@manya-hael/cortex
v1.0.0
Published
A reasoning fabric that routes tasks to the tools that already own them, gates dispatch through access control, and chains every decision into an auditable trail — composed from memory, unify, shield, and stamp.
Downloads
142
Maintainers
Readme
@manya/cortex
A reasoning fabric, not another agent.
Cortex doesn't answer questions itself. It decides who should — by asking
@manya/unify's mesh which already-registered tool owns the capability a
task needs, gating dispatch through @manya/shield, remembering every
decision through @manya/memory, and chaining every decision into its own
tamper-evident trail with @manya/stamp.
Usage
import { memory } from '@manya/memory';
import { unify } from '@manya/unify';
import { cortex } from '@manya/cortex';
// Tools register themselves with the mesh as they come online.
unify.registerTool({
manifest: { id: 'forge', owns: ['keyDerivation'] },
api: { deriveKey: (passphrase) => derive(passphrase) },
});
const store = memory.createMemoryStore('ara');
const brain = cortex.createCortex(store);
// Let atlas provision devices.
cortex.authorize(brain, 'atlas', 'operator', [
{ resource: 'device:*', actions: ['provision'] },
]);
const result = cortex.plan(brain, 'derive a key for the new device', {
capability: 'keyDerivation',
requesterId: 'atlas',
resource: 'device:phone-1',
action: 'provision',
});
// { intent: 'execution', owner: 'forge', authorized: true, confidence: 0.8, ... }
if (result.authorized) {
cortex.dispatchPlan(brain, result, 'keyDerivation', 'deriveKey', ['a passphrase']);
}Why this composition
- "Who should solve this?" is answered by
unify.route/unify.dispatchagainst the mesh's existing capability registry — Cortex doesn't keep a second list of who-does-what. - "What confidence do we have?" blends whether a capability is actually registered right now, whether the requester is authorized, and whether memory already has relevant episodes — not a guess.
- Every decision is remembered twice, for different reasons: once in the
bound memory store's episodic log (so other reasoning can recall why a
plan was made), and once in Cortex's own
stamp-chained decision trail (so the planning process itself can be audited independent of memory). - Dispatch never bypasses authorization, even if a caller forgets to
check
plan.authorized—dispatchPlanrefuses on its own.
