Summary

Gateway warm sessions can accumulate merged \n\n-joined user blobs in the model payload while platform inbound stays single-message per turn.

Observed behavior (reporter session 20260612_085708_c4403c61, 2026-06-12 weixin)

Minimal repro segment 13:15–13:17 CST:

• 13:15:41 — Turn ended last_msg_role=assistant, reply sent to WeChat; AxonHub request feat(agent): support full OpenRouter provider routing options #246 tail is correct (user → assistant → tool → assistant).
• 13:17:20 — Next inbound single user message; AxonHub fix(agent): skip reasoning param for Mistral API to prevent 422 errors #252 tail shows assistant/tool chain 321–323 gone, last user already two messages merged.
• turn_context history count: 342 → 343 (+1 only) — only the inbound user row landed in state.db, not the assistant/tool chain.

Cache eviction after ~1h idle was an amplifier, not the root cause (blob already 5 segments before evict).

Root cause chain

1. turn_context early-persists inbound user at turn start → _last_flushed_db_idx advances past pre-turn length.
2. First API call runs repair_message_sequence() which merges consecutive users (orphan + new) and shrinks messages in place.
3. Flush cursor stays at the pre-compaction index → turn-end _flush_messages_to_session_db computes flush_from past the first post-repair assistant row and skips persisting the assistant/tool chain.
4. Next turn loads orphan user → new user → repair merges again → blob grows.

#44327 (reset _last_flushed_db_idx on cached-agent fresh turns) is necessary but not sufficient when step 2 compacts within the same turn.

Proposed fix

• After repair_message_sequence compacts the list, clamp _last_flushed_db_idx to len(messages).
• In _flush_messages_to_session_db, if cursor overshoots list length, fall back to start_idx.
• Include Gateway cached-agent reuse can leak _last_flushed_db_idx across turns and skip assistant transcript rows #44327 cached-agent cursor reset.

Community PR incoming from @JerryLiu369 fork.