[MacOS] Gateway destroy kills sandbox session with Broken pipe instead of graceful error

[zNeill]

Description

Test Plan https://docs.google.com/document/d/14bIihtww1Jtzz9eU-FdmFixKRFO9WI_OuGw9yrEP9Cw/edit?pli=1&tab=t.0
4.4 - test case 4

When the OpenShell gateway is destroyed while a sandbox session is active, the sandbox connection is abruptly terminated with a Broken pipe error and exit code 255. The user receives no graceful error message indicating the gateway is restarting, and there is no automatic recovery once the gateway comes back. Active inference requests that were in-flight complete successfully, but subsequent requests and the sandbox session itself are killed without warning.

Steps to reproduce

1. Complete nemoclaw setup successfully
2. Set inference route: openshell inference set --provider ollama-local --model llama3.2:1b
3. Connect to sandbox in Terminal 1: nemoclaw connect
4. Start a looping inference request inside the sandbox:

   for i in $(seq 1 20); do
   echo  --- Request $i at $(date +%H:%M:%S) --- 
   curl -s -m 5 https://inference.local/v1/chat/completions \
   -H  Content-Type: application/json  \
   -d '{ model : llama3.2:1b , messages :[{ role : user , content : say hi }]}' \
   | head -c 120
   echo   
   sleep 2
   done

5. In Terminal 2 (host), destroy and restart the gateway:

   export DOCKER_HOST= unix://$HOME/.colima/default/docker.sock 
   openshell gateway destroy -g nemoclaw && openshell gateway start --name nemoclaw

Actual Result

• No graceful error message was shown inside the sandbox
• No automatic recovery.

Steps to reproduce

No steps provided.

---

[NVB# 5982642]

[NVB#5982642]

[wscurran]

✨
Possibly related open PRs:

• #1245 fix(cli): clear sandbox registry when gateway is destroyed during onboard
• #1167 fix: show macOS launchctl instructions instead of systemctl on darwin (#1067)
• #1246 fix(security): redact secret patterns from CLI log and error output

[jyaunches]

Architectural Context — Active Session Detection

Picking this up as part of a broader onboard.ts decomposition effort.

The fix for this bug, done right, creates reusable infrastructure that improves destroy, rebuild, stop, status, list, connect, and uninstall. The core issue isn't just the Broken pipe UX — it's that no destructive operation in the CLI checks for active SSH sessions before proceeding.

Proposed approach

Create a shared sandbox-session-state.ts module with an getActiveSandboxSessions(sandboxName) utility that detects active SSH connections (via openshell forward list output and/or process detection). This module would be consumed by:

Consumer	File	Behavior
sandboxDestroy()	nemoclaw.ts	Warn: "⚠️ Active SSH session detected. Destroy anyway?"
sandboxRebuild()	nemoclaw.ts	Same warning pattern
cleanupGatewayAfterLastSandbox()	nemoclaw.ts	Check before automatic cleanup
destroyGateway()	onboard.ts	Graceful disconnect before teardown
stopAll()	services.ts	Warn about active sessions before stopping
sandboxStatus()	nemoclaw.ts	Show "Connected: yes" in status output
nemoclaw list	nemoclaw.ts	Show active indicator next to connected sandboxes
sandboxConnect()	nemoclaw.ts	Warn "Already connected in another terminal"

This earns its place in the refactoring plan not because the destroy function itself needs extraction from onboard.ts, but because the solution creates a new typed, testable shared module that addresses a cross-cutting concern across 9 call sites in 4 files — exactly the architectural pattern we're driving toward.