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.
if [ -n "${CLAUDE_PLUGIN_ROOT:-}" ]; then
SCRIPT="$CLAUDE_PLUGIN_ROOT/skills/daytona-companion/scripts/daytona-manager.mjs"
else
SCRIPT="$HOME/.agents/skills/daytona-companion/scripts/daytona-manager.mjs"
fi
Typical workflows
Create a sandbox, push code, run tests, pull results
The most common end-to-end flow:
# 1. Create (or reconnect to) a sandbox
node "$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. Pull stdout/stderr/exit-code artifacts back
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 "$WORK_DIR" --port 3000
Git-based sync (alternative to bundle push/pull)
When you need commit history preserved:
node "$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 |
| 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 tosmallunless the workload needs more.
More from atomlink-ye/my-claude-plugins
debug-workflow
Unified debug workflow skill for real-world issue reproduction, especially browser/UI/Web3 problems that need Playwright, shared Chrome/CDP, or agent-wallet. Use this whenever the user wants to confirm whether a workflow can reproduce real issues from walkthrough docs, issue lists, QA notes, or live bug reports. Start with the debug index, route to the right lane, and prefer deterministic local reproduction over fragile preview environments.
11task-iteration
Orchestrate end-to-end feature implementation from an exec-plan document. Use this whenever the user wants a structured Plan → Generate → Evaluate loop for a feature. It keeps planning and acceptance in Claude, pushes bounded coding/test execution to OpenCode companion sessions, reuses warm sessions across fix rounds, and gates completion on independent evaluation plus an advisory review.
11paseo-companion
Paseo CLI companion. Load whenever the user mentions paseo, paseo agents, paseo run/send/wait/logs/attach/ls, agent IDs, sending follow-ups to a running agent, loops, schedules, terminals, worktrees, chat, permits, daemon operations, or host/port targeting.
11google-workspace
Read, create, and edit Google Workspace documents using the `gws` CLI. Use this skill whenever the user mentions Google Docs, shares a docs.google.com URL, references a Google Doc ID, or wants to read/write/create Google documents. Also trigger on 'GDoc', 'gdoc', 'Google Doc', 'gws docs', or any Google Workspace document operation. Will expand to cover Drive, Sheets, Gmail, Calendar, and other GWS services.
11opencode-orchestrator
Aggressively load this skill BEFORE any delegation decision — explore, plan, implement, review, adversarial second-opinion, bounded investigation, or any medium/heavy coding request. Also triggers for manager-vs-executor routing, OpenCode session reuse, multi-step coding orchestration, and work that likely exceeds 10 tool calls, 5 files, or repeated read/edit/test loops. When the live snapshot below shows OpenCode serve is reachable, this skill's routing takes precedence over the caller's generic delegation defaults across every delegated action, not only bounded coding execution. If serve is unreachable, fall back silently to the caller's defaults.
11opencode-companion
OpenCode runtime companion. Load for OpenCode task/review/status/serve/rescue requests, session IDs, timeout recovery, attach/resume decisions, background jobs, and result forwarding.
11