Hands-on test: worktrunk v0.81.0 with five parallel AI agent worktrees
A hands-on report tested worktrunk v0.81.0 on an arm64 Mac running macOS 26.6.2 with git 2.56.0. Creating five worktrees took 0.45s with worktrunk versus 0.11s with native git. Copying a 120MB gitignored test file with wt step copy-ignored took 0.02s with a df difference of -32KB, against 0.20s and roughly 120MB of extra disk for cp -R; the author attributes this to APFS copy-on-write and states only that no measurable disk increase was observed, not that zero bytes were used. Removing a worktree holding a 120MB dependency with wt remove -y --foreground took 0.09s. The five workspaces were set up with wt switch -c -x as a workflow demonstration; the author did not run agents to completion and did not test conflict handling, rebase, the interactive picker or LLM-generated commit messages. The figures are single manual comparisons rather than benchmarks, and the 120MB file was generated with dd instead of a real node_modules directory.
The report offers concrete guidance for developers running parallel coding agents: wt list surfaces branch, ahead/behind and remote status in one view, wt remove handles worktree and branch together and refuses to discard uncommitted changes without --force, lifecycle hooks can install dependencies or start dev servers (with a hash_port template and first-run command approval), and reflinked dependency copying only works on APFS, btrfs and XFS, not ext4. It also documents costs: worktree creation was slower than native git in the tested setup, and the two binaries take about 44MB.
五个 Agent 工作区，worktrunk 实测
下午三点，最容易上头的事不是写代码，是同时开五个 AI 改一个项目。
五个窗口、五个分支、五次 git worktree add、五次 cd ../repo.feat。真正让人放弃并行的不是模型不够聪明，是目录切来切去。所以我今天装了一个专门管这件事的小工具：worktrunk，截至 11 月 4 日 10,672 star；当天我看到的 GitHub trending 第 3，这个排名事后没法复核。
先说清楚这篇测了什么、没测什么，免得标题比证据走得更远：
- 测了：建 5 个 worktree 的耗时；
copy-ignored复制一个 120MB 测试文件的时间和df差值；5 份普通复制的磁盘占用（df实测 600MB）；remove --foreground删完的耗时（0.09s，43 个文件 / 120MB）；一个pre-starthook 的耗时。 - 没测：五个 AI 真的跑完五个任务、冲突怎么解、rebase 怎么走、reflink 之后写入的块分裂、交互式 picker、LLM 生成 commit message。
所以下文里的「快」「慢」「省」，只对应上面被量过的那几个动作。