Strands MigrationBench Agent

An RL-trainable code-migration agent that upgrades Java 8 projects to Java 17, as introduced in MigrationBench. Given a repo URI, the agent iteratively edits pom.xml and source files with shell + editor tools until mvn clean verify passes on Java 17 and the test suite is preserved.

Built on the official JavaMigrationAgent baseline, rewired to use open-source LLMs through Strands.

Quickstart

# Prerequisite: Java 17 + Maven 3.9.6 installed (see full README)
cd examples/strands_migration_agent
uv venv --python 3.13 && source .venv/bin/activate
uv pip install -e .
uv pip install -e ../../ --force-reinstall --no-deps

# Preprocess 2 repos as a smoke test
python preprocess.py --s3-bucket-name my-migration-bench-data \
    --max-repos-per-split 2 --skip-s3-sync

# Start the RL app (needs a local vLLM server running on :4000)
uvicorn rl_app:app --port 8080 --reload --reload-dir ../..

Deploy via the Prepare agent for RL → Deploy flow, then evaluate with evaluate.py / evaluate_async.py.

What’s in the example

• rl_app.py — the rollout entrypoint; loads the target repo from S3, sets up the Maven env, runs the agent with dynamically composed tools.
• reward.py — MigrationReward: build success + test-case preservation vs. the original commit.
• preprocess.py — convert the MigrationBench dataset into per-repo tarballs + metadata + JSONL rollout requests.
• evaluate.py / evaluate_async.py — batch / async evaluation via RolloutClient.
• Dockerfile + deploy.py + config.toml — container build + programmatic ACR deploy.
• models.py, utils.py, eval_utils.py — shared helpers (S3 fetch, Maven setup, build verification).

For Java/Maven install, Docker + ECR setup, CodeArtifact maven mirror (avoids Maven Central rate limits), and full evaluation config, see the full README on GitHub.