Abnormal termination of multi-round dialogues for complex tasks using agent-framework and playwright/mcp. #2657
Description
Using agent-framework and Playwright/MCP, when the input task is complex, such as searching for keywords and then clicking on each search result in a loop to enter the detail page and scrape data, it often leads to the end of multiple rounds of dialogue. Recently, the model interface returned a finish_reason value of "stop" or "tool_calls," but at this point, the task was not actually completed.
The MCP library used by agent-framework printed a log message before termination: "GET stream max reconnection attempts 1000ms." This code is located in site-packages\mcp\client\streamable_http.py.