| name | add-benchmark |
| description | Add a new SWE benchmark task from a real GitHub bug-fix. Use when the user provides a GitHub issue or PR URL and wants to add it to the bench-swe pipeline. |
| argument-hint | <github-issue-or-pr-url> <language> |
| disable-model-invocation | true |
Add SWE Benchmark
Add a new benchmark task to the bench-swe pipeline from a real GitHub bug-fix.
The human provides the GitHub issue or PR URL; the agent handles extraction,
validation, and file creation.
Arguments
- url (required): GitHub issue or PR URL (e.g.
https://github.com/gorilla/mux/issues/534 or
https://github.com/gorilla/mux/pull/585)
- language (required): One of: go, python, typescript, javascript, rust,
ruby, java, c, cpp, php, csharp
Repository selection criteria
Good benchmark repos are focused libraries with a clear bug — not large
applications. Before submitting a URL, prefer repos that are:
- Size: < 50 MB and < 800 source files (excludes vendor/node_modules)
- Dependencies: < 50 direct dependencies (go.mod, package.json, etc.)
- Scope: a library or small service, not a monorepo or full application
The agent will reject repos that exceed these limits.
Steps
-
Dispatch the task-curator agent with the provided arguments. The agent
will:
- Validate inputs (URL, language)
- Check repository size and dependency count (rejects oversized repos)
- Resolve the fix PR (from issue or directly)
- Clone the repo, extract base/fix commits, and generate the gold patch
- Determine the test command from repo conventions
- Write task JSON to
bench-swe/tasks/{language}/ and patch to
bench-swe/patches/
- Run 5 inline verification checks (patch applies, files match, no leaks,
schema completeness, no test files in patch)
- Fix any issues found during verification
-
Report the result including:
- Task ID, repo, issue URL
- Files and lines changed
- Verification table