ADK Java OTel spans do not expose tool-call execution correctly in Langfuse/OpenInference

🔴 Required Information

Describe the Bug:
ADK Java traces exported to Langfuse via OTLP show call_llm spans, but tool-call execution is not represented correctly: tools appear as "not called", and execute_tool visibility/payload mapping is missing or inconsistent even when tools were actually invoked.

Steps to Reproduce:

1. Create an ADK Java agent with MCP tools (for example, list_premerge_failed_tests_by_job and list_unhealthy_tests).
2. Configure OpenTelemetry export to Langfuse using OtlpHttpSpanExporter.
3. Run an agent prompt that should invoke those tools.
4. Open the resulting trace in Langfuse UI and inspect the call_llm observation and tool section.
5. Observe that tools are displayed as "not called" and no reliable execute_tool observation/payload is shown.

Expected Behavior:
When tools execute, Langfuse should show them as called, include execute_tool observation visibility, and render tool request/response payloads correctly.

Observed Behavior:

• call_llm is visible in timeline.
• Tools are shown as "not called" in the same trace.
• Tool-call visibility/payload mapping is missing or inconsistent.
• No crash, but observability is incorrect/incomplete for tool execution.

Environment Details:

• ADK Library Version (see maven dependency): com.google.adk:google-adk:1.2.0 (google-adk-spring-ai:1.2.0)
• OS: macOS
• TS Version (tsc --version): N/A (Java integration)

Model Information:

• Which model is being used: qwen-3.5-122b

---

🟡 Optional Information

Regression:
Unknown.

Logs:

Langfuse UI observation:
- call_llm span present
- Tools section shows entries as "not called"
- Agent output text indicates tools were expected/executed in flow

Screenshots / Video:

Additional Context:

• ADK Java appears to emit ADK-specific attributes (e.g. gcp.vertex.agent.tool_call_args, gcp.vertex.agent.tool_response, gcp.vertex.agent.llm_request, gcp.vertex.agent.llm_response).
• Langfuse/OpenInference expects standard tool/IO mapping on spans (e.g. input.value, output.value) for reliable rendering.

Minimal Reproduction Code:

// Pseudocode reproduction shape:
// 1) Build LlmAgent with MCP tools
// 2) Enable OTLP export to Langfuse
// 3) Trigger prompt that should call tools
// 4) Inspect Langfuse trace: call_llm visible, tools shown as "not called"

How often has this issue occurred?:

• 100%