devboxes

Imported from namespacelabs/agent-skills (skills/devboxes/SKILL.md) via skills.sh. Install upstream with npx skills add namespacelabs/agent-skills --skill devboxes. Copyright stays with the author.

Namespace devboxes

Namespace devboxes are remote linux/amd64 or macos/arm64 machines. Linux devboxes have Docker available. This skill describes the general building blocks for working with devboxes - creating them, transferring files, hydrating a workspace, installing toolchains, running commands with devbox exec, sharing web work with teammates through devbox.so, and managing the devbox lifecycle.

Throughout this skill:

• devbox - a Namespace machine, identified by its name.
• <name> - the devbox name passed to all devbox subcommands.

The devbox may be long-lived, reused across tasks, or torn down immediately depending on the caller's intent. Pass --ephemeral at creation time to mark it as disposable.

For run-once-then-destroy workflows (single-script or test runs, optional sharding), see references/run-tests.md.

0. Establish where you are running

Everything below assumes you are driving a devbox from somewhere else. First check whether you are already ON one - the tooling and the auth story are different:

[ -n "${NAMESPACE_DEVBOX_WORKSPACE_DIR:-}" ] && echo "on a devbox"

The Devbox agent injects that variable into every process on the machine, so it is a reliable one-line check. If it is set, read references/inside-a-devbox.md before doing anything else. Two things are different there:

• Ordinary commands need no wrapper. You are on the machine. Run go test, git and file edits directly, and skip the devbox preflight below for work scoped to the devbox you are in.
• Both CLIs are already usable - do not run devbox login. boxctl and devbox both ship in ~/.devbox/bin. The devbox CLI is authenticated by a workload credential, which acts as the tenant rather than as a person: it can drive workspace-access devboxes, but not ones private to their creator - including, usually, the machine you are on. That is why boxctl, which reaches the local agent over a socket and needs no credential, is the tool for the current devbox rather than merely the quicker one. Beware devbox auth check-login here: it reports "Not logged in" even while the CLI works, and acting on that leads to a browser login flow you cannot complete. The reference covers the rest, including --access_mode workspace for creating sibling boxes.

Preflight Resolve devbox from PATH or its platform's default install location; use the absolute path if PATH is stale, and install the Devbox CLI yourself only if it is absent. Run the resolved CLI's auth check-login; if login is required, run devbox login, let the user complete the browser flow, and wait for the command to return. See references/cli-setup.md for exact steps, including session expiry and workspace switching. Do not try to mint or borrow credentials from other tooling.

1. Create a devbox

If the task needs a project toolchain (Go, Node, Xcode, etc.), pick an image that already includes it to avoid reinstalling dependencies on every run: run devbox image list -o json ONCE to discover existing project images, and use one if it fits. Cache that result - repeating the call does not tell you anything new.

Skip image discovery entirely for scratch work that needs no project toolchain (a one-off command, a throwaway VM, a quick platform check) and go straight to builtin:default. On Linux, builtin:default already ships common toolchains - Go, for instance - so for many tasks it is the whole answer. For simpler Linux cases generally, fall back to builtin:default and install dependencies directly.

Important Pass the short name field from devbox image list -o json to --image (e.g. <org>/<image> or builtin:default), not the full repository URL - full references typically fail. Ensure the image matches the requested platform.

devbox create \
  --name <name> \
  [--platform <linux/amd64|macos/arm64>] \
  --image <image> \
  --size <size> \
  [--ephemeral] \
  [--checkout <repo-url>] \
  --purpose "<purpose>"

Linux/amd64 is the default when --platform is omitted. Pass --platform macos/arm64 when the task requires macOS or Apple Silicon.

Add --ephemeral when the devbox does not need to persist after a task is complete; omit it when persistence is required.

Important (workspace checkout default) If --checkout is unset, Namespace uses the workspace's configured default repo and clones it automatically. When no repository is needed (e.g. a scratch VM for a one-off command), pass --no_checkout to skip cloning entirely.

