voyager-contract

Voyager's cognitive nervous system. One universal Voyager for the user; inside, a model-independent substrate that lets the reasoning brain (any LLM) decompose intent, discover and orchestrate modular senses, link evidence into causal chains across browser → code → host → network → cloud → data, act only through consent, verify, and get better each mission.

The senses stay separate, testable organs (@dir-ai/voyager, voyager-repo, voyager-net, …). This is not a monolith — it's the connective tissue that makes them one coherent agent. The model is the brain and can be swapped without losing memory, capabilities, or state.

npx @dir-ai/voyager-contract demo        # the canonical end-to-end mission

Three foundational layers

1. Universal Cognitive Contract (CognitiveClaim) — one versioned, mergeable object for every step on every sense: observe / infer / simulate / act (consent-gated) / verify / learn. Carries evidence, provenance, entities, relationships, causal chains, hypotheses, contradictions, unknowns, next probes (with expected information gain), predicted impact, and memory candidates. Streams and refines over time.
2. Capability Graph — a queryable graph of what each sense/skill/tool/model can do, with cost, environments, and a learned reliability updated from outcomes. The agent picks by expected utility, and knows its own blind spots (detectGaps).
3. Mission Graph — a live plan, not a fixed list: goals, the claims that inform them, the entities/relationships/causes they surface. It re-plans after every claim — picking the most informative next probe, flagging cross-sense contradictions, and assembling the causal chain.

The end-to-end mission it proves

demo runs "the orders page is slow — find out why and fix it" through the real machinery: the web sense measures it and finds a slow API; the net sense rules the network out (a refuted hypothesis); the host sense finds the time is in the DB (N+1); the repo sense binds the API to the file and the commit that introduced it; the brain assembles the causal chain commit → N+1 → DB → API → page; it proposes a reversible fix that is withheld pending consent (the hands never mutate production autonomously); and it verifies the result and harvests durable memory.

Non-negotiable

Sensing can be autonomous; acting is always consent-gated, reversible, and attested. The act operation carries a blast class, a rollback, and a consent requirement — an unconsented action stays withheld.

Library

import { CognitiveClaim, CapabilityGraph, MissionGraph, newClaim } from '@dir-ai/voyager-contract'

Roadmap

Adapters that turn the shipped senses' output into CognitiveClaims; the voyager-browser sense; the World Model backed by Repotector (episodic / semantic / procedural / temporal / causal memory); a verification + evaluation loop that calibrates confidence and synthesizes new skills; counterfactual simulation.

License

MIT