openclaw - 💡(How to fix) Fix [Bug]: ACP oneshot sessions leave orphaned processes — session reset does not clean up child ACP session keys [1 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Summary

When sessions_spawn creates ACP sessions with mode: "run", the spawned claude-agent-acp + claude processes remain alive indefinitely after the run completes. Additionally, when the parent session is reset via /new or /reset, the child ACP sessions are not cleaned up — the gateway only calls closeAcpRuntimeForSession() for the parent session key, not for child session keys spawned from it.

This results in progressive memory exhaustion. On a 16 GB machine, 11 orphaned pairs accumulated ~3.5 GB within 2 hours of normal use.

Related issues

This is closely related to #52708, #61895, and #35886 but identifies a specific additional gap: the parent session reset path does not enumerate and close child ACP sessions.

Steps to reproduce

1. From a main session, spawn several ACP runs:

   sessions_spawn({ runtime: "acp", mode: "run", task: "..." })

2. Wait for all runs to complete (or not — the behavior is the same).
3. Reset the parent session via /new or /reset.
4. Check OS processes: all claude-agent-acp + claude pairs are still alive, reparented to systemd.

Root cause analysis

Two independent gaps in the cleanup chain:

Gap 1: Oneshot close failure is silently swallowed

In the ACP session manager runTurn(), the oneshot cleanup runs in a finally block:

if (meta.mode === "oneshot") {
  try {
    await runtime.close({ handle, reason: "oneshot-complete" });
  } catch (error) {
    logVerbose(`acp-manager: ACP oneshot close failed...`);
  }
}

If runtime.close() fails (e.g., stdin already closed, connection lost, timeout), the error is logged but no PID-level fallback kill is attempted. The process remains alive.

Gap 2: Parent session reset does not clean up child ACP sessions

performGatewaySessionReset() calls cleanupSessionBeforeMutation() which calls closeAcpRuntimeForSession() — but only for the parent session key. It does not:

• Query the store for child sessions spawned by this parent
• Iterate child ACP session keys to close their runtimes
• Kill processes associated with child ACP sessions

stopSubagentsForRequester() is called inside ensureSessionRuntimeCleanup(), but it only handles subagent-type sessions, not ACP sessions. The child ACP session entries have spawnedBy / parentSessionKey set, but these are never queried during parent cleanup.

Suggested fix

Fix 1: PID-level fallback in terminateAgentProcess()

After close() → SIGTERM → SIGKILL sequence, if the process still has not exited, store the PID in the session record and use process.kill(pid, "SIGKILL") as a last resort:

async terminateAgentProcess(child) {
  // ... existing SIGTERM/SIGKILL logic ...
  if (isChildProcessRunning(child)) {
    // Last resort: kill by PID directly
    try { process.kill(child.pid, "SIGKILL"); } catch {}
  }
}

Fix 2: Enumerate and close child ACP sessions on parent reset

In cleanupSessionBeforeMutation(), after handling the parent session, query the store for child ACP sessions:

// After closing parent ACP runtime:
const childAcpKeys = listAcpSessionEntries({ cfg })
  .filter(e => e.entry?.parentSessionKey === canonicalKey 
            || e.entry?.spawnedBy === canonicalKey);

for (const child of childAcpKeys) {
  await closeAcpRuntimeForSession({
    cfg, sessionKey: child.key, entry: child.entry, reason: params.reason
  });
}

Fix 3: Gateway startup orphan reaper

On gateway startup, scan for claude-agent-acp processes whose session keys no longer exist in the store, and kill them:

// In gateway startup sequence:
const activePids = getActiveAcpSessionPids(); // from session records
const osPids = findProcessesByName("claude-agent-acp");
for (const pid of osPids) {
  if (!activePids.has(pid)) {
    process.kill(pid, "SIGTERM");
  }
}

Expected behavior

1. After a mode: "run" ACP session completes, all child processes should be terminated.
2. When the parent session is reset, all child ACP sessions should be closed and their processes terminated.
3. On gateway startup, orphaned ACP processes from previous runs should be reaped.

Actual behavior

Child processes accumulate indefinitely. Each orphaned pair consumes ~300 MB RSS. subagents list shows 0 entries because the subagent registry does not track ACP sessions, making the leak invisible to the orchestrating agent.

OpenClaw version

2026.4.14

Operating system

Linux (arm64)

Logs, screenshots, and evidence

$ ps -eo pid,ppid,lstart,rss,args | grep claude-agent-acp
# 11 pairs, all with PPID=systemd (orphaned), started at regular intervals
# Total RSS: ~3.5 GB

$ # All pairs have identical parent structure:
# node claude-agent-acp → claude
# PPID is systemd --user (reparented after intermediate processes exited)

Impact and severity

On memory-constrained machines (≤16 GB), this can exhaust available RAM within a few hours of normal use with ACP agents, causing OOM or severe swap thrashing.