OOM crash under multi-subagent load: subagent CLI processes grow to ~6.5 GB RSS each; CliGovernor logs critical memory pressure for 8 min but evicts nothing (Linux, Desktop 1.17377.0 / CLI 2.1.187)

Summary

Under a multi-subagent workload (5–7 background agents over ~35 min), individual Claude Code CLI child processes grew to ~6.5 GB RSS each (plus 3–7.6 GB swapped, each). Four concurrent processes exhausted a 31 GB RAM + 20 GB swap Linux machine. The kernel OOM-killer selected the desktop app's NetworkService utility process (oom_score_adj 300), which fatally broke the Electron main process ("networking is broken until relaunch") — the user experiences this as the desktop app crashing. The app's own memory governor logged pressure at critical repeatedly for ~8 minutes before the kill and took no action each time.

Two distinct defects:

1. CLI subagent memory ballooning — CC CLI subagent/session processes reaching ~6.2–6.7 GB RSS each (17 GB total-vm on the largest). Four concurrent = ~25.6 GB RSS + ~18.6 GB swap.
2. CliGovernor takes no action at critical pressure — from 15:50:48 to 15:57 it logged a warning/critical pair every ~40–60 s, each time reporting "would evict 0 idle session(s), 1 effective", and never evicted, throttled, or surfaced anything to the user before the kernel intervened.

Environment

• Claude Code Desktop 1.17377.0, Electron 42.5.1, bundled CC CLI 2.1.187
• Pop!_OS (Linux 7.0.11-76070011-generic), 31 GB RAM + 20 GB swap, COSMIC desktop
• Workload: interactive session spawning 5–7 background subagents (Task tool fan-out) doing HTTP QA + file writes

Timeline (2026-07-03, CDT)

• 15:23 — app launch; heavy subagent fan-out session begins
• 15:50:48 — first [CliGovernor] memory pressure (warning): would evict 0 idle session(s), 1 effective
• 15:51:28 → 15:57 — memory pressure (critical) + same "would evict 0" line repeating every ~40–60 s
• 15:58:21 — kernel OOM (free swap 180 kB of 20 GB); kills claude-desktop NetworkService (oom_score_adj 300)
• 15:58:22 — app log: Network Service process gone (reason=killed, exitCode=15) outside app quit — main-process networking is broken until relaunch; Sentry event emitted (should exist server-side)
• 15:58:42–15:59:09 — app auto-relaunches
• 15:58:51 — kernel kills the largest CLI process (6.8 GB anon-rss)

Key evidence (verbatim)

Kernel journal:

Jul 03 15:58:21 pop-os kernel: tokio-rt-worker invoked oom-killer: gfp_mask=0x140cca(...), order=0, oom_score_adj=0
Jul 03 15:58:21 pop-os kernel: Free swap  = 180kB
Jul 03 15:58:21 pop-os kernel: Total swap = 20970996kB
Jul 03 15:58:21 pop-os kernel: [  13123]  1000 13123  4253598  1663156  1662700  456  0 29573120  1987257  0 claude
Jul 03 15:58:21 pop-os kernel: [  14644]  1000 14644  3204926  1679146  1678681  465  0 23130112  1166198  0 claude
Jul 03 15:58:21 pop-os kernel: [  14852]  1000 14852  2680590  1625919  1625454  465  0 19800064   810017  0 claude
Jul 03 15:58:21 pop-os kernel: [  14974]  1000 14974  3204910  1749861  1749398  463  0 21512192   908963  0 claude
Jul 03 15:58:21 pop-os kernel: Out of memory: Killed process 8799 (claude-desktop) total-vm:1530612248kB, anon-rss:318980kB, ... oom_score_adj:300
Jul 03 15:58:51 pop-os kernel: Out of memory: Killed process 13123 (claude) total-vm:17014392kB, anon-rss:6806964kB, ...

(rss column is pages: 1,663,156 pages ≈ 6.35 GB RSS; swapents 1,987,257 ≈ 7.6 GB swapped — per process.)

App log (~/.config/Claude/logs/main.log):

2026-07-03 15:50:48 [warn] [CliGovernor] memory pressure (warning): would evict 0 idle session(s), 1 effective
2026-07-03 15:51:28 [warn] [CliGovernor] memory pressure (critical): would evict 0 idle session(s), 1 effective
   (warning/critical pair repeats every ~40–60s, 15:50 → 15:57)
2026-07-03 15:58:22 [error] Network Service process gone (reason=killed, exitCode=15) outside app quit — main-process networking is broken until relaunch
2026-07-03 15:58:53 [info] Starting app {

systemd user journal:

Jul 03 15:58:21 pop-os systemd[1531]: app-cosmic-claude-desktop-8422.scope: Failed with result 'oom-kill'.

Notes for triage

• No crashpad minidump (all Crashpad dirs empty) — expected, SIGKILL is uncatchable. A Sentry event fired at 15:58:22 with the NetworkService error; should be retrievable server-side.
• The 2.1.199 "background-agent daemon self-kill after unclean shutdown" bug is NOT implicated (daemon.log shows clean idle exit the prior evening, no respawn churn).
• Suggested fixes: (a) memory ceiling / recycling for CLI subagent processes; (b) CliGovernor should act at critical (evict, pause spawns, or at minimum surface a user-visible warning) rather than logging "would evict 0" — the governor had 8 minutes of runway; (c) consider oom_score_adj on CLI children so the kernel prefers them over the app's NetworkService (killing networking takes down the whole app; killing one subagent is recoverable).

Repro guidance

Long interactive session on Linux; fan out 4+ background subagents each doing sustained tool work (HTTP calls, file reads/writes) for 30+ min; watch ps -o pid,rss,etime,args -C claude — RSS climbs into multiple GB per process without release.