save-trajectory

Save Trajectory

Overview

This skill saves the current session's conversation history as a JSON file in OpenAI chat completion format. The trajectory is saved to .evolve/trajectories/ in the project root. This enables trajectory analysis, fine-tuning data collection, and session review.

Workflow

Step 1: Walk Through Conversation Messages

Review all messages in the current conversation from start to finish. For each message, identify its type:

• User text messages
• Assistant text responses (may include thinking)
• Assistant tool calls
• Tool results

Step 2: Convert to OpenAI Chat Completion Format

Convert each message to the appropriate format:

User text message:

{"role": "user", "content": "the user's message text"}

Assistant text response (no thinking):

{"role": "assistant", "content": "the assistant's response text"}

Assistant text response (with thinking):

{"role": "assistant", "content": "the assistant's response text", "thinking": "the thinking/reasoning text"}

Assistant tool call (no visible text):

{
  "role": "assistant",
  "content": null,
  "tool_calls": [
    {
      "id": "tool_call_id_here",
      "type": "function",
      "function": {
        "name": "ToolName",
        "arguments": "{\"param\": \"value\"}"
      }
    }
  ]
}

Assistant tool call with text:

{
  "role": "assistant",
  "content": "text before/after the tool call",
  "tool_calls": [
    {
      "id": "tool_call_id_here",
      "type": "function",
      "function": {
        "name": "ToolName",
        "arguments": "{\"param\": \"value\"}"
      }
    }
  ]
}

Tool result:

{"role": "tool", "tool_call_id": "tool_call_id_here", "content": "the tool output text"}

Important Details

• Tool call arguments must be a JSON string, not a nested object. Use json.dumps() on the arguments object.
• Tool call IDs: Use the actual tool call ID from the conversation. If not available, generate a unique ID like call_001, call_002, etc.
• Multiple tool calls: If the assistant made multiple tool calls in one turn, include all of them in a single assistant message's tool_calls array, followed by separate tool result messages for each.
• Thinking blocks: If the assistant had both thinking and text in the same turn, combine them into one message with both content and thinking fields.