# Barn 0.9.3 Changes > Barn 0.9.3 changes covering resource resize, recovery, paths, sharing, node selection, images, and limits. --- LLMS index: [llms.txt](/llms.txt) --- This page records the user-visible changes shipped in **Barn 0.9.3** on 2026-10-31. The workflow changes began in source commit `27f17bb` and were completed by the commits leading to the release. Check `barn version` when auditing another build. See the [Barn 0.9.3 release notes](/blog/release/0.9.3/) for the upgrade summary and the complete bilingual release note. ## Resource changes and recovery {#resize-recovery} Edit `vm_cpu` or `vm_mem` and run `barn resize [node...]`. QEMU hot-plugs supported increases when the running VM has reserved capacity; decreases, older VMs, and unsupported backends use a stop/start fallback. The VM identity, disks, and SSH keys remain unchanged. `resize` refuses image, disk, network, share, creation, removal, and deployment-wide changes; use `recreate`, `up`, `destroy`, or a full selected operation as indicated by its `next:` field. Linux and macOS lifecycle, persistent-disk, and replacement operations journal their transition state. If a process is interrupted or receives `SIGINT` or `SIGTERM`, the next lifecycle command can recover the operation; cancelled commands report the structured `cancelled` result. On macOS, `barn mac up --no-wait` returns when the VM process starts and may leave first-boot SSH and sudo provisioning running. Run `barn mac up NAME` again to wait for readiness. ## Paths {#paths} `barn init` and `setup` work through directory symlinks. `init -o ./lab/config/barn.yml` also creates missing parent directories. Existing files still require `--force`, and writes never follow or replace an output-file symlink. User-selected paths now accept normal development layouts: | Input | Resolution | |---|---| | `BARN_HOME` | Relative paths, `~/`, and directory symlinks; may also be the working directory | | Inventory and `provision --script` | Symlinks to regular files; each file retains its 4 MiB limit | | Linux `vm_shares.host` | Relative to the inventory's real file, or an absolute/`~/` path; directory symlinks work | | Local image repository and `image import` | Relative paths, `~/`, and symlink inputs | | `catalogexport` output | Relative paths and `~/`, directory aliases, and automatic parent creation; target must be absent | | `BARN_MAC_RUNNER` and local `socket_vmnet` archive | Relative paths and `~/`; archive inputs may be symlinks and still undergo digest verification | {.wide .expand} CLI-relative paths, including a relative `BARN_HOME`, are interpreted from the current working directory. Use an absolute `BARN_HOME` to reach the same deployment from different directories. Barn still rejects `/` and the user's home directory as data roots and checks the ownership and structure of its managed state. An owned `~/.ssh` directory may be reached through a directory symlink; a linked `~/.ssh/config` is still left for manual `Include` integration. ## Shared folders and mount paths {#sharing} Linux guests use **QEMU Virtio-9p**, now supported on macOS hosts as well as Linux hosts. QEMU and the guest kernel must support 9p. Native macOS guests continue to use **VirtioFS** through `barn mac --share`. - Linux share sources use the invoking user's actual filesystem permissions. A directory may belong to another user, overlap another source, or be writable by multiple guests. Coordinate concurrent writes as you would between host processes. - Source paths support spaces, Unicode, and commas. Linux guest mount paths are normalized and escaped for shell commands and `fstab`, including spaces, Unicode, and shell punctuation. - Mac share names support spaces and Unicode: valid UTF-8, 1–255 bytes, excluding `.`, `..`, `/`, and ASCII control characters. Share sources may be directory symlinks. - The eight-share limit is removed. Backend device and host resource limits still apply. Linux sources must not overlap `BARN_HOME`, and mount targets must not overlap each other, system directories, data-disk mounts, or the guest SSH directory. See [Storage and Access](../../start/storage/) for Linux shares and [Mac Commands](../../reference/mac/) for native macOS shares. Barn never changes host ownership to make a share writable. ## Apply only the selected node changes {#selection} Explicit selections in `plan`, `up`, `reload`, and `recreate` preserve unselected peers' applied definitions, even if those peers were edited or omitted in the current inventory. Uncreated, unselected nodes do not require image resolution. This allows a multi-node edit to be applied in batches: ```bash barn plan node-1 -f barn.yml barn recreate node-1 -f barn.yml barn plan node-2 -f barn.yml ``` The entire inventory must still parse correctly. Conflicts with applied addresses, aliases, control-node roles, or shared deployment settings are checked before stopping or deleting nodes. Omitting selectors evaluates the whole inventory; removing a YAML entry never deletes its VM. See [Daily Operations](../../start/operations/#selected-nodes). ## Generated template image selection {#template-images} `barn init [template] --image ` writes an explicit `vm_image` under `all.vars` for every generated node. `barn setup [template] --image ` uses the same behavior while preparing a generated template. References may be aliases, channels, or version selectors; the flag does not download an image. Existing inventories selected with `-f` or discovered automatically are never rewritten by this option. ## Images and development limits {#limits} | Area | New behavior | |---|---| | Linux node count | No separate 20-node ceiling; nodes need distinct usable addresses in the deployment's `/24` and sufficient host resources | | External inventory hosts | `vm_skip: true` entries may use hostnames or IPv6 keys; managed VMs still use IPv4, and at least one managed node is required | | Linux capacity strings | Case-insensitive units; `K/M/G/T` and `KiB/MiB/GiB/TiB` are binary, while `KB/MB/GB/TB` remain decimal | | Provisioning | `--parallel` accepts positive integers and uses at most the selected node count; `--timeout` accepts any positive duration, without the old four-worker or 24-hour ceilings | | Image size | No 8 GiB artifact ceiling for local imports or catalog artifacts with declared sizes; downloads without a declared size retain the legacy 8 GiB bound | | Image imports | Arbitrary source filenames; new cache entries use `local/.qcow2`, so same-name files with different contents coexist and identical contents reuse the cache | | Local aliases | No 256-alias ceiling; updates exceeding the 1 MiB registry metadata limit fail before writing. Reimporting an existing alias preserves its old cache path | | Local repositories | Relative or linked roots and shared writable directories work; `repo.yaml` may be a symlink to a regular file | | Mac logs | Any positive `--lines` count; requested lines are no longer silently clipped at 1 MiB or 10,000 lines | | Fallback ports | Try `preferred + n × 10000` through port 65535, rather than stopping after five candidates | {.wide .expand} Provisioning still defaults to one worker and a one-hour deadline, with 1 MiB of captured stdout and stderr per node. These user-facing limits are separate from the internal VM startup scheduler. Image digests, declared sizes, `qcow2` structure, retained-disk compatibility, and process identity are still verified. The contributor image pipeline retains its own recipe bounds.