1 00:00:00,300 --> 00:00:06,320 Right, so a few days ago we talked about forcing every workflow agent onto one cheap 2 00:00:06,380 --> 00:00:06,680 model. 3 00:00:07,520 --> 00:00:08,640 This one is the follow up. 4 00:00:09,220 --> 00:00:13,420 This is The Claude Code Changelog, brought to you by Jellypod AI. 5 00:00:13,960 --> 00:00:14,480 I'm Lachlan. 6 00:00:15,120 --> 00:00:15,720 And I'm James. 7 00:00:16,400 --> 00:00:20,440 Today it's Claude Code 2.1.300, released October 14th. 8 00:00:21,040 --> 00:00:25,360 The headline is that the agent.spawn mod hook can now rewrite a workflow agent's 9 00:00:25,420 --> 00:00:26,540 prompt and its model. 10 00:00:27,320 --> 00:00:31,059 Before this, per the changelog, that hook could only refuse workflow agents. 11 00:00:31,280 --> 00:00:32,600 It could say no, and that was it. 12 00:00:33,258 --> 00:00:37,938 And that matters because of what we covered in 2.1.296. 13 00:00:38,738 --> 00:00:44,378 CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL puts every workflow agent on one model. 14 00:00:45,138 --> 00:00:47,458 Handy for cost, but it's a blunt instrument. 15 00:00:48,178 --> 00:00:54,498 Your simple discovery step runs fine on Haiku 5.5, then your heavy synthesis step, 16 00:00:54,998 --> 00:00:59,258 or the one that needs strict structured output, falls off a cliff because it's on 17 00:00:59,298 --> 00:01:00,438 the same cheap model. 18 00:01:01,302 --> 00:01:03,282 So now you get step aware routing. 19 00:01:03,782 --> 00:01:06,982 In a TypeScript mod, you register on agent.spawn. 20 00:01:07,502 --> 00:01:10,942 For workflow agents, the event carries the run ID and the step index, 21 00:01:11,142 --> 00:01:11,782 plus the prompt. 22 00:01:12,362 --> 00:01:14,182 You look at those, and you return the changes. 23 00:01:14,662 --> 00:01:19,422 Roughly: an object with model set to haiku, and prompt set to the original event 24 00:01:19,482 --> 00:01:23,582 prompt plus a new line saying, rule, return strict JSON only. 25 00:01:23,950 --> 00:01:25,290 Where does that mod live? 26 00:01:25,810 --> 00:01:30,370 Your project's mods folder, or you can ship it inside a custom plugin. 27 00:01:30,850 --> 00:01:31,990 Same idea either way. 28 00:01:32,710 --> 00:01:37,510 And on the model value, you can use a short alias like haiku or sonnet, 29 00:01:37,970 --> 00:01:39,370 or a full model ID. 30 00:01:39,850 --> 00:01:43,210 Both resolve now, and that ties into a fix in this release. 31 00:01:43,943 --> 00:01:44,143 Right. 32 00:01:44,603 --> 00:01:47,683 Agent hooks used to ignore their model setting entirely. 33 00:01:48,343 --> 00:01:54,803 In 2.1.300, aliases like haiku resolve properly, a subagent's own frontmatter model 34 00:01:54,903 --> 00:01:59,203 no longer overrides it, and an unavailable model falls back to the caller's model 35 00:01:59,263 --> 00:02:00,883 instead of letting the action through. 36 00:02:01,503 --> 00:02:05,243 Before that fix, a hook that set a model wasn't something you could lean on. 37 00:02:05,923 --> 00:02:06,903 Let's make it concrete. 38 00:02:07,503 --> 00:02:10,703 Say you've got a workflow that fans out across a repo. 39 00:02:11,283 --> 00:02:14,343 Early steps are scanning files and listing candidates. 40 00:02:14,903 --> 00:02:18,343 Your mod checks the step index, sends those to Haiku. 