[BUG] In AgentTool the input of the first tool will override the manual input of following tools.

[qianheng-aws]

What is the bug? If we create an AgentTool containing several tools and some tool may have their own input parameter, when ChatAgentRunner selects this AgentTool to execute, it will generates its input parameter passing to the first tool of AgentTool. However, this input parameter will also override the default manual input parameter of following tools. We should not let it to override since it's only for the input parameter of the first tool.

How can one reproduce the bug? Steps to reproduce the behavior:

1. Register an flow agent with 2 tools like:

POST /_plugins/_ml/agents/_register
{
  "name": "Test_Agent_For_PPL",
  "type": "flow",
  "description": "this is a test agent for invoking PPLTool",
  "tools": [
    {
      "type": "MLModelTool",
      "parameters": {
        "model_id": "xxxxxxxx",
        "prompt": """You are given a context of alert definition and a nature language question, summarize a final question as the input to generate PPL based that."""
      }
    },
    {
      "type": "PPLTool",
      "parameters": {
        "model_id": "xxxxxxxx",
        "model_type": "FINETUNE",
        "execute": true,
        "input": {
          "question": "${parameters.MLModelTool.output}"
        }
        
      }
    }
  ]
}

2. register a chat agent with AgentTool

OST /_plugins/_ml/agents/_register
{
  "name": "Test_Agent_For_ReAct_ClaudeV2",
  "type": "conversational",
  "description": "this is an agent to help analysis alert, answering follow-up question of the current alert including search data, create new alert.",
  "llm": {
    "model_id": "xxxxxxxx",
    "parameters": {
      "max_iteration": 5,
      "stop_when_no_tool_found": true
    }
  },
  "memory": {
    "type": "conversation_index"
  },
  "tools": [
    {
      "type": "AgentTool",
      "name": "AgentTool",
      "parameters": {
        "agent_id": "xxxxxx",
         "prompt": """xxxxx"""
       }
    }
  ]
}

3. update the current root agent

PUT .plugins-ml-config/_doc/os_chat
{"type":"os_chat_root_agent","configuration":{"agent_id":"xxxxxxxxxx"}}

6. In chat window, ask question for providing more data.
7. In debug mode, you will see the input of PPLTool will be overridden by the input of MLModelTool.

What is the expected behavior? The input of PPLTool should be our manual set input { "question": "${parameters.MLModelTool.output}"} with placeholder substituted.

What is your host/environment?

• OS: mac
• Version: 3.0.0-SNAP
• Plugins: ml-commons, skills, sql

Do you have any screenshots? If applicable, add screenshots to help explain your problem.

Do you have any additional context? Add any other context about the problem.