1
00:00:00,000 --> 00:00:01,600
Welcome to the show everybody!

2
00:00:01,840 --> 00:00:03,920
I'm Ethan Park, here with Maya.

3
00:00:04,060 --> 00:00:08,960
And first off, we have to give a quick
shout-out and thank you to Jellypod for

4
00:00:09,004 --> 00:00:10,000
supporting the show.

5
00:00:10,200 --> 00:00:14,720
Now, Maya, let's look at something that
caught my eye in the latest Codex upstream

6
00:00:14,780 --> 00:00:15,280
cleanup.

7
00:00:15,600 --> 00:00:18,320
PR number 27801.

8
00:00:18,432 --> 00:00:24,160
They completely ripped out the
WebRTC-based real-time voice and dictation system

9
00:00:24,256 --> 00:00:26,000
from the Codex TUI.

10
00:00:26,160 --> 00:00:28,120
It's just gone.

11
00:00:29,399 --> 00:00:31,919
Wait, they completely deleted it?

12
00:00:32,539 --> 00:00:38,000
So the actual `/realtime` session command
and all those audio options in `/settings`

13
00:00:38,059 --> 00:00:38,920
are just...

14
00:00:38,979 --> 00:00:40,659
wiped from the terminal interface?

15
00:00:42,000 --> 00:00:43,040
Completely wiped.

16
00:00:43,671 --> 00:00:47,609
If you try to run `/realtime` now, you'll
get an error.

17
00:00:47,789 --> 00:00:51,529
The main driver here was streamlining the
binary footprint.

18
00:00:51,709 --> 00:00:57,289
WebRTC brings a massive amount of
overhead, and shipping those heavy media libraries

19
00:00:57,335 --> 00:01:02,089
inside a terminal-focused CLI tool was
starting to look like a massive case of

20
00:01:02,139 --> 00:01:02,969
feature creep.

21
00:01:03,380 --> 00:01:06,079
I mean, that actually makes a lot of
sense.

22
00:01:06,679 --> 00:01:09,699
WebRTC is notorious for bloating binaries.

23
00:01:10,500 --> 00:01:15,380
But if you're a developer who actually
used native terminal-first dictation with

24
00:01:15,399 --> 00:01:19,600
your own API keys, this is a pretty abrupt
breaking change.

25
00:01:20,279 --> 00:01:21,539
What are they supposed to do now?

26
00:01:22,000 --> 00:01:26,480
The official recommendation is to shift
those workflows over to the visual desktop

27
00:01:26,547 --> 00:01:32,160
application-the `codex app`-which is
better suited to handle heavy media pipelines.

28
00:01:32,400 --> 00:01:38,000
Or, if you absolutely must stay in the
terminal, you have to route your voice input

29
00:01:38,040 --> 00:01:41,840
through native operating system dictation
utilities instead.

30
00:01:42,500 --> 00:01:49,379
Right, so instead of Codex handling the
microphone stream, you let macOS or Windows

31
00:01:49,419 --> 00:01:52,380
do the heavy lifting of speech-to-text at
the system level,

32
00:01:52,779 --> 00:01:55,599
and just feed the plaintext into the
terminal.

33
00:01:55,659 --> 00:01:59,619
It's less integrated, but honestly, it
keeps the terminal client lean.

34
00:02:01,000 --> 00:02:01,720
Exactly.

35
00:02:01,820 --> 00:02:03,480
And that brings us to the latest

36
00:02:03,720 --> 00:02:10,680
0.140.0-alpha.15 hotfixes,

37
00:02:10,720 --> 00:02:15,480
which really show this strategic shift
toward tightening the CLI's core competency

38
00:02:15,583 --> 00:02:18,360
around raw text and tool execution.

39
00:02:18,540 --> 00:02:23,880
For starters, they had to roll back a
buggy URL parser that broke clickable terminal

40
00:02:23,960 --> 00:02:25,480
paths containing tildes.

41
00:02:27,439 --> 00:02:29,419
Oh, the tilde regression!

42
00:02:30,279 --> 00:02:32,439
That must have been incredibly annoying.

43
00:02:32,859 --> 00:02:39,659
If you can't click a relative path like
`~/projects/src` because the parser

44
00:02:39,699 --> 00:02:44,860
doesn't recognize the tilde, your
terminal-to-editor navigation completely breaks.

45
00:02:46,000 --> 00:02:46,720
Precisely.

46
00:02:46,853 --> 00:02:53,120
It was a major papercut for developer
velocity, so rolling that back in alpha 15 is

47
00:02:53,160 --> 00:02:54,160
a welcome fix.

48
00:02:54,307 --> 00:03:00,000
Along with that, they've also improved how
code-mode web search handles nested tool

49
00:03:00,080 --> 00:03:00,800
calls.

50
00:03:00,907 --> 00:03:04,640
It now returns highly optimized plaintext
context.

51
00:03:07,119 --> 00:03:12,779
Ah, so instead of dumping raw HTML or
heavily formatted markdown that bloats the

52
00:03:12,899 --> 00:03:16,279
LLM's context window, it strips all the
noise out?

53
00:03:17,000 --> 00:03:18,440
Yes, exactly.

54
00:03:18,600 --> 00:03:24,120
It prevents context window bloat by
keeping the payload strictly to clean,

55
00:03:24,182 --> 00:03:25,880
relevant plaintext.

56
00:03:25,987 --> 00:03:28,240
It really shows where their focus is.

57
00:03:28,296 --> 00:03:33,480
They are offloading the heavy media
components like voice to the desktop app,

58
00:03:33,520 --> 00:03:38,840
and hyper-focusing the terminal CLI on
what it does best: fast,

59
00:03:38,920 --> 00:03:43,400
low-overhead text manipulation and precise
tool execution.

60
00:03:44,319 --> 00:03:46,939
It's a classic Unix philosophy move.

61
00:03:47,439 --> 00:03:49,519
Do one thing and do it well.

62
00:03:50,119 --> 00:03:55,799
Let the desktop app be the heavy,
multi-modal playground, and keep our terminal fast

63
00:03:55,840 --> 00:03:56,460
and dependable.

64
00:03:57,000 --> 00:03:57,960
Couldn't agree more.

65
00:03:58,072 --> 00:04:01,320
That is all the time we have for this
quick take.

66
00:04:01,354 --> 00:04:04,480
Thanks for listening, and we'll catch you
in the next one.

67
00:04:05,079 --> 00:04:05,080
See ya!