Prometheus Bug Reproduction Agent

An automated bug reproduction research agent powered by large language models (LLMs) and a code knowledge graph. It clones target repositories, builds a knowledge graph, analyzes and attempts to reproduce issues inside a container, and outputs reproduction commands and results.

Requirements

• Python 3.11+
• Neo4j 5.x (local or remote)
• Docker
• Git

Installation

• Method 1 (recommended): pinned dependencies via requirements.txt

pip install -r requirements.txt

• Method 2: install via pyproject.toml

pip install hatchling
pip install .

• Create a working directory (for logs and cloned repositories cache):

mkdir -p working_dir

Configuration

Copy the example env file and adjust as needed:

cp example.env .env

Key settings (from example.env):

• PROMETHEUS_NEO4J_URI, e.g. bolt://localhost:7687
• PROMETHEUS_NEO4J_USERNAME / PROMETHEUS_NEO4J_PASSWORD
• PROMETHEUS_WORKING_DIRECTORY, e.g. working_dir/
• PROMETHEUS_OPENAI_FORMAT_API_KEY and other LLM keys (Anthropic/Gemini as needed)

If you need to access private repositories, prepare a GitHub token:

export GITHUB_TOKEN="github_pat_xxxx"

Docker prerequisites

Before running, ensure Docker is available on the host and start the following services.

Neo4j (required)

Default ports: HTTP 7474, Bolt 7687.

Minimal startup (ports consistent with example.env):

docker run -d \
  --name neo4j_prometheus \
  -p 7474:7474 \
  -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/password \
  neo4j:5

Recommended (enable APOC and set memory):

docker run -d \
  --name neo4j_prometheus \
  -p 7474:7474 \
  -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/password \
  -e NEO4J_PLUGINS='["apoc"]' \
  -e NEO4J_dbms_memory_heap_initial__size=2G \
  -e NEO4J_dbms_memory_heap_max__size=4G \
  -e NEO4J_dbms_memory_pagecache_size=2G \
  neo4j:5

If you map different host ports (e.g., 7475:7474, 7688:7687), update PROMETHEUS_NEO4J_URI in .env accordingly, e.g., bolt://localhost:7688.

Postgres (optional, for checkpoint storage)

Default port: 5432.

docker run -d \
  --name postgres_prometheus \
  -p 5432:5432 \
  -e POSTGRES_USER=app2_user \
  -e POSTGRES_PASSWORD=app2_password \
  -e POSTGRES_DB=app2_db \
  postgres:16

You can skip this if you don't use Postgres checkpoints.

Usage

Command-line execution (required argument: --dataset_file_path):

python -m app.main --dataset_file_path projects/swe-polybench-verified.txt --github_token "$GITHUB_TOKEN"

Optional arguments:

• --file, -f: output path for predictions, defaults to projects/predictions_YYYYmmdd_HHMMSS.json

Argument description:

• --dataset_file_path, -d: path to the project list file (required). Each line should look like <name> <git_https_url>. Lines starting with # are treated as comments.
• --github_token, -g: token for accessing private repositories (optional).

Workflow

1. Parse the project list file and iterate projects
2. Clone/update repositories and build the knowledge graph
3. Start a Docker container (with volume mapping for real-time sync)
4. Auto-generate testsuite commands and attempt to run them
5. Auto-generate/execute environment setup commands to fix env/dependency issues
6. Record and export reproduction results and related information

Project structure

• app/main.py: CLI entry (arguments: --dataset_file_path/--github_token/--file)
• app/configuration/: configuration loading
• app/container/: container management
• app/lang_graph/: language graphs/subgraphs
• app/services/: knowledge graph, repository, LLM, Neo4j services
• projects/: sample project lists and outputs

Notes

• Ensure Docker is running and you have permission to build images
• Ensure Neo4j is reachable and PROMETHEUS_NEO4J_* are set correctly
• Building the knowledge graph for large repos may take time and memory

Example

export GITHUB_TOKEN='github_pat_xxxx'
python3 -m app.main \
  --dataset_file_path projects/swe-polybench-verified.txt \
  --github_token "$GITHUB_TOKEN"

Run output and files inside the container

After the program starts, a general-purpose build container is launched and it prints the command to enter the container, for example:

To enter container, run: docker exec -it <container_short_id> /bin/bash

Inside the container (workdir /app), two key files are generated:

• /app/prometheus_setup.sh: initial environment setup script auto-generated by EnvAgent
• /app/prometheus_testsuite_commands.txt: the extracted/generated testsuite command list (one command per line)

Example: enter the container and run the environment script:

# Enter the container (use the command printed in logs)
docker exec -it <container_short_id> /bin/bash

# Run the environment setup script (idempotent/re-runnable)
bash /app/prometheus_setup.sh



Tip: Edits you make on the host are reflected in `/app` inside the container in real time, which makes it easy to iterate on the environment script and testsuite commands.