AGENT SKILL

remote-machine-access

Use for native SSH work across Macs and remote hosts.

License 许可
MIT
最近更新
2026-08-25
Stars
★ 1
Forks
0

01 前言 Front Matter

name
remote-machine-access
description
Use for native SSH work across Macs and remote hosts.

02 资源文件 1

03 SKILL.md 正文

Remote Machine Access

Operate a user-owned machine through its existing SSH trust path while keeping the requested effect native to the target application. Prefer direct host and application interfaces over GUI imitation.

Scope

Use this skill when work crosses machine boundaries: running a command on another host, opening live remote files in a local editor, checking a service on a Mac from Linux, or coordinating an application on one machine with files on another.

For native desktop clicking, accessibility trees, or pixel input, use computer-use. This skill owns the SSH topology, native CLI route, and end-to-end verification.

Workflow

  1. Name both legs. Identify the machine executing the application and the machine owning the data. State the requested effect in those terms before acting.
  2. Resolve existing trust. Inspect the configured SSH alias rather than inventing an address, user, key, or port. Use ssh -G <alias> for effective non-secret routing fields and ssh -o BatchMode=yes -o ConnectTimeout=<seconds> <alias> <probe> for reachability.
  3. Discover the native interface. On the application host, locate the app-provided CLI or automation interface and read its live --help. Prefer that interface over simulated typing or clicks.
  4. Preflight from the application host. If the application will reach a second machine, verify that exact second hop and every requested target path from the application host. A first-hop success proves only the first hop.
  5. Apply one minimal action. Use live remote paths when the user asked to edit live state. Do not create local copies unless copying is the requested outcome.
  6. Read back the effect. Verify the exact application state: window or tab names, workspace identity, remote connection process, or another native state query. A zero exit code from a launcher proves invocation, not that the user can see the requested result.
  7. Report topology and boundary. Say which machine owns the files, which machine runs the app, whether changes are live, and what verification observed.

Guardrails

  • Reuse configured aliases and keys; never print private key material, tokens, passwords, or full secret-bearing configuration.
  • Keep SSH probes non-interactive. Stop rather than guessing credentials or opening an authentication prompt.
  • Do not mutate SSH configuration when an existing route works.
  • Do not bring a desktop window to the foreground unless the user asked to see it or the native interface requires it and the effect is expected.
  • Verify file reachability without dumping private file contents unless their contents are needed for the task.
  • Keep editor workspaces on one remote authority. Mixed local and remote arguments can produce ambiguous workspace behavior.

Completion criteria

The task is complete only when:

  • every requested target resolved on the correct machine;
  • the native application accepted the action;
  • an independent readback shows the requested files, window, workspace, or service state;
  • the report distinguishes live remote state from copied local state.

References

  • For opening several live server files in one Zed window on macOS, including the validated sequential CLI pattern and verification probes, read references/zed-ssh-workspaces.md.

04 同仓库更多 Skills