--checkout <repo-url> (e.g. git@github.com:org/repo.git) clones the repository into /workspaces/<repo-name>. If creation fails because checkout is unavailable for the repo or org:

1. Expire the just-created devbox (devbox expire <name> --force) to free capacity.
2. If gh is installed locally, recreate the devbox WITHOUT --checkout, then run devbox setup-github <name> to forward your gh token into the devbox. Once that succeeds, run git clone --depth=1 <repo-url> /workspaces/<repo-name> inside the devbox via devbox exec <name> --. Always use --depth=1 for large monorepos - full clones can run for minutes with sparse progress output and look identical to a hang. Important The forwarded gh token authenticates HTTPS only - clone via https://github.com/<org>/<repo>.git, not git@github.com:....
3. If gh is not installed, or devbox setup-github / the in-devbox git clone fails, instruct the user to link their GitHub organization to Namespace at https://cloud.namespace.so/workspace/workspace/integrations?integration=github - checkout cannot proceed without it.

Important (case-sensitive org slug) The Namespace association lookup is case-sensitive on the GitHub org/owner segment of --checkout. Use the org/owner casing EXACTLY as it appears in the repo URL the user gave you (or as returned by git remote -v). NEVER try alternative casings (e.g. lowercasing the org) as a "fix" when checkout fails.

Linux sizes: s (4 vCPU / 8 GB), m (8 vCPU / 16 GB), l (16 vCPU / 32 GB), xl (32 vCPU / 64 GB). macOS sizes: m (6 vCPU / 14 GB) and l (12 vCPU / 28 GB). Use the smallest size only for narrow commands or small targeted workloads; prefer bigger when the workload has large dependencies or builds multiple packages, to avoid OOM. If a process is killed with signal: killed, upgrade to a larger size and retry.

Overlap creation with local prep devbox create blocks until the machine is ready, which takes tens of seconds. Use that time: run devbox create as a background task, then immediately compute any local git diff and write the setup and run scripts locally. Once creation completes, upload and execute. This eliminates idle waiting.

Note Devbox creation might fail due to Namespace plan limits on resource tiers or instance counts. If this happens: PROMPT THE USER BEFORE PROCEEDING run devbox list and ask to expire or reuse any of the available devboxes, then retry; As a last resort, consolidate work across fewer devboxes rather than blocking.

For test or script runs that benefit from sharding across multiple devboxes, see references/run-tests.md.

2. Hydrate the workspace and install toolchains

Upload and download files

devbox upload <name> <local-path>  <remote-path> [--mkdir]
devbox download <name> <remote-path> <local-path>

• <remote-path> MUST be a full file path. A trailing / fails.
• --mkdir creates missing parent directories; it does NOT make the target a directory.
• The executable bit is NOT preserved. Run chmod +x with devbox exec after uploading scripts.
• Write under /workspaces or $HOME. The remote user is unprivileged (devbox, uid 1001), not root, so an upload to a system path fails on the parent directory - mkdir: cannot create directory '/srv/www': Permission denied - and --mkdir cannot rescue it, because the problem is permission, not absence. Serving a page? /workspaces/<name>/ is the natural home, not /srv or /var/www.

Hydrate the workspace

Namespace devboxes configured with a repo in the UI auto-clone it to /workspaces/<repo-name> on startup. Always check there first before syncing:

devbox exec <name> -- ls /workspaces/

If the repo is already present, use that path directly. The auto-cloned repo reflects the default branch, so it does NOT have your local work on it - you must get that across before running anything.

Pick the cheapest sync that covers your files. Two approaches, and the simple one is usually right:

(a) Default: ship the working tree. For a small or medium repo, copy the files as they are on disk. This captures tracked edits and untracked files in one step, needs no reasoning about commits or bases, and cannot silently omit anything. Upload a handful of files directly, or tar the tree for more:

# A few files - upload them directly.
devbox upload <name> ./calc.go /workspaces/<repo-name>/calc.go --mkdir

# More than a handful - tar the tree, excluding .git and build output.
tar --exclude .git -czf /tmp/tree.tgz -C <repo-root> .
devbox upload <name> /tmp/tree.tgz /tmp/tree.tgz
devbox exec   <name> -- mkdir -p /workspaces/<repo-name>
devbox exec   <name> -- tar -xzf /tmp/tree.tgz -C /workspaces/<repo-name>

(b) Large repo: align the commit, then patch. Once the tree is big enough that copying it is slower than cloning it, let the devbox's own clone do the heavy lifting and send only a diff. This is strictly more machinery, so reach for it only when (a) is genuinely too slow.

