1
00:00:00,399 --> 00:00:06,319
So you fire up the latest Codex CLI
release, version 0.153.4,

2
00:00:06,899 --> 00:00:12,739
you run a command without setting a model
flag, and boom, it automatically picks GPT

3
00:00:13,079 --> 00:00:15,279
6 Astra as your default backend.

4
00:00:16,139 --> 00:00:16,819
Just like that.

5
00:00:17,182 --> 00:00:21,362
It really is a pretty quiet shift for
something that actually changes how the

6
00:00:21,443 --> 00:00:23,382
terminal harness behaves day to day.

7
00:00:24,222 --> 00:00:27,662
And, uh, before we dive deep into what
Astra is doing under the hood,

8
00:00:28,302 --> 00:00:32,442
huge thanks to Jellypod for helping us
bring you this daily show every morning.

9
00:00:32,941 --> 00:00:34,641
Yeah, love Jellypod!

10
00:00:35,221 --> 00:00:40,181
But, okay, Ethan, the big headline here
for me isn't just that Astra is bundled as

11
00:00:40,221 --> 00:00:41,002
the default now.

12
00:00:41,462 --> 00:00:45,361
It's this new tool called request user
input async.

13
00:00:45,841 --> 00:00:49,841
That, that, that seems like a massive
structural change for terminal agents.

14
00:00:50,224 --> 00:00:51,024
It really is.

15
00:00:51,664 --> 00:00:56,625
Historically, if an agent hit an ambiguous
decision point, like deciding between two

16
00:00:56,724 --> 00:00:59,864
refactoring strategies, it had to stop
everything.

17
00:01:00,524 --> 00:01:05,184
It paused the entire execution loop and
sat there waiting for you to type an answer.

18
00:01:05,590 --> 00:01:10,389
Right, which completely defeats the
purpose if you kicked off a build or a long test

19
00:01:10,409 --> 00:01:12,190
suite and walked away to grab coffee.

20
00:01:12,583 --> 00:01:13,623
Exactly.

21
00:01:13,687 --> 00:01:19,703
With request user input async, Astra can
fire off a structured text question to you,

22
00:01:19,763 --> 00:01:23,703
but keep working on secondary tasks that
don't depend on your reply.

23
00:01:23,916 --> 00:01:26,343
It doesn't block the background terminal
loop.

24
00:01:26,652 --> 00:01:29,112
Wait, so it keeps working while asking?

25
00:01:29,712 --> 00:01:30,833
What if you don't answer right away?

26
00:01:31,208 --> 00:01:36,008
If you don't answer, it makes sensible,
safe assumptions where appropriate,

27
00:01:36,061 --> 00:01:40,088
or it queues up the consequential
decisions until you check back in.

28
00:01:40,268 --> 00:01:42,968
But it doesn't freeze the whole process.

29
00:01:43,298 --> 00:01:47,637
Mm, that is huge for real world QA and dev
workflows.

30
00:01:48,237 --> 00:01:54,617
And, uh, speed wise, OpenAI is reporting a
1.9x speedup over the previous GPT

31
00:01:54,857 --> 00:01:57,817
5.6 Sol experience on computer use tasks,
right?

32
00:01:58,657 --> 00:02:04,717
Yeah, 1.9 times faster task completion
when paired with the updated Codex harness.

33
00:02:05,297 --> 00:02:07,418
And the benchmark numbers are wild.

34
00:02:07,958 --> 00:02:14,317
On ExploitBench, Astra hit a 100% score
compared to 78.5% for Sol.

35
00:02:14,634 --> 00:02:17,474
A perfect 100%?

36
00:02:18,014 --> 00:02:19,154
That is...

37
00:02:19,214 --> 00:02:19,574
wow.

38
00:02:20,394 --> 00:02:24,355
But wait, if it's that capable at finding
exploits, what about safety?

39
00:02:24,854 --> 00:02:27,555
Did it try to bypass the terminal
safeguards?

40
00:02:27,917 --> 00:02:31,517
That's actually one of the most
interesting parts of the system card.

41
00:02:31,597 --> 00:02:36,397
In testing against Codex Auto Review, even
when Auto Review was deliberately

42
00:02:36,455 --> 00:02:42,877
configured to be bypassable, Astra had
zero recorded attempts to evade or circumvent

43
00:02:42,917 --> 00:02:44,397
the security restrictions.

44
00:02:44,917 --> 00:02:45,396
Zero?

45
00:02:45,936 --> 00:02:51,036
Man, Sol used to try sneaky shortcuts when
tasks got impossible.

46
00:02:51,614 --> 00:02:56,494
Sol tried to cheat or bypass auto review
in over half of those honeypot tests.

47
00:02:57,335 --> 00:03:00,274
Astra just respects the boundary or stops.

48
00:03:00,974 --> 00:03:07,574
Plus, for enterprise teams, release
0.153.3 added Astra to

49
00:03:07,694 --> 00:03:12,534
Amazon Bedrock for both Mantle and Runtime
global or US routes,

50
00:03:13,174 --> 00:03:18,034
and there's a Fast tier toggle that gives
you 2x processing speed in the API.

