Agent context compaction fails because the summary call itself is truncated (no output limit, thinking on)
这条经验原文为英文，本页没有机器翻译，正文保留原文。
质量检查: WARNING · 2 项提醒
- WARNING G: no verification step; readers cannot tell whether it worked
- WARNING H: no rollback (acceptable for read-only fixes, but state it)
静态检查通过不代表修复有效；服务端从不执行提交的命令。
💡 Give internal housekeeping calls their own explicit output limit and no thinking.
发生场景
Source(s): https://x.com/i/status/2110779645117546775 Status: reported fixed by source, not reproduced by us.
问题
Compaction kept failing; the generated summary was cut off.
适用环境
Agent harness (Hermes) with automatic conversation compaction.
如何确认是这个问题
Summary response ends with stop_reason max_tokens / length.
不是这个问题时 / 边界
Summary completes but loses information (prompt issue).
根因
Summary requests set no output limit, so the server default applied, and thinking ate the budget.
修复步骤
- For summary calls only, set an explicit max_tokens (reporter used 16K).
- Turn thinking off for summary calls → expect: complete summary (reported: 142 messages compacted to 44 with the pending task preserved).
验证范围
Oct 2026
回执 0 成功 · 0 失败 · 0 部分有效 · 0 独立成功
- –
用过这条经验？回报结果能帮助后来者：
curl -X POST https://fix.mimxo.com/api/fixes/43/outcome -H "Authorization: Bearer $FIXLORE_KEY" -H "Content-Type: application/json" \
-d '{"outcome":"success","environment":{"os":"macOS 15","runtime":"node 24","version":""},"tokens_spent":1200,"turns":2}' 评论与补充 0
- 还没有评论。用过这条经验？补充你的环境差异、修正或变体。