Streaming thought chunks render as separate bubbles instead of consolidating

[ferponse]

Description

When using streaming mode (SSE) with models that emit thinking/reasoning content (e.g., Claude with extended thinking via LiteLLM, or Gemini with thinking_config), each partial ReasoningChunk renders as a separate "Thought" bubble in the Dev UI instead of being consolidated into a single thought block.

Non-streaming mode correctly shows a single consolidated "Thought" bubble.

Steps to Reproduce

1. Configure an ADK agent with a model that supports thinking/reasoning (e.g., vertex_ai/claude-sonnet-4-6 via LiteLlm with thinking={"type": "enabled", "budget_tokens": 10000})
2. Run adk web and open the Dev UI
3. Send a message to the agent
4. Observe the event timeline — each streaming thought chunk appears as a separate "Thought" bubble (e.g., remove more unused comments #2 "The user said", add more comment #3 '"hola" (', first commit on README #4 "hello).", etc.)

Expected Behavior

Consecutive partial thought events should be consolidated into a single "Thought" bubble that progressively accumulates text, similar to how text response chunks are merged into a single response bubble.

Actual Behavior

Each ReasoningChunk partial generates a separate "Thought" bubble in the timeline, resulting in dozens of fragmented thought events for a single reasoning block.

Analysis

The consolidation logic in chat.component.ts (lines ~1237-1257) attempts to merge thought partials by searching for an existing event with partial=true and thought=true. However, buildUiEventFromEvent creates a new UiEvent with only the current chunk's content, and the replacement at line ~1285 (newEvents[existingIndex] = uiEvent) replaces rather than accumulates the text. This causes either:

• The consolidation search to fail (resulting in multiple bubbles), or
• The previous thought to be replaced instead of appended to

Environment

• ADK Python: v1.32.0
• ADK Web: latest (cloned from main)
• Model: vertex_ai/claude-sonnet-4-6 via LiteLlm with extended thinking enabled
• Also reproducible with native LiteLlm (not just custom BaseLlm subclasses)