1 00:00:00,000 --> 00:00:05,120 Picture a test runner subagent forty turns deep, still hauling every stack trace it 2 00:00:05,160 --> 00:00:07,680 has ever seen into each new request. 3 00:00:07,840 --> 00:00:13,760 That's the mess Claude Code 2.1.296 goes after, released October 9th. 4 00:00:13,952 --> 00:00:17,200 This episode's brought to you by Jellypod AI. 5 00:00:17,560 --> 00:00:19,680 I'm Lachlan, and with me is James. 6 00:00:19,708 --> 00:00:25,708 Last time we covered 2.1.292 and the effort parameter for subagents, 7 00:00:25,735 --> 00:00:29,708 which lets you dial down reasoning on mechanical jobs like grepping logs. 8 00:00:29,836 --> 00:00:32,348 That handled how hard a subagent thinks. 9 00:00:32,828 --> 00:00:36,508 2.1.296 handles how much it carries. 10 00:00:36,648 --> 00:00:42,108 The release adds `autoCompactWindow` to subagent frontmatter and to the agents flag 11 00:00:42,168 --> 00:00:42,908 definitions. 12 00:00:43,068 --> 00:00:48,028 The changelog describes it as letting a subagent auto compact earlier than the main 13 00:00:48,083 --> 00:00:49,388 conversation's window. 14 00:00:49,417 --> 00:00:50,457 Here's why that matters. 15 00:00:50,571 --> 00:00:55,657 Worker subagents tend to do noisy, repetitive work: running test suites, 16 00:00:55,737 --> 00:00:57,977 linters, parsing build logs. 17 00:00:58,077 --> 00:01:00,857 All that terminal chatter piles up in context. 18 00:01:00,957 --> 00:01:05,577 And every new turn re-sends that whole pile, which shows up as cache reads in your 19 00:01:05,657 --> 00:01:09,977 usage, so cost per turn grows with how full the window is. 20 00:01:10,057 --> 00:01:11,897 There's also the accuracy side. 21 00:01:12,057 --> 00:01:16,857 Long agent sessions on a big window can rot before compaction kicks in, 22 00:01:16,897 --> 00:01:21,097 and one GitHub feature request put it as sessions rotting before they reach the 23 00:01:21,137 --> 00:01:23,177 roughly 95 percent default. 24 00:01:23,310 --> 00:01:27,177 Treat that as a user's description, not an Anthropic benchmark. 25 00:01:27,277 --> 00:01:31,257 But the logic is sound: a lean worker beats a bloated one. 26 00:01:31,500 --> 00:01:32,780 So the setup is decoupling. 27 00:01:32,892 --> 00:01:37,340 Your orchestrator can sit on a huge window with the full architectural history, 28 00:01:37,380 --> 00:01:39,340 while workers compact way sooner. 29 00:01:39,484 --> 00:01:40,780 Here's the config. 30 00:01:40,967 --> 00:01:45,900 In your agent file under the dot claude agents folder, say test-runner.md, 31 00:01:45,920 --> 00:01:50,220 the frontmatter has the name test-runner, a description like runs the test suite and 32 00:01:50,280 --> 00:01:55,100 inspects stack traces, the model set to Haiku 5.5, effort low, 33 00:01:55,200 --> 00:01:58,060 and then `autoCompactWindow` set to 50000. 34 00:01:58,188 --> 00:01:59,700 That's the new line. 35 00:01:59,790 --> 00:02:04,620 Effort low is the 2.1.292 feature, so the two stack nicely. 36 00:02:04,625 --> 00:02:09,025 And if you're spinning agents up on the fly, same key works in the JSON you pass to 37 00:02:09,045 --> 00:02:10,145 the agents flag. 38 00:02:10,257 --> 00:02:16,545 Something like claude, the agents flag, then a JSON object with a key called linter, 39 00:02:16,589 --> 00:02:22,625 containing the Haiku 5.5 model ID and `autoCompactWindow` of 60000. 40 00:02:22,812 --> 00:02:27,665 Worth knowing: the value is an absolute token count, an integer. 41 00:02:27,785 --> 00:02:28,945 Not a percentage. 42 00:02:29,045 --> 00:02:32,945 Writing 0.5 or fifty percent is not what it wants. 43 00:02:32,958 --> 00:02:36,878 Now the caveats, and the first one's a judgment call, so take it as guidance. 