1 00:00:00,220 --> 00:00:04,860 G'day, this is The Claude Code Changelog, brought to you by Jellypod AI. 2 00:00:05,520 --> 00:00:09,080 Quick one for anyone running Claude Code through Bedrock, Vertex, 3 00:00:09,260 --> 00:00:10,240 Foundry or Mantle. 4 00:00:10,860 --> 00:00:17,440 Version 2.1.312 landed on October 27th, and it changes what prompt your custom model 5 00:00:17,560 --> 00:00:19,700 names get without you touching anything. 6 00:00:20,420 --> 00:00:20,600 Right. 7 00:00:20,780 --> 00:00:26,860 Before 2.1.312, a custom deployment name or an inference profile ARN was something 8 00:00:26,900 --> 00:00:28,560 Claude Code couldn't identify. 9 00:00:29,180 --> 00:00:33,080 Now, for model names it doesn't recognize, those four providers use the shorter 10 00:00:33,180 --> 00:00:33,840 system prompt. 11 00:00:34,340 --> 00:00:38,000 The changelog is explicit that this applies even when the deployment fronts Sonnet 12 00:00:38,060 --> 00:00:38,720 or Haiku. 13 00:00:39,020 --> 00:00:43,060 So if your team aliases a model behind a gateway name, your sessions may have 14 00:00:43,120 --> 00:00:45,400 quietly switched to the short prompt after the update. 15 00:00:46,028 --> 00:00:49,408 And that's the bit worth checking, because nothing on screen shouts about it. 16 00:00:49,928 --> 00:00:53,968 A bigger prompt is more tokens on every request, so I'd guess the shorter one is 17 00:00:54,028 --> 00:00:54,828 cheaper and quicker. 18 00:00:55,228 --> 00:00:57,768 But that's my read, not something the changelog claims. 19 00:00:58,388 --> 00:01:00,848 What the changelog does give you is the escape hatch. 20 00:01:01,547 --> 00:01:05,947 Set CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT to zero and you keep the full prompt. 21 00:01:06,387 --> 00:01:09,987 You can export it in your shell, or put it under the env key in your settings dot 22 00:01:10,067 --> 00:01:10,547 json. 23 00:01:11,027 --> 00:01:15,727 Something like env, then CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT with the string zero. 24 00:01:16,307 --> 00:01:19,687 If you want it for the whole team, put it in the project's dot claude settings file 25 00:01:19,727 --> 00:01:20,367 so it's checked in. 26 00:01:21,028 --> 00:01:21,788 Practical test. 27 00:01:22,588 --> 00:01:27,448 If your agent behaved well last week and feels a bit off this week on a custom 28 00:01:27,488 --> 00:01:33,408 deployment name, run one session each way with the variable set to zero and not set. 29 00:01:33,808 --> 00:01:35,188 Same task, same repo. 30 00:01:35,768 --> 00:01:38,008 If the full prompt fixes it, you've got your answer. 31 00:01:38,488 --> 00:01:42,748 If you don't notice a difference, leave the short one and enjoy the lighter load. 32 00:01:43,808 --> 00:01:45,188 Second change in the same area. 33 00:01:45,708 --> 00:01:49,968 Claude Code can automatically switch models after a message gets flagged. 34 00:01:50,588 --> 00:01:56,308 In 2.1.312 that behavior now also covers model names it doesn't recognize, 35 00:01:56,768 --> 00:01:58,288 same custom deployment names. 36 00:01:58,848 --> 00:02:03,188 To turn it off, set CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK to one. 37 00:02:04,162 --> 00:02:05,322 Why would you want that off? 38 00:02:06,182 --> 00:02:09,861 Say you've pointed Claude Code at a specific deployment on purpose, 39 00:02:10,262 --> 00:02:12,062 maybe for data handling reasons. 