openRouterDecider drops usage.cost, and decide() meta cannot surface the Decisions response id/provider

[daveycodez]

Summary

openRouterDecider / createOpenRouterDecider (@tanstack/ai-openrouter, used with decide() from @tanstack/ai) drops data the OpenRouter Decisions API returns:

1. usage.cost is never attached to result.meta.usage, even though the endpoint returns it on every response and the same package already has extractUsageCost() which the text/image adapters use.
2. The response's id (OpenRouter generation id) and provider are not surfaced anywhere, because EvaluateAdapterResult / EvaluateResultMeta only carry model, answers and usage.

I'm billing per request off usage.cost and reconciling with GET /api/v1/generation?id=, so both are needed.

Environment

• @tanstack/ai@0.58.0
• @tanstack/ai-openrouter@0.19.16
• @tanstack/ai-event-client@0.12.1
• Node v24.20.0, server-side (no framework)

Reproduction

import { decide, boolean } from '@tanstack/ai'
import { openRouterDecider } from '@tanstack/ai-openrouter'

const result = await decide({
  adapter: openRouterDecider('typesafe/jev-1.13'),
  state: 'Customer says the invoice total is wrong after applying a coupon.',
  questions: {
    isBug: boolean({ instructions: 'Is this a software bug report?' }),
  },
})

console.log(result.meta.usage)
// { promptTokens: 30, completionTokens: 3, totalTokens: 33 }
// no `cost`, although the raw response body had usage.cost

console.log(result.meta)
// { model: 'typesafe/jev-1.13-20260917', usage: { ... } }
// no `id`, no `provider`

The raw POST https://openrouter.ai/api/alpha/decisions response (from OpenRouter's API reference) is:

{
  "answers": { "...": "..." },
  "id": "gen-dec-1789738314-X5e5eKGQdvR9rblyX250",
  "model": "typesafe/jev-1.13-20260917",
  "provider": "TypeSafe",
  "usage": { "cost": 0.000019992, "input_tokens": 476, "output_tokens": 70 }
}

Expected

• result.meta.usage.cost is 0.000019992 (and costDetails when present), the same as chat() / generateImage() with the OpenRouter adapters. TokenUsage already declares cost?: number and costDetails?: UsageCostBreakdown (@tanstack/ai-event-client index.d.ts, TokenUsage).
• result.meta exposes the response's id and provider (or a providerMetadata bag) so the request can be reconciled against GET /api/v1/generation?id=.

Actual

• result.meta.usage has only promptTokens / completionTokens / totalTokens; cost is undefined.
• result.meta has only model and usage; id and provider are unreachable from decide() without writing a custom adapter.

Where it happens

@tanstack/ai-openrouter/dist/esm/adapters/evaluate.js, mapUsage() (lines 51–65) reads only the token keys and calls buildBaseUsage with them:

function mapUsage(usage) {
	if (!isRecord(usage)) return buildBaseUsage({ promptTokens: 0, completionTokens: 0, totalTokens: 0 });
	const promptTokens = readNumber(usage, ["input_tokens", "prompt_tokens"]) ?? 0;
	const completionTokens = readNumber(usage, ["output_tokens", "completion_tokens"]) ?? 0;
	const totalTokens = readNumber(usage, ["total_tokens"]) ?? promptTokens + completionTokens;
	return buildBaseUsage({ promptTokens, completionTokens, totalTokens });
}

and evaluate() (lines 104–108) returns only model / answers / usage:

return {
	model: json.model ?? model,
	answers: json.answers,
	usage: mapUsage(json.usage)
};

extractUsageCost() lives in the same package (adapters/cost.js, exported with cost.d.ts) and is already spread into usage in adapters/text.js (lines 155, 386, 735), adapters/responses-text.js (133, 439, 1083) and adapters/image.js (121). evaluate.js does not import it.

On the core side, @tanstack/ai/dist/esm/activities/evaluate/adapter.d.ts lines 103–108:

export interface EvaluateAdapterResult {
    /** Resolved model id from the provider. */
    model: string;
    answers: Record<string, WireAnswer>;
    usage: TokenUsage;
}

and activities/evaluate/index.d.ts lines 50–54:

export interface EvaluateResultMeta {
    /** Resolved model id from the provider. */
    model: string;
    usage: TokenUsage;
}

so even a custom adapter has no typed place to put id / provider.

Suggested fix

1. @tanstack/ai-openrouter adapters/evaluate.ts: spread the existing helper into the mapped usage, matching the other adapters:

   import { extractUsageCost } from './cost'
   
   return {
     ...buildBaseUsage({ promptTokens, completionTokens, totalTokens }),
     ...extractUsageCost(usage),
   }

2. @tanstack/ai activities/evaluate: add optional id?: string and provider?: string (or providerMetadata?: Record<string, unknown>) to EvaluateAdapterResult and pass them through to EvaluateResultMeta in decide(). The OpenRouter adapter would then return id: json.id, provider: json.provider.

Both are additive; existing adapters and callers keep working.

I'm happy to open a PR for either or both if that's welcome.