1 00:00:00,000 --> 00:00:03,600 So, I- I- I was spinning up a git worktree the other day, right? 2 00:00:03,720 --> 00:00:07,440 And normally, with CLI tools like Claude Code-this episode, 3 00:00:07,467 --> 00:00:11,520 by the way, is brought to you by Jellypod AI-the moment you bounce between 4 00:00:11,573 --> 00:00:16,720 directories or open a monorepo, you- you have to completely restart the session just 5 00:00:16,773 --> 00:00:21,440 so the Model Context Protocol, the MCP servers, actually know where they are. 6 00:00:22,169 --> 00:00:25,022 Right, which completely wipes your conversational history. 7 00:00:25,209 --> 00:00:28,062 It's a massive pain when you're deep in a debugging flow. 8 00:00:28,345 --> 00:00:29,065 Exactly! 9 00:00:29,185 --> 00:00:34,585 But in v2.1.203 and .204, they finally fixed this. 10 00:00:34,805 --> 00:00:39,785 Now, when Claude Code registers a new directory, it just appends it to the MCP 11 00:00:39,852 --> 00:00:41,865 roots/list dynamically. 12 00:00:42,005 --> 00:00:48,425 And it triggers notifications/roots/list_changed under the hood without a session 13 00:00:48,495 --> 00:00:49,145 restart. 14 00:00:49,912 --> 00:00:50,787 Oh, that's clever. 15 00:00:51,027 --> 00:00:57,187 So the workspace indexers can just listen for that notifications/roots/list_changed 16 00:00:57,240 --> 00:00:59,827 event and update their indexing path on the fly. 17 00:00:59,987 --> 00:01:03,587 No more tearing down the whole process just because you checked out a new branch or 18 00:01:03,644 --> 00:01:04,547 jumped folders. 19 00:01:04,858 --> 00:01:05,498 Precisely. 20 00:01:05,605 --> 00:01:06,938 It just keeps running. 21 00:01:07,029 --> 00:01:13,418 Though, speaking of things tearing down, that headless session bug in v2.1.204 is a 22 00:01:13,530 --> 00:01:16,698 huge save if you're running this in CI/CD pipelines. 23 00:01:16,771 --> 00:01:21,098 Apparently, if you had a SessionStart hook running in a remote environment, 24 00:01:21,125 --> 00:01:22,378 it was dead silent. 25 00:01:23,126 --> 00:01:24,113 Oh, let me guess. 26 00:01:24,173 --> 00:01:28,433 The silent stdout meant the remote orchestrator assumed the container was dead or 27 00:01:28,526 --> 00:01:30,353 stuck, and just reaped it? 28 00:01:30,650 --> 00:01:31,450 Spot on. 29 00:01:31,477 --> 00:01:34,730 It- it- it would get idle-reaped because of a silent startup hook. 30 00:01:35,130 --> 00:01:40,890 v2.1.204 now forces streaming of stdout during those SessionStart hooks. 31 00:01:41,103 --> 00:01:45,690 So the orchestrator sees the traffic, knows the container is actually doing work, 32 00:01:45,730 --> 00:01:47,930 and doesn't murder your build agent mid-run. 33 00:01:48,364 --> 00:01:51,030 That is going to save so many pipeline minutes. 34 00:01:51,190 --> 00:01:54,630 But tell me they did something about the macOS memory leaks. 35 00:01:54,790 --> 00:01:59,910 I've been getting these weird, random freezes since v2.1.196. 36 00:02:00,228 --> 00:02:02,948 Ah, the- the dreaded twenty-second stall! 37 00:02:03,044 --> 00:02:05,668 Yeah, that was a false-positive on a low-memory warning. 38 00:02:05,788 --> 00:02:09,348 The tool thought it was saving you from an out-of-memory crash when the RAM was 39 00:02:09,384 --> 00:02:10,468 actually fine. 40 00:02:10,708 --> 00:02:16,708 v2.1.203 completely bins that false threshold check, so no more random pauses on 41 00:02:16,815 --> 00:02:17,508 macOS. 42 00:02:18,347 --> 00:02:19,378 Thank goodness. 43 00:02:19,518 --> 00:02:21,178 I thought my machine was giving up on me. 44 00:02:21,378 --> 00:02:22,858 And what about Windows? 45 00:02:23,098 --> 00:02:26,778 I saw some devs complaining about auth errors on background agents. 46 00:02:27,083 --> 00:02:31,643 Yeah, they squashed a bug where background agents would throw a 401 unauthorized 47 00:02:31,723 --> 00:02:36,363 because they weren't inheriting the ANTHROPIC_BASE_URL environment variable properly 48 00:02:36,390 --> 00:02:37,243 on Windows. 49 00:02:37,403 --> 00:02:41,643 So if you were using a custom gateway or a proxy, it just fell over. 50 00:02:41,932 --> 00:02:45,532 Ah, that makes sense why my local proxy kept dropping connections! 51 00:02:45,652 --> 00:02:50,012 But the biggest CPU saver has to be the fix to the context-usage indicator. 52 00:02:50,103 --> 00:02:54,732 Before this, Claude Code was literally re-analyzing the entire conversation 53 00:02:54,797 --> 00:02:55,612 transcript... 54 00:02:56,035 --> 00:02:57,442 ...on every single token! 55 00:02:58,210 --> 00:03:00,322 Like, talk about burning cycles. 56 00:03:00,509 --> 00:03:04,882 No wonder my laptop fan sounded like a jet taking off during long chats. 57 00:03:05,182 --> 00:03:05,982 Exactly! 58 00:03:06,102 --> 00:03:10,782 Now it lazily evaluates that transcript state, so your terminal doesn't melt while 59 00:03:10,809 --> 00:03:12,542 it's actively streaming a response. 60 00:03:12,729 --> 00:03:17,262 If you're writing custom MCP servers for your monorepos, make sure they actually 61 00:03:17,308 --> 00:03:21,502 handle that notifications/roots/list_changed payload. 62 00:03:21,689 --> 00:03:25,982 If they do, your tools can dynamically index your code without you ever having to 63 00:03:26,032 --> 00:03:27,102 restart your session.