1 00:00:00,280 --> 00:00:05,780 If you've ever pointed Codex at a gateway that isn't OpenAI and watched a long 2 00:00:05,920 --> 00:00:09,560 session fall over right as it approached the context limit, 3 00:00:10,100 --> 00:00:11,420 today's episode is for you. 4 00:00:12,380 --> 00:00:16,340 And thanks to Jellypod for helping make this daily show a reality. 5 00:00:17,340 --> 00:00:24,140 The change is in Codex CLI 0.162.0, released October 8th, 2026. 6 00:00:24,940 --> 00:00:29,920 It adds a capabilities block for custom model providers that speak the Responses 7 00:00:30,060 --> 00:00:30,540 API. 8 00:00:31,280 --> 00:00:37,860 It lives in your config.toml under model_providers, then your provider ID, 9 00:00:38,260 --> 00:00:39,480 then capabilities. 10 00:00:40,235 --> 00:00:41,135 Here's why it matters. 11 00:00:41,795 --> 00:00:44,795 Codex only speaks the Responses wire protocol now. 12 00:00:45,315 --> 00:00:50,575 The old chat wire API is gone, and if you configure it you get a deserialization 13 00:00:50,635 --> 00:00:50,855 error. 14 00:00:51,555 --> 00:00:55,235 For first party providers, meaning OpenAI and Azure Responses, 15 00:00:55,735 --> 00:00:58,055 remote compaction defaults to version two. 16 00:00:58,655 --> 00:01:01,575 That means when a conversation gets close to the context window, 17 00:01:02,075 --> 00:01:08,615 Codex sends a compaction_trigger item over the slash v1 slash responses protocol and 18 00:01:08,635 --> 00:01:10,215 lets the server handle the truncation. 19 00:01:10,875 --> 00:01:15,655 The problem is that local gateways, private proxies, and third party engines often 20 00:01:15,715 --> 00:01:16,655 can't parse that item. 21 00:01:17,375 --> 00:01:23,475 People running things like LiteLLM, vLLM, or Ollama behind a Responses shim reported 22 00:01:23,675 --> 00:01:29,035 HTTP 400 errors, dropped WebSockets, or aborted tasks right when a long session 23 00:01:29,075 --> 00:01:29,875 needed compaction. 24 00:01:30,435 --> 00:01:33,035 Short sessions looked fine, which made it painful to debug. 25 00:01:33,673 --> 00:01:35,353 So there are two switches. 26 00:01:36,013 --> 00:01:41,013 The first is external_web_access, a boolean. 27 00:01:41,853 --> 00:01:46,593 It controls whether hosted web search, initiated by or for that provider, 28 00:01:47,133 --> 00:01:49,013 is allowed to reach the live internet. 29 00:01:49,913 --> 00:01:56,033 Set it to false and an on premises or air gapped provider can't leak requests out 30 00:01:56,093 --> 00:01:56,833 through web search. 31 00:01:57,653 --> 00:02:03,233 The second is remote_compaction, and it's a string with two valid values. 32 00:02:03,953 --> 00:02:09,933 Version two, written as the string v2, sends those compaction_trigger items to the 33 00:02:09,992 --> 00:02:10,293 server. 34 00:02:11,173 --> 00:02:16,333 Unsupported, written as the string unsupported, tells Codex to skip that and do the 35 00:02:16,373 --> 00:02:18,533 compaction locally on your machine instead. 36 00:02:19,357 --> 00:02:20,657 Let me walk through a config. 37 00:02:21,237 --> 00:02:24,617 At the top level, you set model to whatever your gateway serves, 38 00:02:24,817 --> 00:02:30,057 say mistral large latest, and model_provider to a provider ID. 39 00:02:30,437 --> 00:02:31,917 I'll call it corpgateway. 40 00:02:32,737 --> 00:02:37,637 Then you add a table called model_providers dot corpgateway with a name, 41 00:02:37,977 --> 00:02:44,417 a base_url pointing at your proxy's slash v1 endpoint, wire_api set to the string 42 00:02:44,457 --> 00:02:49,577 responses, and env_key naming the environment variable that holds your key, 43 00:02:49,937 --> 00:02:52,777 for example CORP_AI_KEY. 44 00:02:53,697 --> 00:02:59,377 Then comes the new part, a second table called model_providers dot corpgateway dot 45 00:02:59,437 --> 00:03:00,097 capabilities. 46 00:03:00,777 --> 00:03:07,357 Inside it, external_web_access equals false, and remote_compaction equals the 47 00:03:07,437 --> 00:03:08,497 string unsupported. 48 00:03:08,997 --> 00:03:09,757 That's the whole fix. 49 00:03:10,537 --> 00:03:12,997 You can also test it ad hoc without editing the file. 50 00:03:13,537 --> 00:03:19,617 Run codex with the c flag and pass model_providers dot corpgateway dot capabilities 51 00:03:19,917 --> 00:03:23,457 dot remote_compaction equals quoted unsupported. 52 00:03:24,137 --> 00:03:28,017 And codex doctor will confirm your provider configuration and connection health. 53 00:03:28,673 --> 00:03:29,753 Now the caveats. 