Scripts
May 7, 2026 ยท View on GitHub
This repository keeps most operational entry points in scripts/.
The public surface is intentionally centered on a small set of reproducible commands. Historical or experimental helpers still exist, but they should not be treated as the first thing a new user runs.
Start Here
Most users only need these commands:
run_release_checks.pyEnd-to-end local validation used by quickstart and CI.aggregate_metrics.pyAggregate normalizedruns.jsonlintosummary.jsonl.validate_records.pyValidatejsonorjsonlfiles against SIP-Bench schemas.run_eval.pyImport benchmark outputs or build and execute single-run plans.run_protocol.pyExecute or import multi-run protocol suites from config files.evidence_gate.pyEvaluate whether a completed summary meets the repository's evidence thresholds.check_plan_matrix.pySanity-check protocol suite configs and expected artifact paths.
Core Categories
Validation And Aggregation
run_release_checks.pyaggregate_metrics.pyvalidate_records.pyevidence_gate.pycheck_plan_matrix.py
Benchmark Execution And Import
run_eval.pyrun_protocol.pysmoke_adapters.py
Result Rendering
build_strategy_comparison.pybuild_results_gallery_artifacts.pybuild_task_family_ablation.py
EvoAgentBench Retained Tooling
These scripts are kept because the repository still ships real OmniMath result packages and fixtures:
generate_evoagentbench_omnimath_demo.pyanalyze_evoagentbench_omnimath_results.pybuild_evoagentbench_raw_phase_comparison.pybuild_evoagentbench_failure_taxonomy.pybuild_evoagentbench_verifier_sensitivity_audit.pyrun_evoagentbench_omnimath_codex_real.py
They are retained artifacts and utilities, not the primary public story of the repository.
First Commands
python3 -m venv .venv
. .venv/bin/activate
python -m pip install -e .
python3 scripts/run_release_checks.py
python3 scripts/check_plan_matrix.py --protocol-dir protocol
python3 scripts/aggregate_metrics.py --runs results/dryrun/sample_runs.jsonl --out /tmp/sip_summary.jsonl
Notes
Linux-firstis the maintained public support target.- Some scripts require local benchmark checkouts under
benchmarks/. - Some optional paths require provider credentials, but the default validation path does not.
- Experimental account-auth bridge scripts remain in the repository for historical completeness, but they are not part of the recommended public workflow.