1 00:00:00,300 --> 00:00:01,080 Morning, everyone. 2 00:00:01,680 --> 00:00:04,920 Thanks to Jellypod to help make this daily show a reality. 3 00:00:05,720 --> 00:00:12,100 Today's topic is Codex CLI 0.165.0, which shipped on October 4 00:00:12,300 --> 00:00:13,880 27th, 2026. 5 00:00:14,800 --> 00:00:19,280 The headline is that Google Cloud Vertex AI is now a built in provider. 6 00:00:20,040 --> 00:00:22,900 No gateway in front of it, no custom proxy script. 7 00:00:23,580 --> 00:00:30,080 It's pull request 56265, and it lands right after the Amazon Bedrock support we 8 00:00:30,120 --> 00:00:32,759 covered in 0.164.0. 9 00:00:33,440 --> 00:00:35,280 Same pattern, a second cloud host. 10 00:00:36,168 --> 00:00:38,568 Let's start with the config, because it's small. 11 00:00:39,248 --> 00:00:43,908 In your config.toml under the dot codex folder in your home directory, 12 00:00:44,428 --> 00:00:47,488 you set model_provider equals gcp. 13 00:00:47,948 --> 00:00:48,728 That's the whole switch. 14 00:00:49,548 --> 00:00:55,268 Internally the provider ID is gcp and the display name is Google Cloud Vertex AI. 15 00:00:56,008 --> 00:00:59,548 If you'd rather not touch the file, there's also a command line override for the 16 00:00:59,588 --> 00:01:03,628 model provider, so you can try it on a single run before you commit it to config. 17 00:01:04,482 --> 00:01:07,002 Authentication is where the Google side shows up. 18 00:01:07,502 --> 00:01:10,842 It uses standard Google Cloud Application Default Credentials. 19 00:01:11,282 --> 00:01:16,162 So if you've run gcloud auth application default login on your laptop, 20 00:01:16,622 --> 00:01:20,942 or your CI job or workload runs under a service account that ADC can find, 21 00:01:21,442 --> 00:01:22,462 Codex picks that up. 22 00:01:22,922 --> 00:01:24,542 You don't paste a key into Codex. 23 00:01:25,165 --> 00:01:27,785 Two environment variables matter next. 24 00:01:28,685 --> 00:01:34,125 GOOGLE_CLOUD_PROJECT is your project ID, so billing and quota land on the right 25 00:01:34,225 --> 00:01:34,645 project. 26 00:01:35,665 --> 00:01:41,025 GOOGLE_CLOUD_LOCATION is the region, and if you leave it unset it defaults to 27 00:01:41,085 --> 00:01:41,425 global. 28 00:01:42,285 --> 00:01:44,625 That default is fine for trying things out. 29 00:01:45,185 --> 00:01:49,925 But if your compliance team cares where inference runs, set a regional location 30 00:01:50,065 --> 00:01:50,765 explicitly. 31 00:01:51,445 --> 00:01:54,545 Don't rely on the default and assume it matches your policy. 32 00:01:55,480 --> 00:01:57,740 So a minimal setup is three things. 33 00:01:58,540 --> 00:02:05,100 Export your project ID, export a location if you need one, and set model_provider to 34 00:02:05,160 --> 00:02:05,640 gcp. 35 00:02:06,440 --> 00:02:08,180 Then run codex as usual. 36 00:02:08,980 --> 00:02:15,440 Install the release with npm install, dash g, at openai slash codex 37 00:02:16,060 --> 00:02:21,460 at 0.165.0, or just update however you installed it. 38 00:02:22,275 --> 00:02:23,055 Why does this matter? 39 00:02:23,535 --> 00:02:25,235 Think about who was blocked before. 40 00:02:25,915 --> 00:02:28,615 If you're an engineering org that's committed to Google Cloud, 41 00:02:28,875 --> 00:02:33,175 your constraints are things like Model Garden quotas, private network boundaries, 42 00:02:33,395 --> 00:02:35,475 IAM controls, and consolidated billing. 43 00:02:36,195 --> 00:02:41,055 Before this, routing Codex through that meant standing up something like a LiteLLM 44 00:02:41,135 --> 00:02:44,655 gateway, or writing your own translation layer, and then owning it. 45 00:02:45,355 --> 00:02:49,435 Now the routing is in the CLI itself, and the thing your security team has to review 46 00:02:49,535 --> 00:02:52,495 is a config line and a credential flow they already understand. 47 00:02:53,132 --> 00:02:57,252 Now the caveats, and they're the part to read before you roll this out. 48 00:02:58,092 --> 00:03:02,572 First, model availability is whatever you've provisioned in your Model Garden 49 00:03:02,652 --> 00:03:03,032 project. 50 00:03:03,712 --> 00:03:08,372 If a model isn't enabled or you don't have quota, Codex can't conjure it. 51 00:03:09,232 --> 00:03:14,752 Second, features tied to a direct OpenAI account don't carry over to the raw gcp 52 00:03:14,792 --> 00:03:15,212 provider. 