@agenticprimitives/intent-resolver

Part of Agentic Primitives — the open-source trust substrate for agentic applications: identity that can sign, authority checked at act time, evidence the owner carries. Developer kit · All packages

Status: STUB / SKELETON (Wave 0.5 of the W1 implementation wave). W1 ships types + PassThroughResolver only; the full resolver engine lands in W2 per the w1 implementation wave plan.

Between "what a user said" and "what the system can match" sits resolution: turning an opaque expressed intent into a normalized canonical order. ERC-7683 made the resolver a named role in cross-chain intents; Anoma made it the heart of its architecture. Both treat resolution as separate from matchmaking — and so do we, because the resolver is where AI enters the trust chain. When a model infers a constraint a user never typed, that inference needs provenance: which model, which prompt, what confidence, and whether the user confirmed it. In this substrate that provenance is a typed ResolutionReceipt, and resolution happens under the same canonical identity, delegation, and audit trail as everything else the agent does.

This package is the designed Layer 4 of that spine — interface reserved now, engine landing in W2 — so the resolver grows in its own bounded home instead of being bolted onto the marketplace later.

Part of agenticprimitives — the trust substrate for the agent economy: one canonical Smart Agent identity with custody, delegation, naming, credentials, and audit evidence designed as one system.

What ships today

The complete W1 surface is three exports:

• IIntentResolver — resolve(intent): Promise<ResolvedOrder | null>. The stable contract every future resolver implements.
• ResolvedOrder — resolvedFromIntentId + canonicalConstraints + expandedAssumptions + validationRequirements, with an optional erc7683Order slot reserved for cross-chain interop.
• PassThroughResolver — the deterministic W1 implementation: returns the intent's existing ConstraintSet / AssumptionSet unchanged. No LLM, no normalization, no surprises.

import { PassThroughResolver } from '@agenticprimitives/intent-resolver';

const resolver = new PassThroughResolver();
const order = await resolver.resolve(intent); // constraints pass through unchanged

What lands in W2 (designed, not shipped)

Constraint normalization (colloquial scope → canonical codes), credential-requirement expansion, LLM-driven resolution with full ResolutionReceipt provenance (RR-INV-01..05 — every resolution produces a receipt, and inferred constraints cannot bind without user confirmation), and ERC-7683 order emission. Per-domain resolvers live in apps or W2+ resolver packages, never here.

Where this is heading / market context

• ERC-7683 defines the resolver-assumption pattern for cross-chain orders; this package follows its resolver/matchmaker separation and targets its order shape for interop.
• Anoma treats intent resolution as constraint solving; the ConstraintSet types this resolver consumes (from intent-marketplace) are CSP-shaped for the same reason.
• x402 / Google AP2 machine-payment rails make the stakes concrete: a resolved order can bind a payment mandate, so the resolution step must be attributable and auditable under the same identity substrate as the spend it triggers. That is the combination — resolution with receipts, inside one delegation + custody + audit chain — that stitched stacks do not offer.

Boundary with the engagement layer (spec 336)

Resolution normalizes the requester's side; it never speaks to candidates. Downstream, a candidate ranking is a resolver's assertion, not a fact — apdisc:MatchCandidate carries named-resolver provenance, multi-dimensional evidence, and stated reasons, and a single similarity score must never decide who receives sensitive data. Whether a matched provider will actually take this intent, on what terms is the engagement layer's question (spec 336): probes carry audience-bound projections (never the canonical intent), and the six response kinds — including decline and temporarily-unavailable — are recordable business facts, not resolution failures.

Authoritative spec: spec 239 — intent spine, §4.5 (see spec.md). Owns spine layer 4 (skeleton); bounded surface in CLAUDE.md and capability.manifest.json.

Build

pnpm --filter @agenticprimitives/intent-resolver typecheck
pnpm --filter @agenticprimitives/intent-resolver test
pnpm --filter @agenticprimitives/intent-resolver build

Status — honest version

STUB / SKELETON. PassThroughResolver is the only implementation and it is intentionally trivial. Everything in "What lands in W2" is a spec'd commitment, not shipped capability. Wave sequencing: w1-implementation-wave-plan.md. Repo-wide security findings: docs/audits/findings.yaml.