The agent frequently terminates before completing its tasks, even when it has clearly already laid out a multi-step plan and is actively executing. There is no error message — it just stops.
Steps to reproduce:
Submit a multi-step task request to Agnes (e.g. "replace the tree component in SimpleMenuTree.vue with vxe-tree, then sync menus and refactor dashboard/menu.vue").
The agent correctly breaks down the work into steps, shows Thinking output, and begins executing shell commands (e.g. Select-String, Get-ChildItem) to explore the codebase.
Partway through — while still in the middle of exploration or implementation — the agent session abruptly terminates. No error is shown, only "Agnes 正在处理..." followed by silence / session end.
None of the planned steps are completed. The user has to restart the conversation and ask to continue.
Expected behavior:
The agent should complete all steps in its plan before terminating, or at minimum surface a clear error message explaining why the session stopped (e.g. token limit, timeout, tool failure). If a task is too large, it should save partial progress so the user can resume without re-prompting the entire plan.
Actual behavior:
The agent stops mid-execution without warning and without saving any meaningful state. The user loses the agent's full chain-of-thought and has to restart.
Screenshots:
(See attached — agent listed a 3-step plan, had already run two shell commands, and was mid-sentence when it terminated.)
The agent frequently terminates before completing its tasks, even when it has clearly already laid out a multi-step plan and is actively executing. There is no error message — it just stops.
Steps to reproduce:
Submit a multi-step task request to Agnes (e.g. "replace the tree component in SimpleMenuTree.vue with vxe-tree, then sync menus and refactor dashboard/menu.vue").
The agent correctly breaks down the work into steps, shows Thinking output, and begins executing shell commands (e.g. Select-String, Get-ChildItem) to explore the codebase.
Partway through — while still in the middle of exploration or implementation — the agent session abruptly terminates. No error is shown, only "Agnes 正在处理..." followed by silence / session end.
None of the planned steps are completed. The user has to restart the conversation and ask to continue.
Expected behavior:
The agent should complete all steps in its plan before terminating, or at minimum surface a clear error message explaining why the session stopped (e.g. token limit, timeout, tool failure). If a task is too large, it should save partial progress so the user can resume without re-prompting the entire plan.
Actual behavior:
The agent stops mid-execution without warning and without saving any meaningful state. The user loses the agent's full chain-of-thought and has to restart.
Screenshots:
(See attached — agent listed a 3-step plan, had already run two shell commands, and was mid-sentence when it terminated.)