1 00:00:00,300 --> 00:00:03,440 Thanks to Jellypod for helping make this daily show a reality. 2 00:00:04,220 --> 00:00:08,119 Today's topic is a small protocol change that matters if you build on the Codex 3 00:00:08,260 --> 00:00:14,480 app-server: PR 59601, which adds command output deferral to thread item listings. 4 00:00:15,220 --> 00:00:21,180 It was merged November 5th, and it lands in the 0.167 and 0.168 alpha line. 5 00:00:21,580 --> 00:00:24,640 One caveat up front: it's experimental, so don't expect it in stable. 6 00:00:25,398 --> 00:00:26,418 Some context. 7 00:00:27,198 --> 00:00:32,238 Codex 0.166.0 is still the stable release on npm. 8 00:00:32,898 --> 00:00:36,938 The 0.167 line shipped alphas but never got a stable release, 9 00:00:37,598 --> 00:00:44,278 and on November 4th the npm alpha tag moved to 0.168.0-alpha.1. 10 00:00:44,898 --> 00:00:47,238 So this feature lives in preview territory. 11 00:00:48,618 --> 00:00:51,658 The GitHub release entries for those alphas are one-line titles, 12 00:00:51,838 --> 00:00:53,938 so the details come from the PR itself. 13 00:00:55,002 --> 00:00:55,742 What changed. 14 00:00:56,342 --> 00:01:02,042 The endpoint is thread/items/list in the app-server JSON-RPC interface. 15 00:01:02,661 --> 00:01:07,862 There's a new parameter, deferOutput, and you pass it as an object: commandExecution 16 00:01:07,982 --> 00:01:08,542 set to true. 17 00:01:09,222 --> 00:01:11,662 You send it while querying a specific turnId. 18 00:01:12,342 --> 00:01:16,162 And because it's experimental, your client has to opt in during the initialize 19 00:01:16,222 --> 00:01:20,342 handshake, with capabilities.experimentalApi set to true. 20 00:01:20,702 --> 00:01:23,102 Without that, the server rejects experimental fields. 21 00:01:23,882 --> 00:01:24,962 Now the behavior. 22 00:01:25,722 --> 00:01:30,402 For a completed command, the server suppresses the raw output only if three things 23 00:01:30,442 --> 00:01:30,802 are true. 24 00:01:31,522 --> 00:01:33,282 It has a recorded completion time. 25 00:01:33,902 --> 00:01:35,342 It has an exit code of zero. 26 00:01:35,982 --> 00:01:38,122 And it has non-empty saved output. 27 00:01:39,022 --> 00:01:43,562 When all three hold, the item comes back with outputDeferred set to true, 28 00:01:44,062 --> 00:01:46,942 and item.aggregatedOutput set to null. 29 00:01:47,833 --> 00:01:48,653 Why care? 30 00:01:49,273 --> 00:01:53,193 Think about what thread/items/list does in a real client. 31 00:01:53,933 --> 00:01:57,833 You call it to render a turn: a timeline, a summary, a progress view. 32 00:01:58,353 --> 00:02:01,833 If that turn ran a big test suite, a build, or a noisy linter, 33 00:02:02,293 --> 00:02:06,713 you were pulling megabytes of stdout and stderr just to draw a list of steps. 34 00:02:07,973 --> 00:02:12,373 That means bigger payloads, latency spikes, and memory pressure in IDE extensions, 35 00:02:12,573 --> 00:02:14,633 CI runners, and custom web frontends. 36 00:02:15,333 --> 00:02:18,293 With deferral, you render the lightweight timeline instantly, 37 00:02:18,533 --> 00:02:21,573 and move the heavy bytes only when somebody clicks View Output, 38 00:02:22,013 --> 00:02:23,833 or when an agent needs to inspect a step. 39 00:02:24,508 --> 00:02:25,228 How do you use it? 40 00:02:25,908 --> 00:02:29,228 First, initialize with the experimental capability on. 41 00:02:29,868 --> 00:02:34,748 In your initialize params, clientInfo as usual, then capabilities with 42 00:02:34,808 --> 00:02:36,828 experimentalApi true. 43 00:02:37,488 --> 00:02:38,848 Then send initialized. 