Stella Runtime

Stella Runtime is an instance-neutral cognitive runtime for OpenClaw. It binds explicit, versioned, traceable, and correctable knowledge to one host-owned Agent Run without turning model inference or generated projections into hidden authority.

Published stable release: 0.3.0. The verified rollback version is published 0.2.1. Compatibility is accepted only for OpenClaw extended-stable 2026.7.1-2 on Node.js 24.18.0; compatibility is exact, not minimum-version based.

What ships

The single package @tower1229/stella-cognitive-runtime contains:

• the OpenClaw Plugin cognitive-runtime and its operational CLI;
• versioned JSON Schemas plus generated JavaScript and TypeScript declarations;
• bounded Router, context packet, Current State, correction/outbox, provenance, generation, recovery, and conformance modules;
• the framework-admission Skill and public test Runner;
• selected exact-host compatibility evidence.

OpenClaw still owns sessions, the Agent loop, native memory tools, and its native audit trajectory. Stella Runtime is not a second Agent loop, general memory system, vector database, persona engine, or action authority.

Install

Requirements:

• Node.js 24.18.0 exactly;
• OpenClaw extended-stable 2026.7.1-2 exactly.

Install the exact stable version through OpenClaw:

openclaw plugins install @tower1229/[email protected]
openclaw plugins inspect cognitive-runtime --runtime --json
openclaw cognitive self-check

Library or Runner consumers can install the exact package with:

npm install --save-exact @tower1229/[email protected]

Begin in off, configure the Instance Runtime Config and Git-external Runtime and Generation storage, then pass exact-host conformance before moving through observe to enforce. Active Run Bindings are compiled from the active Pointer, Activation Receipt, immutable Generation, and State View rather than inline configuration. A configured verified domain projection upgrades this chain to the additive composite v3 Manifest/Receipt/Pointer without changing the published v2 Generation identity. See the operations guide.

Configuration reference

The machine-readable configuration authority is openclaw.plugin.json; unknown properties are rejected. The human reference covers:

• off, observe, and enforce semantics;
• active Run capacity and drain/cleanup limits;
• instance/storage identity, eligible Host scope, and binding proof locations;
• recovery root, active instance, and allowed instance revisions;
• every openclaw cognitive CLI command.

Read Configuration reference.

Data and authority

This repository contains generic Runtime source and synthetic verification fixtures. The npm package contains compiled Runtime code, v2 schemas, selected documentation/evidence, the Skill, and the public Runner. A separate private Authority Repository owns durable knowledge, identity, configuration, and migration intent. Git-external Runtime State owns Current State, unfinished corrections/outbox, protected Candidate admission and Approval Receipt records, minimized overlays, and rebuildable projections.

Private data, credentials, live databases, Recovery Snapshots, conversations, real experience records, and Instance Test Packs must not enter this repository, CI artifacts, or the npm tarball. Read Data boundaries.

CLI

The main operational surface is openclaw cognitive:

• read-only validate, non-activating build, full-barrier sync, and generation show;
• self-check, metrics;
• state, trace get|query;
• backup, read-only verify, and rollback-safe restore.

Structured operational commands require --json where offered. Run openclaw cognitive <command> --help or read the Configuration reference for exact options.

The package also exposes stella-runtime-test for repository-owned tests and an optional external Instance Test Pack. Private packs execute locally and are not copied or uploaded by the Runner.

Development and verification

npm ci
npm run typecheck
npm test
npm run test:pack-install

npm test builds generated JavaScript and runs unit, contract, integration, and pack-install tests. Pack-install creates a real tarball, audits its allowlist and public content, installs it into an isolated OpenClaw environment, checks Plugin/CLI/Skill discovery and typed-hook behavior, rehearses successor and recovery failure paths, restarts the Gateway, uninstalls, and proves config restoration. Fixtures are synthetic.

Source of truth

• Domain language
• Historical 0.1 requirements
• Historical 0.1 architecture
• Data contracts
• Requirement provenance
• Historical 0.1 and completed 0.2 roadmap
• OpenClaw compatibility evidence
• Runtime recovery ADR
• Repository authority ADR
• Changelog

This repository is the complete authority for generic Runtime requirements, architecture, decisions, evidence, implementation, build, test, and release. A consumer repository is optional and never an implementation dependency.

Upgrade, rollback, and recovery

For 0.3.0, the exact package rollback version is published 0.2.1. Use exact package versions and integrity, retain that verified artifact and receipt, create and read-only verify a Runtime Recovery Snapshot, then prove State/outbox and restart continuity before enabling observe or enforce. Package rollback and state recovery are separate operations. The full sequence is in Operations.

Known limitations

The 0.3.x line has one exact OpenClaw matrix row. It does not support runContext, native structured output, direct persisted-session embedded Runs, host next-turn injection, or bundled-only scheduling. Successor attempts are at-least-once with one successful completion, not cross-Run exactly-once delivery. Cloud disaster recovery, bare-metal imaging, generic migration, complete erasure, and retention automation are outside this release line.

Consumer product acceptance and personal usefulness are downstream non-blocking evidence; they are not generic Runtime technical release gates. Read the full support policy and known limitations.