daytona-companion

Daytona Companion

Daytona provides remote cloud sandboxes for running code, tests, and services in isolation. This skill manages the full sandbox lifecycle — create, push code, execute, pull results, preview, and tear down.

This skill is a runtime adapter. It documents how to drive Daytona sandboxes; it does not decide whether a given task should run remotely instead of locally. That decision belongs to the orchestration layer (team-lead-orchestration) and any local routing profile that captures host constraints such as memory pressure, secrets, or quota.

Set the manager script path from this skill path:

SCRIPT="${SKILL_ROOT}/scripts/daytona-manager.mjs"

${SKILL_ROOT} is the path to this daytona-companion skill directory. Set it from the loaded skill path; do not prepend another install root. From the marketplace root, the direct manager entrypoint is skills/daytona-companion/scripts/daytona-manager.mjs.

Typical workflows

Create a sandbox, push code, run tests, pull results

The most common end-to-end flow:

# 1. Create (or reconnect to) a sandbox
node "$SCRIPT" up --directory "$WORK_DIR" --task-id "$TASK_ID" --class small

# 2. Push local code to the sandbox. Relative remote paths resolve under the sandbox user's $HOME.
node "$SCRIPT" push --directory "$WORK_DIR" --path "$WORK_DIR" --remote-path "workspace/$TASK_ID"

# 3. Execute tests remotely
node "$SCRIPT" exec --directory "$WORK_DIR" --cwd "workspace/$TASK_ID" -- pnpm test

# 4. Optional: pull stdout/stderr/exit-code artifacts for later inspection
node "$SCRIPT" pull --directory "$WORK_DIR" --output "$WORK_DIR/artifacts/daytona/$TASK_ID"

# 5. Tear down when done
node "$SCRIPT" down --directory "$WORK_DIR"

Adopt an existing sandbox

If a sandbox was created outside this script (e.g. via Daytona CLI directly):

node "$SCRIPT" adopt --directory "$WORK_DIR" --task-id "$TASK_ID" --sandbox-id "$SANDBOX_ID" --remote-path "/workspace/$TASK_ID"

Check sandbox status

node "$SCRIPT" status --directory "$WORK_DIR"            # local cached state
node "$SCRIPT" status --directory "$WORK_DIR" --refresh   # force network check

Preview a running service

node "$SCRIPT" preview --directory "$WORK_DIR" --port 3000

Run a remote agent runtime inside one sandbox

For a long implementation iteration, use Daytona for sandbox lifecycle and preview, then launch agents through the remote Paseo daemon from the local machine:

# 1. Create/reconnect, then pin the remote workspace path used by this image.
node "$SCRIPT" up --directory "$WORK_DIR" --task-id "$TASK_ID" --class large
node "$SCRIPT" adopt --directory "$WORK_DIR" --task-id "$TASK_ID" --sandbox-id "$SANDBOX_ID" --remote-path "/workspace"

# 2. Bootstrap remote Paseo/OpenCode inside the sandbox.
node "$SCRIPT" exec --directory "$WORK_DIR" --cwd "/workspace" -- \
  sh -lc 'PASEO_DAEMON_LISTEN=0.0.0.0:6767 paseo daemon start --listen 0.0.0.0:6767 --hostnames true --no-relay'

# 3. Expose Paseo and use the returned host:port with local paseo commands.
node "$SCRIPT" preview --directory "$WORK_DIR" --port 6767

Do not run daytona exec ... paseo run ... to spawn remote agents. exec is for sandbox bootstrap, direct commands, and artifact probes; remote agents should be addressable from local via paseo run --host "$REMOTE_PASEO_HOST" ....

Git-based sync (alternative to bundle push/pull)

When you need commit history preserved:

node "$SCRIPT" push --directory "$WORK_DIR" --mode git --branch "daytona/$TASK_ID"
# ... work remotely ...
node "$SCRIPT" pull --directory "$WORK_DIR" --mode git --branch "daytona/$TASK_ID"

Git mode transfers committed history only — uncommitted files are not included.

Run a full smoke test

Validates the entire lifecycle in one command:

node "$SCRIPT" smoke-test --class small --include-git --include-preview

When to read references

You need to...	Read
Understand sandbox create/adopt/exec/delete in detail	references/sandbox-lifecycle.md
Push/pull files, choose bundle vs. git mode, handle artifacts	references/artifact-workflows.md
Bootstrap a remote Paseo/OpenCode runtime in one warm sandbox	references/remote-agent-runtime.md
Find where state is stored, configure env vars and tokens	references/state-and-secrets.md
Debug stale status, command failures, secret leaks	references/troubleshooting.md

Non-negotiables

• State is global. Stored under ~/.daytona/claude-code/projects/, keyed by project directory. Never store state in the project tree or skill source.
• Never print secrets. Env values and API tokens must not appear in output.
• Quote shell arguments. Remote commands go after --: exec --directory "$WORK_DIR" -- pnpm test.
• Sandbox classes: small|medium|large — default to small unless the workload needs more.
• Agent dispatch stays local-addressable. Use Daytona preview plus local paseo run --host "$REMOTE_PASEO_HOST" for remote agent trees; don't hide agents behind daytona exec.