1
00:00:00,260 --> 00:00:05,442
This one's brought to you by Jellypod AI,
and it's a fix for anyone running a cheap

2
00:00:05,562 --> 00:00:07,502
worker model under a big orchestrator.

3
00:00:08,222 --> 00:00:15,004
Version 2.1.325, released November 13th,
fixes how subagents with their

4
00:00:15,084 --> 00:00:16,465
own model get sized.

5
00:00:17,182 --> 00:00:21,763
Right, so the changelog line is this:
subagents on a different model were sizing

6
00:00:21,883 --> 00:00:25,164
file reads and MCP results with the main
session's model.

7
00:00:26,324 --> 00:00:26,924
That's the bug.

8
00:00:27,445 --> 00:00:30,846
If your main session has a big context
window and your worker doesn't,

9
00:00:31,246 --> 00:00:35,987
the worker got handed output cut for the
bigger one and could hit the wall.

10
00:00:36,000 --> 00:00:40,681
And that connects to the two earlier
episodes, the subagent effort one and the

11
00:00:40,761 --> 00:00:42,742
workflow subagent model one.

12
00:00:43,682 --> 00:00:47,283
Those gave you granular control over which
model and how much thinking.

13
00:00:48,784 --> 00:00:53,305
But mixed setups had context sizing
mismatches and model resets,

14
00:00:53,825 --> 00:00:57,266
and 2.1.325 is where that gets fixed.

15
00:00:58,186 --> 00:01:00,907
There's a second fix, and it's the model
reset part.

16
00:01:01,527 --> 00:01:05,108
Subagents with their own model were being
moved to the plan mode model when the

17
00:01:05,188 --> 00:01:07,329
model setting was opusplan or haiku.

18
00:01:08,849 --> 00:01:11,690
So you'd pin a worker to something light,
switch to opusplan,

19
00:01:11,850 --> 00:01:13,690
and the worker quietly changed models.

20
00:01:14,270 --> 00:01:17,071
In 2.1.325, that pin holds.

21
00:01:18,038 --> 00:01:18,939
So how do you use it?

22
00:01:19,639 --> 00:01:22,940
Pin the worker, then run the main session
on something bigger.

23
00:01:23,500 --> 00:01:28,301
You can set
CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL in your environment,

24
00:01:28,781 --> 00:01:31,302
or give a subagent its own model in its
definition.

25
00:01:32,042 --> 00:01:36,763
Then run slash model opusplan, and the
worker should stay on what you picked.

26
00:01:37,809 --> 00:01:42,210
A quick check is to run a task that makes
the worker read a large file or call an

27
00:01:42,350 --> 00:01:43,970
MCP tool with big output.

28
00:01:44,571 --> 00:01:46,591
Before this, that's where the mismatch
would bite.

29
00:01:47,231 --> 00:01:50,112
Now reads and results are sized for the
worker's own model.

30
00:01:50,732 --> 00:01:54,833
I'd test it once before trusting it on a
long job, because it's a fresh fix.

31
00:01:56,279 --> 00:01:56,839
Caveats.

32
00:01:57,399 --> 00:02:00,960
The changelog only says the sizing now
uses the subagent's model.

33
00:02:01,320 --> 00:02:04,362
It doesn't promise a small model will
never run short on room,

34
00:02:04,922 --> 00:02:08,943
so still tell workers to summarize instead
of dumping big payloads.

35
00:02:10,384 --> 00:02:14,145
Also, the same release fixes a session
going back to an earlier permission mode

36
00:02:14,525 --> 00:02:18,847
after a subagent's plan was approved, so
watch your permission mode after approvals.

37
00:02:19,905 --> 00:02:20,765
Now the smaller stuff.

38
00:02:21,346 --> 00:02:25,687
Fixed sessions on many third party
gateways and Foundry stopping with a context

39
00:02:25,807 --> 00:02:28,228
length error on every turn instead of
compacting.

40
00:02:29,128 --> 00:02:33,130
And when your organization's policy blocks
a request, the error now tells you to run

41
00:02:33,290 --> 00:02:36,971
slash rewind and change the message, or
run slash clear.

42
00:02:37,774 --> 00:02:39,915
And there's a plansDirectory fix.

43
00:02:40,395 --> 00:02:44,657
If you've set a custom folder in settings
and it can't take an approved plan edit,

44
00:02:45,077 --> 00:02:49,999
the plan now gets saved in the default
plans folder, with a warning saying where and

45
00:02:50,079 --> 00:02:50,379
why.

46
00:02:51,299 --> 00:02:52,400
No more lost plan.

47
00:02:53,800 --> 00:02:58,862
There's also a security fix for how zstd
compressed responses are decoded,

48
00:02:59,362 --> 00:03:04,304
plus a note for Claude when a sandboxed
git checkout or stash can't update protected

49
00:03:04,344 --> 00:03:06,765
files like the ones in dot claude.

50
00:03:07,813 --> 00:03:12,435
Update, re-test your mixed model setup,
and check that your pinned workers stay

51
00:03:12,555 --> 00:03:12,955
pinned.

