1 00:00:00,000 --> 00:00:05,200 Here's a small thing that quietly wrecks long sessions: Claude Code just stops 2 00:00:05,248 --> 00:00:08,320 searching the web, halfway through a big research job. 3 00:00:08,480 --> 00:00:13,120 This is the Claude Code Changelog, brought to you by Jellypod AI. 4 00:00:13,320 --> 00:00:14,160 I'm Lachlan. 5 00:00:14,500 --> 00:00:15,220 And I'm James. 6 00:00:15,300 --> 00:00:22,180 Today it's versions 2.1.290 and 2.1.291, released October 5th 7 00:00:22,220 --> 00:00:22,980 and 6th. 8 00:00:23,100 --> 00:00:29,060 Two changes: WebSearch now refills over time, and WebFetch no longer silently chops 9 00:00:29,120 --> 00:00:30,180 off long pages. 10 00:00:30,208 --> 00:00:31,248 WebSearch first. 11 00:00:31,397 --> 00:00:37,408 Before 2.1.290, an interactive session got a fixed budget of 200 searches. 12 00:00:37,588 --> 00:00:43,248 The changelog says the session ended after 200 calls, and you could adjust that cap 13 00:00:43,296 --> 00:00:48,448 with CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION, which came in around 2.1.212. 14 00:00:48,588 --> 00:00:54,288 Now the budget refills at 100 calls per hour, and a new variable sets that rate: 15 00:00:54,341 --> 00:00:56,528 CLAUDE_CODE_WEB_SEARCH_REFILLS_PER_HOUR. 16 00:00:56,542 --> 00:00:59,982 The practical effect is for long lived sessions. 17 00:01:00,082 --> 00:01:04,062 Say you've got a session doing deep research on a migration, 18 00:01:04,169 --> 00:01:06,622 or a /loop polling for something all day. 19 00:01:06,942 --> 00:01:10,062 Hitting the wall used to mean exiting and resuming. 20 00:01:10,242 --> 00:01:12,702 Now it just waits for the budget to come back. 21 00:01:12,802 --> 00:01:14,782 You don't have to do anything for the default. 22 00:01:14,942 --> 00:01:21,502 If you want faster, export CLAUDE_CODE_WEB_SEARCH_REFILLS_PER_HOUR equals 200. 23 00:01:21,609 --> 00:01:23,849 If you want a hard lid, set it to 0. 24 00:01:23,902 --> 00:01:26,078 The changelog says 0 turns it off. 25 00:01:26,126 --> 00:01:30,652 SessionWatcher reads that as turning refilling off, and Classmethod points out the 26 00:01:30,682 --> 00:01:34,526 changelog doesn't say whether it's the refilling or the quota itself. 27 00:01:34,590 --> 00:01:35,742 So test it before you rely on it in CI. 28 00:01:35,750 --> 00:01:36,790 And a caveat there. 29 00:01:36,910 --> 00:01:41,190 The announcement doesn't state a maximum the budget can build up to, 30 00:01:41,210 --> 00:01:43,270 and it doesn't split rates by plan. 31 00:01:43,483 --> 00:01:47,830 So if an agent burns a hundred searches in a burst, I'd expect to wait for the 32 00:01:47,887 --> 00:01:52,470 hourly refill to catch up, but that's my read, not something the release notes spell 33 00:01:52,550 --> 00:01:52,870 out. 34 00:01:53,046 --> 00:01:56,470 Also, this is a tool budget for interactive sessions. 35 00:01:56,603 --> 00:01:58,710 It's not your subscription message limit. 36 00:01:58,750 --> 00:02:00,990 Now WebFetch, which I think is the bigger deal. 37 00:02:01,063 --> 00:02:05,390 Previously it silently dropped page text past 100,000 characters. 38 00:02:05,550 --> 00:02:07,070 Silently is the problem. 39 00:02:07,203 --> 00:02:12,190 If the method you asked about lived at character 105,000 of an API reference, 40 00:02:12,236 --> 00:02:16,590 Claude had no signal anything was missing, so it could say the method doesn't exist. 41 00:02:16,750 --> 00:02:23,390 In 2.1.290 it tells Claude how much was unread, and it takes an offset parameter to 42 00:02:23,443 --> 00:02:24,430 continue from there. 43 00:02:24,458 --> 00:02:30,778 So with a huge OpenAPI spec or a long RFC, you can just say: fetch the next section 44 00:02:30,805 --> 00:02:33,098 of this spec using the offset parameter. 45 00:02:33,258 --> 00:02:37,178 Claude already knows how much is left, so it can page through. 46 00:02:37,378 --> 00:02:42,458 One watch out: every chunk is up to a hundred thousand characters going into your 47 00:02:42,548 --> 00:02:43,338 context. 48 00:02:43,482 --> 00:02:48,298 That will probably bring compaction on sooner, so ask for the next offset only when 49 00:02:48,318 --> 00:02:49,018 you need it. 50 00:02:49,042 --> 00:02:54,562 Important one: update straight to 2.1.291, with claude update. 51 00:02:55,122 --> 00:03:00,722 2.1.290 had a regression where cloud sessions could drop answers to permission 52 00:03:00,782 --> 00:03:01,282 prompts. 53 00:03:01,422 --> 00:03:08,402 And 2.1.291 also fixes a 2.1.288 regression where the last 54 00:03:08,455 --> 00:03:11,042 messages of a session could be lost when you quit. 55 00:03:11,042 --> 00:03:12,242 A few smaller ones. 56 00:03:12,422 --> 00:03:18,482 claude attach and claude logs now accept part of a session name instead of the full 57 00:03:18,642 --> 00:03:19,282 id. 58 00:03:19,410 --> 00:03:24,882 Scheduled tasks, like an interval /loop, now survive compaction on conversations 59 00:03:24,982 --> 00:03:27,442 compacted from this version on. 60 00:03:27,458 --> 00:03:33,538 And on permissions: Bash checks now ask approval for read only looking commands like 61 00:03:33,805 --> 00:03:38,898 rg or git grep when the arguments contain wildcards the shell would expand. 62 00:03:39,111 --> 00:03:45,858 So rg TODO star dot txt prompts you, but rg TODO 63 00:03:45,991 --> 00:03:48,658 notes.txt still passes. 64 00:03:48,951 --> 00:03:51,138 pyright also now needs approval. 65 00:03:51,264 --> 00:03:53,698 Expect a few more prompts after you update. 66 00:03:53,708 --> 00:03:54,428 That's the update. 67 00:03:54,495 --> 00:03:57,708 Check your version, set the refill variable if you need it, 68 00:03:57,728 --> 00:03:58,908 and we'll catch you next time.