1 00:00:00,263 --> 00:00:00,958 Welcome to the show. 2 00:00:01,158 --> 00:00:07,519 James, May 4th, 2026 -- Claude Code v2.1.128 3 00:00:07,599 --> 00:00:13,034 ships, and buried in the notes is one of those little Git lines that can absolutely 4 00:00:13,098 --> 00:00:14,394 make your stomach drop. 5 00:00:14,594 --> 00:00:18,392 `EnterWorktree` used to branch from `origin/`. 6 00:00:18,552 --> 00:00:21,591 Now it branches from your local `HEAD`. 7 00:00:21,658 --> 00:00:26,867 Which means before this fix, if you had unpushed commits sitting locally, 8 00:00:26,887 --> 00:00:29,584 the new worktree could quietly ignore them. 9 00:00:31,199 --> 00:00:35,648 Wait -- `origin/` versus local `HEAD` is not a tiny distinction. 10 00:00:35,760 --> 00:00:40,685 That's the difference between "branch from what I ACTUALLY have on my machine" and 11 00:00:40,845 --> 00:00:43,561 "branch from whatever I last shoved up to remote." 12 00:00:43,748 --> 00:00:48,207 If I've got a WIP commit locally, that agent is basically starting from yesterday's 13 00:00:48,303 --> 00:00:48,767 news. 14 00:00:49,000 --> 00:00:49,962 Exactly. 15 00:00:50,082 --> 00:00:51,561 And that's the sneaky bit. 16 00:00:51,701 --> 00:00:53,321 The worktree looked normal. 17 00:00:53,401 --> 00:00:55,240 Nothing was obviously on fire. 18 00:00:55,373 --> 00:00:58,597 No big red warning, no busted checkout. 19 00:00:58,677 --> 00:01:03,244 But unless you were the sort of maniac -- and I say that lovingly as a web dev who's 20 00:01:03,276 --> 00:01:08,192 done this in a panic at midnight -- unless you checked `git log` in both places, 21 00:01:08,252 --> 00:01:12,996 you'd have no clue your fresh agent branch was based on stale remote state. 22 00:01:14,272 --> 00:01:17,281 And that is the WORST category of bug. 23 00:01:17,461 --> 00:01:18,719 Not a crash. 24 00:01:18,799 --> 00:01:20,319 Not a loud failure. 25 00:01:20,559 --> 00:01:21,519 A liar. 26 00:01:21,619 --> 00:01:26,629 Because if the branch opens fine, files are there, everything feels clean... 27 00:01:26,669 --> 00:01:27,592 you trust it. 28 00:01:27,720 --> 00:01:31,671 Then twenty minutes later you're asking, "Why doesn't the code match?" 29 00:01:31,771 --> 00:01:36,235 and the answer is: because the base was wrong before you even started. 30 00:01:37,205 --> 00:01:37,401 Yep. 31 00:01:37,454 --> 00:01:40,520 Multi-agent workflows lean on worktrees for isolation. 32 00:01:40,536 --> 00:01:41,640 That's the whole bargain, right? 33 00:01:41,800 --> 00:01:46,043 One agent here, another over there, everyone's got their own little fenced paddock. 34 00:01:46,183 --> 00:01:50,204 But if those paddocks start from different realities -- one from local truth, 35 00:01:50,244 --> 00:01:54,688 one from stale remote truth -- you're not isolating work, you're splitting reality. 36 00:01:54,757 --> 00:01:56,051 Proper dog's breakfast. 37 00:01:57,000 --> 00:01:59,521 "Splitting reality" is good. 38 00:02:00,061 --> 00:02:03,563 [short pause] Because in practice it's not just one missing commit. 39 00:02:03,723 --> 00:02:05,166 It's interpretation drift. 40 00:02:05,313 --> 00:02:07,726 Agent A sees the local fix you made. 41 00:02:07,846 --> 00:02:11,404 Agent B, launched through the buggy path, doesn't. 42 00:02:11,584 --> 00:02:16,362 Now they generate different edits, different assumptions, maybe even different bug 43 00:02:16,442 --> 00:02:17,082 reports. 44 00:02:17,242 --> 00:02:20,366 And none of that LOOKS like a Git-base problem at first. 