1
00:00:00,260 --> 00:00:05,302
Quick heads up before you run your next
sandboxed script, because it might just sit

2
00:00:05,382 --> 00:00:06,382
there waiting on you.

3
00:00:07,302 --> 00:00:11,383
This is The Claude Code Changelog, brought
to you by Jellypod AI.

4
00:00:13,344 --> 00:00:19,305
Version 2.1.325 changed a default that a
lot of people were leaning on.

5
00:00:19,993 --> 00:00:22,513
Yeah, and it connects to the sandbox
coverage we did earlier.

6
00:00:23,454 --> 00:00:29,095
Sandboxing isolated the OS boundary, macOS
Seatbelt, Linux and WSL2 bubblewrap.

7
00:00:29,715 --> 00:00:36,517
But up to 2.1.323 and 324, turning on
sandbox.enabled true quietly turned on

8
00:00:36,737 --> 00:00:37,678
auto allow too.

9
00:00:39,338 --> 00:00:43,479
Sandboxed Bash commands ran without a
prompt, unless a content scoped ask or an

10
00:00:43,499 --> 00:00:44,680
explicit deny matched.

11
00:00:45,568 --> 00:00:47,769
In 2.1.325, that's flipped.

12
00:00:48,229 --> 00:00:50,789
Sandboxed Bash now goes through the normal
permission flow.

13
00:00:51,130 --> 00:00:55,591
So you need an approval, or a matching
allow rule, or you opt back in explicitly.

14
00:00:56,651 --> 00:00:58,852
The reasoning is separation of concerns.

15
00:00:59,612 --> 00:01:01,772
The sandbox bounds what a command can
reach.

16
00:01:02,313 --> 00:01:04,933
Permissions decide whether Claude Code
should run it at all.

17
00:01:05,473 --> 00:01:07,994
Those are two layers, and they shouldn't
be silently merged.

18
00:01:08,664 --> 00:01:12,445
Which is fair, but if you had a hands off
setup, it's gonna bite.

19
00:01:13,605 --> 00:01:17,066
Scripts that assumed no prompts will stall
waiting for approval,

20
00:01:17,486 --> 00:01:19,887
or just fail in non interactive
environments.

21
00:01:21,287 --> 00:01:21,927
So what's the fix?

22
00:01:22,523 --> 00:01:23,023
One setting.

23
00:01:23,363 --> 00:01:28,204
In your user settings or project .claude
settings file, you write sandbox,

24
00:01:28,424 --> 00:01:32,045
enabled true, and autoAllowBashIfSandboxed
true.

25
00:01:32,545 --> 00:01:33,625
That restores the old behavior.

26
00:01:34,066 --> 00:01:36,926
And if you just want to test it, you can
pass it ad hoc: claude,

27
00:01:37,066 --> 00:01:40,927
then the settings flag, with a JSON string
like sandbox enabled true,

28
00:01:41,447 --> 00:01:43,328
autoAllowBashIfSandboxed true.

29
00:01:44,280 --> 00:01:46,341
And for teams that roll this out
centrally?

30
00:01:47,241 --> 00:01:50,281
Then it needs to be in the highest
priority managed settings tier.

31
00:01:50,802 --> 00:01:54,782
If your org enforces the sandbox but
leaves that flag out, developers get a prompt

32
00:01:54,862 --> 00:01:57,203
on every command that isn't covered by
permissions allow.

33
00:01:57,963 --> 00:02:02,144
And in some tiers, local settings can't
override the managed lock to quiet it.

34
00:02:02,144 --> 00:02:05,885
So check your managed config first,
because that's where the complaints will land.

35
00:02:06,660 --> 00:02:08,720
There's a related one for scheduled work.

36
00:02:09,260 --> 00:02:16,042
In 2.1.325, scheduled tasks and /loop now
prompt before

37
00:02:16,102 --> 00:02:19,703
running any skill that declares allowed
tools or hooks.

38
00:02:20,603 --> 00:02:22,923
If you want those to run in the background
without stalling,

39
00:02:23,344 --> 00:02:27,785
add Skill, then the skill name in
parentheses, to permissions allow.

40
00:02:28,288 --> 00:02:33,649
2.1.325 also fixed three Bash permission
gaps, and they matter here.

41
00:02:34,149 --> 00:02:37,630
Deny and ask rules weren't applied to
certain compound Bash commands,

42
00:02:38,070 --> 00:02:38,990
and now those ask.

43
00:02:39,590 --> 00:02:43,412
A Bash allow rule could run a different
program than the one named in the rule.

44
00:02:44,912 --> 00:02:49,394
And relative deny or ask rules, like Read
dot slash dot env,

45
00:02:50,115 --> 00:02:51,935
failed to cover the project after a cd.

46
00:02:52,636 --> 00:02:53,096
All fixed.

47
00:02:53,916 --> 00:02:54,576
One caveat.

48
00:02:55,877 --> 00:03:00,839
The Bash sandbox is still macOS, Linux and
WSL2 only.

49
00:03:01,499 --> 00:03:04,260
Native Windows runs commands unsandboxed.

50
00:03:04,741 --> 00:03:07,482
So this default flip is about those three
platforms.

51
00:03:08,299 --> 00:03:09,320
A few smaller things.

52
00:03:09,920 --> 00:03:14,342
In VS Code, the integrated terminal now
shows a spinner tip reminding you to re run

53
00:03:14,422 --> 00:03:17,604
notebook cells after Claude reads or edits
a Jupyter notebook.

54
00:03:18,324 --> 00:03:22,726
When a subagent hits its output token
limit, Claude Code now asks it to regenerate

55
00:03:22,746 --> 00:03:27,508
the whole answer concisely, instead of
handing back a clipped tail.

56
00:03:27,542 --> 00:03:27,883
Nice.

57
00:03:28,563 --> 00:03:34,186
Also, pressing Esc on a background MCP
task prompt now marks it answer later.

58
00:03:34,786 --> 00:03:39,728
The question stays in /tasks, and nothing
gets killed or sent as an empty reply.

59
00:03:40,709 --> 00:03:46,471
And when org policy blocks a request, you
now get guidance: /rewind to edit the

60
00:03:46,491 --> 00:03:49,253
prompt, or /clear to reset state.

61
00:03:50,513 --> 00:03:53,474
Also, don't keep persistent files in the
scratchpad folder,

62
00:03:53,895 --> 00:03:55,835
since it can get wiped between operations.

63
00:03:56,482 --> 00:04:01,804
So the takeaway: if your sandboxed
workflow suddenly asks for approval after

64
00:04:01,864 --> 00:04:08,586
updating, add autoAllowBashIfSandboxed
true, deliberately, knowing the sandbox is no

65
00:04:08,686 --> 00:04:09,947
longer deciding that for you.

