1
00:00:00,260 --> 00:00:05,620
G'day, this is the Claude Code Changelog,
brought to you by Jellypod AI.

2
00:00:06,280 --> 00:00:06,820
I'm Lachlan.

3
00:00:07,640 --> 00:00:09,220
Here's a situation you might know.

4
00:00:09,780 --> 00:00:15,400
You flip on auto mode, Claude's cruising
along, and then it stops dead because it

5
00:00:15,420 --> 00:00:19,020
wants to push to a sibling repo in your
own GitHub org.

6
00:00:19,600 --> 00:00:20,660
Nothing wrong with the push.

7
00:00:21,080 --> 00:00:24,420
The classifier just has no idea that repo
is yours.

8
00:00:25,335 --> 00:00:27,395
And that's the default behavior, by
design.

9
00:00:27,735 --> 00:00:31,935
Per the auto mode docs, the classifier
trusts only your working directory and the

10
00:00:31,995 --> 00:00:33,595
current repo's configured remotes.

11
00:00:34,215 --> 00:00:37,795
Everything else, your company's other
repos, a team S3 bucket,

12
00:00:38,195 --> 00:00:41,975
an internal API domain, counts as
potentially outside your boundary.

13
00:00:42,675 --> 00:00:49,275
Claude Code 2.1.304, released October 19,
2026, makes the fix for that available in

14
00:00:49,315 --> 00:00:50,155
a lot more places.

15
00:00:50,695 --> 00:00:54,235
The slash auto mode setup command now
works on Amazon Bedrock,

16
00:00:54,455 --> 00:00:58,735
Google Vertex AI, Microsoft Foundry,
Claude apps gateway sessions,

17
00:00:58,895 --> 00:01:00,455
and telemetry disabled sessions.

18
00:01:01,200 --> 00:01:08,140
Which ties back to the episode we did on
2.1.302 and CLAUDE_CODE_AUTO_MODE_MODEL,

19
00:01:08,620 --> 00:01:12,280
pinning the client side classifier model
on Bedrock, Vertex,

20
00:01:12,340 --> 00:01:13,360
and proxy setups.

21
00:01:14,120 --> 00:01:17,680
That gave those teams control over which
model judges the actions.

22
00:01:18,360 --> 00:01:21,240
This gives them a way to tell that
classifier what's trusted,

23
00:01:21,660 --> 00:01:22,900
without hand writing it.

24
00:01:23,487 --> 00:01:23,687
Right.

25
00:01:23,787 --> 00:01:25,647
So what does the command actually do?

26
00:01:26,107 --> 00:01:31,367
You run slash auto mode setup, and Claude
Code drafts autoMode.environment entries,

27
00:01:31,527 --> 00:01:35,227
and sometimes rule entries too, from your
project and your recent sessions.

28
00:01:35,847 --> 00:01:38,287
Those entries are plain prose, not regex.

29
00:01:38,527 --> 00:01:44,187
Things like "Source control:
github.example.com/acme-corp and all repos under it",

30
00:01:44,747 --> 00:01:51,367
or "Trusted cloud buckets:
s3://acme-build-artifacts", or "Trusted

31
00:01:51,427 --> 00:01:54,507
internal domains: star dot corp dot
example dot com".

32
00:01:55,522 --> 00:01:57,022
What does it read to get there?

33
00:01:57,802 --> 00:02:03,282
Always your project's CLAUDE.md, README,
config files, and git remotes.

34
00:02:03,882 --> 00:02:07,282
Your existing autoMode and
permissions.allow settings.

35
00:02:07,822 --> 00:02:12,402
And the hosts, buckets, and command names
from commands Claude ran in recent

36
00:02:12,462 --> 00:02:13,722
sessions in that project.

37
00:02:14,382 --> 00:02:17,022
The docs are explicit that it's never your
messages.

38
00:02:17,842 --> 00:02:20,302
Then there are two optional scans you opt
into.

39
00:02:21,102 --> 00:02:24,202
One reads the first word of each command
in your shell history.

40
00:02:24,742 --> 00:02:28,462
The other reads the remote hosts and names
of repositories under your home

41
00:02:28,502 --> 00:02:28,962
directory.

42
00:02:29,968 --> 00:02:34,388
Then you get a draft, and here's the thing
to know: you accept or discard it as a

43
00:02:34,448 --> 00:02:34,748
whole.

