1
00:00:00,140 --> 00:00:03,939
So I, I was halfway through a massive
codebase refactor yesterday,

2
00:00:04,679 --> 00:00:08,539
and, uh, right when you would expect the
model to start forgetting early files,

3
00:00:09,179 --> 00:00:14,839
Codex zero point one fifty three point
zero drops this new experimental context

4
00:00:14,920 --> 00:00:15,159
mode.

5
00:00:15,799 --> 00:00:18,859
Before we dive into what that actually
changes under the hood,

6
00:00:19,379 --> 00:00:22,819
quick shout out to Jellypod for helping
make this daily episode possible.

7
00:00:23,539 --> 00:00:29,000
But yeah, Maya, this zero point one fifty
three update introduces features dot

8
00:00:29,439 --> 00:00:34,099
context underscore management dot
experimental underscore mode,

9
00:00:34,639 --> 00:00:37,479
and it completely changes how long
sessions handle memory.

10
00:00:38,279 --> 00:00:40,279
It is about time!

11
00:00:40,899 --> 00:00:45,619
I mean, anyone who has used Codex for more
than twenty minutes knows the pain of

12
00:00:45,760 --> 00:00:46,800
turn compaction.

13
00:00:47,340 --> 00:00:50,819
You are deep in a debugging loop, and
suddenly the engine just,

14
00:00:50,920 --> 00:00:54,219
um, aggressively truncates your early
instructions.

15
00:00:54,840 --> 00:00:58,840
But with this flag enabled, instead of
just chopping off old messages,

16
00:00:59,239 --> 00:01:03,579
it uses token budget context allocation
and structured history notes.

17
00:01:04,177 --> 00:01:08,918
Right, and the really neat piece is the
model gets access to a brand new tool called

18
00:01:09,257 --> 00:01:11,017
new underscore context.

19
00:01:11,657 --> 00:01:16,338
So when the context window starts getting
crowded, the agent itself can decide to

20
00:01:16,397 --> 00:01:21,737
prune stale conversation branches while
explicitly preserving your critical session

21
00:01:21,798 --> 00:01:23,657
context and history notes.

22
00:01:24,257 --> 00:01:27,357
It is not just dropping tokens blindly
based on age.

23
00:01:27,736 --> 00:01:33,417
Which sounds fantastic on paper, but,
okay, we, we have to talk about who actually

24
00:01:33,496 --> 00:01:34,876
gets to use this right now.

25
00:01:35,516 --> 00:01:41,337
If you are on ChatGPT Plus, Pro, or Pro
Lite using the native Codex backend,

26
00:01:41,757 --> 00:01:42,716
you can turn it on.

27
00:01:43,297 --> 00:01:47,476
But if you are using API key sessions,
custom provider endpoints,

28
00:01:47,677 --> 00:01:52,476
or temporary structured threads, you are
completely excluded from this experimental

29
00:01:52,536 --> 00:01:52,856
flag.

30
00:01:53,324 --> 00:01:58,904
I get why OpenAI limits experimental
rollouts initially to their own managed

31
00:01:58,984 --> 00:01:59,884
infrastructure, though.

32
00:02:00,544 --> 00:02:05,524
From a product reliability standpoint,
controlling the backend context management

33
00:02:05,585 --> 00:02:10,904
pipeline lets them tune the token budgets
without worrying about third party proxy

34
00:02:10,965 --> 00:02:11,444
weirdness.

35
00:02:12,024 --> 00:02:18,084
And honestly, stopping those mid refactor
hallucinations when context overflows is a

36
00:02:18,125 --> 00:02:20,284
massive win for heavy CLI users.

37
00:02:20,812 --> 00:02:23,852
Sure, Ethan, but think about enterprise
teams!

38
00:02:24,333 --> 00:02:29,133
Most large corporate engineering groups
run Codex CLI routed through private API

39
00:02:29,272 --> 00:02:33,492
gateways or custom internal endpoints for
security and logging.

40
00:02:33,552 --> 00:02:38,892
Those are precisely the developers doing
huge, multi hour refactoring tasks who need

41
00:02:38,972 --> 00:02:43,612
token budgeting the most, and they are the
exact ones walled off from this flag!

42
00:02:44,042 --> 00:02:48,682
That is a fair point, but I bet it is just
an early stage constraint while they

43
00:02:48,746 --> 00:02:50,682
stabilize history note compaction.

44
00:02:50,842 --> 00:02:54,922
The architecture itself clearly points
toward broader support later.

45
00:02:55,276 --> 00:03:00,036
Well, while we wait for that expansion,
let us look at how you actually turn it on

46
00:03:00,156 --> 00:03:01,236
if you are eligible.