45 00:02:21,000 --> 00:02:24,280 Right, and from the dev side, that's the gut-punch. 46 00:02:24,480 --> 00:02:27,554 You trust a tool to just do the sensible thing. 47 00:02:27,649 --> 00:02:31,801 Especially with worktrees, because they're meant to be the tidy option. 48 00:02:31,961 --> 00:02:37,009 Then later you realize it respected the docs less than your actual repo state. 49 00:02:37,289 --> 00:02:41,641 [laughs softly] Mate, that's like asking someone to copy your shopping list and they 50 00:02:41,692 --> 00:02:44,202 faithfully copy last week's receipt instead. 51 00:02:45,704 --> 00:02:47,639 Last week's receipt is perfect. 52 00:02:47,699 --> 00:02:51,885 And I think the horror lands hardest on that phrase "silently skipped." 53 00:02:52,072 --> 00:02:56,365 If an unpushed local fix gets silently skipped, it means the important information 54 00:02:56,445 --> 00:02:57,885 wasn't missing from Git. 55 00:02:57,992 --> 00:02:59,166 It was present. 56 00:02:59,246 --> 00:03:01,565 It was right there in your local `HEAD`. 57 00:03:01,585 --> 00:03:02,924 The wrapper just didn't use it. 58 00:03:04,000 --> 00:03:06,080 And to be fair, now it does. 59 00:03:06,267 --> 00:03:13,045 In v2.1.128, `EnterWorktree` branches from local `HEAD`, so unpushed commits 60 00:03:13,105 --> 00:03:16,009 are finally included instead of being left behind. 61 00:03:16,169 --> 00:03:20,575 But I do think this is one of those bugs that's worth talking about because loads of 62 00:03:20,626 --> 00:03:24,814 developers would've had that weird, low-grade feeling of "hang on... 63 00:03:24,854 --> 00:03:26,575 why is this branch off?" 64 00:03:26,615 --> 00:03:28,894 without ever catching the exact cause. 65 00:03:29,142 --> 00:03:30,120 Let me try to say it back. 66 00:03:30,257 --> 00:03:35,397 Before May 4th, 2026, if I spawned a worktree through `EnterWorktree`, 67 00:03:35,437 --> 00:03:39,156 I might THINK I was branching from my current working reality, 68 00:03:39,176 --> 00:03:41,875 but I was actually branching from `origin/`. 69 00:03:41,928 --> 00:03:46,111 So if I hadn't pushed, the agent missed my latest local commits. 70 00:03:46,204 --> 00:03:52,348 After v2.1.128, it finally branches from local `HEAD`, which is what most developers 71 00:03:52,375 --> 00:03:53,788 would've assumed in the first place. 72 00:03:54,268 --> 00:03:54,801 That's it. 73 00:03:55,041 --> 00:04:00,798 Not a new superpower -- more like the floorboards have been screwed down properly. 74 00:04:01,118 --> 00:04:06,554 And if you've ever nearly tanked a site with one bad assumption -- I have, 75 00:04:06,634 --> 00:04:12,310 long story, awful midnight deploy -- you get very twitchy around tools that fail 76 00:04:12,400 --> 00:04:13,110 quietly. 77 00:04:13,377 --> 00:04:17,353 Quiet bugs are the ones that make you doubt your own memory. 78 00:04:18,000 --> 00:04:24,000 So the key framing here is: this is less a feature change than a correctness repair. 79 00:04:24,020 --> 00:04:26,000 The documented behavior is restored. 80 00:04:26,213 --> 00:04:31,281 If you WANT remote-state branching now, you need to push first or pick a different 81 00:04:31,345 --> 00:04:32,476 base explicitly. 82 00:04:32,556 --> 00:04:37,439 That's a very different mental model from "the tool will secretly choose remote for 83 00:04:37,492 --> 00:04:37,757 me." 84 00:04:38,000 --> 00:04:40,081 [questioning tone] And that's an important distinction, hey, 85 00:04:40,121 --> 00:04:44,716 because plain `git worktree add` still behaves like standard Git. 86 00:04:44,929 --> 00:04:47,517 So the bug wasn't Git being weird. 