ssh-experiment

SSH Experiment Execution

Boundary

Read experiment first. Use this Skill only after Research Assistant configured the project's single easyresearch.ssh object and verified the SSHFS mapping. Experiment may test and run the connection through ssh-bash; it must not reconfigure host, port, username, authentication, credential files, or mounts.

Never read credential files with Agent-visible tools. ssh-bash reads their contents inside the daemon and returns only remote command output.

Freshness Guard

Before the first edit or launch, after reconnect, and after a long idle period:

1. call ssh-bash with action: "test";
2. confirm exact-cwd experiment_ssh/ remains the configured mount;
3. repeat the owner-only marker round trip through ssh-bash run and the local mount; and
4. query current compute availability.

If any check fails, preserve work and return blocked. Research Assistant must repair/reconfigure the connection and mount before this child continues.

Editing And Artifact Paths

Edit through the verified local mount using Read/Edit/Write:

experiment_ssh/src/
experiment_ssh/external/
experiment_ssh/datasets/
experiment_ssh/outputs/
experiment_ssh/results/
experiment_ssh/logs/
experiment_ssh/experiment-record.md

This verified mount is the sole experiment root for SSH mode. Never create, read, or write local-only experiments/ during the remote task and never create an experiments/ child inside the mount. Keep raw/failed artifacts in outputs/, formal promoted evidence in results/, and every command/status in experiment-record.md.

Remote Commands

Use only the configured tool:

ssh-bash { action: "run", command: "<remote command>", timeout: <1-7200> }

The tool is cross-platform and provides complete bounded remote Bash; do not invoke ssh, sshpass, a POSIX wrapper, PowerShell remoting, or a second executable. Commands execute on the configured server and are not confined by the tool to the project root, so every experiment command must explicitly enter the configured remote project root first. Use conservative shell quoting and put free-form experiment values in reviewed config files on the mounted workspace rather than interpolating paper text or user-authored shell fragments.