1 00:00:00,000 --> 00:00:02,160 Hey everyone, welcome to the show! 2 00:00:02,400 --> 00:00:07,760 I'm Ethan Park, here with Maya, and a quick shoutout before we dive in: thanks to 3 00:00:07,840 --> 00:00:10,960 Jellypod for helping make this daily show a reality. 4 00:00:11,133 --> 00:00:18,000 Today we are looking at the Codex CLI 0.139.0 release, and they've 5 00:00:18,058 --> 00:00:21,360 introduced a feature that solves a massive headache for me. 6 00:00:22,352 --> 00:00:27,280 Code mode can now call a standalone web search directly, even from deep inside 7 00:00:27,349 --> 00:00:29,120 nested JavaScript tool calls. 8 00:00:30,340 --> 00:00:33,079 Wait, nested JavaScript tool calls? 9 00:00:33,619 --> 00:00:38,800 So, if you're running a script inside the CLI and it hits a wall with an external 10 00:00:38,880 --> 00:00:42,340 API, the script itself can trigger a search query? 11 00:00:43,000 --> 00:00:44,040 Exactly. 12 00:00:44,093 --> 00:00:47,560 Think about when you're writing a script to automate an integration, 13 00:00:47,620 --> 00:00:51,960 and suddenly you get a 404 because the API endpoint changed yesterday. 14 00:00:52,858 --> 00:00:57,008 Instead of the LLM throwing its hands up or forcing you to drop out of your tool 15 00:00:57,075 --> 00:01:01,808 chain, Codex can issue a direct web search to grab the latest docs, 16 00:01:01,840 --> 00:01:04,368 pull the correct schema, and keep executing. 17 00:01:05,379 --> 00:01:10,480 Okay, that's incredibly useful for things like undocumented npm updates. 18 00:01:10,980 --> 00:01:13,260 But how is it handling the data payload? 19 00:01:13,720 --> 00:01:16,079 Web pages are notoriously noisy. 20 00:01:16,500 --> 00:01:21,779 If it's sucking down raw HTML or even heavy markdown, you're going to blow through 21 00:01:21,819 --> 00:01:23,999 your token context window in seconds. 22 00:01:24,000 --> 00:01:25,200 They actually thought of that. 23 00:01:25,320 --> 00:01:29,040 The engine bypasses heavy markdown parsing entirely. 24 00:01:29,253 --> 00:01:33,760 It streamlines the query and returns flat, clean plaintext. 25 00:01:34,000 --> 00:01:40,080 No massive HTML tables, no complex nested lists -- just the text. 26 00:01:40,548 --> 00:01:44,160 But the flip side is you have to write incredibly precise queries. 27 00:01:44,427 --> 00:01:49,680 If you search for something generic like "react error," you'll just get a wall of 28 00:01:49,760 --> 00:01:50,880 unhelpful text. 29 00:01:51,439 --> 00:01:56,759 Right, because you don't have visual cues or layout structure to help filter the 30 00:01:56,819 --> 00:01:58,039 signal from the noise. 31 00:01:58,879 --> 00:02:02,079 It makes you realize how much we rely on formatting. 32 00:02:02,920 --> 00:02:08,260 So, if I'm feeding this stack traces, I need to make sure the CLI query targeting is 33 00:02:08,440 --> 00:02:13,479 super specific -- like targeting the exact error code and package version. 34 00:02:14,000 --> 00:02:14,920 Precisely. 35 00:02:14,990 --> 00:02:20,240 You want to pass things like the exact npm package version and the specific method 36 00:02:20,304 --> 00:02:20,920 name. 37 00:02:21,040 --> 00:02:25,200 If you do that, the plaintext return is incredibly efficient. 38 00:02:25,253 --> 00:02:28,080 It's basically a surgical strike for documentation. 39 00:02:29,401 --> 00:02:34,019 That efficiency focus seems to carry over to how they're handling schemas in this 40 00:02:34,079 --> 00:02:34,759 release too. 41 00:02:35,399 --> 00:02:39,079 I saw they did some serious hardening on the Model Context Protocol, 42 00:02:39,340 --> 00:02:41,000 or MCP, schemas. 