87 00:04:47,597 --> 00:04:52,711 It was Claude Code's internal `EnterWorktree` wrapper doing something specific in 88 00:04:52,731 --> 00:04:53,910 the agent workflow. 89 00:04:54,000 --> 00:04:57,675 Exactly -- `git worktree add` is standard Git territory. 90 00:04:57,835 --> 00:05:00,310 `EnterWorktree` is the convenience layer. 91 00:05:00,430 --> 00:05:04,789 And when convenience layers drift from expected behavior, that's where trust gets 92 00:05:04,853 --> 00:05:05,749 expensive. 93 00:05:05,789 --> 00:05:08,069 Because developers blame themselves first. 94 00:05:08,165 --> 00:05:10,426 They think, "Did I forget a stash? 95 00:05:10,476 --> 00:05:11,901 Did I checkout wrong? 96 00:05:11,921 --> 00:05:13,260 Did I imagine that commit?" 97 00:05:14,135 --> 00:05:16,956 "Have I become a drongo, or is the tool haunted?" 98 00:05:17,116 --> 00:05:20,395 Those are the two options at 11:47 p.m. 99 00:05:21,519 --> 00:05:22,119 Usually both. 100 00:05:22,319 --> 00:05:25,879 But this release has a broader pattern, and I like that. 101 00:05:26,019 --> 00:05:29,395 The reliability story isn't only the worktree fix. 102 00:05:29,462 --> 00:05:32,360 There's also subprocess OTEL isolation. 103 00:05:32,536 --> 00:05:39,399 Bash, hooks, MCP servers, LSP processes -- they stop inheriting `OTEL_*` variables 104 00:05:39,519 --> 00:05:42,677 and accidentally shipping traces to Claude Code's collector. 105 00:05:43,000 --> 00:06:15,160 `OTEL_*` 106 00:06:15,160 --> 00:06:15,160 is the 107 00:06:15,160 --> 00:08:15,800 memorable 108 00:08:15,807 --> 00:08:16,083 token there for me. 109 00:08:16,098 --> 00:08:19,488 Because environment-variable leakage is another silent one, 110 00:08:19,541 --> 00:08:20,126 right? 111 00:08:20,246 --> 00:08:21,686 Not dramatic. 112 00:08:21,758 --> 00:08:26,278 Just the wrong subprocess inheriting tracing config it shouldn't have, 113 00:08:26,318 --> 00:08:29,630 and suddenly telemetry goes somewhere you never intended. 114 00:08:31,000 --> 00:08:34,765 Yeah, and that phrase "accidentally shipping traces" matters. 115 00:08:35,858 --> 00:08:38,079 We're not talking about some flashy UI bug. 116 00:08:38,092 --> 00:08:43,077 We're talking about background processes -- Bash, hooks, MCP servers, 117 00:08:43,157 --> 00:08:46,521 LSPs -- inheriting state they didn't ask for. 118 00:08:46,734 --> 00:08:49,564 Isolation is the whole game in toolchains like this. 119 00:08:49,740 --> 00:08:53,969 Same theme as worktrees, honestly: keep boundaries real. 120 00:08:55,000 --> 00:08:56,044 That's a nice connection. 121 00:08:56,188 --> 00:08:58,766 Worktree isolation for code state. 122 00:08:59,046 --> 00:09:01,244 OTEL isolation for process state. 123 00:09:01,377 --> 00:09:05,804 In both cases, the fix is basically: stop the invisible bleed-over. 124 00:09:06,000 --> 00:09:11,209 And then there's the other one I think agent users will FEEL immediately: the 125 00:09:11,280 --> 00:09:13,208 parallel tool-call fix. 126 00:09:13,345 --> 00:09:17,926 Before, a failing read-only Bash command like `grep`, `git diff`, 127 00:09:17,979 --> 00:09:20,488 or `ls` could cancel sibling calls. 128 00:09:20,628 --> 00:09:21,927 Now it doesn't. 129 00:09:21,991 --> 00:09:24,651 That is huge for multi-step agent work. 130 00:09:25,146 --> 00:09:28,354 The examples there are so grounded too -- `grep`, `git diff`, 131 00:09:28,434 --> 00:09:28,994 `ls`. 