issuebenchkit 0.1.0

Turn real GitHub issues into small, reproducible coding-agent benchmark tasks.

中文文档

IssueBenchKit

Turn a real GitHub issue, pull request, or local bug into a small coding-agent benchmark task.

SWE-bench is great when you want a public leaderboard. Most teams need something smaller: a repeatable task built from the bugs they actually care about, with a clear test command and a report that says whether a candidate patch really fixed it.

IssueBenchKit is that local builder. It does not try to invent tests for you. It packages the issue context, base commit, reproduction command, and scoring result so you can evaluate coding agents on your own repositories.

Quick Start

pip install issuebenchkit

Create a benchmark task:

issuebench init tasks/qwen-copy \
  --repo ./qwen-code \
  --issue https://github.com/QwenLM/qwen-code/issues/4716 \
  --base 8b4f3b2 \
  --test "npm test -- copyCommand.test.ts"

Run the task against a candidate checkout:

issuebench run tasks/qwen-copy --repo ./candidate-qwen-code --out after.json

Compare before and after:

issuebench score tasks/qwen-copy --before before.json --after after.json

Export a report:

issuebench export tasks/qwen-copy --format html --out report.html

What It Stores

Each task directory contains one issuebench.json manifest:

• source repo path and optional GitHub issue URL
• base commit or version marker
• reproduction / validation command
• expected signal, notes, and tags

Run results are plain JSON files with exit code, duration, command, stdout tail, stderr tail, and the pass/fail verdict. They are easy to archive, diff, or attach to a PR.

Why Not Just Use SWE-bench?

Use SWE-bench for public comparison. Use IssueBenchKit when you need:

• a benchmark task for a private or small repo
• a tiny task that can run in CI
• a before/after report for one real bug
• a dataset of issues that reflects your own engineering workflow

Current Scope

The first version is intentionally small:

• generic shell test commands
• JSON manifest files
• before/after scoring
• JSONL and single-file HTML export

It does not generate tests automatically, mutate repositories, or claim that one command can evaluate every language ecosystem.

License

MIT