Description of the Problem

When MCP (Model Context Protocol) is configured and a thinking-enabled model is used in Lightspeed, the model's thinking content leaks outside of the designated "Show Thinking" collapsible box during response streaming. The thinking text appears inline with the main response content instead of being contained within its dedicated UI section.

When MCP is not configured, the thinking content is correctly displayed only within the "Show Thinking" box as expected.

Prerequisites

• rhdh-plugins Lightspeed workspace

• MCP servers configured in Lightspeed settings

• Thinking-enabled model selected

• Thinking feature enabled in Lightspeed settings

Steps to Reproduce

1. Open the Lightspeed plugin with MCP configured.

1. Start a new chat conversation.

1. Send a message that triggers tool use or any response.

1. Observe the response as it streams in real-time.

1. Note how the thinking content appears during streaming.

Actual Results

During response streaming, the model's thinking content appears outside the "Show Thinking" box, mixing with the main response text. The thinking text "leaks" into the visible response area before being properly contained.

Expected Results

The model's thinking content should remain contained within the "Show Thinking" collapsible box at all times during streaming. The main response area should only display the final response content, not the intermediate thinking process.

Reproducibility

• Always

Build Details

• Dev mode

Additional Information

• The issue only occurs when MCP is configured in combination with thinking models.

• Streaming behavior appears to incorrectly parse/route thinking content when MCP responses are involved.

• The issue is timing-related during the stream; the final rendered state may appear correct, but the streaming experience shows leaked content.

• Comparing streaming behavior with MCP vs Without MCP clearly shows the difference.

• Screen Recording 2026-01-28 at 1.11.35 AM.mov