41 00:02:19,083 --> 00:02:24,063 The final step that merges findings and writes a plan, you send to Sonnet or Opus. 42 00:02:24,783 --> 00:02:29,603 Same workflow, two price tiers, no env var forcing everything down. 43 00:02:30,175 --> 00:02:32,135 The prompt side is just as useful, I think. 44 00:02:32,775 --> 00:02:35,815 Since you can rewrite the prompt before launch, you can inject your team's 45 00:02:35,855 --> 00:02:40,215 formatting rules, a repo specific schema, or strip out sensitive text from the 46 00:02:40,255 --> 00:02:40,615 input. 47 00:02:41,175 --> 00:02:44,635 It's context you'd otherwise have to repeat in every workflow definition. 48 00:02:45,135 --> 00:02:46,555 One mod, applied everywhere. 49 00:02:47,298 --> 00:02:48,358 Now, caveats. 50 00:02:48,978 --> 00:02:53,478 First, the hook's model beats the model in a subagent definition file. 51 00:02:54,158 --> 00:02:59,918 The docs say that when a mod sets a model in agent.spawn, Claude Code uses it in 52 00:02:59,998 --> 00:03:02,138 place of the per invocation choice. 53 00:03:03,078 --> 00:03:07,218 But your organization's availableModels allowlist still applies. 54 00:03:08,058 --> 00:03:11,978 If you assign a model that isn't permitted, it falls back to the caller's model, 55 00:03:12,678 --> 00:03:16,318 and it won't warn you loudly, so check which model actually ran. 56 00:03:18,150 --> 00:03:21,890 Second, validation is stricter in 2.1.300. 57 00:03:22,410 --> 00:03:27,850 The changelog says updatedInput from PreToolUse and PermissionRequest hooks is now 58 00:03:27,910 --> 00:03:29,950 validated like Claude's own input. 59 00:03:30,490 --> 00:03:34,430 Unknown keys get dropped, and input the tool rejects denies the call. 60 00:03:34,910 --> 00:03:36,930 So don't get creative with extra fields. 61 00:03:37,750 --> 00:03:42,230 And third, and this is my own read of how hooks behave, treat errors seriously. 62 00:03:42,670 --> 00:03:46,890 If your hook throws or denies the spawn, that workflow step doesn't run. 63 00:03:47,190 --> 00:03:50,470 So wrap your logic and return the original values as a default. 64 00:03:51,013 --> 00:03:52,573 Quick fire on the smaller stuff. 65 00:03:53,213 --> 00:03:59,693 Mods can now draw virtualText in the prompt box, so ghost text or shortcut hints 66 00:03:59,733 --> 00:04:02,573 that never end up in your draft or the submitted prompt. 67 00:04:03,593 --> 00:04:08,373 OpenTelemetry api_request events now split cache writes into 68 00:04:08,793 --> 00:04:15,753 cache_creation_5m_tokens and cache_creation_1h_tokens, which 69 00:04:15,793 --> 00:04:17,853 is gold if you're watching cache costs. 70 00:04:18,988 --> 00:04:24,368 Also fixed: ENABLE_TOOL_SEARCH=auto no longer switches itself off mid conversation 71 00:04:24,428 --> 00:04:26,988 when MCP servers reconnect after a resume. 72 00:04:27,868 --> 00:04:32,108 Exiting now force kills background commands still running 1.5 seconds after being 73 00:04:32,168 --> 00:04:32,768 asked to stop. 74 00:04:33,508 --> 00:04:37,588 And /autocompact stops warning about high token usage when you pick a smaller 75 00:04:37,668 --> 00:04:37,948 window. 76 00:04:38,630 --> 00:04:44,070 If you've already got workflow pipelines running, update to 2.1.300, 77 00:04:44,630 --> 00:04:48,550 write a tiny mod that logs the step index and prompt first, 78 00:04:48,970 --> 00:04:50,170 and then start routing. 79 00:04:50,870 --> 00:04:53,830 Log before you route, so you know what you're actually sending.