1 00:00:00,000 --> 00:00:04,000 Ever spun up five subagents to grep some logs and watched your token meter climb 2 00:00:04,032 --> 00:00:06,080 like they were all solving proofs? 3 00:00:06,112 --> 00:00:10,240 That's the itch version 2.1.292 scratches. 4 00:00:10,272 --> 00:00:14,080 This episode of The Claude Code Changelog is brought to you by Jellypod AI. 5 00:00:14,360 --> 00:00:15,360 I'm James. 6 00:00:15,375 --> 00:00:16,175 And I'm Lachlan. 7 00:00:16,287 --> 00:00:22,735 Last episode we covered 2.1.290 and 2.1.291, the WebFetch 8 00:00:22,800 --> 00:00:26,015 pagination and the Bash wildcard permission prompts. 9 00:00:26,655 --> 00:00:29,135 2.1.291 was mostly cleanup. 10 00:00:29,615 --> 00:00:34,575 2.1.292 is the bigger one for anyone running multi agent workflows, 11 00:00:34,605 --> 00:00:37,935 because the Agent tool now accepts an effort parameter. 12 00:00:37,958 --> 00:00:38,358 Right. 13 00:00:38,472 --> 00:00:43,638 Before this, when the main session delegated work, it could set the subagent type 14 00:00:43,758 --> 00:00:46,918 and the model, so opus, sonnet, or haiku. 15 00:00:47,058 --> 00:00:50,198 But reasoning effort wasn't something you could set per call. 16 00:00:50,358 --> 00:00:52,998 Subagents inherited the session's effort level. 17 00:00:53,089 --> 00:00:57,558 People tried the effort field in subagent frontmatter, and the GitHub issue tracking 18 00:00:57,590 --> 00:01:03,638 this request, number 43083, has several users reporting that on Task tool spawns it 19 00:01:03,698 --> 00:01:04,918 did basically nothing. 20 00:01:05,078 --> 00:01:09,878 Now the Agent tool takes effort directly: low, medium, high, 21 00:01:10,038 --> 00:01:11,478 xhigh, or max. 22 00:01:11,500 --> 00:01:14,060 And why that matters comes down to cost. 23 00:01:14,247 --> 00:01:19,900 In that same issue, one user measured the same prompt on Sonnet going from 616 24 00:01:19,957 --> 00:01:24,940 output tokens at low effort to 5,556 at max. 25 00:01:25,036 --> 00:01:26,940 That's roughly nine times. 26 00:01:27,153 --> 00:01:33,980 On Opus it was 471 to 2,694, about five and a half times. 27 00:01:34,156 --> 00:01:38,780 Same model, same correct answer, the difference is reasoning depth. 28 00:01:38,967 --> 00:01:44,220 So if your session runs at max, every mechanical subagent, the ones running tests or 29 00:01:44,260 --> 00:01:46,700 grepping logs, pays that tax. 30 00:01:46,780 --> 00:01:51,180 Run the session at low, and the subagent debugging a race condition gets starved. 31 00:01:51,208 --> 00:01:55,048 So now the orchestrator decides per delegation. 32 00:01:55,128 --> 00:01:57,848 The simplest way to use it is just to say so in your prompt. 33 00:01:57,912 --> 00:02:03,448 Something like: spawn a low effort subagent to run the test suite and collect the 34 00:02:03,498 --> 00:02:09,128 failing test names, then spawn a high effort subagent to analyze the root cause of 35 00:02:09,168 --> 00:02:09,928 the deadlock. 36 00:02:10,054 --> 00:02:13,768 Claude passes the effort value into the Agent call for each one. 37 00:02:13,792 --> 00:02:17,872 If you've got skills or a CLAUDE.md that drive an orchestrator, 38 00:02:18,000 --> 00:02:20,272 bake the policy in. 39 00:02:20,365 --> 00:02:25,152 Write it plainly: when delegating code exploration or lint checks, 40 00:02:25,248 --> 00:02:27,792 call Agent with effort low. 41 00:02:27,984 --> 00:02:34,352 When delegating security reviews or schema design, call Agent with effort high. 42 00:02:34,612 --> 00:02:39,632 One of the use cases in that issue was a review fan out, where eight narrow 43 00:02:39,688 --> 00:02:44,832 reviewers run at high and one or two adversarial verifiers get xhigh, 44 00:02:44,872 --> 00:02:47,552 because that's where extra depth actually pays off. 45 00:02:47,583 --> 00:02:53,103 And to check it's working, the brief for this release says you can run slash tasks, 46 00:02:53,143 --> 00:02:58,143 and each active subagent row shows its name, model, and effort level. 