buzz-acp leaks ACP agent process trees after cancel_drain_timeout #5849

Summary

When an ACP agent turn fails with cancel_drain_timeout, the spawned agent process tree is never reaped. Repeated failures accumulate orphaned processes until memory pressure becomes significant.

Environment

• Buzz Desktop v0.5.11
• Windows 11 (build 26200)
• Agent runtime: Hermes Agent v0.20.0 / v0.20.1 via buzz-acp

Observed

After a series of failed agent turns in a single channel:

All 20 hermes-acp processes were spawned in a five-second burst (06:09:19–06:09:24) under a single buzz-acp parent, and none ever exited.

Each failing turn ended with:

Turn error - cancel_drain_timeout: Agent did not stop within 5s after
cancellation; the agent process is being replaced.

The replacement process is created as stated, but the process being replaced is not terminated — so each failed turn leaks one tree.

Expected

On cancel_drain_timeout, buzz-acp should escalate to terminating the agent's full process tree (children first) before spawning a replacement, rather than abandoning the previous process.

Impact

Users hitting any persistent agent-side failure accumulate orphaned process trees indefinitely. Recovery currently requires manually enumerating and killing descendant trees outside Buzz. The leak also makes the underlying agent failure harder to diagnose, because each retry adds another set of processes writing to the same runtime logs.

Note on scope

The agent-side hang that triggered these cancellations is a separate defect in the Hermes runtime and is filed separately against NousResearch/hermes-agent. This report covers only the process lifecycle handling in buzz-acp, which is independent of why any given agent stops responding — any agent that fails to drain within 5s will leak in the same way.