1
00:00:00,000 --> 00:00:06,720
So I was looking at the pull requests for
Codex version 0.146.0 last night,

2
00:00:06,760 --> 00:00:10,800
and thanks to Jellypod for helping us
bring this daily break down to life every

3
00:00:10,850 --> 00:00:11,280
morning.

4
00:00:11,460 --> 00:00:16,960
But Maya, they completely reworked how
tools actually get loaded into the model

5
00:00:17,050 --> 00:00:17,760
context.

6
00:00:18,109 --> 00:00:23,869
Wait, you mean instead of just dumping
every single Model Context Protocol skill

7
00:00:23,929 --> 00:00:26,109
into the system prompt at the start of a
session?

8
00:00:26,757 --> 00:00:27,476
Exactly.

9
00:00:28,076 --> 00:00:34,556
Pull requests 34547 and 34581 introduced
Reciprocal Rank Fusion,

10
00:00:34,996 --> 00:00:38,516
or RRF, right into the skill selection
pipeline.

11
00:00:39,156 --> 00:00:43,777
So now, instead of taking every tool in
your workspace and shoving hundreds of tool

12
00:00:43,857 --> 00:00:49,476
definitions into the prompt, it runs a
hybrid search over routing cards and lexical

13
00:00:49,517 --> 00:00:53,257
metadata to rank what you actually need
for that specific turn.

14
00:00:53,591 --> 00:00:55,311
Oh, that is huge.

15
00:00:55,931 --> 00:01:00,851
Because if you have a massive enterprise
set up with, say, fifty different MCP

16
00:01:00,951 --> 00:01:06,411
servers attached, those skill definitions
were just eating up thousands of tokens

17
00:01:06,651 --> 00:01:09,111
before you even typed a single word of
code.

18
00:01:09,738 --> 00:01:10,037
Right!

19
00:01:10,517 --> 00:01:16,797
And pull request 34626 ties those metadata
budgets directly to the active model

20
00:01:16,977 --> 00:01:17,798
context window.

21
00:01:18,497 --> 00:01:23,338
So if you are on a smaller context limit,
it tightens the budget dynamically.

22
00:01:24,077 --> 00:01:25,817
But here is the really clever part.

23
00:01:26,638 --> 00:01:32,138
Pull request 34738 adds a multi stage
fallback behavior.

24
00:01:32,507 --> 00:01:33,506
A fallback?

25
00:01:34,167 --> 00:01:37,746
Like, what happens when you start pushing
right up against that cap?

26
00:01:38,292 --> 00:01:41,012
It does not just drop the skill entirely
right away.

27
00:01:41,119 --> 00:01:46,372
First, Codex strips out the verbose
descriptions and long form documentation from

28
00:01:46,392 --> 00:01:50,372
the tool definitions, keeping just the raw
function signatures.

29
00:01:50,516 --> 00:01:55,812
That way it can retain more available
skills under tight context budgets before it

30
00:01:55,852 --> 00:01:57,492
has to resort to dropping anything.

31
00:01:57,939 --> 00:02:03,619
Wow, so it degrades gracefully instead of
just chopping off half your toolkit.

32
00:02:04,319 --> 00:02:07,039
That actually explains something I ran
into last week.

33
00:02:07,659 --> 00:02:10,340
You know that custom database query tool I
was building?

34
00:02:11,266 --> 00:02:15,167
The one with the three paragraph
description explaining how to join four legacy

35
00:02:15,186 --> 00:02:15,506
tables?

36
00:02:16,821 --> 00:02:19,141
Yes, that exact one!

37
00:02:19,781 --> 00:02:24,741
I kept wondering why the agent would
randomly pretend the tool did not exist

38
00:02:25,042 --> 00:02:26,701
whenever my context got deep.

39
00:02:27,382 --> 00:02:31,901
I ended up rewriting the routing card
keywords to be super tight and punchy,

40
00:02:32,261 --> 00:02:38,161
like three words instead of a essay, and
boom, it stayed in the active set every

41
00:02:38,222 --> 00:02:38,861
single time.

42
00:02:39,246 --> 00:02:45,506
And that brings us right to pull request
34785, which fixes that exact silent drop

43
00:02:45,587 --> 00:02:46,667
issue you were dealing with.

44
00:02:47,387 --> 00:02:51,906
Codex now explicitly warns when skill
catalogs must be truncated.

45
00:02:52,486 --> 00:02:56,787
You get a clear terminal warning right in
the interface showing you that a tool was

46
00:02:56,886 --> 00:02:58,847
left out due to space constraints.

47
00:03:00,056 --> 00:03:02,776
Man, where was that warning three weeks
ago?

48
00:03:03,156 --> 00:03:05,496
That would have saved me two hours of head
scratching.

49
00:03:06,336 --> 00:03:07,696
No more mystery drops.

50
00:03:08,224 --> 00:03:08,745
Seriously.

51
00:03:09,304 --> 00:03:15,744
And speaking of keeping your workflow
clean, pull request 34605 updated session

52
00:03:15,824 --> 00:03:16,345
management.

53
00:03:16,965 --> 00:03:21,244
You can now pass string names straight
into slash new or slash clear.

54
00:03:21,884 --> 00:03:26,364
So you can type slash new feature test or
slash clear bugfix,

55
00:03:26,844 --> 00:03:31,164
and it labels or resets the branch right
there without breaking your flow.

56
00:03:31,629 --> 00:03:32,969
Oh, I love that.

57
00:03:33,409 --> 00:03:38,170
I am constantly spinning up quick side
experiments and forgetting which temporary

58
00:03:38,289 --> 00:03:39,529
session was doing what.

59
00:03:40,149 --> 00:03:44,590
Being able to name them on creation is
such a nice quality of life bump.

60
00:03:44,917 --> 00:03:46,117
Same here.

61
00:03:46,197 --> 00:03:52,917
And under the hood in pull request 34952,
they added automatic MCP server

62
00:03:52,997 --> 00:03:53,797
reconnection.

63
00:03:54,064 --> 00:03:59,797
If an underlying server drops or needs to
re authenticate, Codex automatically

64
00:03:59,853 --> 00:04:04,757
reconnects in the background without
dropping your active user interface state or

65
00:04:04,813 --> 00:04:07,797
interrupting parallel connections that are
working fine.

66
00:04:08,151 --> 00:04:13,011
That is so important for reliability when
you are chaining multiple remote agents

67
00:04:13,051 --> 00:04:13,431
together.

68
00:04:14,131 --> 00:04:20,792
And speaking of multi platform workflows,
pull request 35105 added native support

69
00:04:20,831 --> 00:04:22,511
for Agent Plugin manifests.

70
00:04:22,991 --> 00:04:27,251
So you can publish workspace configs
across Amazon Bedrock and Claude Code

71
00:04:27,312 --> 00:04:29,851
toolchains without rewriting your
definitions.

72
00:04:30,286 --> 00:04:35,147
It really feels like the team is focusing
on making the whole ecosystem feel like a

73
00:04:35,226 --> 00:04:39,746
polished product rather than a bunch of
loose developer scripts held together with

74
00:04:39,786 --> 00:04:40,026
tape.

75
00:04:41,122 --> 00:04:42,523
Absolutely.

76
00:04:42,662 --> 00:04:45,902
Alright, I am going to go go prune some
more tool descriptions.

77
00:04:46,322 --> 00:04:47,182
Talk to you tomorrow, Ethan.

78
00:04:47,583 --> 00:04:48,383
Catch you later.