47
00:03:01,857 --> 00:03:08,337
In your dot codex slash config dot toml
file, you add bracket features dot context

48
00:03:08,377 --> 00:03:13,496
underscore management bracket, and then
set experimental underscore mode equals

49
00:03:13,576 --> 00:03:13,856
true.

50
00:03:14,316 --> 00:03:14,876
Simple enough.

51
00:03:15,576 --> 00:03:20,396
Though speaking of TOML configurations,
zero point one fifty three point zero also

52
00:03:20,476 --> 00:03:22,476
cleaned up some TUI settings, right?

53
00:03:22,907 --> 00:03:28,427
Yeah, they moved disable underscore paste
underscore burst under the TUI header,

54
00:03:28,987 --> 00:03:33,927
so now it lives under bracket tui bracket
instead of being a top level setting.

55
00:03:34,628 --> 00:03:38,867
Though the old location still works as a
fallback so your build will not break.

56
00:03:39,587 --> 00:03:43,187
And honestly, the terminal UI polish in
this release is stacked.

57
00:03:43,767 --> 00:03:47,987
The plugin CLI now supports remote
marketplaces directly from your shell.

58
00:03:48,507 --> 00:03:50,287
Oh, I loved seeing that!

59
00:03:50,767 --> 00:03:56,787
Being able to run codex plugin list, codex
plugin install, or remove directly from

60
00:03:56,807 --> 00:04:02,187
the command line without manually editing
local JSON manifest files is huge for

61
00:04:02,247 --> 00:04:03,208
developer workflow.

62
00:04:03,708 --> 00:04:07,208
It, it finally feels like a real package
manager inside the tool.

63
00:04:07,640 --> 00:04:08,600
Exactly!

64
00:04:09,079 --> 00:04:15,519
And for Vim mode power users in the
terminal composer, undo with u and redo with

65
00:04:15,619 --> 00:04:21,960
Control R now work smoothly across drafts,
even preserving pasted content and file

66
00:04:22,039 --> 00:04:22,679
attachments.

67
00:04:23,299 --> 00:04:28,959
Plus, if you find automatic turn recaps
distracting, you can now set tui dot auto

68
00:04:29,020 --> 00:04:32,159
underscore recap equals false in your
config.

69
00:04:32,587 --> 00:04:38,628
Wait, so if you turn off auto recap, you
can still run slash recap manually when you

70
00:04:38,708 --> 00:04:40,287
actually want a summary, right?

71
00:04:40,708 --> 00:04:41,428
Precisely.

72
00:04:41,565 --> 00:04:45,988
Manual slash recap stays available
whenever you need to catch up,

73
00:04:46,028 --> 00:04:50,068
but your terminal stays clean during fast,
iterative edits.

74
00:04:50,402 --> 00:04:54,102
See, I tested that setting during a multi
file edit session,

75
00:04:54,583 --> 00:04:59,562
and skipping that auto recap step makes
the terminal feel noticeably snappier

76
00:04:59,602 --> 00:05:00,843
between turns.

77
00:05:01,003 --> 00:05:05,503
You are not waiting for the TUI to format
a summary block before you can type your

78
00:05:05,543 --> 00:05:06,103
next prompt.

79
00:05:06,542 --> 00:05:08,142
It just get out of your way.

80
00:05:08,500 --> 00:05:12,660
And on top of that speed feel, they fixed
TUI reconnection resilience.

81
00:05:12,847 --> 00:05:17,540
If your external app server connection
drops mid session, it automatically

82
00:05:17,604 --> 00:05:21,700
reconnects, preserves your current draft
and full transcript,

83
00:05:21,760 --> 00:05:27,220
and keeps any queued or uncertain commands
safely paused for your review instead of

84
00:05:27,277 --> 00:05:28,580
firing them off blindly.

85
00:05:28,923 --> 00:05:31,362
That safety buffer is huge.

86
00:05:31,922 --> 00:05:37,322
Losing a draft or having a command execute
blindly after a brief network hiccup used

87
00:05:37,362 --> 00:05:39,022
to be so nerve wracking.

88
00:05:39,822 --> 00:05:45,003
Zero point one fifty three point zero
really balances deep context features with

89
00:05:45,083 --> 00:05:46,882
essential reliability tweaks.

90
00:05:47,167 --> 00:05:48,207
It really does.

91
00:05:48,271 --> 00:05:52,287
That is zero point one fifty three point
zero in a nutshell.

92
00:05:52,431 --> 00:05:57,327
Give experimental context mode a shot if
you are on a supported plan,

93
00:05:57,347 --> 00:05:59,007
and we will talk to you next time!