1
00:00:00,000 --> 00:00:01,280
Welcome to the show everyone!

2
00:00:01,360 --> 00:00:05,440
I'm Lachlan Reed, coming to you from the
backyard shed, brought to you by Jellypod

3
00:00:05,627 --> 00:00:06,160
AI.

4
00:00:06,380 --> 00:00:09,280
And look, I've got James Turner here with
me.

5
00:00:09,413 --> 00:00:14,080
James, I have a quick question for you:
have you ever been working in the terminal,

6
00:00:14,100 --> 00:00:17,120
you run a complex command, and the screen
just...

7
00:00:17,168 --> 00:00:18,320
goes dead silent?

8
00:00:19,000 --> 00:00:19,800
Oh, absolutely.

9
00:00:19,914 --> 00:00:24,287
It's that classic developer panic where
you're staring at a blinking cursor,

10
00:00:24,319 --> 00:00:27,807
wondering if your network dropped, or if
the API is just having a massive

11
00:00:27,860 --> 00:00:29,247
existential crisis.

12
00:00:29,367 --> 00:00:32,607
You're just hovering your finger over
Control-C, ready to kill it.

13
00:00:33,000 --> 00:00:34,120
Exactly!

14
00:00:34,360 --> 00:00:36,840
[exhales sharply] You're sweating on the
keys.

15
00:00:37,040 --> 00:00:42,120
And with agentic tools like Claude Code,
that silence usually means the model is

16
00:00:42,180 --> 00:00:44,440
chewing through a massive thinking block.

17
00:00:44,600 --> 00:00:49,160
But in older versions, if the stream
stalled for just ten seconds,

18
00:00:49,200 --> 00:00:54,520
the CLI would panic, assume the connection
was dead, and fire off a hard retry.

19
00:00:54,707 --> 00:00:56,680
It was a proper dog's breakfast.

20
00:00:57,000 --> 00:00:58,280
Wait, ten seconds?

21
00:00:58,560 --> 00:01:01,809
[questioning tone] That is incredibly fast
for a complex agentic task.

22
00:01:01,996 --> 00:01:06,049
If it retries after just ten seconds of
silence, isn't it just spinning up a

23
00:01:06,089 --> 00:01:09,089
duplicate API call while the first one
might still be processing?

24
00:01:10,000 --> 00:01:10,640
Spot on!

25
00:01:11,187 --> 00:01:15,040
It was absolutely burning through tokens
and breaking the agent's state.

26
00:01:15,180 --> 00:01:21,360
But with the new Claude Code v2.1.185
release, they have doubled that stream-stall

27
00:01:21,406 --> 00:01:23,040
window to twenty seconds.

28
00:01:23,140 --> 00:01:27,920
And instead of a scary "No response" error
that triggers an automatic retry,

29
00:01:27,947 --> 00:01:32,480
it now shows a much softer "Waiting for
API response" status message.

30
00:01:33,000 --> 00:01:34,120
Oh, thank goodness.

31
00:01:34,637 --> 00:01:36,843
Twenty seconds makes so much more sense.

32
00:01:36,976 --> 00:01:41,003
It gives those deep thinking blocks room
to breathe without throwing a wrench in the

33
00:01:41,070 --> 00:01:41,563
gears.

34
00:01:41,763 --> 00:01:46,923
I mean, duplicating those API calls
doesn't just waste tokens; it completely breaks

35
00:01:46,943 --> 00:01:50,363
the continuity of what the agent was
trying to solve in your workspace.

36
00:01:51,000 --> 00:01:51,720
Exactly.

37
00:01:52,253 --> 00:01:53,880
It keeps things steady as she goes.

38
00:01:54,040 --> 00:01:57,480
Now, speaking of keeping things from
breaking, they also dropped some great

39
00:01:57,530 --> 00:02:02,280
quality-of-life updates in version 2.1.183
just before this.

40
00:02:02,360 --> 00:02:07,560
They added a new slash-config double-dash
help command so you don't have to guess

41
00:02:07,580 --> 00:02:09,160
the configuration flags anymore.

42
00:02:10,000 --> 00:02:10,560
Finally!

43
00:02:11,400 --> 00:02:15,040
I hate having to pull up the online docs
just to remember how to toggle a local

44
00:02:15,100 --> 00:02:15,600
setting.

45
00:02:15,780 --> 00:02:19,640
But wait, wasn't there also a big safety
change in that release?

46
00:02:19,676 --> 00:02:21,120
Something about Auto Mode?

47
00:02:22,000 --> 00:02:22,720
Yes!

48
00:02:23,080 --> 00:02:27,618
[genuinely surprised] This is huge for
anyone who lets Claude run wild in Auto Mode.

49
00:02:27,794 --> 00:02:32,658
They've built in new protections that
explicitly block destructive commands.

50
00:02:33,058 --> 00:02:38,338
We are talking about things like
accidental hard git resets or destructive

51
00:02:38,386 --> 00:02:39,938
infrastructure teardowns.

52
00:02:40,000 --> 00:02:41,120
Oh, wow.

53
00:02:41,980 --> 00:02:43,612
That is a massive guardrail.

54
00:02:43,702 --> 00:02:47,532
Because in Auto Mode, you're essentially
handing the keys of your codebase to the

55
00:02:47,652 --> 00:02:48,172
LLM.

56
00:02:48,385 --> 00:02:53,692
If it decides a clean slate means running
"git reset --hard" on your uncommitted

57
00:02:53,772 --> 00:02:54,732
work...

58
00:02:54,748 --> 00:02:56,172
that's a nightmare scenario.

59
00:02:57,000 --> 00:02:57,960
Too right, mate.

60
00:02:59,127 --> 00:03:03,087
I've nearly lost a whole night's work
before by being too trusting with automated

61
00:03:03,147 --> 00:03:03,887
scripts.

62
00:03:03,978 --> 00:03:09,247
Having the tool itself say "Whoa there,
I'm not allowed to destroy this" is a

63
00:03:09,297 --> 00:03:10,127
massive relief.

64
00:03:10,367 --> 00:03:15,007
It lets you run Auto Mode without feeling
like you need to watch it like a hawk.

65
00:03:16,000 --> 00:03:22,320
It really shifts the dynamic from
"dangerous experiment" to "reliable junior

66
00:03:22,460 --> 00:03:23,040
dev."

67
00:03:23,140 --> 00:03:27,760
Between the twenty-second stall window and
the destructive command blocks,

68
00:03:27,813 --> 00:03:32,080
it feels like Claude Code is finally
growing up and respecting our environments.

69
00:03:33,000 --> 00:03:33,960
It definitely is.

70
00:03:34,667 --> 00:03:40,600
It's the small, unglamorous stability
fixes that actually make these AI tools usable

71
00:03:40,700 --> 00:03:41,560
day-to-day.

72
00:03:41,800 --> 00:03:44,360
That's all we have for this lightning
update.

73
00:03:44,493 --> 00:03:45,720
We'll catch you next time!

74
00:03:46,000 --> 00:03:46,400
See ya!