DesireCore v10.0.127
Release Date: 2026-08-28

New Features
- Scheduled backup task management: Added a scheduled backup task section in Settings where you can create, edit, enable/disable, and delete daily backup tasks, run them immediately, and view the next run time, last success, and failure reason.
- Automatic pre-backup before restore: When restoring from a backup, a rollback point is created for the current data by default. If successful, the restore proceeds; if it fails, the restore is blocked and can only continue by manually clearing the checkbox.
- Agent update risk confirmation: After pre-update risks are detected for an agent, you can click "Update anyway" in the risk notice and menu to continue to the existing update and conflict resolution flow.
- Delegation delivery boundaries: Delegated subtasks only deliver final results and authorized files/media to the main conversation. Their reasoning process and intermediate outputs are accessible only in "View full process."
- Deterministic operation guard: Long-running tasks such as delegations and subtasks are guaranteed to execute at most once, avoiding duplicate execution from retries or concurrency, and the overly low implicit task round limit has been removed.
Fixes
- Fixed an issue where models such as Zhipu GLM treated the built-in "Tool Search" as a lazy-loading flag and hid the remaining tools, preventing file creation. The tool has been renamed to "Tool Catalog."
- Fixed an issue where after a browser session was terminated due to resource quota, the agent still reported it as "opened." It now accurately reports that the session was terminated, the reason, and whether it can be recovered or should be recreated.
- Fixed an issue where clearing the conversation window could delete manual topics by mistake due to a cross-channel race condition, or messages could be lost after reconnecting.
- Fixed an issue where automatic retries after reasoning violations were displayed as English technical alert cards. Automatic recovery prompts are now shown with friendly copy and as regular informational messages.
- Fixed an issue where when the local marketplace cache lagged behind the remote, it was misidentified as many local commits, preventing the skill cache from syncing and updating automatically.
- Fixed an issue where late results for AI approval requests were discarded after 30 seconds. Late AI approvals or rejections can now still become valid final states.
- Fixed an issue where approval cards and interruption results still displayed Chinese in the English UI. Approval cards now switch with the interface language, and interruption results are saved in the language used when they were generated.
- Fixed an issue where subtasks inherited the main session canvas, current app context, and conversation/UI interaction capabilities. Subtasks can now only return governed file references.
- Fixed an issue where old tool rounds in model history polluted new requests, causing some compatible gateways to reject sessions.
- Fixed an issue where subsequent refresh requests were dropped when refreshing the model list, or cloud model requests were blocked by configuration sync. Refreshing is now more timely and smoother.
Improvements
- Multiple instances now share a single background agent process, with production builds and development environments sharing separately. This reduces resource usage and improves startup and connection stability across multiple instances.
- Settings entry is now unified under "Instances & Connections." New instance creation and system permission status are centralized in that panel, and the old entry in "About" has been removed.
- The wallet indicator at the bottom-right corner of avatars has been changed to a compact light gray secondary badge to improve readability on light-colored avatars.
- Topic summaries, conversation summaries, and read-only deep summaries now consistently use the lightweight summary model designated in the config center, making summaries more stable and lightweight.