47 00:02:58,303 --> 00:03:02,303 That's worth doing once, because people in that thread had trouble verifying what 48 00:03:02,349 --> 00:03:04,303 effort a subagent actually ran at. 49 00:03:04,333 --> 00:03:05,213 Now the caveats. 50 00:03:05,333 --> 00:03:07,133 First, precedence. 51 00:03:07,320 --> 00:03:12,573 If you've exported CLAUDE_CODE_EFFORT_LEVEL in your shell, that acts as an absolute 52 00:03:12,635 --> 00:03:16,013 override, so it beats whatever the Agent call asks for. 53 00:03:16,253 --> 00:03:20,413 If your effort settings seem to be ignored, check your environment first. 54 00:03:20,573 --> 00:03:23,133 Second, thinking inheritance. 55 00:03:23,346 --> 00:03:27,853 If extended thinking is off in the parent session, subagents inherit that. 56 00:03:28,026 --> 00:03:30,253 Third, model support. 57 00:03:30,493 --> 00:03:34,573 Haiku doesn't support the extended thinking levels the bigger models do, 58 00:03:34,626 --> 00:03:37,933 so asking for max on a Haiku subagent won't buy you anything. 59 00:03:37,958 --> 00:03:39,398 There's also a related fix. 60 00:03:39,558 --> 00:03:44,118 Subagents with permissionMode auto used to try entering auto mode even when it was 61 00:03:44,165 --> 00:03:48,598 unavailable, whether that was disabled by settings, tripped by the circuit breaker, 62 00:03:48,678 --> 00:03:50,198 or the model didn't support it. 63 00:03:50,358 --> 00:03:53,718 In 2.1.292 they no longer do that. 64 00:03:53,750 --> 00:03:55,830 A few smaller things worth grabbing. 65 00:03:55,887 --> 00:03:59,030 Claude plugin install now takes a marketplace flag. 66 00:03:59,126 --> 00:04:03,030 Pass it a source, and it adds the marketplace under the usual policy checks if 67 00:04:03,076 --> 00:04:05,590 needed, then installs the plugin in one step. 68 00:04:05,723 --> 00:04:08,470 Handy for Dockerfiles and CI scripts. 69 00:04:08,500 --> 00:04:11,860 Also in CI land, there's a new environment variable, 70 00:04:11,913 --> 00:04:15,300 CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS. 71 00:04:15,487 --> 00:04:20,580 It sets a longer base delay for the backoff when a request gets a 529 overloaded 72 00:04:20,642 --> 00:04:24,660 response, so your runs don't burn through retries during a busy stretch. 73 00:04:24,667 --> 00:04:25,147 Two more. 74 00:04:25,414 --> 00:04:30,347 At mentioned text files over 256 kilobytes used to be left out silently. 75 00:04:30,547 --> 00:04:34,747 Now Claude is told the file's size and to read it in portions. 76 00:04:34,927 --> 00:04:40,027 And local stdio MCP servers now negotiate the newer protocol version, 77 00:04:40,080 --> 00:04:46,027 dated July 28, 2026, by default, including on Bedrock, Vertex and Foundry. 78 00:04:46,160 --> 00:04:51,547 If that causes trouble, set MCP_PROTOCOL_NEGOTIATION to legacy to opt out. 79 00:04:51,647 --> 00:04:55,147 Servers that ignore the newer check are remembered for seven days, 80 00:04:55,254 --> 00:04:57,627 so you don't wait on the slow connect every time. 81 00:04:57,667 --> 00:05:02,947 And one security fix to note: PreToolUse hook approvals and auto mode could bypass 82 00:05:02,967 --> 00:05:06,547 the permission prompt for file reads from network UNC paths. 83 00:05:06,643 --> 00:05:10,227 That's closed in this version, so update if you're on Windows shares. 84 00:05:10,250 --> 00:05:17,050 So the takeaway: update to 2.1.292, put an effort policy in your orchestrator 85 00:05:17,096 --> 00:05:20,090 prompt, and let the cheap work stay cheap.