Important The devbox clone is on the default branch, so the diff and the checkout MUST name the same commit or unrelated files get clobbered:

BASE=$(git merge-base origin/main HEAD)
git diff "$BASE" > /tmp/changes.patch   # add --binary if the diff touches generated/binary files

# For git, prefer `git -C <path>` over `cd <path> && git ...` - no cwd to get wrong.
devbox exec   <name> -- git -C /workspaces/<repo-name> fetch --depth=1 origin "$BASE"
devbox exec   <name> -- git -C /workspaces/<repo-name> checkout "$BASE"
devbox upload <name> /tmp/changes.patch /tmp/changes.patch
devbox exec   <name> -- git -C /workspaces/<repo-name> apply /tmp/changes.patch
devbox exec   <name> -- git -C /workspaces/<repo-name> status --short

Important (the failure mode that makes (b) risky) git diff only covers TRACKED files. A brand-new file you have not git added is invisible to it, so a patch-based sync will quietly leave it behind - and if that file is a new test, you will confidently report a result from code that was never there. Requirement (b) is therefore incomplete on its own: send untracked files too.

git ls-files --others --exclude-standard    # exactly what a patch will MISS

Upload those directly, add them to the tarball, or rsync them if there are many:

# for rsync - configure native ssh
devbox configure-ssh <name>
# After running this, you can connect with:
ssh <name>.devbox.namespace
# Then list untracked files and feed them to rsync
git ls-files --others --exclude-standard | \
  rsync -av --files-from=- . \
  <name>.devbox.namespace:/workspaces/<repo-name>/

Verify before you trust a result. Confirm the files you meant to send are actually there, so a sync gap surfaces now rather than as a wrong answer later:

devbox exec <name> -- ls -la /workspaces/<repo-name>

If the repo has no remote at all (a scratch or local-only repo), (b) is not available - use (a).

Install toolchains

If a tool is missing, install only what the workload needs.

Important Check whether a tool is already present before installing it - on Linux, builtin:default ships more than you might expect, including languages such as Go. Try the command before reaching for a manual install.

Probe with the tool's own flag, which is a real executable:

devbox exec <name> -- go version

command -v will NOT work this way. It is a shell builtin, and devbox exec runs the command directly with no shell, so devbox exec <name> -- command -v go fails with exec: "command": executable file not found in $PATH - which looks like the tool is missing when it is not. If you want command -v, give it a shell:

devbox exec <name> -- bash -lc 'command -v go >/dev/null && go version || echo MISSING_GO'

Important For every CLI tool the script uses, check if it exists first and install only if missing. Use platform-appropriate installers and artifacts. For example, on Linux/amd64:

devbox exec <name> -- apt-get install -y --no-install-recommends \
    git ca-certificates curl tar build-essential
devbox exec <name> -- curl -fsSL -o /tmp/go.tgz https://go.dev/dl/go<version>.linux-amd64.tar.gz
devbox exec <name> -- tar -xzf /tmp/go.tgz -C /usr/local

3. Run commands with devbox exec

Important Run devbox exec invocations against different devboxes in parallel. Do NOT serialize independent work.

devbox exec <name> -- <cmd> <args...>

devbox ssh <name> remains available for interactive shells and TTY-dependent workflows. Prefer devbox exec for non-interactive commands.

Alternative: native ssh If a workflow uses a tool, which relies on standard SSH tooling (scp, rsync, etc.), run devbox configure-ssh <name> once to write an entry into your ~/.ssh/config. After that, the devbox is reachable as <name>.devbox.namespace. Prefer devbox exec for commands; reach for native ssh when you specifically need standard tooling.

Important (exec does not give you a shell - but you can ask for one) devbox exec runs the command directly, so shell operators such as &&, ;, redirections and variable expansion are NOT interpreted by default. devbox exec <name> -- cd /src && make will not do what you want.

Two ways to get multi-step work done, and they trade off differently:

Short multi-step work - invoke a shell explicitly. One round trip, and a mistake costs one re-run:

devbox exec <name> -- bash -lc 'cd /workspaces/<repo-name> && go build ./... && go test ./...'

