1
00:00:00,119 --> 00:00:05,299
If you have been running the Codex CLI
against Amazon Bedrock in a big enterprise

2
00:00:05,359 --> 00:00:08,920
environment, you, you know the exact wall
we have all been hitting.

3
00:00:09,539 --> 00:00:13,179
You get five or six turns into a multi
agent refactoring job,

4
00:00:13,699 --> 00:00:14,359
and suddenly...

5
00:00:15,739 --> 00:00:20,939
boom, context buffer exhausted, or your
turn latency climbs through the roof.

6
00:00:21,659 --> 00:00:28,519
Before we dig into how version 0.147.0
fixes this, a quick shoutout to Jellypod

7
00:00:28,579 --> 00:00:30,539
for helping make this daily show a
reality.

8
00:00:31,260 --> 00:00:35,360
But yeah, Maya, the, the history payload
problem on Bedrock was real.

9
00:00:35,837 --> 00:00:38,158
Oh, it, it was brutal, Ethan.

10
00:00:38,397 --> 00:00:44,117
Like, every single turn, local CLI
truncation was basically re streaming the full

11
00:00:44,258 --> 00:00:46,077
conversation history over the wire.

12
00:00:46,577 --> 00:00:51,697
If you had subagents doing background
passes, you were paying full token price over

13
00:00:51,737 --> 00:00:52,277
and over.

14
00:00:52,917 --> 00:00:57,777
But 0.147.0 drops two massive pull
requests for Bedrock.

15
00:00:58,237 --> 00:01:04,757
Pull request 36938, which enables cached
web search, and 36981 for remote

16
00:01:04,817 --> 00:01:06,037
conversation compaction.

17
00:01:06,536 --> 00:01:09,296
And that compaction piece is the real
shift here, right?

18
00:01:09,856 --> 00:01:13,816
Because instead of your local machine
trying to truncate and assemble this huge

19
00:01:13,917 --> 00:01:20,156
prompt payload, it actually offloads the
context window management directly to AWS

20
00:01:20,276 --> 00:01:21,216
Bedrock endpoints.

21
00:01:21,811 --> 00:01:23,191
Right, exactly!

22
00:01:23,711 --> 00:01:28,792
Bedrock now handles the prompt caching
layer and server side compaction natively.

23
00:01:29,371 --> 00:01:33,731
So when a long running agent session hits
a threshold, Bedrock compacts the

24
00:01:33,792 --> 00:01:36,651
conversation history right on the server
side.

25
00:01:36,731 --> 00:01:42,171
On multi agent workflows, AWS benchmark
tests showed turn latency dropped by up to

26
00:01:42,431 --> 00:01:43,292
60 percent.

27
00:01:44,767 --> 00:01:48,706
Sixty percent faster turns is huge when
you are sitting there watching a terminal.

28
00:01:49,346 --> 00:01:53,946
But for engineering managers, the real eye
opener is the AWS token billing

29
00:01:54,026 --> 00:01:54,386
statement.

30
00:01:55,026 --> 00:02:00,566
You are no longer re sending hundreds of
thousands of identical tokens on turn ten

31
00:02:00,607 --> 00:02:02,427
that you already sent on turn nine.

32
00:02:02,857 --> 00:02:04,297
Mm, totally.

33
00:02:05,037 --> 00:02:09,877
And from a developer machine perspective,
what were you noticing with local memory?

34
00:02:10,491 --> 00:02:13,711
Well, when you have three subagents
spinning in the background,

35
00:02:14,191 --> 00:02:19,752
keeping full uncompacted histories in
local memory was causing noticeable CLI

36
00:02:19,871 --> 00:02:20,891
stutter.

37
00:02:20,991 --> 00:02:26,151
Now that Bedrock owns the compaction state
remotely, the local CLI process stays

38
00:02:26,252 --> 00:02:29,411
lightweight, even during heavy code
refactoring runs.

