Full tool trajectories

Build the CLI

npm install
npm run build

Generate a receptionist dataset

node dist/cli/index.js simulate-dataset \
  --config examples/receptionist/scenario.json \
  --out outputs/tutorial-receptionist.jsonl \
  --limit 3 \
  --mode full_tool_trajectory

For this sample, expect three rows, three tool calls, three tool results, and three rows with tools.

Validate the dataset

node dist/cli/index.js validate-dataset outputs/tutorial-receptionist.jsonl

The validator checks JSONL parsing, message shape, assistant tool-call arguments, tool result references, tool result names, and summary counts.

Inspect one row

Each full_tool_trajectory row follows this exact order:

1. system
2. user
3. assistant with tool_calls
4. tool with the matching tool_call_id
5. final assistant text response

The generated OpenAI-format row looks like this:

{
  "messages": [
    { "role": "system", "content": "..." },
    { "role": "user", "content": "..." },
    {
      "role": "assistant",
      "content": null,
      "tool_calls": [
        {
          "id": "call_sample_receptionist_1",
          "type": "function",
          "function": {
            "name": "search",
            "arguments": "{\"query\":\"sample query\"}"
          }
        }
      ]
    },
    {
      "role": "tool",
      "tool_call_id": "call_sample_receptionist_1",
      "name": "search",
      "content": "{\"answer\":\"...\"}"
    },
    { "role": "assistant", "content": "..." }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "search",
        "description": "...",
        "parameters": {
          "type": "object",
          "properties": {
            "query": { "type": "string" }
          }
        }
      }
    }
  ]
}

Tool definitions are included when the selected export mode contains assistant tool calls and the trajectory has tool schemas.

Compare export modes

The CLI supports three export modes:

• plain_chat: system, user, and assistant text only; no tool calls
• tool_decision: stop immediately after the assistant tool-call message
• full_tool_trajectory: include tool calls, tool results, and the final assistant response

Use full_tool_trajectory when you want the model to learn the complete tool loop. Use tool_decision only when you intentionally want examples to end at tool choice.

Try another domain

node dist/cli/index.js simulate-dataset \
  --config examples/retail-support/scenario.json \
  --out outputs/tutorial-retail-support.jsonl \
  --limit 2 \
  --mode full_tool_trajectory

node dist/cli/index.js validate-dataset outputs/tutorial-retail-support.jsonl

The same commands work for non-receptionist domains because scenario definitions carry business context, tool inventory, and stopping behavior.