1 00:00:00,000 --> 00:00:02,640 Right, so this one's brought to you by Jellypod AI. 2 00:00:02,920 --> 00:00:08,980 A couple of episodes back we skated past CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS, 3 00:00:09,000 --> 00:00:13,413 and then the max delay setting got a quick mention too, just tossed out as a closing 4 00:00:13,520 --> 00:00:14,080 bullet. 5 00:00:14,147 --> 00:00:16,640 Turns out those were pieces of a set. 6 00:00:16,800 --> 00:00:23,280 Versions 2.1.295 and 2.1.296, released October 8 and 9, 7 00:00:23,520 --> 00:00:25,573 finish it. 8 00:00:25,640 --> 00:00:31,973 2.1.295 added CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS, 9 00:00:32,267 --> 00:00:38,544 and 2.1.296 added CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS. 10 00:00:38,576 --> 00:00:44,349 Between them you can finally shape how Claude Code behaves when the API throws a 529 11 00:00:44,400 --> 00:00:46,480 overloaded error during an unattended job. 12 00:00:46,500 --> 00:00:51,540 And the reason you'd care is the default window is shorter than people think. 13 00:00:51,780 --> 00:00:56,660 At your desk, a 529 is no big deal, you wait and hit enter again. 14 00:00:56,873 --> 00:01:02,660 In headless mode, a CI runner, or a loop, it kills the process with exit code 1 15 00:01:02,720 --> 00:01:03,620 halfway through. 16 00:01:03,800 --> 00:01:07,380 One write up tested this against a mock 529 server. 17 00:01:07,480 --> 00:01:14,260 Default backoff starts around 0.55 seconds, then 1.26, 2.05, 18 00:01:14,340 --> 00:01:21,300 4.04, 9.5, 17.7, and levels off around 34 to 38 seconds. 19 00:01:21,440 --> 00:01:25,220 Ten retries burn through in 176.2 seconds. 20 00:01:25,300 --> 00:01:26,500 Under three minutes. 21 00:01:26,740 --> 00:01:30,180 So any capacity crunch longer than that fails your pipeline. 22 00:01:30,208 --> 00:01:34,768 Now, the obvious fix was CLAUDE_CODE_RETRY_WATCHDOG set to 1. 23 00:01:34,896 --> 00:01:40,688 That retries 429 and 529 capacity errors indefinitely, backing off up to five 24 00:01:40,738 --> 00:01:42,048 minutes between attempts. 25 00:01:42,158 --> 00:01:46,048 Problem is, indefinitely means indefinitely. 26 00:01:46,178 --> 00:01:49,488 Picture a GitHub Actions job with a 30 minute limit. 27 00:01:49,608 --> 00:01:53,808 The watchdog's still patiently waiting when the platform sends a SIGKILL. 28 00:01:53,995 --> 00:01:58,048 No error log, no debugging details, nothing salvageable. 29 00:01:58,192 --> 00:02:00,208 That's what the new wait cap fixes. 30 00:02:00,395 --> 00:02:05,008 It bounds how long the watchdog waits per request, and then the job fails cleanly 31 00:02:05,104 --> 00:02:09,488 with an explanatory 529 message, before the runner pulls the plug. 32 00:02:09,500 --> 00:02:10,900 Here's the CI pattern. 33 00:02:11,020 --> 00:02:17,340 In your job definition, export CLAUDE_CODE_RETRY_WATCHDOG equals 1, 34 00:02:17,388 --> 00:02:23,900 then export CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS equals 35 00:02:24,300 --> 00:02:27,340 1,800,000, which is 30 minutes. 36 00:02:27,452 --> 00:02:33,820 Then run claude with the p flag, your prompt, say run tests and fix failures, 37 00:02:33,920 --> 00:02:38,860 and add the fallback model flag with Sonnet 5.5 and Haiku 5.5. 