A perpetual agentic runtime for Artificial General Intelligence research built on the RuV Stack in Rust, WebAssembly, TypeScript and napi-rs.

rGi keeps an agent's work alive across individual prompts and clean restarts. It records observations, asks a planner for actions, checks authority and budgets, executes approved work and retains outcomes. Ambiguous outcomes stop execution for reconciliation instead of risking duplicated effects.

Status: experimental runtime foundation, not demonstrated AGI. The native and WASM policy kernel is functional. The persistent supervisor runs in Node 24. Upstream adapters are source grounded and fixture tested; live services, native upstream packages and physical sensors are not certified by these tests.

Requires Node 24 or newer, npm and a local writable disk.

git clone https://github.com/ruvnet/rGi.git
cd rGi
npm ci --ignore-scripts
npm run demo -- --cycles 5
node src/cli.ts status --db .rgi/runtime.db

The demo runs five simulated actions, persists five receipts of success and charges five accounting micros. These micros are a configured accounting unit, not a claim of provider billing. Run the demo again to continue from the stored sequence without replaying completed action IDs.

npm start -- --config examples/rgi.json
# In another terminal:
node src/cli.ts stop --db .rgi/runtime.db

The second command durably stops the loop. SIGINT and SIGTERM shut down the current process; an interrupted execution becomes uncertain. Nothing is installed as a background service automatically. The example plugin is a simulation, not an LLM or a robot controller.

The TypeScript policy is the portable host guard. An optional native evaluator can further restrict it, never widen it. This initial release does not move the entire daemon into Rust or run persistent SQLite in a browser. WASM is a portable policy kernel, not a browser background service.

examples/local-plugin.ts exports a Planner and Executor. Replace them with reviewed implementations and explicitly allow each required capability in the configuration. Plugins are trusted executable code, not sandboxed model output.

import { Runtime } from './src/runtime.ts';
import { MetaHarnessExecutor } from './src/adapters/metaharness.ts';

const executor = new MetaHarnessExecutor([{
  capability: 'inventory.read',
  tool: {
    name: 'inventory', server: 'local', description: 'Read inventory',
    inputSchema: { type: 'object' },
    handler: async () => ({ available: 3 }),
  },
  validate: payload => Object.keys(payload).length === 0,
  costMicros: 1,
}]);
const runtime = new Runtime({
  dbPath: './inventory.db', executor,
  config: { allowedCapabilities: ['inventory.read'], budgetMicros: 100 },
});
runtime.enqueue({ id: 'inventory:001', capability: 'inventory.read',
  payload: {}, estimatedCostMicros: 1, confidence: 1 });
await runtime.step();
runtime.close();

Default authority is empty and default budget is zero. No generic shell executor, arbitrary URL executor or public HTTP API is provided. Do not register side-effecting handlers unless their service implements independent idempotency and outcome reconciliation.

See integration provenance for exact upstream commit pins, method signatures and remaining live validation.

Rust is pinned in rust-toolchain.toml. On a host with rustup:

rustup toolchain install 1.90.0 --profile minimal --target wasm32-unknown-unknown
cargo test --workspace --locked
cargo build -p rgi-napi --release --locked
cargo build -p rgi-wasm --release --target wasm32-unknown-unknown --locked
cargo install wasm-bindgen-cli --version 0.2.104 --locked
node scripts/check-bindings.ts

The last command creates artifacts/rgi.node and WASM JS glue, loads both and compares actual results with the host guard. Set nativeBinding to ../artifacts/rgi.node in examples/rgi.json to use the native kernel in the daemon. See Rust boundary notes.

npm run typecheck
npm test
cargo install cargo-audit --version 0.22.2 --locked
npm run harness

The strict harness fails if a mandatory test, build, binding execution or advisory check fails or is unavailable. npm run harness -- --local permits unavailable tools for development but still reports an incomplete release gate. Reports include source hashes, tool versions, commands, statuses and timings in artifacts/. CI runs the same gate with read-only repository permissions and no package publication.

npm run benchmark measures portable policy checks. node scripts/benchmark-runtime.ts measures full durable no-op dispatch. Neither measures agent intelligence, useful human work equivalence or production model throughput.

1. Actions are journaled and budget is reserved before dispatch.
2. Completed IDs deduplicate; reused IDs with different content fail.
3. Unknown outcomes keep reservations and stop execution. No automatic retry.
4. Confidence is input metadata, not independently calibrated truth.
5. Observations remain distinct from prediction envelopes and model snapshots.
6. Online adaptation is a bounded scalar experiment; unrestricted weight or permission mutation is not enabled.
7. Policy checks do not sandbox trusted native or JS code. Use OS isolation for untrusted executors.
8. Budget estimates are not an external billing firewall. Provider-side limits are still required.

Read operations, threat model, architecture decisions and scope and acceptance before continuous deployment.

Source publication is separate from npm/crates.io release. Package publication remains disabled. A project license has not yet been selected; dependency licenses remain their respective authors' licenses.

The September research review maps eight primary sources to implemented mechanisms and explicit limits. The runtime now supplies durable execution feedback, filters expired observations at planning time, checkpoints learned planner state with queued actions, and requires a new runtime instance after a planning failure.

src/planning.ts provides bounded predictive beam search through an injected dynamics model. src/adaptive.ts connects outcomes to a scalar online learner for simulation. src/evaluation.ts checks paired audit results, retention, cost and permissions; src/promotion-ledger.ts persists audit consumption and statistical budgets across evaluations. None automatically installs a candidate or expands execution authority.

Run node scripts/benchmark-adaptation.ts for the fixed synthetic comparison. With 128 equal calibration interactions per model and parameters frozen during audit, adaptation completed 120/120 cases versus 80/120 for the frozen baseline. All gains came from the negative dynamics regime; the other two regimes retained 40/40. This tests instances of one scalar family, not general intelligence or transfer to unseen domains. See ADR 006.