Slow Claude Opus responses with low VPS usage

[mesajohn]

Hi Paperclip team,

I am running Paperclip on a Hostinger KVM 4 VPS.

During slow Claude Opus agent runs, the server does not appear to be resource constrained:

CPU: usually around 2% to 5%
RAM: around 2.1 GB out of 15.6 GB
Disk: around 39 GB out of 200 GB
Bandwidth: very low compared with the plan limit

Even when I send a large file to be analyzed, the VPS still shows low CPU and RAM usage. The experience feels slow, almost like Paperclip is waiting for something external rather than being limited by server hardware.

My logs repeatedly show entries like:

GET /heartbeat-runs/.../log 200
GET /issues/.../live-runs 200
GET /companies/.../live-runs 200
WARN: GET /issues/TRIAL-1 404
WARN: GET /issues/STAGE-2 404

My questions are:

1. Does Paperclip stream Claude output directly, or can some agent steps be buffered before the UI shows output?

2. Does Paperclip use Anthropic prompt caching with cache_control for repeated large context, agent instructions, file content, issue history, or tool definitions?

3. Does Paperclip resend the full issue history, file content, and agent instructions on every Claude call?

4. Can some agents be configured to use Claude Sonnet or Haiku for routine work, while reserving Claude Opus only for final reasoning?

5. Could the repeated TRIAL-1 and STAGE-2 404 warnings indicate stale workspace references, broken issue links, or stuck run state?

6. Is there a recommended configuration to improve response speed for large file analysis and multi-agent workflows?

Based on my server usage, upgrading the VPS does not seem likely to help because the current VPS resources are barely being used. I would appreciate any guidance.

Thank you.