This is the official GitHub repository for the paper "SWE-Edit: Rethinking Code Editing for Efficient SWE-Agent".

SWE-Edit is a dual-optimization framework that improves both the performance and efficiency of code editing in LLM-based software engineering agents.

SWE-Edit addresses the context coupling problem in traditional code editing interfaces through:

1. Scaffolding-Level Optimization: Decomposing the code editing interface into specialized subagents

   • Viewer Subagent: Extracts task-relevant code snippets on demand, eliminating context pollution
   • Editor Subagent: Executes modifications from natural language instructions, decoupling reasoning from format-sensitive generation

2. Model-Level Optimization: Training an adaptive editor with GRPO that learns to select between editing modes (find-replace vs. whole-file rewrite) based on task characteristics

Figure 1: Overview of the SWE-Edit framework showing the dual optimization mechanism with Viewer and Editor subagents.

Table 1: Main results comparing baseline vs. SWE-Edit on SWE-bench Verified (500 instances, 3 runs).

SWE-Edit achieves both higher performance and lower cost, breaking the typical accuracy-cost trade-off.

We use uv for Python environment management.

git clone https://github.com/microsoft/SWE-Edit.git
cd SWE-Edit
uv sync

For SWE-bench evaluation:

uv sync --extra eval

Create a .env file in the project root:

OpenAI:

API_TYPE="OPENAI"
OPENAI_API_KEY=<your-api-key>
MODEL=gpt-4o

Anthropic:

API_TYPE="ANTHROPIC"
ANTHROPIC_API_KEY=<your-api-key>
MODEL=claude-sonnet-4-5-20250929

Azure OpenAI:

API_TYPE="AZURE_OPENAI"
AZURE_API_KEY=<your-api-key>
AZURE_ENDPOINT=<your-endpoint>
DEPLOYMENT=<deployment-name>

The editor and viewer subagents can use different (typically smaller and cheaper) models. Configure them in one of two ways:

Option 1: Environment variables (recommended for simple setups)

Add to your .env file with the LLM_EDITOR_ or LLM_VIEWER_ prefixes:

# Main agent uses GPT-4
API_TYPE="OPENAI"
OPENAI_API_KEY=<your-api-key>
MODEL=gpt-4o

# Editor uses GPT-5-mini
LLM_EDITOR_API_TYPE="OPENAI"
LLM_EDITOR_OPENAI_API_KEY=<your-api-key>
LLM_EDITOR_MODEL=gpt-5-mini

# Viewer uses GPT-5-mini
LLM_VIEWER_API_TYPE="OPENAI"
LLM_VIEWER_OPENAI_API_KEY=<your-api-key>
LLM_VIEWER_MODEL=gpt-5-mini

Option 2: JSON configuration files

Create JSON files under ./api_configs/ directory. For example, api_configs/editor/gpt-5-mini.json or api_configs/viewer/gpt-5-mini.json:

{
  "api_type": "OPENAI",
  "api_key": "your-api-key",
  "model": "gpt-4o-mini"
}

Then specify the model type during evaluation (see SWE-bench Evaluation section).

For large-scale evaluation, you can provide multiple API keys by appending _N to environment variables:

API_TYPE_1="OPENAI"
OPENAI_API_KEY_1=<key-1>
MODEL_1=gpt-4o

API_TYPE_2="ANTHROPIC"
ANTHROPIC_API_KEY_2=<key-2>
MODEL_2=claude-sonnet-4-5-20250929

The agent will randomly select a configuration to distribute rate limits.

Run the minimal agent on a simple task to verify your setup:

uv run -m sweedit.main "First run pwd to check the current directory, then create a file called hello.py in that directory with a function that prints 'Hello, World!'"

This runs a minimal agent with basic tools (bash execution, string replacement editor, and finish) to verify your API configuration is working properly. Trajectories are saved to .traj/ for analysis.

We provide complete evaluation scripts for SWE-bench Verified.

First, build the package:

uv build

Then run evaluation (see evaluation/swebench-verified/README.md for details):

uv run python evaluation/swebench-verified/main.py \
  --local_traj_path <path-to-trajectories> \
  --run_name <experiment-name> \
  --agent_type llm-editor \
  --llm_editor_model gpt-5-mini \
  --max_workers <num-workers>

The --llm_editor_model flag specifies which model to use for the editor/viewer subagents (uses LLM_EDITOR_ prefixed env vars or JSON config files in ./api_configs/).

.
├── src/sweedit/              # Core framework
│   ├── core/                 # Agent scaffolding
│   │   ├── agent.py          # Base agent
│   │   ├── agenthub/         # Agent implementations
│   │   └── llm.py            # LLM interface
│   └── tools/                # Agent tools
│       ├── llm_editor.py           # LLM-backed editing tool
│       ├── llm_file_editor.py      # Editor subagent
│       ├── execute_bash.py         # Shell execution
│       └── ...
└── evaluation/
    └── swebench-verified/    # SWE-bench evaluation scripts

If you use SWE-Edit in your research, please cite:

@article{zhang2026sweedit,
  title={SWE-Edit: Rethinking Code Editing for Efficient SWE-Agent},
  author={Zhang, Yikai and Pei, Jiaxin and Li, Kenan and Wang, Maoquan and Pan, Jin and Kang, Yu and Fu, Shengyu and Nallipogu, Elsie and Hu, Junjie and Huang, Yufan and Jin, Zijian},
  journal={arXiv preprint},
  year={2026}
}

This project is licensed under the MIT License - see the LICENSE file for details.

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft’s Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party’s policies.