Large memory spike when spawning helper processes

[jamesmurdza]

Since upgrading to Claude Agent Python SDK 0.1.18, I am commonly seeing the following while running client.query(query) followed by client.receive_response():

ERROR:claude_agent_sdk._internal.query:Fatal error in message reader: Command failed with exit code -9 (exit code: -9)
Error output: Check stderr output for details

After which the method fails with the error:

Command failed with exit code -9 (exit code: -9)

After this has occurred once, subsequent calls to client.query fail with the error:

Cannot write to terminated process (exit code: -9)

Exit code 9 is often caused by OOM Killer, meaning the process ran out of memory. These errors don't always happen immediately. Sometimes the SDK works for one or two calls, and then this happens.

This could potentially be caused by a memory leak in Claude Code.

Original issue: daytonaio/daytona#3208

[jamesmurdza]

I was able to dig a little deeper and find out what's happening.

The baseline RSS (physical RAM) of the Claude process increased from 372 MB to 433 MB after upgrading to SDK 0.1.18. When the SDK spawns short-lived helper processes (e.g. for filesystem search), each child inherits the parent’s address space via fork(). When two helpers run concurrently, the observed RSS was approximately:

Parent: ~433 MB
Child 1: ~290 MB
Child 2: ~270 MB

I'm running Claude on a sandbox with 1GB of RAM. The combined RSS reached ~1.0 GB, triggering the OS OOM killer, which terminated the parent process (exit code -9).

[Uzay-G]

this is actually horrible, it makes running a bunch of agents in parallel, not even doing very resource intensive things, kinda cooked from a mem usage point of view