You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit aac7be6
Browse filesBrowse the repository at this point in the historyBrowse files
feat(server): continue long idle threads in a fresh provider session
A long thread re-sends its whole history on every call. Once its prompt cache
has expired, the next request is uncached anyway, so that is the cheapest moment
to drop old history.
With the new setting on, a person's message to a thread over 100k tokens, after
an hour idle, skips resume and continues through the existing context handoff.
Commands, agent messages, and threads with background work are left alone. The
setting is off by default.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
description="After an hour away, continue a thread over about 100k tokens in a new provider session with recent context. Lowers cached input cost, though the agent may re-read some files."
Copy file name to clipboardExpand all lines: docs/user/portable-handoffs.md
+6Lines changed: 6 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -21,6 +21,12 @@ to retrieve saved messages and activity, including the remainder of a long item.
21
21
constraint, you can still repeat it in your next message. Unless the project has conversation
22
22
summaries turned on, a handoff is a budgeted selection, not a summary.
23
23
24
+
## Fresh sessions after a long break
25
+
26
+
Long threads get more expensive with every message, because the provider re-reads the whole history each time. With **Settings > Advanced > Fresh session after a long break** turned on, a message you send to a thread over about 100,000 tokens, after an hour or more idle, continues in a new provider session. It receives the same handoff described above, and your thread and its history in T3 Code stay as they were.
27
+
28
+
The first request costs more because nothing is cached, and the agent may re-read some files, so this suits threads you keep returning to. Runs with background work still in progress are left alone. The setting is off by default.
29
+
24
30
## Conversation summaries
25
31
26
32
With conversation summaries on, older history that would not fit in a handoff can arrive as short
0 commit comments