[BUG] Desktop app artifact-preview renderer balloons to 7.6 GB, driving the macOS memory pressure that makes the Claude Code daemon evict background sessions

Follow-up — recovery is self-defeating under pressure. As an experiment I restarted ~10 previously-evicted sessions in sequence. Each spawn logged bg: low memory — retiring settled workers before spawning <id> and killed the sessions restarted moments before (they settle within seconds of resuming, making them immediately eligible). All but the last few were dead again in under a minute:

[2026-07-03T12:39:15] bg: low memory (163MB free) — retiring settled workers before spawning 2d1bd440
[2026-07-03T12:39:18] bg: low memory (170MB free) — retiring settled workers before spawning 2edabea6
[2026-07-03T12:39:20] bg: low memory (166MB free) — retiring settled workers before spawning f71f426c
[2026-07-03T12:39:23] bg: low memory (157MB free) — retiring settled workers before spawning 851b7308
[2026-07-03T12:39:29] bg: low memory (145MB free) — retiring settled workers before spawning 77175487
[2026-07-03T12:39:33] bg: low memory (148MB free) — retiring settled workers before spawning 42de252c
[2026-07-03T12:39:37] bg: low memory (165MB free) — retiring settled workers before spawning 41dcca8e
[2026-07-03T12:39:42] bg: low memory (152MB free) — retiring settled workers before spawning 182e357d

So once a machine sits at pressure level 2, a bg-session fleet cannot be restored at all: restarting N sessions kills ~N-1 of them. Aggregate cost of the resumed CLI processes was ~7 GB (~300–700 MB each), which itself sustains the pressure — the mechanism feeds itself.

The macOS eviction reproduces with the desktop app closed entirely — so it's driven by general system memory pressure, not only the preview renderer. Worth decoupling in the fix: the renderer leak is one source of pressure; the retiring settled workers eviction is the shared bug, and fixing only the renderer leaves it in place for any other source.

From the 2.1.201 binary, the macOS trigger never looks at free memory:

function Gve() {               // gate run before each bg spawn
  let e = o_f();               // threshold flag, default 1024 MB
  if (e <= 0) return false;
  if (!macos) return os.freemem() < e;   // Linux/Windows: free-bytes based
  return a_f() >= 2;           // macOS: kern.memorystatus_vm_pressure_level >= 2
}

It fires whenever kern.memorystatus_vm_pressure_level >= 2 ("warn"). The (NN MB free) in the log line is computed separately and is display-only — which is why evictions fire across such a wide free-RAM range. On Apple Silicon, level 2 is routine under normal compression, so this reaps healthy fleets.

Three asks:

1. Escape hatch for the worker reaper — the bg-shell reaper has CLAUDE_CODE_DISABLE_BG_SHELL_PRESSURE_REAP; the session/worker path (Gve) has none.
2. Trigger on level ≥ 4 (critical), not ≥ 2 (warn) — warn isn't a safe signal to kill in-progress work on Apple Silicon.
3. Grace period before a worker becomes evictable — a session resumed seconds ago settles immediately and dies on the next spawn, so restarting a fleet kills N−1 (per the existing follow-up).

Regression landed in 2.1.193 (25 Jun) ("automatic memory-pressure reaping"); before it, pressure caused swapping rather than dropped sessions.

Reproduced on two consecutive days on app 1.21459.0 and 1.21459.1 (macOS, Apple Silicon, 24 GB) — with several additional data points that narrow the trigger beyond what's in the OP:

The artifact pane is NOT required — merely focusing a session that has an artifact is enough

Both occurrences line up to the second with setFocusedSession on a session that had previously produced an artifact. The app immediately registers an html-preview context and spawns the renderer that then leaks:

2026-07-16 11:54:59 [info] [CCD] LocalSessions.setFocusedSession: sessionId=session_01P3Gw…
2026-07-16 11:54:59 [info] handoff: publishing session_01P3Gw…
2026-07-16 11:54:59 [info] [Launch] Updating active servers store { count: 1, servers: '[{"serverId":"html-preview-1784174099415-0",…,"externalUrl":"https://claude.ai/code/artifact/6631d458-…","status":"running"}]' }

The renderer's spawn time (ps -o lstart) matches this log line to the second in both cases. The artifact side panel was never opened on day 2.

Leak characteristics (matching OP)

• ~1.5 MB/s (~100 MB/min) steady growth, 0% CPU, measured via repeated top -l 1 -pid <PID>
• Day 1: 7.4 GB after ~4 h; Day 2: 5.5 GB after ~1.5 h
• Almost fully compressed by macOS (CMPRS ≈ MEM), so ps RSS showed ~100 MB — invisible in Activity Monitor's default view; filled 4 GB of swap
• Renderer has zero open network sockets (lsof), so the growth is a local loop (timer/poll accumulation), not stream buffering

Additional data points

1. Leaks continue after the session ends — and even for archived sessions. Day 1's renderer was still leaking hours after the underlying session had finished and been archived. On one occurrence the view was restored at wake-from-sleep with no visible window at all (hidden webview; setFocusedSession fired during wake restore).
2. Both sessions were remote — Claude Code sessions running on another device (Mac Mini) in the same account, viewed from this machine's desktop app. Their transcripts don't exist locally; the viewer window alone produces the leak.
3. Kill behavior: kill <renderer PID> releases the memory; the view respawns once (crash-reload) and resumes leaking at the same rate; a second kill makes it stay dead. Killing the renderer doesn't affect the underlying CLI session.
4. Corroborating renderer-side warning at the exact spawn second in claude.ai-web.log:

``
MaxListenersExceededWarning: Possible EventEmitter memory leak detected. 11 $eipc_message$…_claude.settings_$_AppPreferences_$_preferencesChanged listeners added.
``
(Same EIPC listener-accumulation family as #67667 / #66127 / #67780.)

Happy to provide fuller logs or run an instrumented build if useful.

Asks #2 and #3 shipped — the retiring settled workers evictions have stopped. Decompiling 2.1.211:

function XRs(){                                    // was Gve()
  let e = et("tengu_bg_low_mem_mb", 1024) * 1024 * 1024;
  if (e <= 0) return { lowMem: !1, level: void 0 };
  if (qt() !== "macos") return { lowMem: C2d.freemem() < e, level: void 0 };
  let t = _jy();
  return { lowMem: t !== void 0 && t >= gjy, level: t };   // gjy = 4, was 2
}
function w2d(){ return et("tengu_bg_retire_grace_bridged_min", 480) * 60000; }  // 8h grace

• #2: trigger constant is now 4 (critical), was 2 (warn).
• #3: new tengu_bg_retire_grace_bridged_min, default 480 min, before a settled worker is evictable.