Remote Cluster Agent

remote-cluster-agent

Local control plane for remote GPU clusters: a Go daemon + rca CLI. All commands go through the daemon's Unix socket to a cluster-side Python agent.py over persistent SSH, ~0.1s/command latency.

What to do when triggered

1. Check if rca is installed: run which rca.
   • If not found → run the first-time install flow (below), then continue to step 2.
2. Check daemon status: run rca daemon status.
   • If daemon request failed → run rca daemon start.
3. Check node connectivity: run rca nodes.
   • If the target node is dead → rca connect <node>. Still dead → re-establish whatever SSH tunneling / VPN / jump host you depend on, then retry.
   • Second round still dead → read rca daemon logs -f for error detail.
4. Run the user's command.

Don't write raw ssh commands, always use rca exec / rca batch.

First-time install (auto, no user action needed beyond SSH info)

When which rca returns not found, execute in order:

cd <skill_dir>
make install              # builds rca into ~/go/bin (requires Go 1.21+)
rca config init           # migrates legacy ~/.config/remote-cluster-agent/*.md, or generates blank template

Then ask the user for each node's SSH command, and write them into ~/.config/rca/config.toml under [nodes.*]. Example:

[nodes.train]
ssh = "ssh gpu-train"

[nodes.eval]
ssh = "ssh -p 2222 gpu-eval"
dir = "/home/user/project"
agent_path = "/shared/.agent/agent.py"

Optional per-node overrides: dir (default working directory) and agent_path (where agent.py lives on that node). Both fall back to globals if unset.

Finally start and verify:

rca daemon start
rca daemon status         # confirm running
rca nodes                 # confirm nodes connected
rca agent check           # if agent missing → rca agent deploy

Common operations

Single-node exec / heredoc

rca exec -n train "nvidia-smi"
rca exec -n train -d /home/user/project "git pull"
rca exec -t 600 -n train "python train.py"

# heredoc (bypasses shell escaping, recommended for multi-line or special chars)
rca exec --stdin -n train <<'EOF'
cd /home/user/project
python -c "import json; print(json.dumps({'k':'v \"q\"'}))"
EOF

Multi-node parallel (batch)

rca batch "nvidia-smi | head -20"
rca batch -n train,eval "df -h /home"
rca batch --json "hostname" | jq -r '.results[] | "\(.node): \(.output)"'

Node status / connection management

rca nodes                   # current connection state
rca nodes --check           # deep ping (with latency)
rca nodes --health          # latency history (monitor-tracked)
rca connect train           # manually reconnect a dead node
rca disconnect train        # actively close a node's connection

File transfer (cp)

rca cp moves files through the agent's JSON-Lines channel (base64-encoded, 50 MB/file limit). Works on any SSH setup — no separate SCP/rsync path needed.

rca cp train:/home/user/logs/train.log ./
rca cp ./config.yaml train:/home/user/project/config.yaml
rca cp -r train:/home/user/checkpoints/exp07 ./local-copy/

For very large files (multi-GB checkpoints), use shared filesystem or object storage — rca cp peaks around 2–3 MB/s.

Streaming / long tasks

rca exec -s -n train "tail -f /var/log/train.log"
rca exec -s -n train "python train.py"

Cluster health inspection

Triggered when the user says "cluster status", "which node is free", "cluster health", "GPU usage". Use rca batch for parallel sampling:

rca nodes --check
rca batch "nvidia-smi --query-gpu=name,memory.used,memory.total,utilization.gpu --format=csv"
rca batch "df -h /home | tail -1"
rca batch "uptime && tmux ls 2>/dev/null | wc -l"

Summarize into a table and recommend the most idle node. See reference/cluster-health.md for detailed probe commands, parse rules, and report format.

Reading files vs running commands

rca exec is for executing commands (starting training, killing processes, checking process state, installing deps) — not reading file contents.

If mutagen real-time sync is configured, the cluster's outputs/ appears locally automatically. For logs, JSON results, CSV — use the native Read tool on the local path. That's ~20x faster than piping through the SSH connection.

Decision rule:

• Want to read file content (logs, JSON, CSV, config files) → Read("outputs/...") locally
• Want to run an operation (ps, nvidia-smi, start/stop, pip install) → rca exec
• Unsure if the file is synced locally → ls outputs/ or Glob("outputs/**/<pattern>") first

Typical mistake: rca exec -n train "cat /home/user/outputs/.../log.txt" to read a log. That file is already synced — just Read it.

