Pre-install security gate for npm lifecycle scripts using TypeSafe System One. Evaluates preinstall, install, and postinstall hooks across intent, threat severity, secret access, and remote execution to intercept supply-chain attacks before execution.
pkg-gate uses `@typesafe-ai/sdk` to evaluate lifecycle scripts in parallel across intent (Choice), threat severity (Score), secret exfiltration (Noul), and remote execution (Noul).
Adds structured choices or scores to the workflow; performance and cost benefits have not been independently verified.
Pinned lifecycle policy and simulator reviewed. Missing keys or provider failures can produce simulated answers; the final report keeps an offline-simulator model marker but drops the intermediate fallbackReason. The site did not install packages, test threat detection or run live Jev.
Review npm lifecycle scripts before installation without confusing a local simulator with a hosted security check.
Material analysed：
gate.js adds fallbackReason to an intermediate result, but drops it when building findings/structured output. The simulator's model marker identifies offline simulation; callers cannot recover the failure reason from the final report.
Inspect the supporting source ↗The policy tracks maximum severity/secret/remote signals and minimum intent/severity confidence. A low minimum confidence routes to warn rather than treating a top label as sufficient.
Inspect the supporting source ↗The fast path allows a manifest with no lifecycle hooks. It does not examine every dependency or establish that installing the full dependency tree cannot execute code.
Inspect the supporting source ↗The aggregation code defaults missing intent/severity confidence to 1 and missing severity/secret/remote signals to 0. The low-confidence warning only catches actual low values; an integrator must verify the SDK or injected client's full response validation before relying on this policy.
Inspect the supporting source ↗Use it to inspect supply-chain rubrics alongside deterministic package inspection. Preserve result provenance and failure reasons in the integrator; simulated reports should not authorize installation.
Trace evaluations into findings and report.structured.scripts. Check the model='jev-latest (offline simulator)' marker and verify that fallbackReason is absent; use a fake failed provider rather than untrusted scripts. Inject an incomplete response in a local test and require a strict shape error before aggregation; do not treat numeric defaults as a model verdict.