Skip to main content

AgentOS v10.0.186

Release Date: 2026-10-22

v10.0.186 更新概览

New Features​

  • Simplified model selection: The model selector in the chat input area now shows the full model list directly, supports cross-source search and source/vendor filtering, and saves filter settings locally. Auto-select can be clicked directly just like a model; its settings are provided in a separate popover where you can check sources and drag to reorder them.
  • Layered reading for thinking and tool processes: Thinking records are collapsed by default. During generation, “Thinking...” and a continuous shimmer are shown. Consecutive work in the same round is grouped into action summaries, which can be expanded layer by layer to view per-item summaries and scrollable details. Detail panels are height-limited and scroll independently from the main chat. The first thinking segment creates a message anchor.

Fixes​

  • The login page now correctly displays password, SMS verification code, registration, and password recovery entry points based on the server authentication configuration, and temporarily disables email verification code login.
  • When switching models or selecting the current agent again, the input box is no longer unmounted, preserving the content being typed, focus, and IME state.
  • After window recreation, the interface can no longer remain stuck on “Running”; it automatically restores the terminal state and history. When synchronization fails, it provides a prompt and manual retry, and does not resend model requests or tools.
  • In enterprise intranet environments or environments that block public network probe addresses, ongoing inference no longer aborts or is misjudged as offline because of a public network probe failure.
  • Fixed an issue where built-in agents or teams were not installed on first open on some distributions, and existing data could be overwritten. First open is now determined by a flag; existing content is only supplemented when missing, and original data is left untouched on conflict.
  • Fixed an issue where, when the model service returned rate limiting, the concurrency slot was released before the cooldown was published, causing queued requests to repeatedly hit the rate limit. It now waits correctly during the shared cooldown period.
  • Fixed multiple issues, including: a server heartbeat event with no data causing the conversation to terminate with a non-retryable error; direct requests hanging before response headers without a controlled timeout; and immediate follow-up after stopping being rejected.
  • Fixed an issue where tool error-stage facts were not loaded and were lost during cross-user turn recovery, which could cause the model to deny the actual receipt from the previous turn.
  • Fixed inaccurate classification when the model returned truncated content or an in-stream quota error: truncation missing the end marker is now retried as a stream interruption, and in-stream quota errors are no longer mistaken for HTTP errors, making retry prompts more accurate.
  • Fixed the boundary between worktree preparation and task planning preconditions, avoiding duplicate plan creation and correctly reading the verified reference plan.
  • When a native app is shown again, it no longer displays a blank screen; instead it shows a loading indicator and restores the original editor after success.

Improvements​

  • Optimized file watcher maintenance during skill synchronization and installation. Leases are acquired on demand only at the first actual publish, reducing unnecessary watcher rebuilds and lowering the risk of jank.
  • The model selector keeps list brightness and selection marker placeholders during saving and configuration re-reads, avoiding flicker and text shift when switching. When a pinned model becomes invalid, the selector provides a recovery entry while preserving drafts and attachments.