SSH connection failures

When rca nodes --check shows dead, or rca exec reports connect failed:

1. Try rca connect <node> once — most transient issues resolve here.
2. Still dead? The SSH transport underneath is broken. rca itself doesn't manage tunnels / VPN / jump hosts — re-establish those through your usual workflow (ssh config, tunnel script, corporate VPN, etc.), then retry step 1.
3. Inspect daemon errors with rca daemon logs -f.

Report failures grouped by cause — don't just say "all nodes failed". Which nodes recovered? Which need tunnel re-establishment? Which look like SSH config issues?

Mutagen session health

When the user mentions "skill updated" or "redeploy", check mutagen session config:

mutagen sync list 2>/dev/null

For each session:

1. Sync mode: should be one-way-replica. If one-way-safe or two-way-resolved, prompt to rebuild.
2. Ignore list: .git should NOT be ignored (v0.4.0+ syncs .git to keep cluster-side git status clean).

Rebuild (only after user confirms):

mutagen sync list <name> --long    # record current params first
mutagen sync terminate <name>
bash <skill_dir>/mutagen-setup.sh <ssh_host> <local_dir> <remote_dir> <session_name>

Container restart: mutagen fails to reconnect

Symptom: mutagen sync list shows Beta Connected: No; mutagen sync resume reports server magic number incorrect.

Root cause: the cluster container restarted and lost ~/.mutagen/ (container home is ephemeral). Mutagen's agent binary and staging data live in a persistent path (typically mounted storage like /shared/.mutagen/).

Fix:

# Recreate the symlink on the cluster node
rca exec -n <node> "ln -sf /shared/.mutagen ~/.mutagen"

# Then resume locally
mutagen sync pause <session-name>
mutagen sync resume <session-name>

Prevention: have the container init script create this symlink on startup.

See reference/mutagen-troubleshooting.md for more failure modes.

Config

Main config: ~/.config/rca/config.toml (edit via rca config edit).

First-time install flow is in "What to do when triggered → First-time install". Day-to-day, just keep the daemon running (rca daemon status).

Agent deployment / updates

rca agent check               # show agent version per node
rca agent deploy              # copy local agent.py (only where missing)
rca agent deploy --force      # force overwrite

Note: rca agent deploy uses direct SSH (bypasses the daemon) and iterates all nodes in config.toml serially. Disconnected nodes will block for SSH's default connect timeout (minutes). Workaround — use a temp config with only live nodes:

cp ~/.config/rca/config.toml /tmp/rca_deploy.toml
# manually delete [nodes.*] sections for dead nodes
rca --config /tmp/rca_deploy.toml agent deploy --force

Auto-start (optional, macOS)

rca daemon register           # register launchd agent (requires App Management permission)

Without this, start the daemon manually with rca daemon start whenever you reboot.

Troubleshooting

Safety boundaries

• Read shared storage restrictions from your config and follow them strictly — files on shared paths often belong to other teams, there's typically no recycle bin, and deletion can destroy days of someone else's work.
• If the cluster has no public internet, use internal PyPI mirrors — don't attempt external URLs (GitHub, PyPI.org, etc.).
• Don't auto-push to master/main — avoid shipping unreviewed code to teammates.
• pkill -f must use the bracket trick: pkill -f "[s]glang.launch_server" not pkill -f "sglang.launch_server" — because the SSH process command line contains the kill pattern, pkill -f would match SSH itself and tear the connection down. Same applies to pgrep -f, grep over process lists, etc.
• Long-running processes must be backgrounded: if a process doesn't exit (inference server, training loop), use nohup ... & or tmux new-session -d, with echo placed after the background command using ;:

  # Correct: nohup background + echo outside
  nohup python -m sglang.launch_server ... > /tmp/log 2>&1 & echo "PID=$!"
  # Correct: tmux detach + echo outside
  tmux new-session -d -s sglang "python -m sglang.launch_server ..."; echo "started"
  # Wrong: echo inside tmux's && chain, never executes
  tmux new-session -d -s sglang "python ... && echo started"
  

Architecture

The daemon runs independently of Claude Code (Unix socket), so every session shares one connection pool. The CLI is standalone too — usable in shells, scripts, or cron. Cluster-side agent.py speaks JSON-Lines v2.1.0 (streaming, cancel, batch, file transfer).

Upgrading from MCP v0.3.x: remote_bash(node, cmd) → rca exec -n node "cmd", remote_bash_batch → rca batch. Legacy ~/.config/remote-cluster-agent/*.md markdown config auto-migrates on rca config init.