Skip to content

quickstart.md

Latest commit

 

History

History
105 lines (78 loc) · 3.57 KB

quickstart.md

File metadata and controls

105 lines (78 loc) · 3.57 KB
Raw

Quick start

This guide runs one CPU task with the Codex coding agent. It keeps the first setup small; other tasks, Pi providers, GPU requirements, optional submission APIs, Claude Code, proxies, and custom Codex providers are covered in the full evaluation guide.

Run every command below from the Search-SWE repository root.

1. Install the launcher

Search-SWE requires Python 3.12+, Docker, and enough CPU, memory, and storage for the selected task. In any existing Python 3.12+ environment, install the launcher and asset-downloader dependencies directly:

python -m pip install -r scripts/requirements.txt

An isolated venv or Conda environment is recommended if you do not already use one, but Search-SWE does not require a particular environment manager. Task-specific Python packages are installed in Docker images, not in this host environment. Confirm that Docker is available before continuing:

docker info

2. Restore one task

Start with task-1-1, a CPU task with a small asset bundle. Do not download every task for a first run:

python scripts/download_assets.py --task task-1-1
python scripts/download_assets.py --task task-1-1 --verify-only

The second command verifies file sizes and SHA-256 checksums. See the asset guide for cache, offline, and per-kind options.

3. Configure the two services

Copy the complete environment template, then fill only the Codex-agent and RewardKit-verifier sections used by this walkthrough:

cp .env.example .env

For task-1-1, fill in these two credential pairs:

  • AGENT_OPENAI_BASE_URL and AGENT_OPENAI_API_KEY run the coding agent.
  • VERIFIER_OPENAI_BASE_URL=https://api.deepseek.com/ and VERIFIER_OPENAI_API_KEY run the independent deepseek-flash trajectory judge through RewardKit 0.2.0.

Set AGENT_MODEL in .env, or pass --model on the command line. The agent and verifier may use the same DeepSeek account, but they remain separate configuration groups and only the verifier receives the VERIFIER_* values. task-1-1 also permits OpenRouter and Jina as submission resources; their keys are optional and are not needed for an implementation that uses only the provided corpus and local runtime.

The local .env is ignored by Git. Do not commit or print credentials. On a multi-user Unix host, restrict it after adding credentials:

chmod 600 .env

4. Preview, then run

First inspect the Harbor command without starting containers or making API calls:

bash scripts/run_task.sh --task task-1-1 --dry-run

The preview checks command construction only; it does not validate credentials, assets, Docker, hardware, or service availability. Start a fresh evaluation after those prerequisites are ready:

bash scripts/run_task.sh \
  --task task-1-1 \
  --reasoning-effort high \
  --output jobs/task-1-1-codex

Omit --reasoning-effort when the selected model or provider does not support it. Results are written below the output directory as Harbor job records. Inspect the job reward, verifier logs, and transferred artifacts before treating the run as successful.

Next steps

  • Read the full evaluation guide before selecting another task or agent. It documents Codex, Pi, and the pinned Claude Code 2.1.273 launcher, and its task matrix lists exactly which additional credentials and hardware each task uses.
  • Read the selected task's README.md and instruction.md for its resource budget, submission contract, and scoring rules.
  • Use the asset guide to restore only that task's fixed inputs.