Exec 大输出防上下文爆炸：落库前截断与摘要化保护
用于“单次 exec 输出过大导致会话超 token 上限并不可恢复”场景。前置：能修改工具结果写入链路。步骤：在写入 session 前执行长度截断/摘要→保留原始输出到文件路径→在 transcript 仅记录摘要+引用→用 100KB+ 输出做回归。关键点：截断必须发生在会话写入前，而非仅 UI 事件流。验证：超大输出不再触发 prompt too long。风险：过度截断可能丢失排障细节。
GITHUBDiscovered 2026-02-15Author QiushiWu95
Prerequisites
- You can patch or configure tool-result persistence behavior in your OpenClaw deployment.
- You have a reproducible large-output command for regression testing.
Steps
- Reproduce with a known large-output exec call and capture the token-overflow failure signature.
- Apply truncation/summarization before session transcript write, not only in stream rendering.
- Persist full raw output outside transcript (file artifact/path) and store only pointer + summary in session.
- Run stress tests with 100KB-1MB outputs and confirm session remains usable.
Commands
openclaw statusopenclaw gateway statusVerify
Large exec output no longer corrupts session context; follow-up prompts still execute normally.
Caveats
- If only UI stream is truncated while transcript stores full text, this issue still reproduces.
- Retention policy for full raw artifacts should be reviewed for privacy/compliance（需验证）.
Source attribution
This tip is aggregated from community/public sources and preserved with attribution.
Open original source ↗