Prism-agent

A coarse-to-fine multi-solution bug-fixing framework

🔍 Diverse Repair Strategies  |  🌿 Heuristic Local Branching  |  🤝 Multi-agent Collaboration

---

Introduction

Prism is a coarse-to-fine multi-solution bug-fixing framework that explores diverse repair strategies, refines them through heuristic local branching, and synthesizes complementary solutions via multi-agent collaboration.

---

Install

Install from source

Latest features, recommended for development.

git clone https://github.com/prism-agent-code/prism-agent.git
cd prism-agent
conda env create -f environment.yml

---

Quick Start

cd src/workflow
python workflow.py \
  --instance_id <instance_id> \
  --tjs_repo <tjs_repo> \
  --dataset <dataset> \
  --tjs_test_date <tjs_test_date>

---

Parameter Description

This script accepts four required command-line arguments to locate the dataset, specify the trajectory storage path, define the test date, and select the specific instance to process.

1. --dataset

• Type: str
• Required: Yes

Specifies the directory where the dataset is stored.

By default, the script reads the following file from this directory:

test-swebench_verified.parquet

For example, if you pass:

--dataset /data/swebench

then the actual file path read by the script will be:

/data/swebench/test-swebench_verified.parquet

---

2. --tjs_repo

• Type: str
• Required: Yes

Specifies the storage path for trajectory results.

This argument is assigned to the variable REPO. Although the variable name is REPO, in this context it actually refers to the output directory or file path for trajectory data rather than the name of a code repository.

It is typically used for:

• Saving trajectory data generated during execution
• Recording intermediate results, execution logs, or reasoning traces
• Providing the trajectory file path for later analysis, reproduction, or debugging

---

3. --tjs_test_date

• Type: str
• Required: Yes

Specifies the test date.

This argument is assigned to the variable TEST_DATE, and is generally used to:

• Mark the date of the current test run
• Distinguish different batches of test results
• Serve as a time identifier in result folders, logs, or reports

Example:

--tjs_test_date 2026-03-25

---

4. --instance_id

• Type: str
• Required: Yes

Specifies the exact sample ID (instance ID) to process.

This argument is assigned to the variable require_instance, and is typically used to:

• Process only one specific sample
• Filter the target instance from the full dataset
• Debug, reproduce, or verify a single SWE-bench case

Example:

--instance_id astropy__astropy-14182

---

Usage Example

python your_script.py \
  --dataset /data/swebench \
  --tjs_repo /path/to/trajectory \
  --tjs_test_date 2026-03-25 \
  --instance_id astropy__astropy-14182

---

Before Running

Please make sure that:

1. The directory specified by --dataset exists
2. The file test-swebench_verified.parquet exists in that directory