1
00:00:00,159 --> 00:00:07,079
So, version 2.1.234 of Claude Code just
landed, and, uh, the

2
00:00:07,139 --> 00:00:12,500
big shift here is that you no longer have
to kill a running turn just to change your

3
00:00:12,539 --> 00:00:13,119
environment.

4
00:00:14,087 --> 00:00:17,067
Mate, you have no idea how much time that
saves me in the shed.

5
00:00:17,767 --> 00:00:20,527
I, I, I was halfway through a multi repo
refactor yesterday,

6
00:00:20,668 --> 00:00:20,787
right?

7
00:00:21,367 --> 00:00:25,567
Claude was hammering away on the backend,
and suddenly it needed context from dot

8
00:00:25,607 --> 00:00:27,708
slash packages slash shared.

9
00:00:27,807 --> 00:00:31,347
Old days, you had to cancel the whole
execution, re run it,

10
00:00:31,427 --> 00:00:33,048
burn those tokens all over again.

11
00:00:33,224 --> 00:00:37,484
But now you can just invoke slash add dir
mid execution?

12
00:00:37,641 --> 00:00:38,241
Exactly!

13
00:00:38,542 --> 00:00:43,621
You hit slash add dir, point it to dot
slash packages slash shared,

14
00:00:44,261 --> 00:00:49,042
or pop open slash permissions or slash
config right while the tool is running.

15
00:00:49,701 --> 00:00:53,901
It just dynamically attaches the new
directory to the scope and keeps rolling.

16
00:00:54,581 --> 00:00:56,882
No restart needed, flat out brilliant.

17
00:00:57,159 --> 00:01:02,199
That, that, that whole idea of not wasting
context connects directly to what they

18
00:01:02,259 --> 00:01:04,819
did with the built in claude api skill,
too.

19
00:01:05,239 --> 00:01:09,739
Previously, if you enabled that skill, it
would do an eager full spec prompt

20
00:01:09,779 --> 00:01:10,599
injection.

21
00:01:10,659 --> 00:01:15,199
That was dumping over 200,000 tokens into
your context window right off the bat!

22
00:01:16,490 --> 00:01:19,610
200,000 tokens before you even type a
prompt?

23
00:01:20,789 --> 00:01:23,229
Crikey, that is a proper memory hog.

24
00:01:23,616 --> 00:01:23,836
Right!

25
00:01:24,056 --> 00:01:28,456
But in 2.1.234, they switched to on demand
documentation loading.

26
00:01:28,857 --> 00:01:33,436
So that initial context overhead drops
from over 200,000 tokens down to roughly

27
00:01:33,677 --> 00:01:35,236
25,000 tokens.

28
00:01:35,257 --> 00:01:38,337
That is nearly an eightfold reduction in
idle context cost.

29
00:01:38,995 --> 00:01:43,494
That is massive for token budgets,
especially when you are running longer sessions.

30
00:01:44,214 --> 00:01:49,195
And, uh, speaking of long sessions, they
added auto resume for daily usage limits,

31
00:01:49,394 --> 00:01:49,534
yeah?

32
00:01:49,786 --> 00:01:50,506
Yeah, exactly.

33
00:01:51,086 --> 00:01:54,986
If you hit your daily limit on claude.ai,
the session doesn't just crash out

34
00:01:55,006 --> 00:01:55,746
anymore.

35
00:01:55,806 --> 00:02:00,246
It pauses, waits for the reset window, and
then automatically continues where it

36
00:02:00,246 --> 00:02:01,286
left off.

37
00:02:01,346 --> 00:02:03,786
You can toggle that on or off in slash
config.

38
00:02:03,952 --> 00:02:07,692
That is so handy if you leave a huge batch
job running overnight.

39
00:02:08,952 --> 00:02:12,312
But, wait, there was also a security patch
tucked into this release,

40
00:02:12,392 --> 00:02:12,793
wasn't there?

41
00:02:13,213 --> 00:02:15,152
Something about Windows path handling?

42
00:02:15,552 --> 00:02:16,672
Yeah, critical fix there.

43
00:02:16,833 --> 00:02:21,452
It rejects Windows NT namespace paths, the
ones starting with backslash question

44
00:02:21,472 --> 00:02:22,912
mark question mark backslash.

45
00:02:23,532 --> 00:02:28,572
Turns out attackers could use those paths
in remote file reads or included CLAUDE.md

46
00:02:28,632 --> 00:02:31,933
files to leak NTLM hash credentials over
SMB.

47
00:02:32,537 --> 00:02:37,017
Oof, leaking NTLM credentials just from
loading a repository file,

48
00:02:37,777 --> 00:02:39,117
that is a nasty vector.

49
00:02:39,857 --> 00:02:41,497
Good on them for patching that out fast.

50
00:02:41,882 --> 00:02:42,461
Definitely.

51
00:02:43,101 --> 00:02:48,261
Now, there is one caveat to keep in mind
if you are messing with slash TUI mode

52
00:02:48,321 --> 00:02:48,741
switches.

53
00:02:49,561 --> 00:02:53,061
If you launched your session with explicit
tool restriction flags,

54
00:02:53,521 --> 00:02:59,241
like allowed tools or disallowed tools,
slash TUI will outright refuse to switch

55
00:02:59,302 --> 00:02:59,621
modes.

56
00:02:59,951 --> 00:03:04,371
Ah, so it keeps you locked into your
security sandbox so you don't accidentally

57
00:03:04,551 --> 00:03:06,571
bypass your own rules mid stream.

58
00:03:07,090 --> 00:03:07,669
Precisely.

59
00:03:08,230 --> 00:03:12,369
Safety guardrails stay intact, but you get
way more control mid turn.

60
00:03:12,690 --> 00:03:13,909
Good update overall.

61
00:03:14,057 --> 00:03:14,677
Too right.

62
00:03:15,417 --> 00:03:18,477
Well, I am off to run some more mid turn
add dirs.

63
00:03:19,197 --> 00:03:20,037
Catch you later, James.