~/.claude and ~/.codex mix config with logins and plaintext transcripts. Back up what you wrote, exclude credential files, and log in again on restore.
A coding agent in VS Code got "No route to host" for every host on the subnet except the router, and blamed the Wi-Fi. macOS was blocking VS Code, which lacked the Local Network permission that Terminal gets automatically.
Every ansible-playbook run from a coding agent failed before reading the playbook, even --syntax-check. The agent shell hands Ansible non-blocking file handles; `&1 | cat` gives it blocking ones.
rsync into an encrypted APFS sparse bundle stored on an SMB share failed on every write with error 45. The files were fine. APFS inside the image needs F_FULLFSYNC on the band files, the NAS's SMB server doesn't support it, and a short Python probe proves it.
A backup job on a macOS SMB share kept dying with EBADF. The NAS stalled under load, macOS reconnected without a word, and because the old SMB 2.1 server granted no durable handles, every open file on the share went bad, including the bash script that was running. Only the kernel log says so.
Three times a coding agent read a result from a check that never ran: zsh's `log` builtin (`(eval):log:1: too many arguments`) hid `/usr/bin/log`, `cargo test --exact` matched no test and passed, and an empty variable from "argument list too long" made 400 of 400 runs fail.
A macOS FSKit extension mounted from the command line could stat its source folder but not read it. The extension sandbox was denying file-read-data on a plain path. Reinstalling the app then switched the module off in a way pluginkit does not undo.
A coding agent ran pnpm test and got ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY from pnpm 11, in a repo pinned to pnpm 10. The agent's runtime ships its own pnpm and turns off the packageManager check. Following the CI=true hint reinstalled node_modules, filled the disk, edited pnpm-workspace.yaml and lost a native Next.js binary.