandboxed Bash follows the permission flow unless
sandbox.autoAllowBashIfSandboxedis true - Scheduled tasks and
/loopask before a skill withallowed-toolsor hooks runs - Codex 0.169.0:
codex execignores project config and MCP servers in untrusted directories - Copilot CLI 1.0.112: subagents count against
--max-ai-credits - Test each unattended agent job after the upgrade, and pin versions in CI
An agent pipeline needs the same change control as the code it writes. We build coding agent workflows with pinned versions, narrow permission rules and a test run before each upgrade. See how we build agent pipelines, or ask us to review your unattended jobs.
Sources: Claude Code v2.1.325 release notes, Codex 0.169.0 release notes, Copilot CLI 1.0.112 release notes.
- #claude-code
- #codex
- #copilot-cli
- #coding-agents
- #permissions
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
TSMC employee data breach traced to a cloud provider
TSMC confirmed an employee data breach in an SEC filing. The leaked data matches records at a cloud service provider. Inventory the HR data your vendors hold.
Read itSignDrop Windows zero-day: SYSTEM access, no patch, no CVE
The SignDrop Windows zero-day uses CLFS and Defender to give a standard user SYSTEM rights. No patch or CVE exists. Keep CLFS authentication on and cut admin rights.
Read it