SocketError: other side closed

[cloudycotton]

Description

I'm facing a problem where if my agent streams for long (we have agent that runs for 100+ steps and write big files), I get error socket connection closed error. It always happens during streaming a big write to a file using a tool.

This seems to happen across providers (anthropic, openai).

AI SDK Version

latest verison of anthropic provider and ai sdk.

Code of Conduct

• [x] I agree to follow this project's Code of Conduct

[cloudycotton]

@gr2m any pointers for me to create a reproducible example?

The only thing I can do to reproduce the problem consistently is to ask it to write 50 page long research.txt file in a loop. That will run into this.

(formatted for readability)

start: {
  "level": 50,
  "time": 1761315667228,
  "pid": 14,
  "hostname": "srv-d2j1d7mmcj7s73eh3kmg-5c9bb5c585-cb5rn",
  "name": "@runable/api",
  "err": {
    "type": "TypeError",
    "message": "terminated: other side closed",
    "stack": "TypeError: terminated ... error stack trace ..."
  },
  "msg": "Got error in agent stream"
}
start: {
  "level": 50,
  "time": 1761315667228,
  "pid": 14,
  "hostname": "srv-d2j1d7mmcj7s73eh3kmg-5c9bb5c585-cb5rn",
  "name": "@runable/api",
  "err": {
    "type": "Error",
    "message": "",
    "stack": "Error ... error stack trace ..."
  },
  "msg": "Got error in agent stream"
}

[cloudycotton]

So I think the reason why it kills the stream and throw error beyond max-retry is because we don't handle the unidici error separately because those error structures contain circular dependencies causing json serialization to fail and not re-try. @gr2m

[cloudycotton]

@lgrammel this is a very frequent issue on openai side. Any time, I use codex model to write a long file. It times out especially in non streaming mode eg - generateText

[gr2m]

It's hard for us to investigate until we have a way to replicate the problem ourselves, unfortunately