agenttool: Run() leaks thinking parts into tool result

[dmora]

Description

agenttool.Run() collects all Part.Text from the sub-agent's final response without checking Part.Thought. When thinking models (Gemini 2.5/3 with IncludeThoughts: true) are used as sub-agents, their internal reasoning tokens leak into the tool result that the calling agent sees.

Root cause

tool/agenttool/agent_tool.go lines 223-228:

for _, part := range lastContent.Parts {
    if part != nil && part.Text != "" {
        textParts = append(textParts, part.Text)
    }
}

No check on part.Thought. Compare with genai.GenerateContentResponse.Text() (types.go:3148-3151) which correctly skips thought parts:

if part.Thought {
    continue
}

Impact

• Thinking tokens leak into tool results — the root agent sees the sub-agent's internal reasoning mixed into the response text
• OutputSchema validation can fail — when thought text is concatenated before JSON output, unmarshalling breaks
• Token waste — the root agent processes thinking tokens it shouldn't see, inflating context

Related

• fix: interleaved thought&text aggregation in remoteagent #655 — same class of bug, fixed for remoteagent (merged)
• session.Events() and separating thinking & final response #313 — request for separating thinking & final response in session events
• Missing ThoughtSignature on adk_request_confirmation parts breaks execution #656 / Add thought signature propagation to avoid hard crashes #686 — ThoughtSignature propagation issues (related thinking-model support gaps)

Fix

Submitted in #696 — adds && !part.Thought to the text extraction loop, matching genai.Text() behavior.

[indurireddy-TF]

Hi @dmora I can confirm that this is indeed a bug. I wrote a test case to verify the behavior, and I was able to reproduce the issue locally exactly as you described. The test clearly showed the model's tokens leaking into the sub-agent's text output, which explains the JSON unmarshalling failures.I see you have already submitted PR #696 with the !part.Thought fix and the accompanying test cases. @mazas-google @dpasiukevich can you please take a look and review the changes.