54 00:03:30,553 --> 00:03:33,013 First, local compaction isn't free. 55 00:03:34,033 --> 00:03:39,073 When Codex compacts on the client, it spends extra prompt round trips and tokens to 56 00:03:39,113 --> 00:03:41,933 condense the history, instead of the server doing it. 57 00:03:42,553 --> 00:03:45,553 So you're trading a crash for some cost and latency. 58 00:03:46,633 --> 00:03:50,813 Second, any capability you omit inherits the provider defaults, 59 00:03:51,333 --> 00:03:55,013 and those assume an OpenAI style Responses feature set. 60 00:03:55,973 --> 00:03:59,593 If you leave remote_compaction out on a gateway that can't handle it, 61 00:04:00,073 --> 00:04:01,233 you're back where you started. 62 00:04:01,933 --> 00:04:03,093 So be explicit. 63 00:04:03,733 --> 00:04:08,053 Don't assume your proxy supports the full surface just because it accepts basic 64 00:04:08,093 --> 00:04:08,573 requests. 65 00:04:09,303 --> 00:04:12,823 One more thing to watch if you use model_catalog_url. 66 00:04:13,523 --> 00:04:20,303 Since 0.160.0, bundled models are no longer merged in with a provider's 67 00:04:20,343 --> 00:04:21,323 remote catalog. 68 00:04:22,023 --> 00:04:27,403 That means the catalog your provider returns has to explicitly list every model you 69 00:04:27,423 --> 00:04:28,063 want to select. 70 00:04:28,683 --> 00:04:31,723 If a model you expect is missing from the picker after this change, 71 00:04:32,203 --> 00:04:35,383 check that catalog before you blame the capabilities block. 72 00:04:35,815 --> 00:04:41,555 The companion fix is 0.162.1, which shipped October 9th. 73 00:04:42,375 --> 00:04:48,995 In 0.162.0, if the shared background daemon was running with a different feature set 74 00:04:49,035 --> 00:04:51,995 than your CLI defaults, startup could fail. 75 00:04:52,815 --> 00:04:57,135 The usual trigger was launching Codex from different folders or worktrees. 76 00:04:57,975 --> 00:05:04,655 The patch, pull request 52648, narrows that compatibility check so it applies only 77 00:05:04,695 --> 00:05:09,415 to explicit command line feature overrides, meaning the c flag, 78 00:05:09,675 --> 00:05:12,075 the enable flag, and the disable flag. 79 00:05:12,995 --> 00:05:18,375 If you hit a silent startup lockout on 0.162.0, upgrade to 80 00:05:18,635 --> 00:05:20,215 0.162.1. 81 00:05:21,115 --> 00:05:25,235 The same release fixes a terminal crash when asynchronous questions contained 82 00:05:25,415 --> 00:05:26,295 multiple lines. 83 00:05:26,855 --> 00:05:29,995 It now keeps the line breaks and the full hyperlink destinations. 84 00:05:30,888 --> 00:05:34,868 A few smaller things from 0.162.0 are worth knowing. 85 00:05:35,388 --> 00:05:39,928 In the agent Command Center, press p on a task and it moves into a shared Pinned 86 00:05:40,028 --> 00:05:41,828 group, when the server supports it. 87 00:05:42,368 --> 00:05:46,888 For transcripts, the slash copy command lets you navigate and copy transcript 88 00:05:46,908 --> 00:05:52,928 blocks, Control plus Insert copies a selection, and tui.mouse_scroll_speed tunes the 89 00:05:52,968 --> 00:05:53,488 mouse wheel. 90 00:05:53,888 --> 00:05:56,548 It takes a positive number, and the default is 1.0. 91 00:05:57,348 --> 00:06:01,168 URLs are now clickable in approval headers, questions, MCP prompts, 92 00:06:01,308 --> 00:06:04,108 warnings, and banners, even when they wrap across lines. 93 00:06:04,788 --> 00:06:10,188 On Windows or DOS files, apply_patch now preserves existing CRLF line endings by 94 00:06:10,228 --> 00:06:12,788 default, so you don't need the old opt in flag. 95 00:06:13,388 --> 00:06:18,608 And Codex now honors the server's Retry After header on retryable Responses and 96 00:06:18,688 --> 00:06:22,108 WebSocket failures, which helps when a provider is shedding load. 97 00:06:22,828 --> 00:06:25,648 One calendar item, and it's today. 98 00:06:26,528 --> 00:06:33,048 October 14th, 2026 is the retirement date for GPT 5.5 in 99 00:06:33,108 --> 00:06:37,628 ChatGPT, ChatGPT Work, and Codex, on all plans. 100 00:06:38,528 --> 00:06:41,388 It doesn't apply to the OpenAI API. 101 00:06:42,568 --> 00:06:49,568 If any config, script, scheduled task, or custom agent still names gpt 5.5, 102 00:06:50,068 --> 00:06:50,868 update it now. 103 00:06:51,888 --> 00:06:58,728 The options are GPT 6.1 Sol, GPT 6 Sol, or GPT 6 Luna. 104 00:06:59,668 --> 00:07:05,828 Grep your repos and your config.toml, and then run codex doctor once to check 105 00:07:05,968 --> 00:07:07,148 everything still resolves.