| name | swebench |
| description | Run and inspect SWE-bench benchmark workflows for alan's coding line.
Use this when:
- The user wants to run SWE-bench Lite or curated SWE-bench subsets
- The user wants to prepare benchmark workspaces or materialize suite manifests
- The user wants to score benchmark predictions with the official harness
- The user wants benchmark operations to remain separate from the repo-coding product contract
|
| metadata | {"short-description":"Operate SWE-bench benchmark workflows","tags":["benchmark","swebench","evaluation","coding"]} |
| capabilities | {"required_tools":["bash"]} |
| compatibility | {"requirements":"Use the package-local `bin/` and `scripts/` entrypoints for SWE-bench preparation, manifest materialization, harness setup or validation, and scoring; supply prediction execution independently."} |
SWE-bench
This first-party package owns SWE-bench benchmark orchestration for alan's
coding line.
Working Model
- Treat SWE-bench as a measurement layer, not as the definition of coding
behavior.
- Keep benchmark orchestration separate from
$repo-coding.
- Use this package to prepare benchmark inputs, materialize suite manifests,
validate the official harness environment, and score predictions produced
by an independent Agent Process execution path.
- Keep execution orchestration outside this package until Alan has a
file-native benchmark Process launcher.
Package Surfaces
bin/: package-local executable entrypoints for deterministic SWE-bench
preparation/materialization helpers
scripts/: official harness environment setup and scoring
references/: package map and operator guidance
evals/files/: benchmark templates and curated instance-id lists
Rules
- Do not encode repository-specific or corpus-specific behavior into alan's
coding product.
- Benchmark findings should be generalized back into reusable coding,
governance, prompt, or harness improvements.
- The benchmark operator flow should remain auditable and reproducible.
repo-coding stays the measured coding skill; this package owns only the
benchmark orchestration layer.