1
00:00:01,021 --> 00:00:05,303
If you route Codex through Google Cloud
Vertex AI, your token bill may have just

2
00:00:05,323 --> 00:00:08,485
gotten a little smaller, and you didn't
touch a thing.

3
00:00:09,345 --> 00:00:11,767
Thanks to Jellypod AI for sponsoring this
episode.

4
00:00:12,727 --> 00:00:13,968
Now, the change itself.

5
00:00:14,468 --> 00:00:19,891
Codex CLI 0.167.2 shipped on November 11,
2026.

6
00:00:20,511 --> 00:00:26,715
That's npm package @openai/codex at
0.167.2, tag

7
00:00:26,915 --> 00:00:33,799
rust-v0.167.2, following PR 60617 and PR

8
00:00:33,979 --> 00:00:35,260
60848.

9
00:00:36,008 --> 00:00:36,789
Here's what changed.

10
00:00:37,889 --> 00:00:43,553
In earlier releases, a multi turn session
routed to Vertex AI resent the complete

11
00:00:43,593 --> 00:00:47,596
catalog of tool and function declarations
on every single request.

12
00:00:48,196 --> 00:00:50,778
Every turn, the same schemas, again.

13
00:00:52,479 --> 00:00:59,164
Now, in 0.167.2, Codex sends tool updates
incrementally by default on Vertex

14
00:00:59,284 --> 00:00:59,584
AI.

15
00:01:00,204 --> 00:01:04,607
The client transmits only the changes to
the tool catalog, not the unchanged

16
00:01:04,707 --> 00:01:05,328
definitions.

17
00:01:06,007 --> 00:01:10,691
If you heard the last episode on 0.167.1,
this will sound familiar.

18
00:01:11,232 --> 00:01:15,015
That one covered the incremental catalog
mechanism on Amazon Bedrock,

19
00:01:15,155 --> 00:01:19,518
where the flag was called
incremental_tools_legacy and it was opt in.

20
00:01:20,999 --> 00:01:23,021
Same mechanism, now on Google Cloud.

21
00:01:23,381 --> 00:01:25,143
But the rollout strategy flips.

22
00:01:25,783 --> 00:01:27,365
On Bedrock you had to turn it on.

23
00:01:27,745 --> 00:01:29,947
On Vertex AI it's already on.

24
00:01:30,635 --> 00:01:31,396
Why does that matter?

25
00:01:32,076 --> 00:01:36,259
Think about a long running agent task, or
a big suite of MCP servers.

26
00:01:37,320 --> 00:01:41,863
Each of those tools carries a schema, and
resending all of them every turn adds

27
00:01:41,923 --> 00:01:46,747
token overhead, burns through quota, and
adds latency to each API call.

28
00:01:47,427 --> 00:01:50,710
Cutting that out shrinks the request
payload and token consumption,

29
00:01:51,110 --> 00:01:52,551
with no configuration from you.

30
00:01:53,232 --> 00:01:57,615
That's the practical win: nothing to do,
and the requests just get lighter.

31
00:01:58,378 --> 00:01:59,459
So how do you use it?

32
00:02:00,420 --> 00:02:01,181
You don't, really.

33
00:02:01,401 --> 00:02:01,821
It's on.

34
00:02:02,442 --> 00:02:04,303
But you should know the escape hatch.

35
00:02:04,904 --> 00:02:08,807
If something between you and Vertex AI
chokes on the differential payloads,

36
00:02:09,327 --> 00:02:12,850
you can restore full transfers in
config.toml.

37
00:02:13,911 --> 00:02:20,116
Under the features table, set
incremental_tools_legacy_gcp equals false.

38
00:02:20,957 --> 00:02:21,738
That's the whole fix.

39
00:02:22,407 --> 00:02:25,650
And watch the naming, because it's easy to
trip on.

40
00:02:26,971 --> 00:02:31,795
Bedrock used incremental_tools_legacy as
an opt in toggle.