53 00:03:15,852 --> 00:03:20,192 Native web search, server side image generation, and certain reasoning summary 54 00:03:20,312 --> 00:03:21,912 streams are not supported there. 55 00:03:22,592 --> 00:03:25,432 That's the same boundary you'd see on other cloud providers, 56 00:03:25,692 --> 00:03:26,492 including Bedrock. 57 00:03:27,092 --> 00:03:31,872 So if a workflow of yours leans on web search, test it on the gcp provider before 58 00:03:31,912 --> 00:03:33,852 you tell a team it's a drop in swap. 59 00:03:34,640 --> 00:03:36,000 One related fix helps here. 60 00:03:36,620 --> 00:03:42,540 PR 56445 makes Codex stop retrying when a provider reports that a model was not 61 00:03:42,600 --> 00:03:45,940 found, and it keeps the provider's original error message. 62 00:03:46,380 --> 00:03:50,140 That's exactly the failure you'll hit if you point at a model your project doesn't 63 00:03:50,200 --> 00:03:50,480 have. 64 00:03:51,000 --> 00:03:53,000 Before, you could watch it retry. 65 00:03:53,540 --> 00:03:56,140 Now it fails once, and the message tells you what's wrong. 66 00:03:56,960 --> 00:04:02,000 One more note on PR 56293: it mentions fixed token fixtures, 67 00:04:02,300 --> 00:04:04,880 but those are test fixtures for the Vertex tests. 68 00:04:05,260 --> 00:04:09,620 They're not a new feature for your CI, so don't read that as a headless auth option. 69 00:04:10,278 --> 00:04:13,458 A few smaller changes in this release are worth a minute. 70 00:04:14,198 --> 00:04:20,958 PR 56163 turns on reasoning effort changes during a conversation by default 71 00:04:21,138 --> 00:04:22,158 for supported models. 72 00:04:23,118 --> 00:04:28,718 In practice you can step effort up when a task gets hard, then step it back down for 73 00:04:28,758 --> 00:04:31,278 routine edits, without starting a new session. 74 00:04:32,038 --> 00:04:35,058 It's on by default, but only for models that support it. 75 00:04:35,950 --> 00:04:42,410 If you run parallel agents across linked git worktrees, PR 56308 is for you. 76 00:04:42,970 --> 00:04:47,870 The session picker now has a repository filter, and the agent command center groups 77 00:04:47,890 --> 00:04:48,510 by directory. 78 00:04:49,010 --> 00:04:52,870 That cuts the clutter when you've got a dozen sessions across several checkouts of 79 00:04:52,910 --> 00:04:53,570 the same repo. 80 00:04:54,437 --> 00:04:57,417 Memory v2 is now the default for fresh profiles. 81 00:04:58,097 --> 00:05:03,297 PR 56414 adds importing memories from external coding agents, 82 00:05:03,777 --> 00:05:09,277 and PR 56462 stops those imports from racing with consolidation. 83 00:05:10,177 --> 00:05:11,557 Note the word fresh. 84 00:05:12,317 --> 00:05:16,937 This is for new profiles, so check your own setup instead of assuming an existing 85 00:05:17,037 --> 00:05:17,857 profile changed. 86 00:05:18,597 --> 00:05:21,557 Two fixes for anyone who interrupts agents a lot. 87 00:05:21,937 --> 00:05:27,557 PR 56330 stops queued agent messages from restarting work after you interrupt, 88 00:05:27,777 --> 00:05:29,337 or after you switch to Plan mode. 89 00:05:29,937 --> 00:05:36,017 And PR 56132 cancels remote MCP tool requests when the local response future is 90 00:05:36,057 --> 00:05:39,057 dropped, so abandoned calls don't keep running on the server. 91 00:05:39,715 --> 00:05:41,135 Last one is security. 92 00:05:41,835 --> 00:05:48,095 PR 56263 rejects backslashes in the user info part of 93 00:05:48,315 --> 00:05:52,435 HTTP and HTTPS marketplace URLs. 94 00:05:53,275 --> 00:05:58,835 The gap was that Git could connect to a different host than the one checked against 95 00:05:58,935 --> 00:05:59,955 your allowed patterns. 96 00:06:00,795 --> 00:06:04,875 If you restrict marketplace sources by pattern, that's a reason to update. 97 00:06:05,732 --> 00:06:07,472 So the practical checklist is short. 98 00:06:07,912 --> 00:06:14,592 Update to 0.165.0, set model_provider to gcp, export your 99 00:06:14,632 --> 00:06:20,372 project and a location, confirm ADC works, and test any workflow that depends on web 100 00:06:20,412 --> 00:06:20,672 search. 101 00:06:21,292 --> 00:06:22,012 That's it for today.