fix(agentspan): MCP/API agents lose every tool result — set _agent_tool_name in enrichToolsScriptDynamic

[ling-senpeng13]

Fixes #1532.

Any agent whose toolset includes an MCP or API tool never receives its own tool results. Every LLM turn is dispatched with the identical bare [system, user] message list, so the model reissues the same call until the ReAct loop hits its iteration cap.

The fix

One line in enrichToolsScriptDynamic, which enrichToolsScript already had:

t.inputParameters._agent_tool_name = n;

Why it fixes it

Join.compactAgentOutput decides whether a fork branch is an agent tool call by reading that marker from the branch's input:

Object agentToolName = forkedTask.getInputData().get("_agent_tool_name");
if (agentToolName != null) {
    compact.put("_agent_tool_name", agentToolName);
    compact.put("_agent_tool_output", output);   // the full payload rides along here
    return compact;
}
// otherwise: only AGENT_PROPAGATED_KEYS = {"_state_updates", "state"}

and the caller only records non-empty results:

if (!compact.isEmpty()) task.addOutput(joinOnRef, compact);

Setting the marker makes compact non-empty and attaches _agent_tool_output, so the whole chain refills.

Verification

Reproduced and fixed with the python-sdk examples 04_mcp_weather.py and 04_http_and_mcp_tools.py

[ling-senpeng13]

#1532 is not reproducible now