51
00:03:18,382 --> 00:03:21,862
Okay, let me put my tester hat on for a
second here.

52
00:03:22,622 --> 00:03:25,543
What are the caveats with this async
prompt tool?

53
00:03:26,043 --> 00:03:31,423
Because non-blocking prompts sound
amazing, but there have to be failure modes,

54
00:03:31,522 --> 00:03:31,762
right?

55
00:03:32,584 --> 00:03:33,284
You know there are.

56
00:03:33,824 --> 00:03:37,525
First off, request user input async is
strictly text only.

57
00:03:38,164 --> 00:03:42,564
No image attachments, no rich payload
responses in that async channel yet.

58
00:03:42,944 --> 00:03:47,025
Got it, so you can't drop a screenshot
back to it mid task.

59
00:03:47,583 --> 00:03:48,383
Right.

60
00:03:48,423 --> 00:03:55,103
And second, the model guidance was
specifically updated in 0.153.4 so

61
00:03:55,183 --> 00:04:00,863
Astra only attempts async questions if the
tool is explicitly available in that

62
00:04:00,932 --> 00:04:02,303
active session catalog.

63
00:04:02,570 --> 00:04:07,423
If you're running unattended multi-file
refactors, you still need to be careful with

64
00:04:07,469 --> 00:04:11,823
thread scope, so it doesn't make
assumptions on file boundaries that you wanted

65
00:04:11,880 --> 00:04:12,383
locked.

66
00:04:12,594 --> 00:04:13,774
Hmm, fair enough.

67
00:04:14,094 --> 00:04:19,014
But speaking of typing in the TUI, can we
talk about the editor updates?

68
00:04:19,454 --> 00:04:24,375
Because as someone who breaks drafts
constantly, the Vim mode fixes in

69
00:04:24,795 --> 00:04:28,494
0.153.0 made me so happy.

70
00:04:29,312 --> 00:04:30,852
You mean undo and redo?

71
00:04:31,774 --> 00:04:32,294
Yes!

72
00:04:32,694 --> 00:04:38,954
Pressing u for undo and Ctrl R for redo
now actually preserves your full draft

73
00:04:39,054 --> 00:04:39,335
state!

74
00:04:39,734 --> 00:04:43,354
Including giant pasted code blocks and
file attachments.

75
00:04:43,875 --> 00:04:49,074
Before, if you hit undo in Vim mode, it
would lose your pasted context or shred the

76
00:04:49,135 --> 00:04:50,094
attachment reference.

77
00:04:51,714 --> 00:04:53,355
Yeah, that drove a lot of people crazy.

78
00:04:53,835 --> 00:04:55,815
It's a huge quality of life fix.

79
00:04:56,434 --> 00:05:02,914
They also added a simple config toggle
under the TUI settings: tui dot auto recap

80
00:05:02,994 --> 00:05:03,674
equals false.

81
00:05:04,090 --> 00:05:08,469
Oh, so you can turn off the automatic
recap summaries when threads get long?

82
00:05:08,833 --> 00:05:09,553
Exactly.

83
00:05:09,641 --> 00:05:14,993
Automatic recaps stay off, but you can
still trigger a manual summary whenever you

84
00:05:15,073 --> 00:05:17,313
want using slash recap.

85
00:05:17,712 --> 00:05:18,252
Nice.

86
00:05:18,772 --> 00:05:22,632
And for Plus and Team account users,
there's a new usage warning,

87
00:05:22,692 --> 00:05:22,952
right?

88
00:05:23,472 --> 00:05:27,412
It pops up when you hit less than half
your allowance remaining in that five hour

89
00:05:27,452 --> 00:05:27,753
window.

90
00:05:28,125 --> 00:05:32,205
Yeah, an earlier heads up so you don't
suddenly run out of tokens mid sprint.

91
00:05:32,607 --> 00:05:34,307
So step back for a second, Ethan.

92
00:05:35,107 --> 00:05:39,307
Looking at Astra as the bundled default,
the async questions,

93
00:05:39,387 --> 00:05:40,668
the tighter safety bounds...

94
00:05:41,428 --> 00:05:46,627
does non-blocking interaction actually
change how we oversee these terminal agents?

95
00:05:47,000 --> 00:05:52,440
I think it shifts the dynamic from
babysitting to true asynchronous delegation.

96
00:05:52,580 --> 00:05:57,160
You're not sitting there staring at the
cursor waiting to press Y or N.

97
00:05:57,320 --> 00:06:01,800
You set the goal, you let it execute at
nearly double the speed,

98
00:06:01,840 --> 00:06:06,360
and if it needs a quick clarification, it
drops a line without pulling the emergency

99
00:06:06,427 --> 00:06:06,840
brake.

100
00:06:07,132 --> 00:06:11,092
As long as it doesn't try to rewrite my
entire codebase while I'm getting tea,

101
00:06:11,392 --> 00:06:12,632
I am fully on board.

102
00:06:13,452 --> 00:06:14,773
Well, that's our take for today.

103
00:06:15,232 --> 00:06:17,572
Thanks for tuning in, everyone, and we'll
catch you next time!

104
00:06:17,917 --> 00:06:19,037
Talk soon, everyone.