40 00:02:12,542 --> 00:02:15,662 A silent switch to a different model is the last thing you'd want. 41 00:02:16,242 --> 00:02:18,422 Flip the variable and the switch doesn't happen. 42 00:02:18,922 --> 00:02:23,802 The changelog doesn't say exactly what you'll see instead, so try it on a throwaway 43 00:02:23,862 --> 00:02:27,922 session first and see how a flagged turn behaves before you roll it out. 44 00:02:28,875 --> 00:02:29,855 Now the guardrails. 45 00:02:30,335 --> 00:02:34,655 Prompt and agent hooks are capped at 24 running at a time per session, 46 00:02:35,135 --> 00:02:36,735 and that count includes subagents. 47 00:02:37,215 --> 00:02:41,435 On top of that, at most 12 hooks that have waited longer than their event's timeout. 48 00:02:42,155 --> 00:02:45,895 And a hook timeout under one second now counts as one second. 49 00:02:46,515 --> 00:02:50,315 So if you've got a swarm of subagents each firing prompt hooks, 50 00:02:50,855 --> 00:02:51,295 audit them. 51 00:02:51,795 --> 00:02:55,655 Anything relying on a sub second timeout is now getting a full second. 52 00:02:56,482 --> 00:02:58,982 There's also a subagent compaction tweak. 53 00:02:59,562 --> 00:03:04,362 A subagent's autoCompactWindow now lowers the window set by 54 00:03:04,442 --> 00:03:09,002 CLAUDE_CODE_AUTO_COMPACT_WINDOW, where before it had no effect. 55 00:03:09,722 --> 00:03:11,062 It still never raises it. 56 00:03:11,562 --> 00:03:16,082 So you can make a chatty subagent compact earlier without touching the session 57 00:03:16,142 --> 00:03:16,442 setting. 58 00:03:17,202 --> 00:03:21,082 And a callback for anyone who followed our earlier episode on setting up the 59 00:03:21,142 --> 00:03:24,862 claude_code.safety_event OpenTelemetry metric. 60 00:03:25,522 --> 00:03:29,602 In 2.1.312, Claude Code has stopped sending that event. 61 00:03:30,122 --> 00:03:33,602 If you built a dashboard or an alert on it, expect it to go quiet, 62 00:03:33,822 --> 00:03:35,662 and that's the version, not your collector. 63 00:03:36,142 --> 00:03:39,062 Worth checking before someone reads silence as all clear. 64 00:03:39,858 --> 00:03:40,878 Few smaller wins. 65 00:03:41,398 --> 00:03:45,898 In fullscreen, you can click the "to send now" hint under a queued message and it 66 00:03:45,958 --> 00:03:47,978 sends right away, same as the key. 67 00:03:48,758 --> 00:03:54,378 There's a new oidc.user_id_claim on the Claude apps gateway, 68 00:03:54,758 --> 00:04:00,358 so you can identify people by a claim other than sub, like Microsoft Entra ID's oid. 69 00:04:01,263 --> 00:04:04,303 The one I'd flag is the CLAUDE.md fix. 70 00:04:04,943 --> 00:04:11,123 Claude could overwrite a very large CLAUDE.md, rules file, or at included file in a 71 00:04:11,183 --> 00:04:13,603 subdirectory after seeing only the start of it. 72 00:04:14,163 --> 00:04:18,523 Now the Write tool asks Claude to read the whole file or edit in place. 73 00:04:19,083 --> 00:04:21,503 If you've got a giant rules file, that's a real save. 74 00:04:22,203 --> 00:04:26,403 Also, usage limit messages now say "five hour limit" instead of "session limit", 75 00:04:26,883 --> 00:04:30,783 and in VS Code you get a reset clock time, like resets at 3:10 PM. 76 00:04:31,525 --> 00:04:36,605 So the takeaway: if you're on a custom deployment name, check which prompt you're 77 00:04:36,645 --> 00:04:41,865 getting and which fallback behavior you want, then set those two variables on 78 00:04:41,925 --> 00:04:42,345 purpose. 79 00:04:43,205 --> 00:04:49,285 Update to 2.1.312 and run claude dash dash version to confirm.