44 00:02:37,022 --> 00:02:42,638 Compaction costs an LLM call, and the harness reserves up to 20,000 tokens for the 45 00:02:42,688 --> 00:02:43,598 summary output. 46 00:02:43,710 --> 00:02:47,998 That figure comes from a third party analysis of Claude Code's internals, 47 00:02:48,051 --> 00:02:50,718 so treat it as background rather than official docs. 48 00:02:50,878 --> 00:02:55,998 The upshot is that if you set the window too low, say under roughly 40,000, 49 00:02:56,038 --> 00:02:59,198 a subagent could spend more turns summarizing than working. 50 00:02:59,305 --> 00:03:03,438 Starting between 40,000 and 80,000 is a sensible range. 51 00:03:03,458 --> 00:03:08,578 Second caveat: compaction summarizes older turns, so exact details can get 52 00:03:08,650 --> 00:03:09,378 condensed. 53 00:03:09,618 --> 00:03:15,218 If your subagent needs a specific git hash, a line number, or a raw diff from way 54 00:03:15,282 --> 00:03:19,218 back, tell it in its prompt to write those into a scratchpad file. 55 00:03:19,351 --> 00:03:21,458 Files survive compaction. 56 00:03:21,578 --> 00:03:23,698 Summaries are lossy by nature. 57 00:03:23,708 --> 00:03:25,508 One more related addition. 58 00:03:25,688 --> 00:03:30,634 `CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL` runs every workflow agent on one model, 59 00:03:30,691 --> 00:03:33,415 while other subagents keep theirs. 60 00:03:33,468 --> 00:03:38,401 So in CI you can export it as Haiku 5.5 and keep your scripted workflow agents 61 00:03:38,455 --> 00:03:41,788 cheap, without touching the specialized reasoning agents. 62 00:03:41,868 --> 00:03:46,988 And a quiet one that pairs well: the debug flag now names unrecognized frontmatter 63 00:03:47,068 --> 00:03:50,908 fields in custom agent files, with a hint for likely typos. 64 00:03:50,961 --> 00:03:54,668 So if you misspell `autoCompactWindow`, you'll find out instead of wondering why 65 00:03:54,668 --> 00:03:54,668 nothing compacts. 66 00:03:54,708 --> 00:03:55,828 A few smaller wins. 67 00:03:55,968 --> 00:04:00,948 The Read tool gets an `allow_large` option, so Claude can read a text file past the 68 00:04:01,001 --> 00:04:05,428 usual size limits in one call, when it needs the whole file and the context has 69 00:04:05,492 --> 00:04:05,988 room. 70 00:04:06,168 --> 00:04:11,508 Sonnet 5.5 cache reads are now priced at ten cents per million tokens, 71 00:04:11,540 --> 00:04:15,188 down from twenty, across the cost command, the status line, 72 00:04:15,228 --> 00:04:18,068 the max budget flag, and the SDK's figures. 73 00:04:18,083 --> 00:04:22,803 The default limit on MCP tool descriptions and server instructions sent up front 74 00:04:22,883 --> 00:04:27,523 doubled, from 2,048 to 4,096 characters. 75 00:04:27,923 --> 00:04:34,643 `CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS` lets you set a longer maximum delay when 76 00:04:34,710 --> 00:04:37,443 retrying overloaded 529 errors. 77 00:04:37,623 --> 00:04:42,323 And debug output for command hooks now logs the command, plugin, 78 00:04:42,383 --> 00:04:47,363 outcome, and duration when they finish, so a slow hook finally has a name attached. 79 00:04:47,375 --> 00:04:53,135 Last one: in auto mode, a tool call skipped because the safety check had no usable 80 00:04:53,192 --> 00:04:58,175 answer now shows as a dim Not run row, not a red error. 81 00:04:58,303 --> 00:04:59,855 Less panic over nothing. 82 00:05:00,228 --> 00:05:02,095 So here's your homework. 83 00:05:02,223 --> 00:05:07,855 Pick your noisiest subagent, add `autoCompactWindow` at 50000, 84 00:05:07,975 --> 00:05:11,695 run it, and watch the context command to see it compact.