1 00:00:00,460 --> 00:00:03,900 Righto, grab a coffee, this one's brought to you by Jellypod AI. 2 00:00:04,520 --> 00:00:09,680 I'm Lachlan, that's James, and today it's Claude Code 2.1.302, 3 00:00:10,160 --> 00:00:12,300 released October 16, 2026. 4 00:00:13,040 --> 00:00:17,100 One new environment variable in there closes a loop we opened last time. 5 00:00:17,745 --> 00:00:17,965 Right. 6 00:00:18,345 --> 00:00:22,925 Last episode we covered why proxies and Bedrock or Vertex setups lose the free 7 00:00:23,025 --> 00:00:26,905 server side safety check in auto mode and fall back to a classifier running on your 8 00:00:26,945 --> 00:00:27,525 own client. 9 00:00:28,125 --> 00:00:29,105 We left that as a trap. 10 00:00:29,765 --> 00:00:31,925 2.1.302 gives you a lever. 11 00:00:32,405 --> 00:00:36,865 It's called CLAUDE_CODE_AUTO_MODE_MODEL, and it chooses the model the client side 12 00:00:36,945 --> 00:00:38,445 auto mode classifier uses. 13 00:00:39,332 --> 00:00:41,592 And the release notes are careful about scope. 14 00:00:42,032 --> 00:00:44,952 It does not change the model for server side auto mode. 15 00:00:45,592 --> 00:00:48,892 So if you're on the Anthropic API and the server side check is running, 16 00:00:49,352 --> 00:00:51,092 this variable does nothing for you. 17 00:00:51,838 --> 00:00:52,558 Why does it matter? 18 00:00:53,158 --> 00:00:56,258 In auto mode, Claude Code checks tool calls before running them. 19 00:00:56,778 --> 00:01:00,878 When that check happens on your client, it's a model call, and that call takes time 20 00:01:00,918 --> 00:01:01,838 and spends tokens. 21 00:01:02,398 --> 00:01:04,318 Until now you couldn't pick which model made it. 22 00:01:05,157 --> 00:01:08,538 My read, and this is interpretation, not something the changelog states, 23 00:01:08,998 --> 00:01:12,938 is that a small fast model is the sensible fit for a yes or no safety check. 24 00:01:13,498 --> 00:01:16,158 That could cut latency on turns with lots of tool calls. 25 00:01:16,727 --> 00:01:18,087 Setting it is dead simple. 26 00:01:18,627 --> 00:01:25,047 In your shell profile, or CI config, you export CLAUDE_CODE_AUTO_MODE_MODEL equals 27 00:01:25,387 --> 00:01:26,587 the model ID you want. 28 00:01:27,267 --> 00:01:30,187 Pick a Haiku class model if you want cheap and quick. 29 00:01:31,207 --> 00:01:35,567 On Bedrock or Vertex, use the identifier your provider uses for that model, 30 00:01:35,927 --> 00:01:38,667 whether that's an inference profile or a model ID. 31 00:01:39,247 --> 00:01:42,147 Don't paste the Anthropic API name and hope for the best. 32 00:01:42,942 --> 00:01:43,802 One caveat there. 33 00:01:44,142 --> 00:01:47,422 The notes don't say what happens with a bad or retired model ID. 34 00:01:47,642 --> 00:01:48,442 So test it. 35 00:01:48,942 --> 00:01:51,962 Set the variable, run something that triggers a checked tool call, 36 00:01:52,102 --> 00:01:54,702 and confirm it behaves before you roll it out to a team. 37 00:01:55,392 --> 00:01:57,352 Now, the one that could actually bite you. 38 00:01:58,052 --> 00:02:04,512 Also in this release, claude in print mode, the p flag, and Agent SDK sessions with 39 00:02:04,552 --> 00:02:07,672 no permission mode set now start in default mode. 40 00:02:08,272 --> 00:02:12,932 That applies on third party providers, with telemetry off, or on a fresh install. 41 00:02:13,672 --> 00:02:16,072 Default mode means it asks for permission. 42 00:02:16,777 --> 00:02:19,497 Which in a headless pipeline means nobody's there to answer. 43 00:02:19,857 --> 00:02:24,417 So if your CI job quietly relied on auto, it can stall or fail after you upgrade. 44 00:02:24,977 --> 00:02:26,737 The fix is to say what you want. 45 00:02:27,017 --> 00:02:30,697 Pass the permission mode flag with the value auto, and auto stays auto. 46 00:02:31,197 --> 00:02:35,497 Something like claude, the p flag, the permission mode flag set to auto, 47 00:02:35,797 --> 00:02:38,037 then your prompt, say, run the build and the tests. 48 00:02:38,888 --> 00:02:43,128 So the combo for Bedrock CI is: set the permission mode explicitly, 49 00:02:43,628 --> 00:02:48,568 and set CLAUDE_CODE_AUTO_MODE_MODEL so the fallback classifier isn't running on 50 00:02:48,608 --> 00:02:49,848 whatever it picked by default. 51 00:02:50,475 --> 00:02:54,035 There's one more auto mode fix, and it's for the server side path. 52 00:02:54,615 --> 00:02:57,635 Before, if a conversation got too long for the safety check, 53 00:02:57,955 --> 00:02:59,915 every checked tool call was denied. 54 00:03:00,535 --> 00:03:03,115 Now Claude Code compacts the conversation instead. 55 00:03:03,375 --> 00:03:05,315 So long sessions shouldn't hit that wall. 56 00:03:06,232 --> 00:03:07,852 A few smaller bits worth knowing. 57 00:03:08,432 --> 00:03:14,192 Fork subagents, that's subagent_type fork, are now on by default in print mode, 58 00:03:14,492 --> 00:03:19,892 the Agent SDK, the desktop app and VS Code, same as the terminal already had. 59 00:03:20,452 --> 00:03:25,132 If you don't want them, set CLAUDE_CODE_FORK_SUBAGENT to 0. 60 00:03:25,652 --> 00:03:28,332 And claude mcp serve never offers them. 61 00:03:29,270 --> 00:03:34,190 Also in the desktop app and VS Code, the main conversation's subagents now always 62 00:03:34,250 --> 00:03:36,030 run in the background, like the terminal. 63 00:03:36,650 --> 00:03:39,790 A subagent's own Agent calls run in the foreground by default. 64 00:03:40,803 --> 00:03:46,943 There's a new Idle compaction row in /config, so you can turn idle compaction off or 65 00:03:47,003 --> 00:03:52,023 back on without digging through files, and the notices after an idle compaction now 66 00:03:52,083 --> 00:03:53,283 tell you how to switch it off. 67 00:03:53,923 --> 00:03:57,583 Prompt and agent hooks are capped at twelve running at once per event, 68 00:03:58,023 --> 00:03:58,803 the rest queue. 69 00:03:59,343 --> 00:04:02,463 If you've got a pile of hooks on one event, expect that queueing. 70 00:04:03,325 --> 00:04:08,605 And the Read tool now refuses image sources over 128 MB instead of loading them 71 00:04:08,685 --> 00:04:09,665 whole into memory. 72 00:04:10,185 --> 00:04:14,065 Plus plugin startup is faster with lots of plugins, especially on Windows. 73 00:04:14,685 --> 00:04:20,345 So the takeaway: after you upgrade to 2.1.302, check your headless jobs, 74 00:04:20,765 --> 00:04:25,425 set the permission mode explicitly, and if you're on a proxy or a cloud provider, 75 00:04:25,745 --> 00:04:27,125 pin that classifier model. 76 00:04:27,705 --> 00:04:28,325 Catch you next time.