Skip to main content

DesireCore v10.0.183

Release Date: 2026-10-15

v10.0.183 更新概览

New Features​

  • Built-in General Assistant: Added a system built-in "General Assistant" agent supporting Q&A, writing, analysis, programming, and tool calling, with manual multi-session enabled by default. The offline marketplace also provides this assistant and a transparent avatar.
  • Browser automation and controlled launch: Added an external browser launch tool. After user approval, the Agent can launch a browser installed on the local machine and open a specified webpage. In offline environments, the runtime environment required for browser automation is automatically preconfigured.
  • Webhook management: Supports configuring webhook owners, enabling/disabling/deleting, and rotating shared secrets, and directly viewing available local model proxy addresses in the compute credential list.
  • Task board discovery: Agents can now discover task boards within the current permission scope through read-only tools, without relying on task board numbers from legacy plans.
  • Intranet deployment login: The enterprise intranet deployment version supports phone number and password login, mandatory password change on first login, session recovery and recovery after session invalidation, and displays available models and features according to deployment policy.

Fixes​

  • Fixed an issue where opening chat history on macOS could cause a gray screen or an inability to exit, and where rendering large or very long histories caused lag.
  • Fixed an issue where after a session was transferred to another Agent, the original session could still continue sending messages and possibly execute tasks repeatedly.
  • Fixed an issue where staged content of child teams was discarded when the parent team was disbanded, and prevented team working directories from being preempted or creating orphaned bindings during concurrent creation, disbanding, or vacancy filling.
  • Fixed an issue where an unresponsive AI approval service was incorrectly shown as "Security Denied". It now clearly shows "AI approval unavailable/not executed".
  • Fixed an issue where the original request in an approval card was truncated by platform reference. The actual request content is now fully preserved.
  • Fixed an issue where small SVG and similar image formats entering the model protocol caused subsequent request failures. A compatible preview is now generated and recovery is supported.
  • Fixed an issue where viewing a data source from a task board only opened the hidden chat file workbench and did not open a visible file page.
  • Fixed an issue where returning after viewing a file from a built-in app lost previously opened subpages, selected objects, and in-memory drafts.
  • Fixed an issue where returning to the reading view from the editing entry lost the previous zoom and pan state.
  • Fixed an issue where a failure to read the runtime environment version was incorrectly reported as "Not installed". It now shows "Read failed" and preserves the last known state.
  • Fixed an issue where nodes in the decision workspace with valid technical evidence were still blocked by repeated review, and where partial business examples were misjudged as infeasible due to non-target constraints.
  • Fixed an issue where after automatic compaction completed, other sessions in the same run lacked summaries.
  • Fixed an issue where usage from settled tool rounds was lost when a subsequent request failed or was interrupted.
  • Fixed an issue where network disconnection caused the Agent to be incorrectly shown as powered off.
  • Fixed an issue where a single operation timeout in the embedded process caused the entire channel to be closed.
  • Fixed an issue where no effective feedback was given when tool parameter parsing failed, and where the system could not wait for recovery within the original budget when the model was temporarily unavailable or in cooldown.
  • Fixed defects in offline Office skills related to Chinese XML validation and offline dependencies.
  • Fixed an issue where packaging continued to generate the client when the offline runtime directory for the target platform was missing. It now blocks packaging directly and prompts for runtime preparation.

Improvements​

  • Optimized chat history viewing: initially only recent records are loaded, scrolling loads paginated history, searches are processed in batches, the message list is virtualized, and window snapshot updates are merged.
  • Optimized offline runtime preparation: first use no longer depends on additional downloads, offline guidance is more complete, and the supported Node version range has been corrected to avoid version conflicts across different toolchains.
  • During chat execution, tool call failure notifications that the Agent is internally recovering from are hidden to reduce user distraction; approval recovery and interruption notifications remain unchanged.
  • Preserve original mathematical evidence, previous summaries, and audit records from before purely derived outputs, ensuring decision workspace results remain traceable and verifiable.