脈絡視窗提醒改幫你顧品質,而且提早發出
之前提醒你清理對話的脈絡視窗(context,AI 能記住的前後文範圍)通知,現在會提早在 70% 就跳出來,而且不再跟你算錢,而是幫你顧品質。因為模型在脈絡被塞滿之前,就會開始忘記長對話中段的內容。所以現在它會趕在品質開始掉之前提醒你:切換任務可以用 /clear,想繼續做可以用 /compact 保留重點。就像讀書一樣,與其等大腦完全塞滿才喊停,不如在裝到七分滿、開始恍神前,就先把重點整理下來。
發布日期:2026-06-14
之前提醒你清理對話的脈絡視窗(context,AI 能記住的前後文範圍)通知,現在會提早在 70% 就跳出來,而且不再跟你算錢,而是幫你顧品質。因為模型在脈絡被塞滿之前,就會開始忘記長對話中段的內容。所以現在它會趕在品質開始掉之前提醒你:切換任務可以用 /clear,想繼續做可以用 /compact 保留重點。就像讀書一樣,與其等大腦完全塞滿才喊停,不如在裝到七分滿、開始恍神前,就先把重點整理下來。
The notification nudging you to clean up your chat's context window (context — the range of previous messages the AI remembers) now jumps out earlier at 70%, and it no longer bothers lecturing you on costs, but focuses on preserving quality. This is because the model starts forgetting the middle parts of long conversations well before the context is actually full. So now it warns you right before quality drops: use /clear when switching tasks, or /compact if you want to continue but keep the focus. It's like studying: rather than waiting until your brain is 100% full to take a break, it's better to pause at 70%—before you start spacing out—and summarize the key points.
**Context-window nudge reframed around quality, and fires earlier (≥70%)**: the status line reminder added in 0.20.0 was framed around cost, but Claude Code (and Codex) auto-compact at ~80% and prompt caching makes resent context cheap — so the cost angle added little. What actually degrades as a conversation grows is quality: models lose the middle of long inputs well before the window fills. The nudge now triggers at 70% — ahead of the lossy automatic compaction — and suggests taking control yourself: `/clear` when switching tasks, or `/compact` to keep the focus you choose. The dollar figure was dropped.