@intelligo-dev/executions

The execution lifecycle (admission, running, settlement) and the model cost registry.

Part of Intelligo, an application framework and operational platform for vertical AI SaaS products. Every @intelligo-dev/* package is released at one version and shares one database schema; pnpm dlx @intelligo-dev/cli create my-app installs the set an application needs. Documentation: intelligo.dev/docs/packages/executions.

Install

pnpm add @intelligo-dev/executions drizzle-orm

drizzle-orm is a peer. The executions table is part of the framework's schema, which intelligo migrate from @intelligo-dev/cli applies.

The boundary

Intelligo records what an AI execution cost and whether it was allowed. It does not wrap the AI framework: prompts, tools and orchestration stay native.

Use

The composition root builds one instance, with the ports bound:

// lib/intelligo.ts
import {
  DEFAULT_MODELS,
  createExecutions,
  registerModels,
} from "@intelligo-dev/executions";

registerModels(DEFAULT_MODELS);

export const executions = createExecutions({
  checkEntitlement, // may this workspace run this model? takes the hold
  settleUsage, // charge what the run used
  findSettlement, // did that charge already commit?
  releaseHold, // the run failed: give the hold back
});

A run is one begin, then exactly one complete or fail:

const run = await executions.begin({
  workspaceId,
  userId,
  capability: "support.reply",
  model: "google/gemini-2.5-flash",
});
if (!run.allowed) return refuse(run.code, run.reason);

try {
  const result = await agent.generate(messages); // your framework, natively
  await run.complete({ usage: result.usage, model: "google/gemini-2.5-flash" });
} catch (error) {
  await run.fail({ error });
  throw error;
}

All four ports are optional, so this package never imports billing; the reference composition root binds them to @intelligo-dev/billing. The lifecycle claims each transition with a compare-and-swap before money is spent, so a late complete racing a fail cannot charge twice or release twice.

executions.reconcile(executionId, { abandonRunningAfterMs? }) repairs a run whose process died: it asks findSettlement whether the charge committed, confirms it onto the row or re-runs settlement from the recorded usage, and fails a running row that outlived the threshold. findStaleExecutions finds the candidates; listExecutions, summarizeExecutions and summarizeExecutionsByDay read the record for a usage page.

Model pricing

@intelligo-dev/executions/pricing is a deliberate zero-import leaf that client bundles can reach. Model ids are registry keys: an unregistered id has no price, and guessing one is how a model runs on the cheapest provider and bills at the most expensive rate. Pricing throws UnknownModelError instead.

Nothing self-registers. DEFAULT_MODELS is the catalogue the framework ships, as data; a deployment registers it, its own contracted rates, or a model the framework has never heard of:

import { registerModel } from "@intelligo-dev/executions/pricing";

registerModel({
  id: "acme/large-1",
  provider: "acme",
  model: "large-1-2026-08",
  displayName: "Acme Large",
  costPerMInputTokens: 1.25, // USD per million, as the provider quotes it
  costPerMOutputTokens: 5,
  maxOutputTokens: 8_192, // sizes the worst-case hold at admission
  capabilities: {
    thinking: false,
    toolCall: true,
    vision: false,
    webSearch: false,
    codeExec: false,
  },
});

providerCost is what a run cost the deployment; chargeFor applies the deployment's currency, rate and margin (BillingRate) and returns both figures as money in micros.

Licence

Apache-2.0