v0.9.5 · MIT · CPU-only · Python 3.10–3.14

Configuration coverage, measured deterministically.

High code coverage does not guarantee high configuration coverage. ConfigReach discovers runtime configuration inputs, values, branches and important combinations, then maps them to test evidence without an LLM, GPU, API key, telemetry or hosted service.

No LLMNo GPUNo telemetryNon-root containerSBOM + provenance
Curated 50K benchmark

50,000 controlled cases. Balanced ground truth. Production-engine scoring.

The committed curated benchmark uses explicit, version-controlled scenario families with deterministic labels: 25,000 positive and 25,000 negative cases across supported languages and configuration formats. It is scored through configreach.engine.scan.

50,000scenarios
25,000true positives
25,000true negatives
0 / 0false positives / false negatives

Precision

100.0000% on the committed controlled curated benchmark.

Recall

100.0000% on the committed controlled curated benchmark.

Claim boundary

Controlled benchmark evidence, not independently human-labelled evidence and not a universal real-world accuracy claim.

Enterprise security

Designed for reviewable, restrictive deployment.

Static scanning is local and offline and does not execute target application code. v0.9.5 adds explicit enterprise software-intake controls and untrusted-repository boundary regressions.

Runtime boundary

Repository-controlled file symlinks are ignored; the release container runs non-root and is tested with egress disabled, a read-only root filesystem and a read-only source mount.

Supply chain

Immutable Action pins, CodeQL, Gitleaks, Dependabot, dependency review, pip-audit, Trivy, reproducible packages, SBOM/provenance and OIDC publishing.

Governance

Validation workflows are read-only reproduction gates. Enterprise teams can pin, rebuild, scan and mirror an approved release internally.

Install

Run the verified release locally.

shell
python -m pip install --upgrade configreach
configreach scan .
configreach coverage .