Skip to main content

DesireCore v10.0.61

Release Date: 2026-05-19

v10.0.61 更新概览

New Features​

  • Team Git Version Management: You can now manage teams as versionable resources, connect them directly to remote repositories on GitHub, Gitee, or Gitea, and push, pull, publish, or install team forks with one click.
  • High-Precision Math Tool: Agents now use a new high-precision math capability for numerical calculations, matrices, unit conversions, fractions, and similar scenarios, reducing errors in complex computations.
  • Model Capability Indicators: The model selector in chat now shows whether a model supports capabilities such as vision, so you can see what it can and cannot do before switching.
  • Paste File References into Chat: You can now paste local files or folders directly into the chat input box and send them as file references, while pasted images continue to use the existing upload behavior.
  • Immersive Voice Interaction: Immersive mode now includes voice input and voice playback, supporting continuous speech recognition, speech synthesis, and more natural conversational feedback.
  • AI Next Message Prediction: After the assistant finishes replying, smart prediction suggestions will appear above the input box, allowing you to accept common next-step inputs with one click.
  • Immersive Decision Overlay: In immersive mode, approval cards and question cards are now displayed as overlays, so key interactions that require your attention no longer disappear when entering immersive mode.
  • /steer Mid-Execution Intervention Command: You can now use /steer during Agent execution to send new intervention instructions, and see messages marked as “Intervened” in the chat history.
  • AskUserQuestion Timeout Mechanism: When an Agent asks you a question, it can now set a timeout countdown so the task does not remain stuck waiting indefinitely after the timeout expires.
  • Skill Directory Context Injection: Skills can now automatically detect their own directories and resource manifests, making the use of dependency scripts, reference files, and attached resources more stable.
  • Enhanced Workflow Validation: Workflows can now detect issues earlier before execution, such as conditional branches with no follow-up path or incorrectly filled node fields, reducing cases where they appear successful but do not actually finish running.
  • Compute Token Plan Support: Compute configuration now supports Token Plan-related fields, providing compatibility for the new team metering scheme.

Fixes​

  • Fixed an issue where the message component could error or crash when some models initially returned an empty reply and only filled it in later through streaming.
  • Fixed an issue where Markdown body content in the notification center was not displayed correctly; bold text, inline code, lists, and other formatting now preview properly.
  • Fixed an issue where inline diff decorations in SuperDoc would sometimes not appear, making document edit markers for additions, deletions, and changes show up more reliably.
  • Fixed an issue where editing .md files could sometimes produce an empty assistant bubble, and now automatically re-sends an openable file card when the model does not proactively send an attachment.
  • Fixed an issue where skill names and descriptions in the runtime status panel did not switch along with the UI language.
  • Fixed an issue where quickly switching languages or agents could cause the skill list to be overwritten by results from an outdated request.
  • Fixed an issue in the merged message view where old versions of a user message could remain after repeatedly editing the same message.
  • Fixed an issue where memory from old topics could still be injected into later conversations after history truncation.
  • Fixed an issue where, after clicking the version number on the settings page to check for updates, the UI could remain stuck on “Checking.”
  • Fixed issues with the manual version switching dialog, including insufficient width, unclear loading states, and confusing download prompts.
  • Fixed an issue where compute configuration and node interfaces could return 500 errors under certain startup paths.
  • Fixed an issue with incorrect path resolution for the default repository zip after packaging, preventing built-in repositories from failing to extract properly on first launch.
  • Fixed an issue where TLS certificate renewal after expiration depended on the startup order of other services, preventing connection failures caused by expired local service certificates.
  • Fixed multiple issues in Windows automatic updates that could cause upgrade failures, significantly improving upgrade success rates under both Program Files and standard installation directories.
  • Fixed an issue where dirty cache state in the marketplace repository could cause sync failures, and improved automatic self-healing for abnormal cache states.
  • Fixed an issue where Unix executable permissions were lost after zip extraction, which could cause repositories to be mistakenly detected as locally modified.
  • Fixed an issue where the main process could fail to build after dependency upgrades, preventing the application from starting.
  • Fixed an issue where running tasks could still be incorrectly shown as “Running” after the app restarted, and also addressed similar webhook status anomalies.
  • Fixed an issue where scheduled tasks could remain stuck on an old run for a long time and stop generating new execution records.
  • Fixed an issue where, after the compatible proxy injected the server-side web search tool by default, some model replies could be polluted by search result text.
  • Fixed a misleading error where interrupted streaming responses due to network issues still prompted you to check the model configuration; the system can now distinguish network interruptions from empty model responses more accurately.
  • Fixed status display and rollback issues in AskUserQuestion during scenarios such as skip, timeout, reconnect after disconnection, and submission failure, preventing question cards from getting stuck or showing inconsistent states.
  • Fixed several race condition issues when submitting or skipping question cards in immersive mode, reducing mistaken rollbacks and unresponsive behavior.

Improvements​

  • AskUserQuestion Interaction Experience: Question cards now show in real time how many questions remain unanswered, support keyboard shortcuts for submission, and automatically jump to incomplete questions, making long forms easier to complete.
  • Models with limited vision capabilities now automatically disable image upload, screenshot, image paste, drag-and-drop, and similar entry points, avoiding situations where you only realize the selected model does not support vision after trying to use those features.
  • Immersive mode now offers more ways to enter, including keyboard shortcuts, entering directly from attachments, and Agent-initiated requests to switch, making it easier to access.
  • The startup sync process for marketplace and built-in repositories is now more stable, making it more reliable to obtain complete default content even during an offline first-time installation.
  • Audit logs now retain complete history by default and once again record streamed response chunks, making later tracing and troubleshooting easier.
  • /steer intervention messages in chat are now persisted to history, so you can still trace which instructions you changed mid-execution after refreshing or restarting.
  • After optimization of the skill auto-sync logic, startup is now more efficient, and scheduled background sync no longer repeatedly performs unnecessary scans of built-in content.
  • After model and dependency upgrades, the overall runtime environment, icon resources, search capabilities, and desktop shell compatibility have all been improved.