cmux-cloud
Use cmux Cloud to inspect or run work on an existing remote machine, monitor durable jobs, transfer task files, or balance work across available compute.
- License 许可
- MIT
- 最近更新
- 2026-11-11
- Stars
- ★ 6
- Forks
- 0
01 前言 Front Matter
- name
- cmux-cloud
- description
- Use cmux Cloud to inspect or run work on an existing remote machine, monitor durable jobs, transfer task files, or balance work across available compute.
03 SKILL.md 正文
cmux Cloud machines
Use the installed cmux CLI and its current vendor guidance (cmux cloud --skill)
for exact syntax. Commands and limits can change; check the relevant
cmux cloud <verb> --help whenever a flag, limit, or exit behavior matters.
Inspect before routing
Start with read-only machine inventory, authentication status, and routing information. Use the vendor's inventory command as the authority for which machines exist and their current capacity. Prefer an already-running suitable machine and pin its identifier on remote execution. A router may provision a machine for an unrecognized path, so inspect its plan before running it.
Do not provision machines, increase a paid plan, or trigger other billable resources unless the user authorized that action. Pause or resume a machine only when the user authorized it and its owner and task impact are clear. Deleting a machine is destructive and always requires explicit user approval, even when the task created it. Never delete a machine outside the explicitly authorized target and scope. Pausing may stop compute charges while retaining its disk; confirm current semantics with vendor help.
Run and observe work
Use a short remote-exec command for a bounded probe. For scripts whose next steps depend on their exit status, use the command that propagates the remote exit code; some versions of a probe command report only a generic local failure. Verify these behaviors in the installed CLI before relying on them.
For long work, start a durable remote terminal or job. Record its identifier, wait for completion, and read its output before reporting a result. A finished wait means the process ended; it does not mean the command passed. Keep the remote work directory and transferred files scoped to the task.
Before transferring files, check current size limits and default excludes. Avoid copying credentials, browser profiles, or unrelated repository state. Pulling a directory may require archiving it on the remote machine first; use the documented transfer mechanism and verify the resulting file.
Browser and credentials
Use only the task-authorized isolated browser and authentication route. Keep personal browser profiles separate from automation profiles. Some provider consent or multi-factor steps require a human gesture; report that exact dependency rather than automating around it.
Never export, copy, print, or place credentials in task files or command-line arguments. Use an existing authorized credential store or supported identity flow. Creating or rotating credentials and granting new access require the appropriate authorization.
Balance work across machines
Compare current available CPU and memory on the candidate hosts, then route the next independent heavy job to the host with more capacity. Re-measure before starting another heavy batch. Keep shared-resource and simulator exclusivity separate from host scheduling: a capacity note does not reserve an account, service, campaign, or device. Do not start competing builds on the same simulator device and app bundle without coordinating its active operator.
Prefer fewer active machines when the current work fits, because provisioned machines may continue consuming resources while idle. Check live pool limits before starting work and pause task-owned idle machines when authorized and safe.
Verify before reporting
- Bind each result to the machine, command, and process/job identifier used.
- Read both the completion status and captured output.
- Report the actual remote exit status and any missing verification layer.
- Recheck inventory before claiming a machine was paused, removed, or available.
Portable skill contract
This seven-element specification covers the three skill packages in this change.
- Outcome: The catalog gains reusable design-comparison, interface-exploration, and cmux Cloud skills.
- Non-goals: No application runtime, live machine, account, profile, or credential is changed; no scripts or unrelated packages are added.
- Acceptance: All three packages have valid frontmatter, are discoverable, omit personal host data, and retain bounded delegation and resource rules.
- User story: As a catalog user, I can find these skills when their tasks apply and use them without inheriting private host configuration.
- Proof plan: Record source hashes and dependencies, validate each package, verify read-only installer discovery, scan for personal data, and inspect the complete base-to-head diff.
- Mock exemption: This instruction-only change does not alter product UI, so screenshots and runtime video do not apply. The design-fidelity skill still requires the approved mock when used for UI implementation.
- Risks: Vendor command behavior can drift, and local discovery checks do not prove installation or runtime success on another host.
04 同仓库更多 Skills
instinct
Use when the user invokes /instinct or /in, or asks to "message Instinct", "ask Instinct", "tell Instinct", "check Instinct", or read replies from the configured peer assistant. iMessage first, WhatsApp Web as fallback.
hr
Compatibility alias for history-resume; produce the same fresh-state, provenance-aware resume briefing.
history-resume
Resume briefing for the current repository branch or worktree using fresh state and admitted history sources.
grok
Alias for the canonical grokbot browser conversation workflow; /grok talks to the Grok bot the same way /dot talks to Dot.
codexa-bench
Compare `codexa` and normal `codex` on a fair split of comparable work by default; use `--strict` for duplicate runs of the exact same task.
frontend-design
Create distinctive, production-grade frontend interfaces with high design quality. Use this skill when the user asks to build web components, pages, artifacts, posters, or applications (examples include websites, landing pages, dashboards, React components, HTML/CSS layouts, or when styling/beautifying any web UI). Generates creative, polished code and UI design that avoids generic AI aesthetics.