Output files

Overview

mini-SWE-agent saves run results in JSON format. This page documents the structure of these output files.

Trajectory files (`.traj.json`)

v2.0 format changes

The output format changed in v2.0 (trajectory_format: mini-swe-agent-1.1). See the v2 migration guide for more information.

Viewing trajectory files

Use the inspector to browse trajectory files interactively.

Trajectory files contain the full history of an agent run, including all messages, configuration, and metadata.

Structure

{
  "info": {
    "model_stats": {
      "instance_cost": 0.05,  // total cost of API calls for this run
      "api_calls": 12  // number of API calls made
    },
    "config": {
      "agent": { ... },  // agent configuration
      "agent_type": "minisweagent.agents.default.DefaultAgent",
      "model": { ... },  // model configuration
      "model_type": "minisweagent.models.litellm_model.LitellmModel",
      "environment": { ... },  // environment configuration
      "environment_type": "minisweagent.environments.local.LocalEnvironment"
    },
    "mini_version": "2.0.0",  // version of mini-SWE-agent used
    "exit_status": "Submitted",  // final status (Submitted, LimitsExceeded, etc.)
    "submission": "..."  // final output/patch submitted by the agent (if any)
  },
  "messages": [  // full conversation history
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."},
    ...
  ],
  "trajectory_format": "mini-swe-agent-1.1"  // format version identifier
}

Messages follow the OpenAI chat format with an additional extra field for mini-SWE-agent metadata. Models may add other fields to messages (e.g., tool_calls, reasoning_content).

Toolcall models

When using toolcall-based models (e.g., LitellmToolcallModel), the roles differ slightly: assistant messages include tool_calls instead of content, and observation messages use role: "tool" with a tool_call_id field.

// System message (agent instructions)
{"role": "system", "content": "You are a helpful assistant..."}

// User message (task description)
{"role": "user", "content": "Please solve this issue: ..."}

// Assistant message (model response with parsed actions)
{
  "role": "assistant",
  "content": "Let me check the files...\n\n```mswea_bash_command\nls -la\n```",
  "extra": {
    "actions": [{"command": "ls -la"}],  // parsed actions to execute
    "cost": 0.003,  // cost of this API call
    "timestamp": 1706000000.0,  // unix timestamp of when this message was created
    "response": { ... }  // raw API response
  }
}

// Observation message (execution result)
{
  "role": "user",
  "content": "<returncode>0</returncode>\n<output>\nfile1.py\nfile2.py\n</output>",
  "extra": {
    "returncode": 0,
    "timestamp": 1706000001.0
  }
}

// Final message (when agent submits)
{
  "role": "user",
  "content": "",
  "extra": {
    "exit_status": "Submitted",
    "submission": "diff --git a/file.py..."
  }
}

`preds.json` format

The predictions file aggregates results from all instances in a format compatible with SWE-bench evaluation:

{
  "owner__repo__123": {  // keyed by instance_id
    "model_name_or_path": "anthropic/claude-sonnet-4-5-20250929",  // model used
    "instance_id": "owner__repo__123",  // SWE-bench instance identifier
    "model_patch": "diff --git a/file.py b/file.py\n..."  // generated patch (unified diff)
  },
  ...
}