remote-coding-agent-operations

# Remote coding agent operations Raw SSH/tmux fallback for `external-agent-terminal-multiplexer-operations`. This skill owns remote transport mechanics, not generic external-control semantics, production run queues, completion-doorbell policy, or multi-slice scheduling. ## Remote identity before control Before any non-read remote action, reconcile the SSH host/user, absolute remote workspace, tmux pane cwd, repository identity, and session ownership. A local sshfs path is read evidence, not proof of the remote control target. If these facts disagree or ownership is unclear, remain read-only. If the verified remote root is dirty, create an isolated worktree only with explicit authorization. ## Remote loop 1. **Probe** — Through `zsh -lic`, verify repository/branch/status, Claude binary/version/auth, and provider/model environment. - Complete when the same login-shell path used by tmux can start the agent. 2. **Launch** — Create or select an explicitly named tmux session in the remote workspace. Preserve sessions that may be user-controlled. - Complete when pane cwd and process command line match the intended target. 3. **Submit** — Stream the prompt through SSH stdin into a tmux buffer, paste it, send Enter, then capture the pane. - Complete when the pane proves the prompt entered processing/history; visible unsubmitted text is not enough. 4. **Inspect** — Treat watcher/pane output as hints. Inspect actual git state, scope, tests, compile/static checks, and smoke...