@idriszade/observe-vercel

Vercel AI SDK v7 usage shim for pipeline-kit. Maps the nested inputTokenDetails.* / outputTokenDetails.* fields from a v7 LanguageModelUsage object into a UsageAccumulator using the OTel GenAI v1.37 key constants from @idriszade/observe. Non-additivity is preserved by construction — cache and reasoning sub-fields are recorded under their own keys and never summed into the parent totals.

Installation

pnpm add @idriszade/observe-vercel

Quick start

import { UsageAccumulator } from '@idriszade/core';
import { recordVercelUsage } from '@idriszade/observe-vercel';

// Typical v7 usage object from generateText / streamText.
const usage = {
  inputTokens: 500,
  outputTokens: 120,
  inputTokenDetails: {
    cacheReadTokens: 350,
    cacheWriteTokens: 0,
  },
  outputTokenDetails: {
    reasoningTokens: 40,
    textTokens: 80,
  },
};

const acc = new UsageAccumulator();
recordVercelUsage(usage, acc);

acc.get('gen_ai.usage.input_tokens');              // 500
acc.get('gen_ai.usage.input_tokens.cache_read');   // 350
acc.get('gen_ai.usage.output_tokens');             // 120
acc.get('gen_ai.usage.output_tokens.reasoning');   // 40

Fields that are undefined are silently skipped — no zero records are emitted.

Non-additivity

Cache and reasoning sub-fields are descriptive — they describe how the parent total was composed. Do not add them to the parent yourself.

See @idriszade/observe — Cache tokens are non-additive for the full explanation, reference, and code example.

v7 vs v6 migration

Vercel AI SDK v7 changed how sub-token counts are surfaced:

• Top-level cachedInputTokens field removed.
• Top-level reasoningTokens field removed.
• Sub-counts now live exclusively under inputTokenDetails.cacheReadTokens, inputTokenDetails.cacheWriteTokens, outputTokenDetails.reasoningTokens, outputTokenDetails.textTokens.
• inputTokens / outputTokens at the top level remain the inclusive totals — unchanged from v6.

recordVercelUsage accepts the v7 shape. If you are on v6 and passing a legacy object with top-level cachedInputTokens, those fields are not read; migrate to v7 inputTokenDetails first.

API

recordVercelUsage(usage, accumulator)

import type { VercelUsageV7 } from '@idriszade/observe-vercel';

function recordVercelUsage(usage: VercelUsageV7, accumulator: UsageAccumulator): void;

VercelUsageV7 mirrors the Vercel AI SDK v7 LanguageModelUsage shape with all fields optional to tolerate partial payloads from streaming responses.

License — MIT