phnx-labs Cloud Skill
Dispatch a task to a cloud agent instead of running it locally. Where agents run executes on this machine, agents cloud run hands the work to a remote backend that clones the repo, runs the agent, and (for Rush) opens a PR. Use it to fan out many tasks past local capacity, or to run work that needs a clean, pre-built environment.
Providers
| Provider | What it is | Dispatch needs | PR? | Multi-repo |
|---|---|---|---|---|
rush (default) |
Rush Cloud — clones a GitHub repo + branch, runs the agent, auto-commits and opens a PR | --repo owner/repo |
yes | yes |
codex |
Codex Cloud — runs in a pre-built Codex environment (repos fixed at env-creation time) | --env <id> |
per env | no |
factory |
Factory pods — Droid + computer-use targets | --computer <name> |
n/a | no |
Check what's signed in before dispatching:
agents cloud providers # human-readable
agents cloud providers --json # {id, name, available, default}
available: false means that backend isn't signed in on this machine — pick another or authenticate first.
Dispatch
agents cloud run "<prompt>" --provider <id> [options]
The prompt can be inline (positional), via -p/--prompt, or a path to a file whose contents become the prompt.
| Flag | Purpose |
|---|---|
--provider <id> |
rush, codex, or factory (defaults to the configured default) |
--agent <name> |
Agent to run: claude, codex, droid |
--repo <owner/repo> |
GitHub repo. Repeatable for multi-repo (Rush only) |
--branch <name> |
Target git branch |
-p, --prompt <text> |
Inline prompt (alternative to the positional arg) |
--env <id> |
Codex Cloud environment ID (required for --provider codex) |
--computer <name> |
Factory/Droid computer target |
--model <model> |
Model override |
--mode <mode> |
Execution mode (plan, edit, full) |
--timeout <dur> |
Kill after a duration (30m, 2h) |
-b, --balanced |
Route a factory run across all healthy accounts |
--strategy balanced |
Account selection strategy (factory) |
--upload-account-tokens |
Consent to upload Claude OAuth creds to Rush Cloud (first dispatch) |
--json |
Structured JSON output |
--no-follow |
Dispatch and exit without streaming (returns the task id) |
Rush Cloud
Runs against a GitHub repo and opens a PR with the result.
# Dispatch and stream the output
agents cloud run "fix the flaky e2e in apps/web/tests/checkout.spec.ts" \
--provider rush --repo acme/example --branch main
# Fire-and-forget — returns the task id, no streaming
agents cloud run "bump tailwind to v4 and fix the breaks" \
--provider rush --repo acme/example --no-follow
# Multi-repo: cloned into /workspace/<owner>/<name>/ each
agents cloud run "rename POST /v1/charge -> /v2/charge across server + extension" \
--provider rush --repo acme/example --repo acme/example-extension
Token-upload consent. On first dispatch (or after your Claude token rotates) Rush Cloud asks to sync your Claude OAuth credentials (accessToken + refreshToken) to its API so the pods can act as your Anthropic account:
agents cloud run "..." --provider rush --repo acme/example --upload-account-tokens
# or: AGENTS_RUSH_UPLOAD_TOKENS=1 agents cloud run ...
Consent is recorded at ~/.agents/.cache/cloud/rush-consent.json (delete it to revoke). This is a credential-upload decision — get explicit OK from the user before passing the flag.
Codex Cloud
Runs in a pre-built environment. The provider wraps the real codex CLI (codex cloud exec/status/list).
agents cloud run "add pytest fixtures for the new billing module" \
--provider codex --agent codex --env <env_id> --timeout 30m
--env <id>is required. Set a default in~/.agents/agents.yamlundercloud.providers.codex.envto omit it.- Environment IDs are not listable from the CLI. Find them in the Codex web settings — open
https://chatgpt.com/codex/cloud/settings/environments, click an environment, and read the id from the URL (.../environment/<id>). The repos a Codex task can touch are fixed at env-creation time, so--repomulti-dispatch is rejected — use--provider rushfor that. - Codex Cloud has no live stream;
logs/statuspoll until the task reaches a terminal state. cancelandmessageare not supported for Codex tasks via the CLI.
Factory pods
Droid + computer-use targets.
agents cloud run "QA the new onboarding flow end-to-end" \
--provider factory --computer linux-vm-1 --agent droid
Manage tasks
agents cloud list # every task you've dispatched, most recent first
agents cloud list --json # machine-readable
agents cloud status <id> # task detail + latest status
agents cloud logs <id> # live-tail output (polls for codex)
agents cloud message <id> "<text>" # follow-up while in needs-review (rush/factory)
agents cloud cancel <id> # cancel a running task (not supported for codex)
Set a default provider
Add to ~/.agents/agents.yaml:
cloud:
defaultProvider: rush
providers:
codex:
env: <env_id> # so codex dispatch needs no --env
Then agents cloud run "refactor auth module" --repo owner/repo uses the default.
When to use cloud vs run vs teams
run— one agent, on this machine, now. Fast iteration, local files.cloud— offload to a remote backend; scale past local capacity; needs a repo (rush) or env (codex). Async by default with--no-follow.teams— multiple local agents collaborating on shared work with boundary contracts.
For local execution see the run skill; for parallel local work see teams.
For everything else, run agents cloud --help or agents cloud run --help.