Skip to main content

达摩院AgentOS v10.0.196

Release Date: 2026-11-02

v10.0.196 更新概览

DAMO Academy AgentOS v10.0.196​

New Features​

  • Agent archive and restore: Added the ability to archive, restore, permanently delete, and view agents while preserving team lead relationships and scheduling state.
  • Local audio/video transcription: Added an audio transcription tool that lets agents convert local audio or video into timestamped text and reuse the configured speech recognition service.
  • Large optimization model files and artifact references: The optimization modeling tool supports submitting large-scale models via file or artifact references, avoiding timeouts caused by generating parameters token by token, so compilation and solving can complete reliably.
  • Mobile file editing and viewing: Mobile supports secure text editing, file diff rollback, tool review, and authenticated read-only directory and screenshot viewing.
  • MCP service updates and upgrade protection: Installed MCP services can prompt for updates when a reviewed new version is available, and upgrades preserve user configuration such as environment variables, working directory, disabled state, and protocol preferences.
  • Skill update awareness: Built-in, inline, and some conversational model skills support content digest-aware updates while preserving local modifications and runtime state.

Bug Fixes​

  • Very long single-turn continuation: Long-running single-turn tasks over 16 MiB can now continue and no longer incorrectly report "history is empty"; for records without proper termination or that are corrupted, the actual reason is reported.
  • History pagination reliability: Fixed issues where session history pagination could skip surviving messages, mix up windows, or jump back to the bottom during concurrent deletion, rollback, or topic switching, and restored keyboard retry at the top.
  • Archived topic write protection: Archived manual topics can no longer be appended to through existing run records; they must be restored before continuing, and error messages no longer appear as generic API errors.
  • Task board recovery and read limit: After an interrupted task board write, creating and viewing tasks no longer hang, and cleanup can complete automatically after restart; the single-commit read limit is increased to 16 MiB, and exceeding it returns a clear error.
  • Team installation receipt: When a team member requires a higher client version, team installation or forking a team by address no longer returns 500, and installation receipts can be recorded normally.
  • Team member status detection: Team settings no longer incorrectly report local members that already exist on this machine as missing, and provide actionable paths to add, resolve, or remove members that are truly missing.
  • Team rules preserved on upgrade: After a distribution seed upgrade, team rules no longer silently stop working; when review is needed, a notification is sent and team settings can be opened directly.
  • Python runtime environment fix: Python source records left by the first offline import are rewritten to the official address, and the runtime environment page can reinstall and delete Python.
  • Tool output recovery: When the original model text or state projection is missing after a tool call ends, it can be recovered from the complete copy of the same run record; local cloud sync refresh errors are no longer misreported as upstream API errors.
  • Speech recognition language: Default automatic speech recognition no longer always recognizes Chinese; when no language is specified, the model determines it automatically, so English audio is no longer transcribed incorrectly.
  • Speech recognition presets made usable: Unavailable presets are removed, usable models are added, and local speech recognition models are treated as credential-free compatible options.
  • Approval policy real-time sync: After an approval policy directory appears, its listener can be created automatically; external changes to approval files sync in real time, and the UI no longer remains stuck on the old setting.
  • MCP process reaping: When an MCP stdio service shuts down, the process tree is fully reaped, preventing server processes started through wrappers from lingering.
  • App startup error messages: When an app lacks a startup contract, the exact cause is reported, and delegation takes the correct repair path instead of repeatedly dispatching useless "check and open" tasks.
  • Task progress tools: During long session wrap-up, task progress tools are loaded, retained, and evicted by group, and the agent no longer keeps creating junk tasks such as "placeholder/ignore".
  • Rollback and search scope: When manual topic rollback fails, the full history and recovery entry are preserved; deleting old messages no longer removes the remaining results in the same search scope from the active window entirely.

Improvements​

  • Tool call hints: Built-in tool facade parameters indicate applicable actions, unknown parameters point to the correct level, and task update tools are narrowed for the model with alternative syntax provided, reducing invalid calls.
  • Whiteboard error details: When whiteboard initialization fails, the specific reason and location in the document are provided, and structure definitions can be correctly retrieved before creating a new one.
  • Task board performance: After backfilling old task board history, the directory is activated automatically; duplicate checks only read the first commit, making very large task boards and startup recovery faster.
  • Tool declaration persistence: Tool declaration recent-use and failure-closed records persist with the session, so after rebuilding, tools are still evicted based on recent use.
  • MCP connection semantics: MCP connection tests and health probes use unified timeout, cleanup, and error classification, with clearer and sanitized failure messages.
  • MCP forward compatibility: Strict validation tolerates unknown fields in MCP configuration, and source summaries exclude unrecognized keys, preventing tools for scheduled tasks and web callbacks from being silently dropped.
  • Approval mode sync: Approval modes are ordered by server epoch and revision number, external changes arrive in real time, and unreadable policy files show a warning and repair entry in the UI.
  • Built-in skill marketplace sync: Fixed outdated path instructions for Python and Node.js environment skills, preventing agents from repeatedly installing environments with the system package manager.
  • Security patch upgrades: Upgraded desktop runtime, HTTP parsing, validation, and chart rendering to security-fixed versions while keeping native capabilities and visual effects unchanged.