mirror of
https://github.com/diegosouzapw/OmniRoute.git
synced 2026-09-15 03:12:36 +03:00
7cfabfc5c9
Validated in a combined 3-PR batch worktree off release/v3.8.51 tip (a sibling PR from the same author, #11495, was held out — a typecheck error in zai-web.ts only reproduced with this PR + #11495 boarded together, and cleared without #11495; isolated this PR alone confirmed clean on its own too, so the interaction belonged to #11495's side — see its comment). - Golden lock: executor-map-golden.test.ts — passes byte-identical (same keys, classes, provider identities, dispatch guards) - Focused tests part of batch's 94/94 node:test run - typecheck:core, file-size, changelog-integrity, complexity, cognitive-complexity — all OK - Full-repo lint: 228 pre-existing dashboard react-hooks/* findings, unrelated to this diff Thanks for the measured, careful methodology here — the golden-lock contract plus the isolated DATA_DIR benchmarking make this an easy PR to trust despite the wide surface (72 files).
45 lines
1.8 KiB
TypeScript
45 lines
1.8 KiB
TypeScript
import { getExecutor } from "@omniroute/open-sse/executors/index";
|
|
import type { ExecuteInput, ProviderCredentials } from "@omniroute/open-sse/executors/base";
|
|
import type {
|
|
ChatTurn,
|
|
ModelCallResult,
|
|
ModelClient,
|
|
} from "@omniroute/open-sse/services/compression/eval/types";
|
|
import { calculateCost } from "@/lib/usage/costCalculator";
|
|
|
|
/**
|
|
* Cost-aware judge ModelClient for the compression playground's fidelity verify.
|
|
* Hard Rule #18 — NOT unit-tested (touches the real executor); the cost math is calculateCost
|
|
* (already covered) and the cap logic is judgeFidelityBatch (unit-tested with a stub). Computes
|
|
* usdCost from FULL usage (prompt + completion tokens) via the canonical pricing engine, so the
|
|
* USD cap actually engages and totalUsd is real.
|
|
*/
|
|
export function createPricedJudgeClient(
|
|
provider: string,
|
|
credentials: ProviderCredentials
|
|
): ModelClient {
|
|
return {
|
|
async complete(model: string, messages: ChatTurn[]): Promise<ModelCallResult> {
|
|
// #11220: getExecutor is async (lazy registry) — resolve per call.
|
|
const executor = await getExecutor(provider);
|
|
const input: ExecuteInput = {
|
|
model,
|
|
body: { model, messages, stream: false },
|
|
stream: false,
|
|
credentials,
|
|
};
|
|
const raw = (await executor.execute(input)) as { response: Response };
|
|
const json = (await raw.response.json()) as {
|
|
choices?: Array<{ message?: { content?: string } }>;
|
|
usage?: { prompt_tokens?: number; completion_tokens?: number };
|
|
};
|
|
const text = json.choices?.[0]?.message?.content ?? "";
|
|
const usdCost = await calculateCost(provider, model, {
|
|
prompt_tokens: json.usage?.prompt_tokens,
|
|
completion_tokens: json.usage?.completion_tokens,
|
|
});
|
|
return usdCost > 0 ? { text, usdCost } : { text };
|
|
},
|
|
};
|
|
}
|