feat: add trajectroy schema for agent

[DhavalRepo18]

Trajectory analysis is important for

• Eval
• Failure Mode analysis

We like to introduce a formal pydentic class to capture the important schema.

[ShuxinLin]

✗ uv run claude-agent --model-id litellm_proxy/aws/claude-opus-4-6 "What sensors are on Chiller 6?"  --show-trajectory --json
{
  "turns": [
    {
      "index": 0,
      "text": "",
      "tool_calls": [],
      "input_tokens": 3,
      "output_tokens": 9
    },
    {
      "index": 1,
      "text": "Let me first find the site, then list the sensors for Chiller 6.",
      "tool_calls": [],
      "input_tokens": 3,
      "output_tokens": 9
    },
    {
      "index": 2,
      "text": "",
      "tool_calls": [
        {
          "name": "mcp__iot__sites",
          "input": {},
          "id": "toolu_bdrk_01SMWMxuJU1mRK4cB24BmNw6",
          "output": "{\"sites\":[\"MAIN\"]}"
        }
      ],
      "input_tokens": 3,
      "output_tokens": 9
    },
    {
      "index": 3,
      "text": "",
      "tool_calls": [
        {
          "name": "mcp__iot__sensors",
          "input": {
            "site_name": "MAIN",
            "asset_id": "Chiller 6"
          },
          "id": "toolu_bdrk_01XEQwsQgEP6TrZ7zfPJYiv3",
          "output": "{\"result\":{\"site_name\":\"MAIN\",\"asset_id\":\"Chiller 6\",\"total_sensors\":11,\"sensors\":[\"Chiller 6 Chiller % Loaded\",\"Chiller 6 Chiller Efficiency\",\"Chiller 6 Condenser Water Flow\",\"Chiller 6 Condenser Water Return To Tower Temperature\",\"Chiller 6 Liquid Refrigerant Evaporator Temperature\",\"Chiller 6 Power Input\",\"Chiller 6 Return Temperature\",\"Chiller 6 Schedule\",\"Chiller 6 Setpoint Temperature\",\"Chiller 6 Supply Temperature\",\"Chiller 6 Tonnage\"],\"message\":\"found 11 sensors for asset_id Chiller 6 and site_name MAIN: Chiller 6 Chiller % Loaded, Chiller 6 Chiller Efficiency, Chiller 6 Condenser Water Flow, Chiller 6 Condenser Water Return To Tower Temperature, Chiller 6 Liquid Refrigerant Evaporator Temperature, Chiller 6 Power Input, Chiller 6 Return Temperature, Chiller 6 Schedule, Chiller 6 Setpoint Temperature, Chiller 6 Supply Temperature, Chiller 6 Tonnage.\"}}"
        }
      ],
      "input_tokens": 1,
      "output_tokens": 24
    },
    {
      "index": 4,
      "text": "**Chiller 6** (site: MAIN) has **11 sensors**:\n\n1. **Chiller % Loaded** \u2013 load percentage\n2. **Chiller Efficiency** \u2013 operational efficiency\n3. **Condenser Water Flow** \u2013 condenser water flow rate\n4. **Condenser Water Return To Tower Temperature** \u2013 water temp returning to the cooling tower\n5. **Liquid Refrigerant Evaporator Temperature** \u2013 evaporator refrigerant temp\n6. **Power Input** \u2013 electrical power consumption\n7. **Return Temperature** \u2013 chilled water return temp\n8. **Schedule** \u2013 on/off schedule\n9. **Setpoint Temperature** \u2013 target temperature setpoint\n10. **Supply Temperature** \u2013 chilled water supply temp\n11. **Tonnage** \u2013 cooling output in tons",
      "tool_calls": [],
      "input_tokens": 1,
      "output_tokens": 1
    }
  ]
}

this is an exemplar trajectory from implemented claude agent

[DhavalRepo18]

@ShuxinLin Thank you for the output. As a user, what happens if we switch to Plan-Execute? Will it produce the same trajectory object? if yes, then this is great as it will help to drive the Eval module design.

[DhavalRepo18]

On a side now, if we use Reasoning Model, they produce two level of details:

• "output_tokens": 1
• "reasoning_tokens": 1

We as user like to have both the information as a result we can study the impact of token size.

[DhavalRepo18]

@nianjunz We now have an example trajectory coming from the new AssetOpsBench agent. Can you now use this as a trajectory for moving forward?

#177

Also please look at the src folder and bring your trajectory analysis work to src.

#205

[DhavalRepo18]

@ShuxinLin We have a few follow-up questions

"turns": [
    {
      "index": 0,
      "text": "",
      "tool_calls": [],
      "input_tokens": 3,
      "output_tokens": 9
    },

Please help to create a one-to-one mapping.

Think-Act-Observe

• Which field can we map to `think', 'Act', and 'Observer', and then some failure?
• Why is input_tokens three? output_tokens is 9?

Our original ReAct used to capture many system metrics such as State, is_loop_detected, additional_feedback (like whenever we detect loop), etc

{
            "step": 1,
            "raw_llm_thought_output": "",
            "raw_llm_action_output": "```python\n# ```",
            "raw_observation_output": " events\n2020-05-29: 8 events\n2020-05-30: 1 events\n2020-05-31: 1 events\n",
            "raw_llm_output": "",
            "thought": "",
            "action": [
                "# Define the equipment and date range\nequipment = "
            ],
            "action_input": "",
            "observation": "2020-05-01: 20 events\n2020-05-02: 25  1 events\n2020-05-31: 1 events",
            "state": "Valid Action",
            "is_loop_detected": null,
            "additional_scratchpad_feedback": ""
        }