You'll noticeTier: how much it should matter to you
2Useful: my rating, 1 to 5
1Signal: worth watching, 1 to 5
CompactionArea: what it touches
ImprovementsKind: in v2.1.292,
ImprovementsSection of the release
What
Compaction is when Claude Code shortens a long conversation by summarising older parts. It now tracks how many requests to the model it has left, through compactionApiAttemptsLeft.count, and acts when that reaches 0:
If the cache-sharing response has no text and the count is 0, the response is returned. Before, it always fell back with the warning Compact cache sharing: no text in response, falling back.
The compaction loop returns { ok: !1, reason: "exhausted" } when the count is 0.
A new compaction reason, compact_larger_model, is also listed.
Why
Compaction no longer keeps making requests after its attempt budget is used up, which avoids extra calls to the model during compaction retries.