达摩院AgentOS v10.0.186

Release Date: 2026-10-22

DAMO Academy AgentOS v10.0.186

New Features

• Layered reading of thinking and tool processes: Long thinking is collapsed by default. While generating, "Thinking..." and a continuous sweep animation are shown. Consecutive tool executions within the same run are merged into an action summary, and users can expand layer by layer to view details, preventing long thinking or tool results from overwhelming the chat window.
• Simplified model selection and automatic selection settings: The chat input area now directly expands the complete model list, with cross-source search and filtering. Automatic selection can be clicked directly, and its source scope and priority order can be configured in separate settings and saved per current agent.

Fixes

• When switching models or reselecting the current agent, the input box is no longer unmounted and rebuilt; unsent drafts and ongoing IME compositions are preserved.
• In restricted network environments, a failed public network probe no longer aborts in-progress inference or misjudges the client as offline; network retries can recover normally.
• Errors or unsent messages no longer occur when some model services return only heartbeat events, remain unresponsive for a long time before response headers, or when a follow-up is sent immediately after stopping generation.
• After the window is rebuilt, the chat interface automatically restores state and history even if the task has already completed, instead of continuing to display "Running".
• When restoring history across user turns, failure notes for tools that failed to load are preserved, and the model no longer denies the actual execution results of the previous turn.
• On first launch, the built-in solver agent is no longer skipped due to leftover files from earlier versions; existing content is only filled in where missing, not overwritten.
• When re-displaying a kept-alive native application, a loading indicator is shown while waiting; after success, the original editor is restored instead of a blank screen.
• The login screen now displays phone verification code, registration, and password recovery entries according to the authentication configuration; email verification code login is temporarily hidden and disabled.

Improvements

• The model management entry at the top of chat has been consolidated into the working directory and resource explorer entries. After model selection, the menu remains expanded, making interaction simpler and switching more stable.
• During model saving and configuration backfill, list brightness and selection markers remain stable, reducing flicker and text jumping.
• Error classification, empty response marking, and rate-limit retries are handled according to the actual situation of each request, making retry prompts and stream interruption explanations more accurate.
• The rate-limit cooldown is published before concurrency slots are released, so queued sessions wait until the cooldown ends before making requests, reducing consecutive rate-limit triggers.
• Built-in skill synchronization and installation now merge operations only when an update is actually needed, reducing extra overhead when there is no update and reduging UI lag on some desktop environments.
• The installer now more accurately identifies blocking processes and installation directories, making installation and upgrade more reliable.