Skip to main content

DesireCore v10.0.92

Release Date: 2026-07-20

v10.0.92 更新概览

New Features​

  • Interactive terminal system: You and the agent can share a single terminal panel and collaborate, with persistent sessions and simultaneous control of multiple terminals running CLI tools like claude, codex, gemini, aider, etc.
  • Manual multi-session mode: Specific agents support creating, switching, renaming, and deleting independent sessions, allowing you to manually manage separate, non-merged conversation histories just like Claude Code.
  • Agent-controlled context: Agents can use the ManageContext tool to proactively reset the context, pin key content, discard specified tool results, or adjust the trimming budget, keeping conversations more focused.
  • Delegation context control: When delegating to a sub-agent, you can precisely narrow its capability declarations, memory scope, and system prompt scope, giving it only the minimal context needed for the subtask.
  • Observable delegation: Delegation cards now include a “View full process” entry, which jumps to the sub-agent’s session so you can review its full execution history.
  • Team member session summaries: The team panel has a new “Recent session summaries” entry that shows member turns, tool call usage, and AI-generated deep summaries.
  • InspectRuns tool: Agents can use a new tool to view historical messages and structured summaries from other sessions, making it easier to synchronize information during multi-agent collaboration.
  • Batch delete messages: After selecting multiple messages, you can batch delete them with one click from the action bar, including across sessions.

Fixes​

  • Fixed an issue where continuing a conversation after undoing or resuming would cause write errors and message loss due to a missing session directory.
  • Fixed an issue where approval replay could occasionally produce duplicate notifications or freeze approvals after app restart or reconnection.
  • Fixed an issue where the heartbeat delivery queue would create corrupt files after an abnormal exit, causing the log to be flooded with continuous errors; it now uses atomic writes and automatically cleans up bad files.

Improvements​

  • Comprehensive approval system upgrade: Supports three AI approval modes: auto AI decision, AI suggestion only, and a foreground human window. All settlements are authoritatively handled on the server side. Background delegation and dispatch chains can be auto-approved quickly, while the foreground retains a human-first window. After disconnection and reconnection, pending approval cards and AI suggestions can be restored. Seven approval preferences come with a full visual interface.
  • More realistic installation status: When installing or uninstalling MCP services, the UI now shows an intermediate “Installing…” / “Uninstalling…” state first, and only switches to the final state after the agent actually verifies success, eliminating false “Installed” indicators.
  • Built-in agent updated to v1.9.0: The installation management flow now uses endpoint write-back and strong validation, improving installation reliability.
  • Context length exceeded hints from some model gateways (without specific numbers) are now correctly recognized and automatically trigger output reduction or context trimming, reducing direct errors caused by non-recognition.
  • Streaming request retry budgets are now calculated independently per failure reason, so failures of different causes no longer consume each other’s budget, reducing the chance of retry exhaustion from transient transport-layer errors.
  • Brief streaming failures during long tasks no longer accumulate toward the conversation’s retry limit; only consecutive failures reaching the threshold trigger an error, improving the success rate of long-running tasks.