@taskproof/core

The shared contract every runner adapter implements. Three things live here so the grader, report generator, and taskproof diff never have to branch on which agent produced a run:

• The run-artifact schema (RunArtifact) — a Zod-validated, versioned record of a single agent run: status, per-step trace (narration + actions + screenshots), network log, and token/cost usage. Claude computer use, browser-use, Gemini, and OpenAI all serialize to this identical shape. That uniformity is the moat.
• The adapter interface (Adapter) — run(input, config) => Promise<RunArtifact>.
• The cost meter (CostMeter) — token→USD accounting with cache-aware pricing and a soft per-run budget cap (wouldExceed() lets a caller stop before the next turn it can't afford; a turn can still overshoot the cap by its own cost since cost is only known once billed).

import { CostMeter, parseRunArtifact, type Adapter } from '@taskproof/core';

const meter = new CostMeter({ model: 'claude-opus-4-8', maxCostUsd: 1.0 });
meter.record({ inputTokens: 2400, outputTokens: 120, cacheReadTokens: 1800 });
if (meter.wouldExceed(estimatedNextTurnCost)) stop('budget_exceeded');

Pre-release; 0.x artifact-schema versions may break between releases.