diff --git a/docs/user/providers-claude.md b/docs/user/providers-claude.md index c94fc3420f875..0ed918a9d0580 100644 --- a/docs/user/providers-claude.md +++ b/docs/user/providers-claude.md @@ -56,7 +56,7 @@ compacts automatically whether the meter is on or off. When you return to a large thread after Claude's prompt cache expires, a **Compact** chip with the thread's token count shows next to the send button. While it is on, Enter summarizes the history first, then -sends your message. Click the chip to switch it to **Full** and keep the full +sends your message. Click the chip to switch it to **Keep** and keep the full history for that message. The cache lasts about an hour on a Claude subscription and 5 minutes with an API key, Bedrock, or Vertex. See [commands and skills](./composer.md#commands-and-skills) for using composer commands. @@ -65,9 +65,12 @@ composer commands. over 200,000 tokens. Each Claude provider's settings can override it: - **When you return** shows the **Compact** chip described above. This is the default. -- **Before the cache expires (beta)** compacts an idle thread shortly before its - prompt cache expires, so coming back does not reprocess the whole conversation. +- **Before the cache expires (beta)** tries to compact an idle thread shortly before its + prompt cache expires, while the history is still cached. The compaction appears in the thread as a `/compact` turn sent by T3 Code. + If the window is missed, such as when you enable this mode after the cache has + expired or the server restarts before compaction runs, it falls back to + **When you return**. Switch the chip to **Keep** to keep the full history for that message. - **Never** skips idle compaction and Claude's resume prompt. **Auto-compact after** still applies, and you can still compact by hand.