41
00:02:32,456 --> 00:02:39,381
Vertex uses incremental_tools_legacy_gcp,
and here it works as an opt out.

42
00:02:40,502 --> 00:02:42,884
Same family of names, opposite default.

43
00:02:43,485 --> 00:02:45,827
Setting the wrong one will do nothing
useful.

44
00:02:46,830 --> 00:02:47,431
Caveats.

45
00:02:48,172 --> 00:02:52,295
If you go through a legacy third party
proxy, a strict API gateway,

46
00:02:52,455 --> 00:02:56,358
or some intermediate Vertex endpoint that
doesn't expect differential tool

47
00:02:56,378 --> 00:03:01,682
definitions, tool call dispatch might
fail, or unrecognized declarations might get

48
00:03:01,722 --> 00:03:02,123
dropped.

49
00:03:03,364 --> 00:03:06,886
If tools suddenly vanish or calls stop
dispatching after upgrading,

50
00:03:07,407 --> 00:03:08,127
check that first.

51
00:03:08,808 --> 00:03:11,150
Flip the flag to false, and you're back to
full payloads.

52
00:03:11,800 --> 00:03:14,241
A few smaller things from the main branch
this past week.

53
00:03:14,802 --> 00:03:18,964
First, PR 61028, merged November 12, adds

54
00:03:19,104 --> 00:03:21,945
otel.headers_helper.

55
00:03:21,945 --> 00:03:28,509
This answers issue 48045, where exporting
to an authenticated OTLP HTTP collector

56
00:03:28,889 --> 00:03:31,950
meant putting a bearer token in plaintext
in your config file.

57
00:03:32,511 --> 00:03:35,472
Now you point the helper at a command, run
from CODEX_HOME,

58
00:03:35,812 --> 00:03:38,374
that returns a JSON object of header
strings.

59
00:03:39,334 --> 00:03:42,436
It refreshes every 29 minutes by default,
tunable with

60
00:03:42,536 --> 00:03:48,519
otel.headers_helper_refresh_interval_ms.

61
00:03:48,519 --> 00:03:50,560
So the secret can live in your OS
keychain.

62
00:03:51,878 --> 00:03:53,459
Details worth knowing.

63
00:03:54,040 --> 00:03:56,762
Helper headers override matching static
headers.

64
00:03:57,242 --> 00:04:01,345
If a refresh fails, affected exports block
until the helper recovers.

65
00:04:02,366 --> 00:04:05,749
HTTP redirects are disabled when helper
headers are used.

66
00:04:06,550 --> 00:04:12,476
And repository local config can't set the
helper, and gRPC exporters and admin

67
00:04:12,476 --> 00:04:14,136
managed exporters don't use it.

68
00:04:15,822 --> 00:04:16,222
Two more.

69
00:04:16,883 --> 00:04:23,067
PR 60872 switches Windows atomic file
writes to std::fs::rename,

70
00:04:23,507 --> 00:04:28,230
so overlapping readers don't cause locking
collisions, and temp file cleanup still

71
00:04:28,270 --> 00:04:29,451
happens if publishing fails.

72
00:04:29,451 --> 00:04:36,235
And PR 60964 trims whitespace around
OPENAI_API_KEY

73
00:04:36,615 --> 00:04:41,318
and rejects anything outside printable,
non space ASCII, so malformed authorization

74
00:04:41,358 --> 00:04:42,419
headers get caught early.

75
00:04:43,079 --> 00:04:45,360
Note that one landed in the issue labeler
workflow.

76
00:04:46,681 --> 00:04:47,582
Small, but handy.

77
00:04:48,278 --> 00:04:55,008
So: on Vertex AI, update to 0.167.2, enjoy
lighter requests,

78
00:04:55,428 --> 00:05:00,595
and remember incremental_tools_legacy_gcp
equals false if a proxy complains.

79
00:05:01,236 --> 00:05:02,377
That's the changelog for today.

