JJarvisv0.1 specification v0.1

Terminology

Protocol terms have fixed meaning.

Jarvis uses worker, actor, session, policy, request, review, evidence, and learning terms to describe portable human-agent collaboration records.

Core terms

Worker

A protocol participant. Worker records identify who or what participates and the authority and accountability attached to that participant.

Actor

An acting identity for protocol events and contributions. Actor records bind accepted state to authority and contribution scope.

HumanWorker

The accountable human participant. HumanWorker owns direction, policy authority, review authority, correction, and takeover judgment.

AgentWorker

The agent participant. AgentWorker contributes planning, research, drafting, evidence collection, and governed improvement proposals through host-owned execution.

Host

The implementation environment. Hosts own UI, identity, auth, storage, execution, tools, models, monitoring, deployment, and workflow.

Compatible implementation

A host or external system that produces, validates, rejects, exports, and preserves Jarvis-compatible protocol records.

Work terms

WorkSession

The durable record of a focused unit of human-agent work. WorkSession is the source of truth for protocol state.

JarvisEvent

An append-only event with sequence, Actor, timestamp, payload, previous hash, and event hash.

Policy

The human-owned boundary for autonomous AgentWorker action.

PolicyDecision

The protocol record that explains why an AgentWorker action is allowed, denied, narrowed, or sent for human review.

Control and proof terms

  • Request records scoped deferral when AgentWorker cannot continue safely without human input.
  • Review records human judgment over a Request, artifact, action, contribution, evidence item, or outcome.
  • ApprovalScope bounds Review-approved continuation to a specific request, action, Actor, WorkSession, expiry, and max use count.
  • Takeover records direct human control over a scope and blocks stale autonomous continuation with a lock epoch.
  • Contribution records who did what and preserves individual refs for shared work.
  • EvidenceManifest exports portable proof without host-private fields.
  • LearningRecord records what the human, agent, or pair learned from source events.
  • MemoryProposal and SkillProposal keep future behavior changes governed.
  • OutcomeReport carries post-session feedback into governed LearningRecords without mutating sealed records.
Source: core objects Source: request protocol