44 00:02:39,908 --> 00:02:46,228 Next, call thread/items/list with your turnId and deferOutput set to 45 00:02:46,288 --> 00:02:48,128 commandExecution true. 46 00:02:48,968 --> 00:02:53,288 Items that qualify come back with the null output and the outputDeferred flag. 47 00:02:54,008 --> 00:03:00,068 When your UI needs the text, call thread/items/read with that same turnId and the 48 00:03:00,148 --> 00:03:02,708 specific itemId, and you get the full item. 49 00:03:03,405 --> 00:03:05,605 The Rust SDK side deserves attention. 50 00:03:06,185 --> 00:03:10,085 Conversion from ThreadItemEntry to ThreadItem is now fallible. 51 00:03:10,625 --> 00:03:11,345 That's deliberate. 52 00:03:12,045 --> 00:03:16,645 Typed consumers have to explicitly detect deferred output, instead of silently 53 00:03:16,725 --> 00:03:18,705 treating an incomplete item as complete. 54 00:03:19,805 --> 00:03:23,905 So if you're a Rust consumer, expect to handle a Result where you used to have a 55 00:03:23,965 --> 00:03:26,225 plain conversion, with TryFrom semantics. 56 00:03:27,045 --> 00:03:30,265 A compile error now beats a quietly empty build log later. 57 00:03:30,925 --> 00:03:31,745 Caveats. 58 00:03:32,725 --> 00:03:33,725 One, scope. 59 00:03:34,265 --> 00:03:37,365 Deferral only applies when you target a single turnId. 60 00:03:37,825 --> 00:03:41,125 Listings that span multiple turns still return full output, 61 00:03:41,525 --> 00:03:44,205 and so does the default call when you don't pass the parameter. 62 00:03:45,165 --> 00:03:46,345 Two, eligibility. 63 00:03:46,925 --> 00:03:51,985 A command with a nonzero exit code is never deferred, so failure diagnostics stay 64 00:03:52,045 --> 00:03:53,065 immediately available. 65 00:03:53,765 --> 00:03:57,125 That means a noisy failing build still sends its whole log, 66 00:03:57,625 --> 00:04:00,025 so don't assume payloads shrink across the board. 67 00:04:01,045 --> 00:04:06,465 Three, your own code: anything that reads aggregatedOutput must now handle null. 68 00:04:07,195 --> 00:04:09,135 A few smaller items from the past week. 69 00:04:09,935 --> 00:04:14,255 PR 59233 fixes Windows sandbox ACL handling. 70 00:04:14,875 --> 00:04:19,715 It runs resolve_for_enforcement when building restricted token and elevated sandbox 71 00:04:19,795 --> 00:04:25,735 overrides, keeps read-only carveouts at volume roots, and expands deny-read ACL 72 00:04:25,775 --> 00:04:27,635 targets to prevent unintended leaks. 73 00:04:28,615 --> 00:04:34,595 PR 59677 removes the legacy summarize_sandbox_policy helper in favor of 74 00:04:34,635 --> 00:04:38,035 summarize_permission_profile, which matters if you call that internally. 75 00:04:38,455 --> 00:04:44,095 And PR 59512 modernizes the Python SDK build, requiring uv_build at least 76 00:04:44,395 --> 00:04:50,555 0.13.0 and below 0.14, with lockfiles refreshed to 0.13.6. 77 00:04:51,258 --> 00:04:55,978 On the wider ecosystem, OpenAI announced on November 4th that it's acquiring Pierre 78 00:04:56,078 --> 00:04:59,998 Computer Company, the makers of Code Storage, Trees, and Diffs, 79 00:05:00,198 --> 00:05:03,158 aimed at Git and source control infrastructure for coding agents. 80 00:05:03,738 --> 00:05:07,238 That's the announcement; what it means for Codex is not yet spelled out. 81 00:05:07,958 --> 00:05:08,658 That's the update. 82 00:05:09,078 --> 00:05:13,818 If you run an app-server client, try deferOutput on a preview build and see how much 83 00:05:13,858 --> 00:05:15,058 your turn listings shrink.