1 00:00:00,119 --> 00:00:03,439 Quick shoutout to our sponsor Jellypod AI before we dive in! 2 00:00:03,539 --> 00:00:08,659 So, if you are running an enterprise proxy like LiteLLM or Envoy in front of your 3 00:00:08,659 --> 00:00:12,680 developer tools, you know the absolute nightmare of inspecting traffic. 4 00:00:13,279 --> 00:00:18,260 Every single time a developer or a subagent hits `/v1/messages`, 5 00:00:18,680 --> 00:00:22,840 the gateway used to have to unpack the entire JSON body just to figure out what kind 6 00:00:22,859 --> 00:00:24,119 of request it even was. 7 00:00:24,706 --> 00:00:25,546 Oh mate, yeah. 8 00:00:25,866 --> 00:00:31,326 Digging through massive payload bodies on every single HTTP POST just to see if it 9 00:00:31,366 --> 00:00:34,346 is a quick interactive prompt or some heavy background task? 10 00:00:34,967 --> 00:00:36,167 It destroys latency. 11 00:00:36,826 --> 00:00:40,767 Your ingress routers get bogged down, token attribution gets super messy, 12 00:00:41,347 --> 00:00:44,767 and suddenly your primary dev sessions are getting rate limited because some 13 00:00:44,806 --> 00:00:46,746 background task is eating all the bandwidth. 14 00:00:46,833 --> 00:00:47,713 Exactly. 15 00:00:47,773 --> 00:00:52,513 But in the latest Claude Code release, version 2 point 1 point 273, 16 00:00:52,561 --> 00:00:55,713 Anthropic fixed this metadata blindspot entirely. 17 00:00:55,809 --> 00:01:01,370 They added a new feature where if you export `CLAUDE_CODE_GATEWAY_HINT_HEADERS=1`, 18 00:01:01,370 --> 00:01:06,833 the client automatically passes five clear HTTP hint headers on every request. 19 00:01:06,992 --> 00:01:08,972 Five headers, right. 20 00:01:09,013 --> 00:01:09,672 What are we looking at? 21 00:01:09,750 --> 00:01:14,590 So you get `x claude code request class`, `x claude code agent type`, 22 00:01:14,770 --> 00:01:19,750 `x claude code prev tool durations`, `x claude code compaction`, 23 00:01:19,910 --> 00:01:23,030 and `x claude code context compacted`. 24 00:01:23,126 --> 00:01:28,310 That means your proxy can instantly read `x claude code agent type` set to subagent, 25 00:01:28,417 --> 00:01:33,190 or see a compaction request, without ever opening or reading private prompt text! 26 00:01:33,327 --> 00:01:34,967 That is huge for DevOps! 27 00:01:35,467 --> 00:01:37,987 Think about NGINX or Envoy routing rules now. 28 00:01:38,487 --> 00:01:43,347 You can take any request where `x claude code compaction` is set to 1 and send it 29 00:01:43,387 --> 00:01:48,027 off to a low priority background queue, or route background subagent traffic over to 30 00:01:48,088 --> 00:01:50,627 cheaper Bedrock or Vertex inference pools. 31 00:01:51,208 --> 00:01:55,967 Main terminal sessions stay lightning fast on high priority tier while long tasks 32 00:01:56,027 --> 00:01:56,867 run in the background. 33 00:01:57,140 --> 00:01:57,459 Right! 34 00:01:57,919 --> 00:02:01,739 Just remember it is totally opt in for now, so you need to set 35 00:02:02,039 --> 00:02:08,199 `CLAUDE_CODE_GATEWAY_HINT_HEADERS=1` in your developer shell scripts or runner 36 00:02:08,239 --> 00:02:10,139 containers to turn those headers on. 37 00:02:10,407 --> 00:02:11,007 Easy enough. 38 00:02:11,787 --> 00:02:15,608 And look, there are a few other crackin quality of life updates in 2 point 1 point 39 00:02:15,668 --> 00:02:16,708 273 as well, yeah? 40 00:02:17,134 --> 00:02:17,694 Oh yeah! 41 00:02:18,154 --> 00:02:22,574 First, if an MCP server drops connection mid session and auto reconnect gives up, 42 00:02:22,974 --> 00:02:27,475 Claude Code now pops a clean notification telling you to jump straight to `/mcp` to 43 00:02:27,514 --> 00:02:27,914 fix it. 44 00:02:28,675 --> 00:02:32,134 Second, if you launch a session with `claude --remote control`, 45 00:02:32,614 --> 00:02:36,614 you can now fork that session directly from the Claude app into a background task on 46 00:02:36,635 --> 00:02:37,395 your local machine. 47 00:02:37,817 --> 00:02:40,877 That forking trick is going to save me so much hassle. 48 00:02:41,537 --> 00:02:46,218 Plus, auto mode on Bedrock and Vertex switched back to using the local classifier by 49 00:02:46,258 --> 00:02:52,117 default, unless you set `CLAUDE_CODE_AUTO_MODE_SERVER=1` to force platform server 50 00:02:52,157 --> 00:02:52,757 side checking. 51 00:02:53,737 --> 00:02:55,057 Good little release all around!