43 00:02:41,579 --> 00:02:46,940 Apparently, large tool and connector schemas now preserve `oneOf` and `allOf` 44 00:02:47,039 --> 00:02:48,399 structures during compaction. 45 00:02:49,000 --> 00:02:50,800 Oh, thank goodness. 46 00:02:51,982 --> 00:02:57,560 Previously, when Codex tried to compact those massive JSON schemas to save tokens, 47 00:02:57,587 --> 00:03:02,560 it would occasionally flatten or strip out `oneOf` and `allOf` blocks. 48 00:03:02,650 --> 00:03:07,720 And if you're relying on strict custom validation for complex API payloads, 49 00:03:07,789 --> 00:03:12,120 losing those logical operators completely breaks your runtime validation. 50 00:03:13,560 --> 00:03:14,620 It's the worst. 51 00:03:15,159 --> 00:03:19,099 You end up with a compacted schema that says any input is fine, 52 00:03:19,579 --> 00:03:24,380 and then your backend crashes because a required mutually exclusive field was 53 00:03:24,439 --> 00:03:24,859 missing. 54 00:03:26,119 --> 00:03:31,440 Preserving those structures means we keep custom MCP validation fully intact, 55 00:03:31,899 --> 00:03:33,100 even after compaction. 56 00:03:34,000 --> 00:03:38,560 And speaking of things breaking under the hood, they also cleaned up some CLI 57 00:03:38,622 --> 00:03:40,240 argument parsing bugs. 58 00:03:40,427 --> 00:03:47,120 If you ran `codex resume --last` or `codex fork --last` with a 59 00:03:47,164 --> 00:03:50,240 trailing string, the parser used to get confused. 60 00:03:50,453 --> 00:03:55,760 It would try to interpret that trailing string as a session ID instead of the new 61 00:03:55,817 --> 00:03:57,120 prompt you wanted to run. 62 00:03:58,600 --> 00:04:01,759 Ah, the classic CLI ambiguity. 63 00:04:02,440 --> 00:04:07,679 So if you typed `codex resume --last "refactor the auth route"`, 64 00:04:08,159 --> 00:04:12,639 it would look for a session called "refactor the auth route" and error out. 65 00:04:13,000 --> 00:04:13,720 Exactly. 66 00:04:13,827 --> 00:04:16,040 It was a frustrating papercut. 67 00:04:16,253 --> 00:04:20,040 In 0.139.0, that's fixed. 68 00:04:20,227 --> 00:04:25,720 It correctly identifies the trailing string as your new input prompt for the resumed 69 00:04:25,770 --> 00:04:26,200 session. 70 00:04:26,360 --> 00:04:32,360 They also updated `codex doctor` -- it now prints out your local editor and pager 71 00:04:32,420 --> 00:04:38,360 details, but they've safely redacted sensitive raw values in the JSON outputs so you 72 00:04:38,400 --> 00:04:42,440 don't accidentally leak API keys when sharing debug logs. 73 00:04:43,379 --> 00:04:44,659 That's a smart touch. 74 00:04:45,079 --> 00:04:49,979 It's those little developer-experience guardrails that make a tool feel mature. 75 00:04:50,460 --> 00:04:53,639 It shows they're thinking about the real-world environments where these tools 76 00:04:53,719 --> 00:04:54,559 actually run. 77 00:04:55,000 --> 00:04:55,960 Absolutely. 78 00:04:56,120 --> 00:05:00,440 It makes you wonder how long before we see these clean plaintext search patterns 79 00:05:00,543 --> 00:05:04,920 become the default for all developer-facing LLM integrations. 80 00:05:05,032 --> 00:05:06,920 That's all for today's quick take. 81 00:05:07,160 --> 00:05:08,360 I'm Ethan Park. 82 00:05:10,339 --> 00:05:10,340 And I'm Maya. 83 00:05:10,379 --> 00:05:11,279 We'll see you in the next one.