ssh-lab（SSH 实验室）

ssh-lab — Remote Server SSH Workbench

Structured SSH operations for managing remote GPU servers. Standardizes what's universal (GPU/disk/process), delegates interpretation to the LLM.

When to Use

• User asks about GPU status, server health, remote processes
• Need to check if training is running, disk is full, GPU is idle
• Running commands on remote servers
• Tailing remote log files, listing checkpoints
• File transfer between local and remote (rsync)
• Comparing hosts side-by-side (GPU, disk, processes)
• Setting up alerts for GPU idle, disk full, process died
• Any mention of: ssh-lab, gpu status, server status, nvidia-smi, remote server

Setup

cd /path/to/ssh-lab && npm run build

After building, use the CLI via any of:

ssh-lab status all          # if npm link'd or installed globally
npx ssh-lab status all      # via npx
node dist/cli.js status all # direct invocation

The CLI reads ~/.ssh/config automatically — zero configuration needed for existing SSH hosts.

Commands

List hosts

ssh-lab hosts [--json]

Server status (GPU + disk + process)

ssh-lab status [host|all] [--json] [--timeout ms] [--heartbeat] [--quiet]

Returns structured probe data: GPU utilization/VRAM/temp, disk usage, active processes. Includes automatic alerts: GPU idle, disk full (>90%), high temp (>85°C). --heartbeat gives compact one-liner output for agent integration. --quiet suppresses all-clear output (for cron).

Run remote command

ssh-lab run <host|all> <command...> [--json] [--timeout ms]

Supports all for parallel execution across all hosts.

Tail remote log

ssh-lab tail <host> <path> [-n lines] [--json]

Default: last 50 lines. Smart truncation at 50KB.

List remote directory

ssh-lab ls <host> <path> [--sort time|size|name] [--json]

Lists files with size, date, permissions. Great for checkpoint directories.

Disk usage

ssh-lab df <host|all> [path] [--json]

Shows disk usage for all filesystems. Optional path runs du -sh on a specific directory.

File sync (rsync)

ssh-lab sync <host> <src> <dst> [--direction up|down] [--dry-run] [--exclude pat1,pat2]

Direction: down (remote→local, default), up (local→remote). Always use --dry-run first!

Watch file (snapshot for polling)

ssh-lab watch <host> <path> [-n lines] [--prev-hash hash] [--json]

Captures file state + metadata. Returns changed: true/false when --prev-hash is provided. Agent calls this repeatedly via cron/heartbeat — the tool itself is stateless.

Compare hosts (side-by-side)

ssh-lab compare <host1> <host2> [--probes gpu,disk,process] [--json]
ssh-lab compare all [--probes gpu]
ssh-lab compare GMI4 GMI5 GMI6 --probes gpu,disk

Side-by-side comparison of GPU, disk, and process data across hosts. Probes default to gpu,disk,process. Supports any number of hosts. Uses bounded concurrency pool for parallel data collection.

Doctor (diagnostics)

ssh-lab doctor <host> [--json] [--timeout ms]

Runs connectivity and health diagnostics on a host. Checks SSH reachability, authentication, command execution, and basic system health.

Alert management

# List configured rules
ssh-lab alert list [--json]

# Add a rule
ssh-lab alert add <kind> <host> [--threshold N] [--process-pattern regex]

# Remove a rule
ssh-lab alert remove <rule-id>

# Evaluate alerts against live status
ssh-lab alert check [host|all] [--json] [--quiet]

Alert kinds: gpu_idle, disk_full, process_died, ssh_unreachable, oom_detected, high_temp

Examples:

ssh-lab alert add disk_full all --threshold 90
ssh-lab alert add process_died GMI4 --process-pattern "torchrun|deepspeed"
ssh-lab alert check all --json

Add custom host

ssh-lab add <name> <user@host> [--port N] [--tags a,b] [--notes "desc"]

Output Modes

All commands support three output modes:

• summary (default): Human-readable text with formatting
• json (--json or -j): Structured JSON for programmatic use
• raw (--raw or -r): Raw command output only

Timeout Tiers

Commands use per-type timeout defaults (user --timeout always overrides):

• quick (5s): hosts, add, doctor — connectivity checks
• standard (15s): status, run, ls, df, alert, compare — normal probes
• stream (30s): tail, watch — streaming/log tailing
• transfer (60s): sync — rsync bulk transfers

Architecture

• Probe system: Pluggable probes (gpu, disk, process) collect structured data
• SSH via child_process: Uses native ssh — inherits user's config, keys, ProxyJump
• ControlMaster reuse: Persistent SSH connections via /tmp/ssh-lab-%r@%h:%p sockets; auto-cleanup detects stale sockets (ssh -O check), gracefully exits (ssh -O exit), then removes dead socket files before retry
• Retry + error classification: Transient failures retried with exponential backoff; errors classified (AUTH_FAILED/TIMEOUT/etc.) with hints
• Tiered timeouts: Per-command defaults — quick (5s) for doctor/add, standard (15s) for status/run/ls/df/compare/alert, stream (30s) for tail/watch, transfer (60s) for sync. --timeout always overrides
• Parallel execution: status all / run all queries all hosts concurrently via withPool (default: 5)
• Alert system: 6 built-in alert rules with configurable thresholds
• Zero npm dependencies: Only devDependencies (typescript, @types/node)
• SSH config hosts use alias: For hosts from ~/.ssh/config, SSH alias is used directly so all ProxyJump/Match/Include rules are respected

Heartbeat Integration

In HEARTBEAT.md:

- [ ] Run: ssh-lab status all --heartbeat
  Check for: GPU idle, disk >90%, processes died
- [ ] Run: ssh-lab alert check all --json
  Act on any critical firings

Config

Custom hosts stored in ~/.config/ssh-lab/config.json (auto-created, XDG-compliant). SSH hosts from ~/.ssh/config discovered automatically via ssh -G. Override config path with SSH_LAB_CONFIG env var.