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
Changed: a summary that is too long is asked to leave room. The retry message told the compressor to cut "just enough" to fit, and many retries missed the limit again by a few bytes. Each miss now asks for a shorter line: 32 bytes under the 512-byte limit after the first miss, 64 after the second, and so on, to at most half the limit. The limit itself, the five tries and the rule that a stored line is never truncated are unchanged.
Fixed: more preambles are cut from summary lines. The guard now recognizes a ⁂ in front of a label (⁂ Compaction:), a message-range label (Msgs 5672–5687 (high, finding):), labels with to, into, one, single, result or ≤512 in them (Merged line for messages 8 to 15 (under 512 bytes):), and a one-line compaction label followed by a blank line. A reply that opens this way is retried, as before; if every try repeats it, the stored line has the preamble cut.
Metrics evidence (live stores, 2026-11-09 to 2026-11-14, no paired control): 41% of compressor calls were retries. Length retries repeated most: of gpt-6.1-terra merges that got a length retry, 43% used all five tries, and for Haiku 5.5 it was 8%. 1,868 stored lines were over 512 bytes, with a median of 520. The new guard matches 1,655 of 40,480 lines stored in that window (4.1%) that the old guard let through; a read of the matched openings found no ordinary summary among them. The effect of both changes on retries, cost and stored sizes is not measured yet.
Checked, no change: chat_search returning 0 hits for a hyphenated term was correct. The term first entered the log in the search call itself. A check of host compactions after 0.16.5 found no compaction rows yet.
Tests: ten head cases failed before the guard change and pass after.