hermes-admin/hermes-agent
1
0
Fork 0
Code Issues Pull Requests Actions 4 Packages Projects Releases Wiki Activity

fix(gateway): never persist a hygiene compression that grows the transcript

Browse Source 
Session hygiene could persist a compressed transcript LARGER than the
original (observed: 427K -> 598K), when the generated summary was bigger
than the middle it replaced. Compare like-for-like (both rough estimates)
before persisting a rotated transcript; on growth, keep the original
unchanged so a failed compression is a strict no-op, never a net increase.
This commit is contained in:
dhruv kejriwal
2026-08-10 16:37:59 +00:00
committed by kshitij
parent bc36d7d6c8
commit e7418f1621
1 changed files with 25 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
gateway/run.py
View File

@@ -19031,6 +19031,31 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
_hyg_in_place = bool(
getattr(_hyg_agent, "_last_compaction_in_place", False)
)
# Anti-regression guard: never persist a
# compression that did not shrink the
# transcript. A summary that is larger than
# the middle it replaces would GROW the
# session instead of reclaiming it (observed:
# 427K -> 598K). Compare like-for-like (both
# rough estimates) so an "actual vs
# estimate" measurement mismatch can't
# produce a false verdict. On growth, keep
# the original transcript untouched.
if (
_hyg_rotated
) and estimate_messages_tokens_rough(
_compressed
) > estimate_messages_tokens_rough(history):
logger.warning(
"Gateway hygiene compression for session %s "
"would grow transcript (~%s -> ~%s tokens); "
"keeping the original transcript unchanged",
session_entry.session_id,
f"{estimate_messages_tokens_rough(history):,}",
f"{estimate_messages_tokens_rough(_compressed):,}",
)
_hyg_rotated = False
_compressed = history
# Only rewrite the transcript when rotation produced
# a NEW session id. In-place compaction does NOT
# need a rewrite: archive_and_compact() has already