39
00:02:29,959 --> 00:02:33,339
So if someone wants to turn this on in
their project today,

40
00:02:33,880 --> 00:02:37,939
what does the config look like in dot
codex slash config dot toml?

41
00:02:38,412 --> 00:02:39,672
It is pretty straightforward.

42
00:02:40,232 --> 00:02:46,273
You set CODEX PROVIDER to bedrock in
quotes, map your target AWS region endpoints,

43
00:02:46,833 --> 00:02:51,352
and then you can specify custom compaction
thresholds if you are running Claude

44
00:02:51,452 --> 00:02:52,432
models on Bedrock.

45
00:02:52,992 --> 00:02:56,132
But there are a couple of operational
caveats to keep in mind.

46
00:02:56,566 --> 00:02:58,907
Like cross region inference?

47
00:02:59,606 --> 00:03:04,206
I know AWS gets picky with Bedrock model
IDs if you are routing across regions.

48
00:03:04,744 --> 00:03:05,644
Yeah, precisely.

49
00:03:06,164 --> 00:03:11,364
You have to ensure your IAM roles permit
cross region inference profiles if your

50
00:03:11,424 --> 00:03:12,805
primary region hits quota.

51
00:03:13,505 --> 00:03:19,164
Also, cached search queries have their own
specific rate limits on the Bedrock side.

52
00:03:19,264 --> 00:03:23,524
And, uh, if you rely on local SQLite
transcript replays for debugging,

53
00:03:24,104 --> 00:03:28,824
remember that remote compaction means the
local log stores the compacted summary

54
00:03:28,965 --> 00:03:32,104
state rather than every raw verbose turn.

55
00:03:32,465 --> 00:03:35,644
Ah, that is a really good callout for
debugging teams.

56
00:03:36,304 --> 00:03:42,885
But speaking of security and logs, 0.147.0
brought some big quality of life

57
00:03:42,925 --> 00:03:43,884
patches too, right?

58
00:03:44,780 --> 00:03:51,399
Yes, pull request 36893 now automatically
redacts bearer tokens from command history

59
00:03:51,500 --> 00:03:51,879
logs.

60
00:03:52,599 --> 00:03:57,900
So if a tool execution outputs an
authorization header, it gets scrubbed before

61
00:03:57,979 --> 00:03:58,619
hitting disk.

62
00:03:59,586 --> 00:04:00,826
Thank goodness.

63
00:04:01,246 --> 00:04:04,366
No more accidental secret leaks in log
files.

64
00:04:05,026 --> 00:04:10,766
And they fixed that annoying terminal
input freeze, too, under pull request 35649,

65
00:04:10,886 --> 00:04:11,146
right?

66
00:04:12,245 --> 00:04:14,304
Man, that bug was frustrating.

67
00:04:14,884 --> 00:04:19,784
If you were starting up Model Context
Protocol servers or switching window focus in

68
00:04:19,864 --> 00:04:22,864
Ghostty, the terminal input would just
freeze up.

69
00:04:23,524 --> 00:04:24,744
That is finally resolved.

70
00:04:25,324 --> 00:04:32,305
Plus, pull request 36960 adds mandatory
project trust prompts whenever you

71
00:04:32,385 --> 00:04:34,984
open Codex in an unfamiliar local
directory.

72
00:04:35,382 --> 00:04:39,242
Which is huge for zero data retention
corporate security.

73
00:04:39,862 --> 00:04:44,522
Getting full Bedrock feature parity with
first party OpenAI endpoints means

74
00:04:44,662 --> 00:04:50,442
enterprise teams can finally run Codex CLI
at scale inside their private AWS

75
00:04:50,543 --> 00:04:51,003
perimeter.

76
00:04:51,591 --> 00:04:52,292
Absolutely.

77
00:04:52,731 --> 00:04:56,631
Alright, that is Codex 0.147.0.

78
00:04:57,232 --> 00:04:58,131
Talk to you next time.