opencode

Experimental agent IDE and runtime for coding tasks, built with LangGraph, MySQL, Redis, a web workbench, and benchmark tooling for SWE-bench Lite.

中文 README

Status

This repository is not a finished product. It is an active engineering sandbox focused on:

• a persisted agent runtime with LangGraph orchestration
• a browser-based IDE shell for sessions, files, and terminal actions
• MiniMax and Claude provider adapters
• a headless SWE-bench Lite runner and official harness workflow

The current state is best described as: implemented in parts, validated in parts, still incomplete as a full agent IDE.

Useful project status documents:

• Reading Guide
• Progress Log
• Definition of Done
• Architecture Notes

What is in the repo

• apps/ide-web Minimal web IDE shell, provider hooks, browser endpoints, and dev server.
• packages/core Core domain entities such as sessions, goals, plans, tasks, memory, and tool policy.
• packages/runtime LangGraph workflow contracts, runtime orchestration, and application services.
• packages/db MySQL repositories, Redis-backed caches, bootstrap, and checkpoint persistence.
• packages/tools Built-in file and shell-oriented tools used by the runtime.
• packages/evals Smoke scripts, benchmark runners, and SWE-bench Lite export helpers.
• docs Architecture, reading order, deep-dive notes, benchmark instructions, and interview notes.

Local setup

Prerequisites

• Node.js
• pnpm 10.x
• Docker Desktop
• Python 3 for the SWE-bench export script

1. Install dependencies

pnpm install

2. Create local environment config

cp .env.example .env

At minimum, configure:

• MySQL / Redis connection values
• one LLM provider
  • MINIMAX_API_KEY, MINIMAX_BASE_URL, MINIMAX_MODEL
  • or the Claude env vars if you are using the Anthropic-compatible path

3. Start infrastructure

pnpm infra:up

The repository ships with:

• MySQL 8.4
• Redis 7

Compose file:

• compose.yaml

4. Run the web IDE

pnpm dev:ide-web

The dev server listens on 127.0.0.1:3440 by default.

Smoke checks

Persistence smoke:

pnpm smoke:persistence

MiniMax smoke:

pnpm smoke:minimax

Type check:

pnpm typecheck

SWE-bench Lite

This repository includes a headless runner that produces predictions.json for the official SWE-bench harness.

Primary files:

• packages/evals/src/swebench-lite.ts
• packages/evals/scripts/export_swebench_lite_subset.py
• docs/swebench-lite.md

Export a small instance set:

pnpm bench:swebench:export --count 5 --output .benchmarks/swebench-lite/instances.json

Run the local Lite runner:

LLM_PROVIDER=minimax pnpm bench:swebench:lite -- \
  --instances-file .benchmarks/swebench-lite/instances.json \
  --run-id swebench-lite-manual-1

Then evaluate the generated predictions in the official SWE-bench repository:

python -m swebench.harness.run_evaluation \
  --dataset_name princeton-nlp/SWE-bench_Lite \
  --predictions_path /path/to/opencode/.benchmarks/swebench-lite/runs/swebench-lite-manual-1/predictions.json \
  --instance_ids <instance ids...> \
  --max_workers 1 \
  --run_id swebench-lite-manual-1 \
  --namespace ''

Recommended reading order

Start here if you want to understand the codebase instead of only running it:

1. docs/reading-guide.md
2. docs/langgraph-in-project.md
3. docs/project-deep-dive.md
4. docs/progress.md

Suggested GitHub description

Experimental agent IDE and LangGraph runtime with MySQL/Redis persistence, MiniMax/Claude providers, a web workbench, and SWE-bench Lite tooling.