1 00:00:00,340 --> 00:00:01,180 Picture this. 2 00:00:01,600 --> 00:00:04,180 You are in your terminal, working with an agent. 3 00:00:04,640 --> 00:00:09,800 You spend three minutes carefully typing out a complex multi line instruction in the 4 00:00:09,840 --> 00:00:15,140 composer, maybe pasting in a specific code fix, while the agent is still running a 5 00:00:15,180 --> 00:00:15,900 background turn. 6 00:00:16,552 --> 00:00:20,272 Right, and then the agent starts heading in completely the wrong direction, 7 00:00:20,852 --> 00:00:23,132 so you hit Control C to interrupt it. 8 00:00:24,003 --> 00:00:25,203 And poof! 9 00:00:26,603 --> 00:00:30,843 Your entire unsubmitted prompt, all those lines you just typed, 10 00:00:31,463 --> 00:00:32,963 vanish into thin air. 11 00:00:33,803 --> 00:00:34,503 Wiped clean. 12 00:00:34,955 --> 00:00:36,375 It is such a brutal feeling. 13 00:00:37,015 --> 00:00:43,775 But in Codex CLI version 0.157.0, that whole invisible draft 14 00:00:43,835 --> 00:00:45,615 wiper issue is finally gone. 15 00:00:46,315 --> 00:00:52,215 And before we dive into how pull request 47340 fixes it, a quick shoutout to 16 00:00:52,255 --> 00:00:54,695 Jellypod for helping make this daily show possible. 17 00:00:55,170 --> 00:00:59,410 So what actually changed under the hood when you hit Control C now? 18 00:00:59,750 --> 00:01:05,270 So previously, the terminal UI treated a turn cancellation as a complete reset of 19 00:01:05,290 --> 00:01:06,950 the composer input state. 20 00:01:07,150 --> 00:01:14,070 But PR 47340 introduces what OpenAI calls conditional turn 21 00:01:14,119 --> 00:01:14,790 interruption. 22 00:01:15,172 --> 00:01:20,072 Okay, meaning it separates stopping the model from clearing what you typed? 23 00:01:20,458 --> 00:01:21,738 Exactly. 24 00:01:21,788 --> 00:01:27,098 Instead of flushing the composer buffer, it intercepts Control C and turn completion 25 00:01:27,168 --> 00:01:31,418 signals, holding your pending keystrokes in a temporary state handle. 26 00:01:31,658 --> 00:01:36,938 So the running turn halts, but your text stays right there in the composer, 27 00:01:36,991 --> 00:01:38,938 ready for you to edit or send. 28 00:01:39,358 --> 00:01:41,038 Ah, that makes so much sense. 29 00:01:41,778 --> 00:01:46,538 But what if the agent raised an interactive question, like asking for clarification, 30 00:01:47,038 --> 00:01:51,258 and you typed an answer into the box, but then the turn ended before you pressed 31 00:01:51,318 --> 00:01:51,598 enter? 32 00:01:51,958 --> 00:01:57,638 That is where PR 47422 and PR 47423 come in. 33 00:01:57,814 --> 00:02:00,678 Maya, you were looking at those edge cases, right? 34 00:02:01,088 --> 00:02:01,548 Yeah! 35 00:02:02,068 --> 00:02:03,768 Those handle draft recovery. 36 00:02:04,328 --> 00:02:09,788 When a turn finishes or gets cut off, Codex now sweeps any unsent question answers 37 00:02:10,148 --> 00:02:12,128 straight back into the main composer box. 38 00:02:12,700 --> 00:02:17,100 Wait, does that disrupt you if you happen to be doing a reverse history search with 39 00:02:17,140 --> 00:02:20,560 Control R or forward slash at that exact second? 40 00:02:21,960 --> 00:02:26,220 No, and that is the clever part in PR 47423! 41 00:02:26,820 --> 00:02:31,240 The restoration happens quietly in the background without kicking you out of your 42 00:02:31,360 --> 00:02:32,640 active history search mode. 43 00:02:33,100 --> 00:02:37,340 So you do not lose your place in your command history while your draft gets saved. 44 00:02:37,808 --> 00:02:40,608 Now, how does this work when subagents are involved? 45 00:02:41,328 --> 00:02:46,548 Say a subagent prompts you for input using request user input async? 46 00:02:47,005 --> 00:02:51,105 If you have unsubmitted drafts from those async clarification questions, 47 00:02:51,505 --> 00:02:56,025 they get queued sequentially right back into the composer as editable text when the 48 00:02:56,105 --> 00:02:56,865 turn wraps up. 49 00:02:57,445 --> 00:03:04,265 Plus, PR 47424 clears out pending async question notifications automatically 50 00:03:04,665 --> 00:03:07,145 so you do not get spammed with stale popups. 51 00:03:07,680 --> 00:03:09,500 Okay, but there has to be a catch. 52 00:03:10,120 --> 00:03:11,940 Where are these drafts actually stored? 53 00:03:12,440 --> 00:03:14,960 That is a crucial limitation to keep in mind. 54 00:03:15,640 --> 00:03:19,460 Draft recovery lives entirely in the terminal client memory. 55 00:03:20,100 --> 00:03:25,280 If your local background server daemon suffers an unrecoverable crash or gets killed 56 00:03:25,320 --> 00:03:30,640 with a hard kill signal mid turn, those uncommitted drafts cannot be restored from 57 00:03:30,680 --> 00:03:31,740 SQLite checkpoints. 58 00:03:32,125 --> 00:03:32,885 Got it. 59 00:03:32,952 --> 00:03:37,565 So it protects against turn cancellations, subagent completions, 60 00:03:37,605 --> 00:03:41,485 and normal process stops, but not a full process crash. 61 00:03:41,900 --> 00:03:42,280 Right. 62 00:03:42,800 --> 00:03:48,420 Still, for day to day workflow, it removes a massive source of accidental work loss. 63 00:03:48,750 --> 00:03:55,550 There are a couple of other really handy quality of life updates in 0.157.0 as well. 64 00:03:55,950 --> 00:04:02,910 PR 47414 adds Shift click selection extension in fullscreen transcripts. 65 00:04:03,217 --> 00:04:04,337 Oh, finally! 66 00:04:04,777 --> 00:04:09,197 So if you have a huge block of code in the transcript, you can click the start, 67 00:04:09,637 --> 00:04:13,677 hold Shift, and click the end to select the whole chunk easily? 68 00:04:14,000 --> 00:04:15,440 Yep, exactly that. 69 00:04:15,560 --> 00:04:22,160 And PR 47122 increases the OpenAI file blob upload timeout from 70 00:04:22,320 --> 00:04:27,840 60 seconds all the way to 5 minutes, complete with automatic retry handling for 71 00:04:27,896 --> 00:04:29,360 transient network drops. 72 00:04:29,835 --> 00:04:35,875 That is huge if you are uploading large context files or diagnostic logs over a 73 00:04:36,075 --> 00:04:36,955 spotty connection. 74 00:04:37,375 --> 00:04:42,015 So if you want that safety net for your unsubmitted text and smoother transcript 75 00:04:42,066 --> 00:04:44,255 selections, the update is live. 76 00:04:44,399 --> 00:04:47,295 Just run npm install g 77 00:04:47,415 --> 00:04:53,695 @openai/codex@0.157.0 in your terminal. 78 00:04:54,068 --> 00:04:55,848 Definitely worth upgrading today. 79 00:04:56,548 --> 00:04:58,588 Alright, good chatting Ethan, talk soon!