1
00:00:00,000 --> 00:00:05,120
Lachlan here from the shed, with
James--this episode of the Claude Code Changelog is

2
00:00:05,150 --> 00:00:07,440
brought to you by Jellypod AI.

3
00:00:08,600 --> 00:00:12,560
Now, mate, you know when you're deep in a
massive, one-million-token streaming

4
00:00:12,610 --> 00:00:17,200
session, and you're just, like, staring at
the terminal waiting for this beautiful,

5
00:00:17,260 --> 00:00:21,120
complex script to finish writing, and
then--bam!

6
00:00:21,184 --> 00:00:25,760
Your Wi-Fi drops for a split second, and
the whole bloody thing vanishes into the

7
00:00:25,827 --> 00:00:26,480
ether.

8
00:00:26,613 --> 00:00:28,640
It's enough to make you chuck a wobbly.

9
00:00:29,533 --> 00:00:30,440
Oh, absolutely.

10
00:00:30,568 --> 00:00:35,640
That total loss of partial code outputs
during a network flicker is a massive

11
00:00:35,688 --> 00:00:36,760
developer pain point.

12
00:00:36,880 --> 00:00:42,040
But in Claude Code v2.1.199, they finally
fixed this.

13
00:00:42,253 --> 00:00:47,560
If the API stream drops mid-flight now, it
actually preserves whatever partial code

14
00:00:47,600 --> 00:00:52,600
was already written, and it prints this
explicit "incomplete-response" notice right

15
00:00:52,627 --> 00:00:53,480
there in your terminal.

16
00:00:53,620 --> 00:00:55,560
You don't lose the whole chunk anymore.

17
00:00:56,042 --> 00:00:57,050
Wait, fair dinkum?

18
00:00:57,263 --> 00:01:01,850
So, let's say it's ninety percent done
writing some massive backend refactor,

19
00:01:02,010 --> 00:01:03,530
and the packet drops.

20
00:01:03,930 --> 00:01:06,010
I actually get to keep that ninety
percent?

21
00:01:06,498 --> 00:01:06,878
Yes!

22
00:01:06,988 --> 00:01:07,838
Exactly.

23
00:01:07,958 --> 00:01:09,598
You keep the ninety percent.

24
00:01:09,668 --> 00:01:14,878
Because in previous versions, if that HTTP
connection timed out or hit a transient

25
00:01:14,926 --> 00:01:19,118
rate limit, the tool would basically
panic, throw an unhandled error,

26
00:01:19,138 --> 00:01:21,118
and discard the entire response.

27
00:01:21,278 --> 00:01:24,078
Now, you get that incomplete-response
notice.

28
00:01:24,158 --> 00:01:28,158
You can inspect the partial code, see
exactly where the connection choked,

29
00:01:28,198 --> 00:01:30,558
and just ask Claude to pick up right from
that line.

30
00:01:30,745 --> 00:01:34,878
It saves you from burning through another
massive context window just to regenerate

31
00:01:34,910 --> 00:01:36,398
what you already had on screen.

32
00:01:36,755 --> 00:01:38,128
That is huge for the wallet.

33
00:01:38,368 --> 00:01:42,208
But, look, what if I'm running some
massive background task?

34
00:01:42,384 --> 00:01:46,368
Like, we talked yesterday about those
hands-off background agents.

35
00:01:46,528 --> 00:01:50,288
If those things are running headless and
the network drops for more than a few

36
00:01:50,338 --> 00:01:52,368
minutes, does it just die?

37
00:01:52,770 --> 00:01:55,858
Well, it used to, because of the hard cap
of fifteen retries.

38
00:01:55,978 --> 00:02:00,178
But v2.1.199 introduces a new environment
variable called

39
00:02:00,365 --> 00:02:02,818
CLAUDE_CODE_RETRY_WATCHDOG.

40
00:02:02,951 --> 00:02:07,298
If you set that, it raises the transient
retry limit all the way up to three

41
00:02:07,338 --> 00:02:10,978
hundred, and it completely lifts that
fifteen-retry cap.

42
00:02:11,085 --> 00:02:13,538
It basically tells the agent, "Hey, don't
give up.

43
00:02:13,570 --> 00:02:14,898
Keep knocking on the door."

44
00:02:15,550 --> 00:02:17,510
Three hundred retries!

45
00:02:17,686 --> 00:02:19,590
That's proper stubborn.

46
00:02:19,718 --> 00:02:23,430
That means if my home internet decides to
take a five-minute nap,

47
00:02:23,470 --> 00:02:27,910
the agent just, uh, keeps on trying in the
background until the line's clear.

48
00:02:28,435 --> 00:02:33,265
And speaking of keeping things alive, this
version also fixes a really nasty Linux

49
00:02:33,311 --> 00:02:34,306
daemon crash loop.

50
00:02:34,396 --> 00:02:39,105
Before this update, background agents on
Linux were literally self-killing every

51
00:02:39,158 --> 00:02:42,145
fifty seconds due to unhandled signal
cleanups.

52
00:02:42,345 --> 00:02:45,185
Now, that daemon loop is completely
stable.

53
00:02:45,345 --> 00:02:49,585
Plus, they added proactive system memory
warnings, so if your machine is starving

54
00:02:49,625 --> 00:02:54,305
for RAM, Claude will warn you before it
tries to spin up a subagent and crashes your

55
00:02:54,358 --> 00:02:54,865
terminal.

56
00:02:55,525 --> 00:02:57,338
No worries about the machine blowing up,
then.

57
00:02:57,518 --> 00:02:59,498
But what about the corporate crowd?

58
00:02:59,638 --> 00:03:04,218
You know, the guys behind those massive,
soul-crushing corporate TLS firewalls where

59
00:03:04,247 --> 00:03:05,658
everything gets intercepted?

60
00:03:05,955 --> 00:03:10,195
Yeah, those corporate TLS-inspecting
proxies are notorious for throwing SSL

61
00:03:10,243 --> 00:03:13,795
handshake errors and burning through your
API retries instantly.

62
00:03:13,982 --> 00:03:17,795
In this update, if Claude Code detects a
custom certificate issue,

63
00:03:17,822 --> 00:03:19,235
it doesn't just crash.

64
00:03:19,368 --> 00:03:22,915
It actually intercept-diagnoses the error
and prints a direct,

65
00:03:22,966 --> 00:03:27,715
actionable hint telling you to use the
NODE_EXTRA_CA_CERTS environment variable

66
00:03:27,786 --> 00:03:30,275
pointing to your company's custom CA
bundle.

67
00:03:30,462 --> 00:03:32,915
It bypasses the proxy headache entirely.

68
00:03:33,015 --> 00:03:37,075
And once you're in, you can even stack
your slash-commands in a single run--like

69
00:03:37,128 --> 00:03:41,235
chaining slash-bug-hunt right into
slash-explain--without the parser getting

70
00:03:41,306 --> 00:03:43,715
confused by the stacked slash-skills.

71
00:03:44,471 --> 00:03:44,852
Beautiful.

72
00:03:44,945 --> 00:03:50,172
So, if you're hitting those TLS walls,
make sure you set NODE_EXTRA_CA_CERTS to your

73
00:03:50,225 --> 00:03:54,892
local CA bundle path, and that'll keep
your streams clean and your retries from

74
00:03:54,932 --> 00:03:55,612
burning out.