44
00:02:35,248 --> 00:02:37,228
No picking single lines in the dialog.

45
00:02:37,788 --> 00:02:42,968
So accept, then open
~/.claude/settings.json and trim anything too generous.

46
00:02:43,588 --> 00:02:49,328
When it saves, it writes to that user
settings file, and it includes "$defaults" in

47
00:02:49,368 --> 00:02:53,908
the allow, soft deny, and hard deny lists
it touches, so the built in rules you

48
00:02:53,948 --> 00:02:55,468
haven't replaced stay active.

49
00:02:56,048 --> 00:02:59,808
It'll also offer to remove
permissions.allow rules that auto mode ignores,

50
00:03:00,048 --> 00:03:01,648
like a blanket Bash star.

51
00:03:02,318 --> 00:03:05,138
After that, run claude auto mode config.

52
00:03:05,558 --> 00:03:09,258
That prints the effective rules so you can
confirm your entries landed.

53
00:03:09,898 --> 00:03:14,738
And if you're curious what you're layering
on top of, claude auto mode defaults

54
00:03:15,158 --> 00:03:16,338
prints the built in entries.

55
00:03:17,397 --> 00:03:19,497
Now the caveats, and there are a few.

56
00:03:19,957 --> 00:03:20,857
First, scope.

57
00:03:21,237 --> 00:03:26,397
The classifier does not read autoMode from
.claude/settings.json or

58
00:03:26,537 --> 00:03:29,617
.claude/settings.local.json in the repo.

59
00:03:30,137 --> 00:03:34,057
That's on purpose, because a checked in
repo or a build step could otherwise inject

60
00:03:34,097 --> 00:03:35,157
its own allow rules.

61
00:03:35,617 --> 00:03:39,477
So for a team, you'd distribute trusted
infrastructure through managed settings

62
00:03:39,517 --> 00:03:39,957
instead.

63
00:03:40,657 --> 00:03:46,537
Second, it needs a Pro, Max, or Team plan
and Claude Code 2.1.228 or later,

64
00:03:46,977 --> 00:03:48,537
and it can't run in a cloud session.

65
00:03:49,480 --> 00:03:52,440
Third, ask and deny rules still win.

66
00:03:52,880 --> 00:03:58,620
Put Bash(git push star) in permissions.ask
and you'll still get a prompt in auto

67
00:03:58,680 --> 00:04:01,480
mode, because those rules run before the
classifier.

68
00:04:02,180 --> 00:04:04,420
Handy if you want a human checkpoint on
pushes.

69
00:04:05,340 --> 00:04:09,480
And if the setup dialog nags you, "Teach
auto mode about your environment?",

70
00:04:10,000 --> 00:04:14,240
you can switch off both the offer and the
command with a skillOverrides entry in

71
00:04:14,300 --> 00:04:17,779
your user settings: auto mode setup, set
to off.

72
00:04:18,729 --> 00:04:20,130
Quick hits from the same release.

73
00:04:20,630 --> 00:04:25,090
The Ctrl+O transcript view now lists the
names of tools called earlier in the

74
00:04:25,130 --> 00:04:26,150
compaction summary.

75
00:04:26,870 --> 00:04:30,890
On Pro and Max, you get a warning when
resuming an inactive conversation would eat a

76
00:04:30,970 --> 00:04:34,750
noticeable share of your five hour limit,
with the option to start fresh and let

77
00:04:34,810 --> 00:04:35,970
Claude look up the old one.

78
00:04:36,867 --> 00:04:37,907
Two fixes worth knowing.

79
00:04:38,467 --> 00:04:45,307
CLAUDE_CODE_AUTO_COMPACT_WINDOW set to
200k used to silently become a 100K

80
00:04:45,347 --> 00:04:45,747
window.

81
00:04:46,187 --> 00:04:50,607
It now accepts k and M suffixes and warns
when a value's out of range.

82
00:04:51,507 --> 00:04:56,847
And self paced loop wakeups scheduled at
the one hour maximum were firing just after

83
00:04:56,887 --> 00:05:02,047
the one hour prompt cache expired, so they
re read the whole conversation uncached.

84
00:05:02,487 --> 00:05:03,147
That's fixed too.

85
00:05:04,027 --> 00:05:08,387
Also, Code Review owners can now pick
Quick, Standard, or Deep levels.

86
00:05:08,987 --> 00:05:10,127
That's the changelog for today.