Longer or reusable logic - upload a script. Worth the extra round trips (upload, chmod, run) when the logic is long, needs careful quoting, or you want the exact steps recorded on the machine. Note that a bug in the script costs a re-upload AND a re-chmod, so prefer bash -lc while you are still iterating:

devbox upload <name> /tmp/script.sh /tmp/script.sh
devbox exec   <name> -- chmod +x /tmp/script.sh
devbox exec   <name> -- bash /tmp/script.sh

Quoting is the thing to watch with bash -lc: wrap the whole program in single quotes and let the remote shell expand $VARS, or they will expand locally before the command ever leaves your machine.

Detached commands and retained logs

Use devbox exec -d for long-running commands. It starts the command, returns immediately, and prints an exec ID. Use that ID to stream retained and live output with devbox logs. devbox logs list shows previous executions and their status.

devbox exec -d <name> -- <cmd> <args...>
# exec_<id>

devbox logs <name> <exec-id>
devbox logs list <name>

Important Preflight every required CLI before set -e so a missing tool exits with a clear message rather than an opaque failure: command -v <tool> >/dev/null || { echo "MISSING_TOOL: <tool>"; exit 127; }. On exit 127 or a MISSING_TOOL marker, install the dependency and re-run - do NOT report the run as failed.

Important Write all script output to a remote log file and return only a tiny summary by default.

cat > /tmp/run.sh <<'EOF'
#!/bin/bash
export PATH=/usr/local/go/bin:$PATH
command -v <tool> >/dev/null || { echo "MISSING_TOOL: <tool>"; exit 127; }
set -euo pipefail
mkdir -p /workspaces/tmp
export TMPDIR=/workspaces/tmp
LOG=/workspaces/run.log

cd /workspaces/<repo-name>
status=0
<workload-cmd> >"$LOG" 2>&1 || status=$?

echo "exit=$status log=$LOG bytes=$(wc -c <"$LOG")"
exit "$status"
EOF
devbox upload <name> /tmp/run.sh /tmp/run.sh
devbox exec   <name> -- chmod +x /tmp/run.sh
devbox exec   <name> -- bash /tmp/run.sh

# Pull the full log only if the summary shows a non-zero exit.
# devbox download <name> /workspaces/run.log /tmp/run.log

Note Capture the workload's exit code (status=0; cmd || status=$?) before printing the summary. Use devbox download only for devboxes whose summary reported a non-zero exit, unless specified otherwise.

For test-runner-specific guidance (sharding, parallel runs, the hydrate-and-test recipe, result reporting), see references/run-tests.md.

4. Lifecycle

devbox create --name <name> [--platform <linux/amd64|macos/arm64>] --image <image> --size <size> [--ephemeral] [--checkout <repo-url>] --purpose "<purpose>"
devbox exec <name> -- <cmd>
devbox exec -d <name> -- <cmd>                                     # run in the background and print an exec ID
devbox logs <name> <exec-id>                                       # stream retained and live output
devbox logs list <name>                                            # inspect retained executions and status
devbox ssh <name>                                                  # open an interactive shell; prefer exec for commands
devbox list [-o json]                                              # inspect running devboxes
devbox upload <name> <local-path> <remote-path>                    # transfer files to devbox
devbox download <name> <remote-path> <local-path>                  # retrieve files from devbox
devbox configure-ssh <name>                                        # add to ~/.ssh/config for native ssh/scp/rsync
devbox url access <name> [--mode <private|workspace>] -o json      # inspect or change devbox.so access
devbox url expose <name> --port <port> [--name <purpose>] -o json  # create a persistent devbox.so URL
devbox url get <name> (--port <port> | --name <purpose>) -o json   # retrieve an exposed URL
devbox url list <name> -o json                                     # list exposed URLs
devbox url unexpose <name> (--port <port> | --name <purpose>)      # remove an exposed URL
devbox port-forward <name> --ports <local:remote,...>              # forward devbox ports to localhost (e.g. a dev server or DB)
devbox shutdown <name>                                             # stop a devbox without destroying it
devbox expire <name> --force                                       # tear down a devbox

Keep a devbox alive through long-running work

A devbox with an idle timeout auto-stops when nothing on it looks busy. The Devbox agent treats a running devbox exec or SSH command as busy for its whole runtime. Creating or connecting to a terminal session and receiving HTTP responses through an exposed devbox.so URL count only as recent activity for a bounded period. Work not covered by an active exec or SSH command needs an explicit marker:

devbox exec <name> -- boxctl task mark <task>          # hold the devbox busy
devbox exec <name> -- boxctl task clear-mark <task>    # release it
devbox exec <name> -- boxctl task list                 # what is currently holding it

boxctl is pre-installed on every devbox, so this works from your machine without installing anything on the box. From inside a devbox, drop the devbox exec <name> -- prefix.

