@habenula-ai/governance

The Habenula governance kernel: two pure evaluators, one for permission and one for spending.

Install

Run this in your project directory, once per project:

npm i @habenula-ai/governance

evaluatePolicy(entries, action) → decision is a hard invariant — a pure function with no side effects: no database calls, no network, no logging. It maps a set of (service, verb, noun) policy entries and a requested action to an allow / deny / hold decision, and it is the one place that authorises a tool call. evaluateSpend(input) → within | exceeds | unavailable holds the same discipline for the spending cap: totals and limits go in as values, so an action that moves money is additionally priced against the user's windows without the evaluator ever reading a ledger or a clock. Because it is pure it is exhaustively unit-testable in isolation, which is why it lives in its own leaf package.

This package decides. It does not record. The audit log's hash chain, its verifier, and the decision-closure check live in @habenula-ai/audit.

It is a leaf package with no runtime dependencies: no engine imports, no Cloudflare bindings, no agents SDK, and no Node built-ins. The engine consumes it as source (an exact-pinned workspace sibling, no build step) and owns everything environmental — the policy entries come from the DO's SQLite, and the spend totals are read there and passed in as values.

Part of the Habenula OSS release. Licensed under AGPL v3.