A coding agent keeps your whole conversation in its head. When that gets too big, it writes a summary and carries on from the summary. Compact too early and you keep paying to rebuild the same context. Compact too late and every request drags a heavy load. These charts show where the sweet spot sits, using real sessions.编程助手会把整段对话都放在脑子里。对话太长时,它会写一份摘要,再从摘要接着干。压缩太早,就要一遍遍花钱重建同样的上下文;压缩太晚,每次请求都背着沉重的负担。这些图用真实的使用记录告诉你,平衡点大约在哪里。
Each page has a chart and a few ready-made picks for the trigger setting.每个页面都有一张图,还有几个现成的触发点建议。
Codex sessions on DeepSeek Flash. Cached input is very cheap, so the cheapest triggers are fairly low and the curve is flat near its minimum.Codex 里 DeepSeek Flash 的会话。缓存输入非常便宜,所以最便宜的触发点相对较低,最低点附近的曲线也很平。
Cheapest near a 116K trigger触发点在 116K 附近最省钱
Open the chart查看图表 → Claude CodeClaude CodeClaude Code transcripts on Opus 5.5, main conversations and subagents together.Claude Code 转录里 Opus 5.5 的使用记录,主对话和子代理一起算。
Cheapest near a 121K trigger
Set autoCompactWindow to 154000触发点在 121K 附近最省钱
把 autoCompactWindow 设为 154000
Claude Code transcripts on Sonnet 5.5. Few compactions of its own, so its timing is borrowed.Claude Code 转录里 Sonnet 5.5 的使用记录。自己的压缩很少,所以压缩耗时是借来的。
Cheapest near a 105K trigger
Set autoCompactWindow to 138000触发点在 105K 附近最省钱
把 autoCompactWindow 设为 138000