1
00:00:00,280 --> 00:00:06,982
This one's brought to you by Jellypod AI,
and it's about Claude Code 2.1.323,

2
00:00:07,442 --> 00:00:09,743
which landed November 12th, 2026.

3
00:00:10,883 --> 00:00:17,305
The top line is Code Mode, an experimental
opt-in where Claude writes JavaScript to

4
00:00:17,345 --> 00:00:21,006
call its tools, instead of sending every
call as a separate tool call.

5
00:00:21,733 --> 00:00:22,733
Here's why that matters.

6
00:00:23,273 --> 00:00:26,614
Say the agent needs to read twenty files
and run three checks.

7
00:00:27,075 --> 00:00:31,276
Today that's twenty three round trips, and
each one has a model response in between.

8
00:00:32,416 --> 00:00:36,057
With Code Mode, it can write one small
script with loops and conditions,

9
00:00:36,097 --> 00:00:38,438
and handle the intermediate results in a
single pass.

10
00:00:39,338 --> 00:00:41,999
The reported benefit is fewer tokens and
less waiting.

11
00:00:43,562 --> 00:00:45,322
Now, the how-to is a bit thin, mate.

12
00:00:45,802 --> 00:00:52,584
The changelog points to code.claude.com
slash docs slash code mode for turning it

13
00:00:52,644 --> 00:00:55,124
on, or off for a whole organization.

14
00:00:55,964 --> 00:01:01,065
But one write-up, from clauding.de,
fetched that page and got a 404.

15
00:01:01,565 --> 00:01:03,665
So the docs page looks like it's still
coming.

16
00:01:04,386 --> 00:01:07,806
It's off by default, so you'll want to
check there before you flip anything.

17
00:01:08,614 --> 00:01:09,635
And the caveat is real.

18
00:01:10,375 --> 00:01:13,615
Code that calls tools is harder to review
than a list of calls.

19
00:01:14,535 --> 00:01:16,516
An allow list tells you which tools can
run.

20
00:01:16,896 --> 00:01:18,816
It doesn't tell you what a script does
with them.

21
00:01:19,296 --> 00:01:22,937
That's the author's interpretation, but
it's why Anthropic marked it experimental

22
00:01:23,057 --> 00:01:24,397
and gave orgs an off switch.

23
00:01:25,024 --> 00:01:29,865
Okay, but the bits more likely to bite you
this week are the defaults.

24
00:01:30,545 --> 00:01:36,287
First one: for a model ID Claude Code
doesn't recognize, the assumed context window

25
00:01:36,427 --> 00:01:40,528
jumps from 200,000 tokens to one million.

26
00:01:40,542 --> 00:01:44,443
Which is nasty if you're routing a smaller
third party model through a gateway.

27
00:01:45,464 --> 00:01:49,605
Once the session passes its real limit, it
can fail on every turn until you run

28
00:01:50,125 --> 00:01:50,445
/clear.

29
00:01:51,045 --> 00:01:57,567
The fix is to set
CLAUDE_CODE_DISABLE_1M_CONTEXT equals 1 in your environment,

30
00:01:58,828 --> 00:02:00,688
and that keeps the 200,000 assumption.

31
00:02:01,168 --> 00:02:02,969
Do it before you update, not after.

32
00:02:03,592 --> 00:02:04,032
Second one.

33
00:02:04,452 --> 00:02:08,573
The number of messages sessions on the
same machine can send each other without

34
00:02:08,734 --> 00:02:13,675
asking you goes from roughly twenty to a
hundred thousand.

35
00:02:13,675 --> 00:02:15,276
That applies to Claude Desktop too.

36
00:02:15,876 --> 00:02:20,597
If you run agents talking to each other,
that old brake after twenty messages is

37
00:02:20,677 --> 00:02:21,497
basically gone.

38
00:02:21,937 --> 00:02:23,058
So set your own limits.

39
00:02:23,998 --> 00:02:25,719
Plugin authors get something useful.

40
00:02:26,299 --> 00:02:31,421
plugin.json now accepts engines, with the
key claude-code.

41
00:02:31,421 --> 00:02:34,262
A plugin outside its stated version range
won't be installed,

42
00:02:34,362 --> 00:02:37,983
updated or loaded, and install picks the
newest version that fits.

43
00:02:38,843 --> 00:02:43,785
Also, claude plugin validate now type
checks a plugin's TypeScript hooks module.

44
00:02:44,325 --> 00:02:47,686
Findings are warnings, and the strict flag
turns them into failures,

45
00:02:47,946 --> 00:02:49,427
so it slots straight into CI.

46
00:02:50,343 --> 00:02:55,805
There's also
CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES, available since

47
00:02:55,905 --> 00:02:58,026
2.1.318.

48
00:02:58,706 --> 00:03:04,008
Set it to 1 and you turn off subagent
progress summaries, and the extra requests

49
00:03:04,048 --> 00:03:04,608
that write them.

50
00:03:05,568 --> 00:03:10,150
If you've been watching tokens vanish
while subagents run, that's your first knob to

51
00:03:10,210 --> 00:03:10,450
try.

52
00:03:11,297 --> 00:03:14,678
And this ties back to the recent sandbox
and permission mode episodes.

53
00:03:15,379 --> 00:03:19,880
In the Bash sandbox, write access no
longer counts when it comes through a symlink

54
00:03:19,960 --> 00:03:21,921
in a writable folder that leads somewhere
else.

55
00:03:23,061 --> 00:03:27,163
git worktree add, move and remove now fail
read only there.

56
00:03:27,743 --> 00:03:31,205
So if a worktree script suddenly errors in
the sandbox, that's why.

57
00:03:32,393 --> 00:03:33,534
Two smaller ones I like.

58
00:03:34,094 --> 00:03:38,416
If you ask for auto mode, with the flag or
defaultMode auto,

59
00:03:39,016 --> 00:03:44,078
and auto mode's unavailable, it now starts
in Manual mode, which asks first.

60
00:03:44,679 --> 00:03:47,180
Before, it could land somewhere that
allowed more.

61
00:03:48,800 --> 00:03:52,202
And the first generation MCP client is
gone.

62
00:03:52,722 --> 00:03:59,565
If you set MCP_SDK_GENERATION to v1, you
get the v2 client

63
00:03:59,685 --> 00:04:01,326
in a compatibility mode instead.

64
00:04:02,245 --> 00:04:03,506
So the checklist is short.

65
00:04:04,006 --> 00:04:06,827
Set the one million context variable if
you run smaller models,

66
00:04:07,307 --> 00:04:11,789
cap your multi session loops yourself, and
treat Code Mode as something to test on a

67
00:04:11,909 --> 00:04:14,190
throwaway project, once those docs appear.

68
00:04:15,471 --> 00:04:17,412
Don't assume it's safe just because it's
fast.

