Index of every page on this site: llms.txt
All skills
scode avatar

/scode-ssh-delegate

@03a9f7f
Run it yourself

Make Ubuntu 26.04 LTS hosts, Fly.io sprites, Tensorlake sandboxes, and boat sandboxes available as remote workers. Use when the user invokes scode-ssh-delegate or $scode-ssh-delegate followed by one or more hostnames, including terse forms such as "$scode-ssh-delegate foo" or "$scode-ssh-delegate foo, bar, baz"; when it is followed by named sprites or sandboxes, such as "$scode-ssh-delegate sprites foo1 foo2", "$scode-ssh-delegate sbx foo1 foo2", or "use the sprites foo1 foo2"; or when it grants a sprite, sandbox, or boat budget, such as "$scode-ssh-delegate sprites:3", "$scode-ssh-delegate sbx:3", "$scode-ssh-delegate boat 10", "$scode-ssh-delegate boat:10", or "it's okay to use up to 3 sprites". Immediately identify each named host's, sprite's, or sandbox's OS, CPU count, and memory, then retain the supported workers (or the budgets) for possible use by another workflow. This invocation does not request delegation. If retained context after compaction or resume says workers or a budget were registered, re-read this skill and its companion files before touching them.

  • 7 files
  • 68.9 KB
  • Updated last week
  • GitHub
Use this Skill: https://skilld.dev/gh/scode/dotfiles/scode-ssh-delegate

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

using-workers.md

≈1.6k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Using an accepted worker

Read this file before the first command, transfer, install, or credential touches a registered worker, whoever asked for that, and again before the first use after compaction or resume. Registration (in SKILL.md, sprites.md, tensorlake.md, and boat.md) only declares capacity; everything here is about what may happen on a worker once work is sent to it: what gets installed, which credentials may move, and how source and results cross the boundary. It applies to SSH hosts, sprites, tensorlake sandboxes, and boat sandboxes alike; the platform-specific mechanics (bootstrap, transport, lifecycle) are in sprites.md, tensorlake.md, and boat.md and take precedence where they differ.

Base packages (all kinds)

Every worker — SSH host, sprite, tensorlake sandbox, or boat sandbox — gets these before anything else is installed on it:

sudo -n apt-get update
sudo -n env DEBIAN_FRONTEND=noninteractive apt-get install -y \
    build-essential ca-certificates curl git rsync

Worker contract for Ubuntu SSH hosts

An accepted SSH worker is either a fresh base Ubuntu 26.04 LTS installation or a machine prepared during an earlier use of this skill. No other Ubuntu release or distribution is supported.

Only after a host has been selected for a task, install whichever missing CLIs that work requires. Confirm the worker's architecture meets the selected CLI's current requirements. The Codex installer is https://chatgpt.com/codex/install.sh; run the downloaded script with CODEX_NON_INTERACTIVE=1 sh. The Claude Code installer is https://claude.ai/install.sh; run it with Bash, not sh. Require HTTPS for the initial URL and redirects, download each installer completely before executing it, and verify each installed command. The vendor-controlled installers are an explicit upstream trust boundary.

Both native installers place their command in ~/.local/bin. Add that directory to PATH explicitly in non-login SSH commands before deciding that a CLI is absent.

Sprites use the same base packages but a different bootstrap, because their image ships the agent CLIs and toolchains preinstalled and stale; see "Bootstrapping a sprite" in sprites.md. Tensorlake sandboxes use the same base packages and the same vendor installers as SSH hosts (their image ships no agent CLIs at all); see "Bootstrapping a sandbox" in tensorlake.md. Boat sandboxes use the same base packages, and their image ships current agent CLIs, so the vendor installers run only when a selected CLI is missing or broken; see "Bootstrapping a sandbox" in boat.md.

Credentials

Only when another workflow has selected a worker for actual use, inspect whether its required CLI and authentication are ready. For Codex, copying ~/.codex/auth.json to the same path on a headless worker is an officially supported fallback when the controlling host actually uses file-backed credentials. Otherwise use Codex device authentication or another supported login path. For Claude Code automation with a Pro, Max, Team, or Enterprise subscription, prefer a token produced by claude setup-token and pass it as CLAUDE_CODE_OAUTH_TOKEN; do not assume its local credential cache is portable. A Console user may instead pass ANTHROPIC_API_KEY to the selected Claude invocation. These are the only agent credentials this skill permits transferring. On a sprite, tensorlake sandbox, or boat sandbox, see "Credentials on a sprite" in sprites.md or "Credentials on a sandbox" in tensorlake.md or boat.md for how they are passed and cleaned up.

Announce the transfer because these are reusable secrets. Copy only the credential needed by the selected provider, keep any persisted secret readable only by the remote user, and verify authentication. Do not overwrite an existing credential merely because an authentication check returned an unexpected error.

Apart from an agent credential intentionally installed as above, treat the worker as having no GitHub, cloud, or package-registry credentials and no access to private repository remotes. Never forward the SSH agent or copy other home-directory state. Project-specific public prerequisites may be installed when the selected task needs them; never add private registry configuration.

Running work

The orchestrating workflow chooses the host, provider, exact local source state, transfer mechanism, remote working directory, prompt, commands, and result format. Whatever mechanism it chooses runs over the transport recorded for that worker during registration. A user-specified commit, revision, or working-copy snapshot remains binding. Resolve private refs locally and treat the local checkout as the source of truth; the remote worker must not contact a private repository remote to clone, fetch, pull, or push.

The orchestrating workflow decides how a worker is launched; galaxy-brain's launch rules come from scode-harness-shellout and have been verified locally only. The two proven remote entrypoints are the ones below.

If the orchestrating workflow chooses unrestricted execution, the proven entrypoints are codex exec --yolo and claude -p --dangerously-skip-permissions. These modes give the agent full access to the remote Unix account. They are worker capabilities, not a request from this skill to run either agent. Use a disposable remote directory, and remember that the remote agent has none of the controlling session's conversation.

Choose what to transfer for the task at hand. Over SSH that is normally rsync with the recorded transport as -e; on a sprite it is sprite file push as described in sprites.md; on a tensorlake sandbox it is tl sbx cp with tarballs for trees as described in tensorlake.md; on a boat sandbox it is boat scp as described in boat.md. Do not transfer credentials as source material; if that conflicts with a requested working-copy snapshot, stop for direction. Do not blindly copy repository metadata, ignored files, or unrelated local state.

In the return direction, treat command output as returned data and keep it bounded. Do not mirror or reverse-rsync the remote workspace. Retrieve only a small, explicitly selected set of patches, changed files, reports, or logs into local staging outside the checkout, then inspect them before applying anything. Remove disposable source and prompts from the worker when they are no longer needed.

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No third-party reports yet.

Signed by skilld at 03a9f7f. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated last week

README badge

README badge for scode/dotfiles/scode-ssh-delegate