OOM crash (heap out of memory) when resuming long session [1.39.2]

cmd --resume <session-id> --yolo の再開パスから始め、コンパクション、大量のツール出力、画像の読み取りを含む長時間のセッションで再現します。再開時にセッションのトランスクリプトとツール結果がどのように読み込まれ、保持されるかを追跡します。セッションがデフォルトヒープの OOM を回避するか、中断してセッションを失う代わりに、明確で回復可能なエラーを報告すれば完了です。

Description

Running cmd --resume <session-id> --yolo on a long-running agent session crashes with a Node.js FATAL ERROR: Ineffective mark-compacts near heap limit Allocation failed - JavaScript heap out of memory. The process aborts (zsh: abort) and the session is lost.

Environment

• Command Code: 1.39.2
• Node: v24.19.0 (/Users/shreyashphakadepawar/.nvm/versions/node/v24.19.0/bin/node)
• OS: macOS 26.6.2 (Build 25G83), arm64
• Invocation: cmd --resume 06c2423e-70fd-4af7-8416-4d6a2f8d8e87 --yolo
• Default Node heap limit on this machine: 2240 MB (v8.getHeapStatistics().heap_size_limit)

Crash output

<--- Last few GCs --->
[20041:0x75280c000] 68684349 ms: Mark-Compact 1938.6 (2098.2) -> 1923.8 (2098.0) MB, pooled: 1 MB, 31.92 / 0.04 ms  (average mu = 0.305, current mu = 0.306) task; scavenge might not succeed
[20041:0x75280c000] 68684395 ms: Mark-Compact 1939.8 (2098.0) -> 1922.8 (2097.1) MB, pooled: 1 MB, 32.96 / 0.04 ms  (average mu = 0.300, current mu = 0.296) allocation failure; scavenge might not succeed

FATAL ERROR: Ineffective mark-compacts near heap limit Allocation failed - JavaScript heap out of memory
----- Native stack trace -----
 1: 0x1043219f4 node::OOMErrorHandler(char const*, v8::OOMDetails const&) [...]
 ...
 39: 0x188ef84e4 start [/usr/lib/dyld]
zsh: abort      cmd --resume 06c2423e-70fd-4af7-8416-4d6a2f8d8e87 --yolo

What happened / steps to reproduce

1. Run a long-lived agent conversation (this one had been active across compaction, with a large session transcript and many tool calls — screenshots, sub-agents, image reads).
2. Resume that session headlessly: cmd --resume <session-id> --yolo.
3. The process climbs to ~1.94 GB, GC runs become ineffective (mark-compact can't reclaim), and Node aborts with OOM.

Suspected cause

The resume path likely loads the full session transcript (which after many turns / compaction / large tool outputs — including read_file image data and agent results — grows large) into memory after the shared/global heap is already near the default Node limit. Because the process was launched by node without --max-old-space-size, it dies at the default 2 GB-ish ceiling rather than degrading gracefully. Possible contributors:

• Session transcript/context payload loaded in full on --resume.
• Tool outputs (screenshots/images/large read results from sub-agents) retained in memory instead of being streamed or evicted.
• No explicit heap sizing or soft-cap/memory-pressure handling in the resume path.

Expected behavior

A long session resume should either:

• not exceed the default heap (stream/evict large payloads), or
• surface a clear, recoverable error (e.g. "session too large, please compact/start fresh") instead of a hard abort that loses the work, or
• raise/self-tune the heap limit (e.g. --max-old-space-size) or chunk the resume.

Additional context

This happened on a session that had already been auto-compacted once and contained a large number of tool calls including image reads. It is the first reproducible-ish OOM in ordinary usage — no custom memory settings were in place.

Impact

• The resumed session crashes and the user loses the working context.
• Error message only points at the JS engine, not at Command Code — no hint to reduce session size or compact first.