Hey 👋 @calm-artist-7206 reviewed github.com/OpenHands/OpenHands/pull/15015 on the OSS side and suggested bringing it here for enterprise eyes, since it touches the shared conversation-metadata service. What it fixes: the conversation-start webhook (on_conversation_update) reads conversation metadata before the create commits, gets a stub with selected_repository=None, and the old merge() then nulls the stored repo. The repository silently vanishes from a running conversation (UI shows "no repo connected"). Same race as #14476, but the data-loss symptom the retry-only PRs (#14579 / #14940) don't cover, since re-merging the stub still writes None. The fix: git fields are write-once on the webhook path only. save_app_conversation_info gained preserve_git_fields_on_null (default False, COALESCE only when set); only the webhook passes True. Explicit update_app_conversation (repository=null to remove) still clears, exactly as before. Why it needs your eyes: it changes the base SQLAppConversationInfoService.save_app_conversation_info, which SaasSQLAppConversationInfoService inherits and calls via super(). So the default-clears / webhook-preserves behaviour applies to SaaS saves too. Would appreciate a sanity check that it plays nicely with the SaaS metadata / user-org consistency path and the existing injector assertions. Reproduced on a self-hosted box (~40% of repos nulled before, 0 after). CI green, 36 unit tests including the explicit-clear contract and the insert-race recovery. 🙏

cc @straight-sundown-84319 I'd love your 👀 on this

Thats correct - Self hosted in docker on my linux server at home