1
00:00:00,099 --> 00:00:04,360
Brought to you by Jellypod AI, if you were
coding with Claude Code around August

2
00:00:04,440 --> 00:00:08,740
eighteen, twenty twenty six, you might
have noticed version two point one point two

3
00:00:08,880 --> 00:00:12,399
three five dropping a fix for a massive
hidden token sink.

4
00:00:12,960 --> 00:00:18,799
It, it, it turns out that whenever your
Language Server Protocol or LSP reconnected

5
00:00:18,859 --> 00:00:22,479
mid session, your whole prompt cache was
getting completely wiped.

6
00:00:23,542 --> 00:00:25,301
Wait, the whole prompt cache?

7
00:00:25,921 --> 00:00:29,061
So if your language server just blinks or
restarts in the background,

8
00:00:29,522 --> 00:00:32,661
it invalidates hundreds of thousands of
context tokens?

9
00:00:32,787 --> 00:00:33,587
Whole thing, mate!

10
00:00:34,047 --> 00:00:35,527
Total cache invalidation.

11
00:00:36,227 --> 00:00:40,387
Because the prompt cache relies on keeping
that long context prefix static on

12
00:00:40,427 --> 00:00:46,048
Anthropic servers, any tiny break from the
LSP triggered a full cache rebuild.

13
00:00:46,128 --> 00:00:50,447
That meant Claude Code had to reindex and
resend two hundred thousand plus tokens

14
00:00:50,487 --> 00:00:51,327
back over the wire.

15
00:00:51,679 --> 00:00:56,179
Wow, that explains those sudden multi
second lag spikes during long sessions.

16
00:00:56,859 --> 00:00:59,960
You think the model is taking forever to
answer, but in reality,

17
00:01:00,019 --> 00:01:04,359
you are paying to resend your entire
codebase history and waiting for it to process

18
00:01:04,439 --> 00:01:05,179
from scratch!

19
00:01:05,409 --> 00:01:05,909
Spot on!

20
00:01:06,289 --> 00:01:10,689
You get hit with massive latency, plus a
surprise jump on your API bill.

21
00:01:11,489 --> 00:01:15,949
But version two point one point two three
five stops those mid session LSP

22
00:01:16,050 --> 00:01:20,369
reconnects from blowing away the cache
key, so your sessions stay fast and cheap.

23
00:01:20,697 --> 00:01:22,157
That is a massive win.

24
00:01:22,777 --> 00:01:25,017
What else landed in two point one point
two three five?

25
00:01:25,517 --> 00:01:27,597
I saw something about terminal spellcheck?

26
00:01:27,750 --> 00:01:28,310
Yeah!

27
00:01:28,390 --> 00:01:31,670
They added an optional spellcheck setting
in slash config.

28
00:01:31,857 --> 00:01:36,470
As you type your prompt in the terminal
UI, it underlines misspelled words in real

29
00:01:36,550 --> 00:01:37,190
time.

30
00:01:37,390 --> 00:01:42,550
And, uh, it does it by leveraging your
local system tools, whether you have aspell,

31
00:01:42,610 --> 00:01:44,710
hunspell, or ispell installed.

32
00:01:44,708 --> 00:01:45,668
Oh, I love that!

33
00:01:45,788 --> 00:01:51,668
Using local binaries like aspell means
zero extra latency or API calls just to catch

34
00:01:51,721 --> 00:01:53,348
typos before sending a prompt.

35
00:01:53,492 --> 00:01:56,388
What about performance for local tools
like grep?

36
00:01:56,375 --> 00:01:57,255
Ah, totally!

37
00:01:57,395 --> 00:02:02,055
The embedded grep in native macOS and
Linux builds got a big improvement.

38
00:02:02,242 --> 00:02:05,815
If you accidentally throw a pathological
regex pattern at it,

39
00:02:05,863 --> 00:02:10,215
grep now fails fast instead of exhausting
system memory and freezing your terminal.

40
00:02:10,407 --> 00:02:11,667
That is huge.

41
00:02:12,127 --> 00:02:15,847
No more killing the process because a
runaway search locked up your machine!

42
00:02:16,467 --> 00:02:21,407
Did they touch background tasks like slash
ultrareview or slash autofix pr as well?

43
00:02:21,667 --> 00:02:22,147
They did!

44
00:02:22,234 --> 00:02:28,067
Background cloud sessions like ultrareview
and autofix pr got a massive CPU and

45
00:02:28,113 --> 00:02:29,267
memory optimization.

46
00:02:29,376 --> 00:02:35,027
Previously, every single update was
rescanning and rerendering the entire event

47
00:02:35,084 --> 00:02:36,467
stream from scratch.

48
00:02:36,627 --> 00:02:42,627
Now it handles background streams way
smarter, so your local CPU stays nice and

49
00:02:42,707 --> 00:02:43,107
cool.

50
00:02:43,984 --> 00:02:46,444
My laptop fans are crying tears of joy
right now.

51
00:02:47,184 --> 00:02:50,125
Any other quality of life tweaks or
caveats we should know about?

52
00:02:50,326 --> 00:02:52,246
A couple of small but sweet ones!

53
00:02:53,286 --> 00:02:58,267
If an Agent tool call omits subagent type
in a session where that subagent isn't

54
00:02:58,326 --> 00:03:03,566
available, it now throws an explicit error
listing all valid available agents

55
00:03:03,646 --> 00:03:06,066
instead of advertising a general purpose
default.

56
00:03:06,966 --> 00:03:11,306
And inside permission prompts, pressing
Shift Tab in the comment field won't

57
00:03:11,366 --> 00:03:15,806
accidentally approve the edit and grant
session wide edit permissions anymore.

58
00:03:15,847 --> 00:03:17,806
It just closes the comment box like it
should.

59
00:03:18,972 --> 00:03:23,333
Man, accidentally approving a session wide
permission because you tried to tab out

60
00:03:23,333 --> 00:03:25,852
of a text field is terrifying!

61
00:03:25,912 --> 00:03:27,472
Glad that one is patched.

62
00:03:27,512 --> 00:03:28,432
Good updates all around.