Claude Code Held Every Subagent Result in Memory Until 14 Releases Ago the Cap Came Off
2.1.238 fixes unbounded memory growth in long sessions by releasing subagent tool results once they scroll out of view. The leak was always there — the 200-subagent cap was hiding it.
NeuroX AI · August 21, 2026
The Claude Code 2.1.238 changelog fixes "unbounded memory growth in long interactive sessions": subagent tool results are now released once they leave the recent display window. Version 2.1.224 removed the 200-subagent-per-session spawn cap — fourteen releases earlier.
That ordering is the whole story. A parent session that retains every subagent's tool results forever is a leak at any scale, but with a hard ceiling of 200 spawns it was a leak with a known worst case. Lift the ceiling and the same code becomes a process that grows until the machine says no.
Look at what the fix keys on: the display window. Retention was coupled to what a human might scroll back to, not to what the model still needs in context. Those two lifetimes have nothing to do with each other, and one of them is unbounded.
Same release, same class of bug: leftover /tmp/claude-*-cwd files when a Bash command is killed, times out, or is interrupted. Cleanup existed — on the path where the command returns normally.
In-house orchestrators are built exactly this way. Fan out twelve agents, each returning a few hundred KB of file contents and command output, hold all of it on the parent because that was never a decision anyone made, and the thing runs fine for an hour of demos. Concurrency limits are load-bearing memory management in most fleets, and nobody wrote them down as such.
Every ceiling you remove promotes a bounded bug to an unbounded one.