132 00:09:29,090 --> 00:09:31,357 That's not some exotic edge case. 133 00:09:31,413 --> 00:09:33,398 That's bread-and-butter terminal stuff. 134 00:09:33,638 --> 00:09:37,714 If one read-only command hits a dead end and takes down its siblings, 135 00:09:37,734 --> 00:09:39,874 the workflow feels brittle as. 136 00:09:40,018 --> 00:09:44,429 Like one wheel hits a pothole and the whole bike decides the ride is over. 137 00:09:45,000 --> 00:09:45,360 Right. 138 00:09:45,370 --> 00:09:49,324 And in agent workflows, parallel calls are often about exploration. 139 00:09:49,444 --> 00:09:53,805 One command checks files, another diffs changes, another searches text. 140 00:09:53,965 --> 00:09:58,284 If `grep` fails because the pattern isn't there, that should be information, 141 00:09:58,324 --> 00:09:59,723 not a global emergency. 142 00:09:59,863 --> 00:10:05,641 The old behavior basically treated "no result" or "that path's wrong" as a reason to 143 00:10:05,694 --> 00:10:07,476 collapse the whole branch of thought. 144 00:10:08,137 --> 00:10:11,915 Which is very unlike how a good dev actually works. 145 00:10:12,006 --> 00:10:13,431 Humans poke around. 146 00:10:13,631 --> 00:10:16,470 One command fizzles, we try the next thing. 147 00:10:16,683 --> 00:10:20,710 We don't throw the laptop into the bin because `ls` had a sad day. 148 00:10:21,000 --> 00:10:24,840 And that's why these fixes belong together in one conversation. 149 00:10:25,048 --> 00:10:30,987 Worktree bases, OTEL inheritance, sibling-call cancellation -- on paper they're 150 00:10:31,031 --> 00:10:31,463 separate. 151 00:10:31,603 --> 00:10:34,903 But the shared theme is reducing silent interference. 152 00:10:35,031 --> 00:10:36,341 Less hidden state. 153 00:10:36,461 --> 00:10:38,421 Fewer invisible cancellations. 154 00:10:38,541 --> 00:10:42,417 Fewer moments where the tool quietly nudges reality sideways. 155 00:10:43,160 --> 00:10:44,000 Yeah. 156 00:10:44,130 --> 00:10:47,080 The dream with agent tools isn't magic. 157 00:10:47,187 --> 00:10:48,203 It's trust. 158 00:10:48,550 --> 00:10:52,044 If I branch from here, branch from HERE. 159 00:10:52,284 --> 00:10:55,007 If a subprocess shouldn't inherit `OTEL_*`, don't leak it. 160 00:10:55,007 --> 00:10:55,007 If `git diff` fails in one lane, 161 00:10:55,007 --> 00:12:42,527 don't 162 00:12:43,647 --> 00:12:48,688 nuke the other 163 00:13:07,551 --> 00:13:09,006 lane. 164 00:13:09,006 --> 00:13:09,006 Basic stuff... 165 00:13:09,006 --> 00:13:14,773 but basic stuff is what keeps your head straight. 166 00:13:15,067 --> 00:13:19,391 And honestly, that's the takeaway I'd leave people with: not "learn this one weird 167 00:13:19,471 --> 00:13:23,786 bug," but notice which tools fail loudly and which ones fail politely. 168 00:13:23,966 --> 00:13:27,145 The polite failures are the ones that rewrite your assumptions. 169 00:13:27,252 --> 00:13:29,067 Those are the dangerous ones. 170 00:13:30,000 --> 00:13:30,559 Too right. 171 00:13:30,607 --> 00:13:32,082 Push if you want remote truth. 172 00:13:32,162 --> 00:13:33,841 Choose a base if you need one. 173 00:13:33,981 --> 00:13:38,635 But if a wrapper promises isolation, make sure it's not doing a sneaky little side 174 00:13:38,702 --> 00:13:40,079 quest behind your back. 175 00:13:40,228 --> 00:13:42,162 Cheers, James. 176 00:13:43,000 --> 00:13:43,721 See you next time.