@taskproof/adapter-claude

The Claude computer-use runner adapter. Given a task spec, it drives a Playwright-managed Chromium instance through Anthropic's computer-use API and emits a @taskproof/core RunArtifact — the same shape every other adapter produces.

import { createClaudeAdapter } from '@taskproof/adapter-claude';
import { parseTaskSpec } from '@taskproof/spec';

const adapter = createClaudeAdapter();
const artifact = await adapter.run(
  { spec: parseTaskSpec(yaml), runId: 'run-1' },
  {
    model: 'claude-opus-4-8',
    apiKey: process.env.ANTHROPIC_API_KEY,
    maxCostUsd: 1.0,
    artifactsDir: 'taskproof-runs',
  },
);

What it does each turn: screenshot → send to Claude with the computer tool → execute the requested actions on the page (clicks, typing, scrolling, drags, key combos, zoom) → return a fresh screenshot. It tracks token cost with a soft --max-cost cap (it stops before a turn it can't afford, so it can overshoot by ≤1 turn's cost), records a per-step trace (narration + actions + screenshots), and logs network requests.

• Models: computer_20251124 + beta computer-use-2025-11-24 for Opus 4.8/4.7/4.6, Sonnet 4.6, Opus 4.5; computer_20250124 + beta computer-use-2025-01-24 for Sonnet 4.5, Haiku 4.5, Opus 4.1, Sonnet 4.0, Opus 4.0. Default display is 1280×800, which keeps screenshot coordinates 1:1 on every model.
• Browser binary: run pnpm exec playwright install chromium once before a live run.

Pre-release. The live API path is not yet covered by automated tests (it costs money and needs a target); the pure pieces — action parsing, key mapping, tool/model resolution, cost accounting — are unit-tested.