SWE-1.6 triggers unbounded worker spawning → load >1100, ~42GB RAM, SSH instability, ECONNREFUSED (stable with GPT-5.4 Mini)

[oib]

Title

SWE-1.6 triggers unbounded worker spawning → load >1100, ~42GB RAM, ECONNREFUSED (stable with GPT-5.4 Mini)

---

Summary

Using SWE-1.6 in Windsurf causes extreme resource usage after sending a prompt. The system becomes unstable and the Windsurf backend enters a reconnect loop with ECONNREFUSED. Switching to GPT-5.4 Mini on the same setup resolves the issue entirely.

---

Environment

• OS: Debian (Trixie)
• CPU: AMD 5950X (16C/32T)
• Container CPU allocation: 12 threads (6 cores)
• RAM: 64 GB (48 GB assigned to the container)
• Runtime: Incus container
• Filesystem: BTRFS
• Concurrent workload: CI runner in another container

---

Steps to Reproduce

1. Start Windsurf with SWE-1.6 selected
2. Open a moderately sized project
3. Send a prompt (e.g., code query / analysis)
4. Observe process spawning and system metrics

---

Observed Behavior

• Dozens of processes spawn:

  language_server_linux_x64
  --enable_index_service
  --enable_local_search
  

• Load average spikes dramatically (observed >1100 on 32-thread CPU)
• Memory usage grows to ~42 GB (container assigned 48 GB)
• Windsurf logs:

  Connection to server got closed. Server will restart.
  windsurf client: couldn't create connection to server.
  Error: connect ECONNREFUSED 127.0.0.1:<port>
  Restarting server failed
  

• Requires manual window reload to recover

---

Comparison (Same System, Same Project)

Model	Load	RAM Usage	Stability
SWE-1.6	100–1100+	up to 42GB	❌ unstable
GPT-5.4 Mini	<10	~7GB	✅ stable

---

Additional Observations

• Trigger occurs after prompt, not at startup
• Behavior resembles unbounded parallelism / worker spawning
• System instability correlates with process explosion
• Increasing RAM improves stability but does not fix the root cause
• Limiting container CPU/processes mitigates but does not eliminate the issue
• Updating SSH client and server resolved connection drops, indicating prior issues were caused by system starvation rather than network failure

---

Expected Behavior

• Bounded worker pool
• Graceful degradation under load
• Stable backend connection
• Resource usage proportional to workload

---

Actual Behavior

• Unbounded worker spawning
• Extreme CPU and memory usage
• Backend becomes unreachable (ECONNREFUSED)
• Requires manual recovery

---

Related Issues

• ECONNREFUSED restart loop (Restarting server failed (since new update from 03.03.25) #141)
• Language server memory explosion (# Bug Report: language_server_macos_arm Unbounded Memory Consumption on Startup #300)

This report identifies a likely root cause: resource explosion triggered by SWE-1.6 after a prompt.

---

Workarounds

• Switching to GPT-5.4 Mini (fully stable)
• Increasing available RAM (partial mitigation)
• Limiting container CPU / processes (partial mitigation)
• Updating SSH configuration prevents disconnects but does not address the root cause

---

SSH Mitigations Applied

• Client (~/.ssh/config):

  Host *
      ServerAliveInterval 10
      ServerAliveCountMax 10
      TCPKeepAlive yes
      ConnectTimeout 10
  

• Server (/etc/ssh/sshd_config):

  ClientAliveInterval 15
  ClientAliveCountMax 10
  TCPKeepAlive yes
  UseDNS no
  MaxStartups 10:30:60
  LoginGraceTime 30
  MaxSessions 4
  

• Result: SSH disconnects no longer occur under load; prior disconnects were due to scheduler starvation rather than network issues

---

Notes

This appears to be a scaling/control issue specific to SWE-1.6 runtime behavior. High-core systems amplify the problem, but the lack of concurrency limits likely affects general use as well.