Skip to main content

DesireCore v10.0.198

Release Date: 2026-11-03

v10.0.198 更新概览

New Features​

  • Built-in plugins and open-source notices: The plugins page now has a “Built-in plugins” group for managing the built-in FFmpeg suite (toggle, uninstall, re-add). The About page now includes entries for open-source licenses and third-party notices. This version does not actually ship with FFmpeg yet.
  • MiniMax speech synthesis adapter: Added a MiniMax speech synthesis protocol adapter. After configuring the API key, you can use its speech synthesis models.
  • Official marketplace inline app definitions: Apps that use inline definitions can now be installed from the official marketplace, fixing an issue where some apps could not be installed from the marketplace.

Fixes​

  • Task board per-board isolation: Task board listing and queries now skip individual task boards that cannot be read instead of failing the entire request with a 500, and they show the number of skipped boards.
  • Streaming disk write performance: Streaming incremental writes no longer open and close the session log file for each item, greatly reducing lag caused by file system filter drivers on Windows and improving throughput by up to thousands of times.
  • Model silent cutoff notice: After a model request is silently cut off, the next request includes a factual hint to help the agent adjust its strategy and reduce pointless retries.
  • MCP tool parameter validation: MCP tool parameters are no longer automatically coerced. Type mismatches now return validation errors, and approved parameters remain consistent with the parameters sent to the server.
  • App start/stop approval: Fixed context propagation for app start/stop approval. Error codes are no longer overwritten, and concurrent waiters receive the correct reason.
  • Skill relative paths: The skill directory is now included in skill loading results, so the model can correctly resolve relative-path scripts bundled with a skill.
  • Team artifact evidence ledger: After team artifacts exceed the previous limit, the evidence ledger no longer fails permanently, and recovery proofs can still be generated.
  • Manual topic recovery: After a page reload, idle execution for manual topics recovers correctly and is no longer invalidated by concurrent writes.
  • Default model voice selector: The speech recognition and speech synthesis selectors now only list models that pass capability checks, preventing selection of unavailable models.
  • Git skill import: Git import of root skills now preserves nested directories. The Agent skill list respects skill directory boundaries and no longer produces extra <id>/<sub> entries.
  • Deep tree projection and preview scrolling: Fixed depth calculation that caused valid deep tree projections to be incorrectly rejected with a 413, and fixed an issue where decision tree preview content could not scroll within the grid.
  • Local model recovery: Improved local model recovery and offline request experience. Fixed recovery races, timeout overrides, and waiting status display, and now distinguishes incomplete output from channel failures.
  • MCP connection stability: Multiple stability fixes for MCP shutdown, health probes, and stdio process reclamation, including normal cancellation, gate generations, Windows shutdown diagnostics, and task termination behavior.

Improvements​

  • Built-in tool naming and loading: Built-in tool names are clearer and have distinct responsibilities. Discovery calls now use explicit loading, message delivery modes are mutually exclusive, and inbox deletion is independent.
  • Script resource protection: Agent Service no longer crashes when a script consumes excessive memory, CPU, or produces large amounts of output. It now lowers script priority, collects output in a bounded manner, provides commit memory guarding on Windows, and explains the reason in tool results or banners.
  • MCP unavailable tool display: During MCP tool discovery, unavailable tools are listed separately with diagnostic codes, and the store service details page shows the specific reason.
  • Voice selector experience: The voice selector shows check status and retry options, selected models are marked with unavailability reasons, and no fake empty list is shown when a check fails.
  • Built-in marketplace skill updates: Built-in marketplace skills were updated. The watch skill now prefers the speech transcription tool when no subtitles are available, and the detection script escapes paths when outputting JSON.