38 00:02:38,972 --> 00:02:43,340 Each model in that chain gets its own retry budget, and since capacity is tracked 39 00:02:43,380 --> 00:02:47,900 per model, falling back from an overloaded model often succeeds straight away. 40 00:02:47,917 --> 00:02:51,517 If you'd rather skip the watchdog, there's a tuned backoff pattern. 41 00:02:51,697 --> 00:02:55,117 Set CLAUDE_CODE_MAX_RETRIES to 5. 42 00:02:55,207 --> 00:02:59,677 Default is 10, and it's capped at 15 unless the watchdog is on. 43 00:02:59,821 --> 00:03:06,077 Then set the base delay to 4,000 milliseconds and the new max delay to 60,000. 44 00:03:06,205 --> 00:03:11,597 Base delay has a range of 500 to 32,000, with a 500 default, 45 00:03:11,637 --> 00:03:14,637 and it arrived in 2.1.292. 46 00:03:14,777 --> 00:03:19,357 The point is a small retry budget now stretches across several minutes, 47 00:03:19,397 --> 00:03:21,517 instead of getting blown through in seconds. 48 00:03:21,682 --> 00:03:22,602 A few gotchas. 49 00:03:23,182 --> 00:03:26,282 First, the wait cap is per request, not per job. 50 00:03:26,742 --> 00:03:30,642 A session with many turns can burn several wait periods, so leave headroom under 51 00:03:30,662 --> 00:03:31,562 your runner timeout. 52 00:03:32,322 --> 00:03:36,602 Second, retries only happen for requests that failed before Claude produced any 53 00:03:36,682 --> 00:03:39,142 completed text or ran any tool calls. 54 00:03:39,842 --> 00:03:43,382 If a stream dies after tools have run, Claude Code won't replay them, 55 00:03:43,862 --> 00:03:47,682 which protects you from duplicate writes or a repeated destructive shell command. 56 00:03:48,402 --> 00:03:54,582 Third, a 429 from a hard spend cap, exhausted credits, or an org policy denial comes 57 00:03:54,602 --> 00:03:58,282 back with a header telling it not to retry, so it fails immediately. 58 00:03:58,962 --> 00:04:03,422 And please don't put the watchdog in your .bashrc or .zshrc. 59 00:04:03,922 --> 00:04:05,662 Your interactive sessions will look frozen. 60 00:04:06,142 --> 00:04:08,582 Scope it to CI containers and automated jobs. 61 00:04:08,667 --> 00:04:10,027 Quick quality of life round. 62 00:04:10,160 --> 00:04:16,987 In 2.1.296, Edit and NotebookEdit now refuse to edit files that aren't valid UTF 63 00:04:17,067 --> 00:04:23,947 8, like Shift JIS, Windows 1252, or GBK, where before they quietly 64 00:04:24,027 --> 00:04:26,667 mangled every non ASCII character. 65 00:04:26,760 --> 00:04:30,907 Turns or bang commands started while a session moves to the background with the left 66 00:04:30,960 --> 00:04:33,787 arrow are now stopped, not run out of sight. 67 00:04:33,947 --> 00:04:38,427 On Windows, PowerShell commands over 1 KB used to always prompt, 68 00:04:38,447 --> 00:04:43,067 and allow rules and read only detection now apply up to 32 KB. 69 00:04:43,083 --> 00:04:43,843 Two more. 70 00:04:43,933 --> 00:04:48,283 Run claude with debug, and command hooks now log their command, 71 00:04:48,352 --> 00:04:53,883 plugin, outcome, and duration when they finish, so a slow hook is easy to spot. 72 00:04:53,923 --> 00:05:00,763 And CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL runs every workflow agent on one 73 00:05:00,830 --> 00:05:05,403 model, say Haiku 5.5, while your regular subagents keep their own. 74 00:05:05,590 --> 00:05:10,683 If a nightly job's been dying on a 529, set the wait cap and a fallback chain this 75 00:05:10,747 --> 00:05:11,083 week.