Client computation

Client computation

Motivation

Keep portable declarations separate from trusted host execution, current access checks and durable effects.

Architectural Diagram

flowchart LR
  Contract[Public contract] --> Host[Verified host scope]
  Host --> Operation[Explicit runtime operation]
  Operation --> Validation[Result and permission verification]

Public import

Install the package; the subpath below is an import target, not a separate package. Registry installation requires a published EAOS release; before publication use the owning workspace’s checksum-verified build artifact.

bun add @alneas/eaos-react

import { createWorkerPluginExecutor } from '@alneas/eaos-react/client/compute'

export function bindCompute(createWorker: Parameters<typeof createWorkerPluginExecutor>[0]) {
  return createWorkerPluginExecutor(createWorker)
}

Run in a browser with Worker support. No React provider or CSS is required by this entry. Supply the Worker factory and operation allowlist, and close the executor when its authenticated session ends.

Trade-offs & Decisions

createPluginInvoker supports explicit wasm, browser-ai, native and server placements. A failure never changes placement or uploads input automatically.

@alneas/eaos-react/client/compute supplies createWorkerPluginExecutor and installComputeWorker. Create a Worker per authenticated client scope, register exact plugin:action handlers, and close the executor on logout or tenant change. Requests are correlated by IDs; cancellation, worker crashes and timeout reject outstanding calls and terminate the Worker. A subsequent call can create a fresh Worker unless the executor was closed. Cancellation affects every call sharing that Worker.

Obtain private model/WASM asset metadata through an authenticated server endpoint, then call fetchVerifiedComputeAsset. The loader fetches a signed URL without cookies, refuses redirects and verifies SHA-256 before returning bytes. Signed storage URLs need suitable CORS configuration. Local HTTP is accepted only on loopback for development.

loadGoCompute instantiates bytes using a host-supplied matching Go wasm_exec.js runtime and explicit readiness/invocation bridge. loadBrowserAiCompute selects WebGPU/WASM providers through a supplied ONNX session factory. Run these loaders inside the registered Worker; they do not turn an arbitrary Go HTTP server into browser-compatible code. Plugins own tensor preprocessing, result serialization and actual WASM/model artifacts. These adapters do not include migrated private calculations or a trained model, and do not promise a fixed execution time.