Mark anything that will run unattended outside an active devbox exec or SSH command: an overnight job started in a terminal session, a migration, a queue worker, or a dev server nobody is hitting yet. A foreground or detached devbox exec needs no marker while its command is still running. Without either form of activity the machine can stop mid-task and take the work with it.

Important Markers have no TTL and nothing cleans them up. One left behind holds the devbox awake and billed indefinitely, defeating the idle timeout entirely - so always pair a mark with the matching clear-mark, on failure paths too. Both commands are idempotent and exit 0 when the marker already exists or is already gone, which makes them safe in traps and retries. See references/inside-a-devbox.md for the trap pattern.

A marker only holds off the idle timeout. devbox shutdown and devbox expire still stop the machine regardless.

Important These devbox url commands require devbox CLI v0.0.182 or newer. Check with devbox version; if the installed version is older, run devbox update before using them.

Choose who can open devbox.so URLs Each devbox has one access mode shared by all of its exposed URLs:

• private - only the devbox owner.
• workspace - members of the same Namespace workspace.

The URL mode initially follows the general devbox access mode, but is configured separately afterward. Changing URL access does not grant or remove SSH, exec, or file access to the devbox.

Inspect or change the Devbox-wide URL mode independently:

devbox url access my-box -o json
devbox url access my-box --mode workspace -o json

Important devbox url expose ... --access <mode> changes this same Devbox-wide setting while exposing a port. It affects every existing and future exposed URL on that devbox; it does not set access only for the new URL. Per-URL access modes are not supported. If URLs need different audiences, ask before changing access or use a separate devbox.

Share web work with teammates When the user intends work running in a devbox to be reviewable by teammates, expose its port with --access workspace and return the resulting devbox.so URL. The URL remains available after the command exits. Pass --name when the port's purpose is known so the URL is easier to retrieve and remove later. Prefer JSON output for agent workflows.

Important (how to verify an exposed URL - and how not to) A successful devbox url expose that returns a URL IS your confirmation. Stop there and hand the URL to the user.

Do NOT try to fetch the URL to "prove" it works. devbox.so URLs are gated on a signed-in user session, which you do not have and cannot mint - curl will return 401 unauthorized, and no API or ID token will change that. That 401 is not a failure and not evidence of anything you need. Chasing a 200 is the single most expensive dead end in this skill: it burns minutes on token endpoints and SSH tunnels to re-confirm a fact the CLI already gave you.

If you want to confirm the server is actually up before sharing, check it from inside the devbox, where no gateway auth applies:

devbox exec <name> -- curl -sS -o /dev/null -w '%{http_code}\n' http://127.0.0.1:<port>/

Bind servers to 0.0.0.0, not 127.0.0.1, or the devbox.so proxy cannot reach them.

devbox url expose my-box --port 3000 --name web --access workspace -o json

If the user has not asked to share, omit --access and preserve the current mode. When sharing intent is unclear, inspect the mode and do not broaden access. If the user explicitly requires every exposed URL on the devbox to be owner-only, set private before exposing the new port:

devbox url access my-box --mode private -o json
devbox url expose my-box --port 3000 --name web -o json

For local-only access or non-web services, use devbox port-forward. Example: devbox port-forward my-box --ports 3000:3000,5432:5432 maps ports to localhost:3000 and localhost:5432. The command blocks the terminal until Ctrl+C, so run it in a dedicated terminal or as a background process.

Whether and when to call devbox expire depends on the caller's intent - the test-suite workflow in references/run-tests.md tears down at the end; other use cases may keep the devbox alive.

Important On failure caused by missing dependencies, install them and retry the failing command. Do the same for any uploaded scripts.