What happened? / 问题描述

近期多个长会话场景会出现 Node/V8 OOM 崩溃，例如：

FATAL ERROR: Ineffective mark-compacts near heap limit Allocation failed - JavaScript heap out of memory

从已有报告和本地排查看，这类崩溃通常发生在长会话、大上下文模型、大量 tool output / diff / file read、或 /compress 附近。一个典型 GC 片段显示 heap 已接近 4GB limit，Mark-Compact 后只能释放很少内存，说明大量 JS 对象仍然 reachable：

Scavenge ... 4085.8 -> 4082.3 MB allocation failure
Mark-Compact (reduce) ... 4088.3 -> 4079.8 MB
FATAL ERROR: Ineffective mark-compacts near heap limit

目前 compaction 主要基于 token/context threshold 触发。对于 1M context window 的模型，70% token threshold 可能非常晚才触发，但 V8 heap 会先被 conversation history、tool result、UI history、compression 临时拷贝、stream/logging accumulator 等对象撑满。

What did you expect to happen? / 期望行为

长会话在 V8 heap 接近 limit 前应该有轻量 safety net：

• 不应只依赖 token/context threshold 才触发 compaction。
• 应该基于 heapUsed / heap_size_limit 做 heap-pressure guard。
• 高 heap pressure 下应尽量避免额外 full-history structuredClone()。
• 对大 tool result / large command output 应有预算或 offload 策略，避免长期保留在 hot JS heap 中。
• /doctor memory 应能辅助判断是 V8 heap pressure、large tool result retention，还是 compression peak。

Analysis / 初步原因分析

这看起来不是单一 bug，而是几类 OOM 场景叠加：

1. token threshold 与 V8 heap pressure 脱节

   例如 1M context window 下，70% compaction threshold 约等于 700K tokens。实际 JS 对象表示会有很高放大，包括 nested Content / Part 对象、React/Ink UI history、tool result、logging buffers 等，V8 heap 可能在 token threshold 前先到 2GB/4GB/8GB limit。

2. getHistory() 的 full-history clone 放大峰值

   GeminiChat.getHistory() 会 structuredClone() 整个 history。部分路径只是需要读最后一条消息或只读 history，但仍然触发全量 deep clone。长会话时，这会显著放大瞬时 heap 峰值。

   一个明显例子是 nextSpeakerChecker 同时调用：

   chat.getHistory(true)
   chat.getHistory()

   第二次 comprehensive history clone 主要用于检查最后一条消息，理论上可以用 O(1) accessor 替代。

3. compression 本身可能制造 peak memory

   /compress 或 auto-compaction 时，流程会读取 curated history、切分数组、构造 compression request、接收 summary response。若 heap 已经很高，compression 过程本身可能成为 OOM 触发点。

4. large tool output / stream/logging retention

   长会话中大 diff、file read、shell output、stream chunks、logging responses 可能长期保留在 live history 或 logging buffers 中。这个方向也与 Diagnose and mitigate large tool-result retention in long sessions #4184 相关。

Related issues / 相关 issue

• problem critical error #4116 — long-session OOM; analysis points to structuredClone() / large context window pressure.
• cli crashed #4167 — crash during or near compression; useful for compression peak memory analysis.
• FATAL ERROR: Ineffective mark-compacts near heap limit Allocation failed — JavaScript heap out of memory #4149 — recent Ineffective mark-compacts near heap limit report.
• 运行出现OOM错误 #4134 — recent OOM report.
• Heap out of memory #2868 — fast heap growth OOM.
• FATAL ERROR: Ineffective mark-compacts near heap limit Allocation failed - JavaScript heap out of memory #2945 — OOM around /resume.
• Reduce memory usage of long-running tasks #2036 — reduce memory usage of long-running tasks.
• fix(core): structuredClone OOM in long sessions #2562 — structuredClone OOM in long sessions.
• [P3] Memory Diagnostics / 内存诊断 #3000 — memory diagnostics parent issue.
• Add baseline /doctor memory diagnostics #4179 / Add interpretation layer to /doctor memory diagnostics #4181 / Add structured /doctor memory output and safe session-scale stats #4182 — /doctor memory diagnostics slices.
• Diagnose and mitigate large tool-result retention in long sessions #4184 — large tool-result retention diagnostics / offload direction.

Related PRs / 相关 PR

• fix(core): memory-based chat compression to prevent heap OOM #4127 — adds memory-based compression trigger using fixed heap thresholds. Similar goal, but may need ratio-based threshold using v8.getHeapStatistics().heap_size_limit.
• feat(core)!: redesign auto-compaction thresholds with three-tier ladder #4168 — redesigns auto-compaction token thresholds with a three-tier ladder. Useful but larger in scope and mostly token/context oriented.
• feat(cli): add baseline /doctor memory diagnostics #4180 / feat(cli): add structured memory diagnostics JSON #3785 — /doctor memory diagnostics.
• feat(telemetry): add interaction span and detailed sensitive attributes #4097 / feat(telemetry): unify span creation paths for hierarchical trace tree #4126 — telemetry/logging span changes; relevant because logging paths may affect retention if enabled.
• feat(core,cli): two-phase session listing for instant /resume first frame #3989 / feat(cli): add fork-session resume flag #4159 / feat(worktree): Phase C — session persistence, hooksPath, Footer + WorktreeExitDialog, three-mode --resume restore #4174 — resume/session-related changes; relevant to /resume memory behavior.

Suggested minimal mitigation / 建议的最小修复方向

A minimal, lower-conflict fix could be:

1. Add a heap-pressure guard in GeminiChat.tryCompress() based on:

   process.memoryUsage().heapUsed / v8.getHeapStatistics().heap_size_limit

   This avoids fixed 2GB thresholds and adapts to 2GB/4GB/8GB heap limits.

2. Avoid full-history clone in nextSpeakerChecker by adding a small accessor such as getLastHistoryEntry() / peekLastHistoryEntry() and cloning only the last entry if mutation safety is required.

3. Treat advanced.autoConfigureMemory / --max-old-space-size as mitigation, not root fix. It gives headroom, but does not address retained history/tool-output/compression peak.

Larger follow-ups can separately handle large tool-result offload, stream/logging accumulator reduction, and post-compaction release of old history references.

Client information / 客户端信息

Qwen Code v0.15.11
Model: qwen3.6-plus
Fast Model: qwen3-coder-flash
Auth: openai
Platform: darwin arm64 (24.1.0)
Node.js: v22.22.0
Git commit: 782403d71