1 00:00:00,140 --> 00:00:03,740 If you have been firing off parallel subagents in Claude Code, 2 00:00:04,259 --> 00:00:06,920 you might have been accidentally burning money on duplicate tokens, 3 00:00:07,359 --> 00:00:13,559 and this episode brought to you by Jellypod AI is all about how version 2.1.229 4 00:00:13,960 --> 00:00:15,719 fixes that exact race condition. 5 00:00:16,129 --> 00:00:18,789 Mate, it was a proper trap, wasn't it? 6 00:00:19,689 --> 00:00:24,170 You run a workflow fan out, spin up five sibling agents at the exact same 7 00:00:24,229 --> 00:00:28,750 millisecond, and boom, every single one sends the full prompt prefix to the 8 00:00:28,809 --> 00:00:31,789 Anthropic API before the cache even registers it. 9 00:00:32,032 --> 00:00:35,992 Right, because prompt caching requires that initial response to write the prefix 10 00:00:36,052 --> 00:00:36,813 cache first. 11 00:00:37,333 --> 00:00:41,313 If all five subagents hit the endpoint simultaneously, none of them see the cache 12 00:00:41,333 --> 00:00:41,833 hit. 13 00:00:41,872 --> 00:00:43,852 They all pay full price for input tokens. 14 00:00:44,076 --> 00:00:44,996 Exactly right. 15 00:00:45,796 --> 00:00:51,516 So what the Claude Code team did in 2.1.229 is dead simple but clever. 16 00:00:52,297 --> 00:00:58,096 They improved workflow fan outs to stagger same prefix sibling agents so subsequent 17 00:00:58,216 --> 00:01:00,597 agents read the cached prompt prefix. 18 00:01:00,765 --> 00:01:04,964 So it puts a tiny delay on the launch of subagent two, three, 19 00:01:05,404 --> 00:01:10,085 four, and so on, just long enough for agent one to warm up that cache prefix? 20 00:01:10,125 --> 00:01:10,685 Spot on. 21 00:01:10,965 --> 00:01:16,045 A tiny millisecond stagger, and suddenly your input token cost on big parallel agent 22 00:01:16,125 --> 00:01:17,805 runs drops off a cliff. 23 00:01:18,045 --> 00:01:23,005 But, if you are running a local test harness and you do not want any delay at all, 24 00:01:23,025 --> 00:01:25,645 you can tune or turn it off with the environment variable 25 00:01:25,805 --> 00:01:29,325 CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS set to zero. 26 00:01:29,512 --> 00:01:34,412 Setting CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS to zero is great for local benchmarks 27 00:01:34,472 --> 00:01:35,793 where latency beats cost. 28 00:01:36,232 --> 00:01:40,732 But man, for production workflows, keeping that stagger on is going to save serious 29 00:01:40,793 --> 00:01:41,032 money. 30 00:01:41,426 --> 00:01:42,946 My wallet certainly thanks them! 31 00:01:43,486 --> 00:01:46,366 And speaking of stopping people from shooting themselves in the foot, 32 00:01:46,846 --> 00:01:49,446 they also locked down the commit push pr tool. 33 00:01:49,759 --> 00:01:50,899 Oh, what did they do there? 34 00:01:51,417 --> 00:01:55,337 Well, if Claude tries to auto approve a pull request workflow, 35 00:01:55,357 --> 00:02:00,297 but it detects destructive git flags like force, amend, or no verify, 36 00:02:00,404 --> 00:02:02,697 it flat out refuses auto approval now. 37 00:02:02,857 --> 00:02:05,017 It forces a human to look at it. 38 00:02:05,239 --> 00:02:06,539 That is huge! 39 00:02:06,979 --> 00:02:11,339 No agent accidentally force pushing to main or bypassing pre commit hooks because it 40 00:02:11,379 --> 00:02:12,799 got a little overambitious. 41 00:02:13,052 --> 00:02:14,292 Exactly. 42 00:02:14,312 --> 00:02:16,872 I know a thing or two about pushing bad updates in a hurry, 43 00:02:17,193 --> 00:02:21,253 so having guardrails around force and amend is a total lifesaver. 44 00:02:21,592 --> 00:02:25,053 What else landed in 2.1.229 for quality of life? 45 00:02:25,083 --> 00:02:26,523 A few really slick ones! 46 00:02:26,603 --> 00:02:31,483 You get claude remote control continue to instantly resume a remote session, 47 00:02:31,603 --> 00:02:37,323 resizable btw side panels, and session grouping in the VS Code sidebar. 48 00:02:38,052 --> 00:02:41,892 Resizable btw panels alone will save my sanity. 49 00:02:41,952 --> 00:02:44,232 Good stuff in 2.1.229 all around.