| 1 | # Output files
|
| 2 |
|
| 3 | !!! abstract "Overview"
|
| 4 |
|
| 5 | mini-SWE-agent saves run results in JSON format. This page documents the structure of these output files.
|
| 6 |
|
| 7 | ## Trajectory files (`.traj.json`)
|
| 8 |
|
| 9 | !!! warning "v2.0 format changes"
|
| 10 |
|
| 11 | The output format changed in v2.0 (`trajectory_format: mini-swe-agent-1.1`). See the [v2 migration guide](../advanced/v2_migration.md) for more information.
|
| 12 |
|
| 13 | !!! tip "Viewing trajectory files"
|
| 14 |
|
| 15 | Use the [inspector](inspector.md) to browse trajectory files interactively.
|
| 16 |
|
| 17 | Trajectory files contain the full history of an agent run, including all messages, configuration, and metadata.
|
| 18 |
|
| 19 | ### Structure
|
| 20 |
|
| 21 | ```json
|
| 22 | {
|
| 23 | "info": {
|
| 24 | "model_stats": {
|
| 25 | "instance_cost": 0.05, // total cost of API calls for this run
|
| 26 | "api_calls": 12 // number of API calls made
|
| 27 | },
|
| 28 | "config": {
|
| 29 | "agent": { ... }, // agent configuration
|
| 30 | "agent_type": "minisweagent.agents.default.DefaultAgent",
|
| 31 | "model": { ... }, // model configuration
|
| 32 | "model_type": "minisweagent.models.litellm_model.LitellmModel",
|
| 33 | "environment": { ... }, // environment configuration
|
| 34 | "environment_type": "minisweagent.environments.local.LocalEnvironment"
|
| 35 | },
|
| 36 | "mini_version": "2.0.0", // version of mini-SWE-agent used
|
| 37 | "exit_status": "Submitted", // final status (Submitted, LimitsExceeded, etc.)
|
| 38 | "submission": "..." // final output/patch submitted by the agent (if any)
|
| 39 | },
|
| 40 | "messages": [ // full conversation history
|
| 41 | {"role": "system", "content": "..."},
|
| 42 | {"role": "user", "content": "..."},
|
| 43 | {"role": "assistant", "content": "..."},
|
| 44 | ...
|
| 45 | ],
|
| 46 | "trajectory_format": "mini-swe-agent-1.1" // format version identifier
|
| 47 | }
|
| 48 | ```
|
| 49 |
|
| 50 |
|
| 51 | Messages follow the [OpenAI chat format](https://platform.openai.com/docs/api-reference/chat) with an additional `extra` field for mini-SWE-agent metadata. Models may add other fields to messages (e.g., `tool_calls`, `reasoning_content`).
|
| 52 |
|
| 53 | !!! note "Toolcall models"
|
| 54 | 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.
|
| 55 |
|
| 56 | ```json
|
| 57 | // System message (agent instructions)
|
| 58 | {"role": "system", "content": "You are a helpful assistant..."}
|
| 59 |
|
| 60 | // User message (task description)
|
| 61 | {"role": "user", "content": "Please solve this issue: ..."}
|
| 62 |
|
| 63 | // Assistant message (model response with parsed actions)
|
| 64 | {
|
| 65 | "role": "assistant",
|
| 66 | "content": "Let me check the files...\n\n```mswea_bash_command\nls -la\n```",
|
| 67 | "extra": {
|
| 68 | "actions": [{"command": "ls -la"}], // parsed actions to execute
|
| 69 | "cost": 0.003, // cost of this API call
|
| 70 | "timestamp": 1706000000.0, // unix timestamp of when this message was created
|
| 71 | "response": { ... } // raw API response
|
| 72 | }
|
| 73 | }
|
| 74 |
|
| 75 | // Observation message (execution result)
|
| 76 | {
|
| 77 | "role": "user",
|
| 78 | "content": "<returncode>0</returncode>\n<output>\nfile1.py\nfile2.py\n</output>",
|
| 79 | "extra": {
|
| 80 | "returncode": 0,
|
| 81 | "timestamp": 1706000001.0
|
| 82 | }
|
| 83 | }
|
| 84 |
|
| 85 | // Final message (when agent submits)
|
| 86 | {
|
| 87 | "role": "user",
|
| 88 | "content": "",
|
| 89 | "extra": {
|
| 90 | "exit_status": "Submitted",
|
| 91 | "submission": "diff --git a/file.py..."
|
| 92 | }
|
| 93 | }
|
| 94 | ```
|
| 95 |
|
| 96 |
|
| 97 | ## `preds.json` format
|
| 98 |
|
| 99 | The predictions file aggregates results from all instances in a format compatible with SWE-bench evaluation:
|
| 100 |
|
| 101 | ```json
|
| 102 | {
|
| 103 | "owner__repo__123": { // keyed by instance_id
|
| 104 | "model_name_or_path": "anthropic/claude-sonnet-4-5-20250929", // model used
|
| 105 | "instance_id": "owner__repo__123", // SWE-bench instance identifier
|
| 106 | "model_patch": "diff --git a/file.py b/file.py\n..." // generated patch (unified diff)
|
| 107 | },
|
| 108 | ...
|
| 109 | }
|
| 110 | ```
|
| 111 |
|
| 112 | {% include-markdown "../_footer.md" %}
|
| 113 |
|
