eltonandClaude Fable 5.1
0d971b81dc
Weiterleitung an OpenRouters Messages-Endpunkt (x-api-key, anthropic-version, anthropic-beta), Effort als reasoning.effort, Usage in Anthropic-Form gelesen (auch im Stream aus message_start/message_delta). Nur openrouter-Ziele; lokale llama-server sprechen kein Messages-Format. Verifiziert: claude -p über http://llm.lan:4010 mit alt/anthropic/claude-sonnet-5 -> glm-5.3-flash. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
78 lines
3.2 KiB
TypeScript
78 lines
3.2 KiB
TypeScript
/**
|
|
* log.ts — ein Eintrag je Anfrage, zweimal: als JSONL (grep-bar, ueberlebt
|
|
* alles) und in SQLite (fuer die Tagesstatistik je Key). Die Kosten kommen
|
|
* von OpenRouter aus der Antwort (usage.cost), fuer lokale Upstreams bleibt
|
|
* das Feld null -- wir erfinden keine Preise.
|
|
*/
|
|
import { appendFileSync, mkdirSync } from "fs";
|
|
import { Database } from "bun:sqlite";
|
|
|
|
export type Eintrag = {
|
|
ts: string;
|
|
key: string;
|
|
alias: string;
|
|
angefragt: string;
|
|
model: string;
|
|
upstream: string;
|
|
effort: string | null;
|
|
pfad: string;
|
|
status: number;
|
|
ms: number;
|
|
stream: boolean;
|
|
prompt_tokens: number | null;
|
|
completion_tokens: number | null;
|
|
cached_tokens: number | null;
|
|
cost: number | null;
|
|
fehler?: string;
|
|
};
|
|
|
|
export class Log {
|
|
private db: Database;
|
|
private jsonl: string;
|
|
constructor(dir: string) {
|
|
mkdirSync(dir, { recursive: true });
|
|
this.jsonl = `${dir}/requests.jsonl`;
|
|
this.db = new Database(`${dir}/requests.sqlite`);
|
|
this.db.run(`CREATE TABLE IF NOT EXISTS requests (
|
|
ts TEXT, key TEXT, alias TEXT, angefragt TEXT, model TEXT, upstream TEXT, effort TEXT, pfad TEXT,
|
|
status INTEGER, ms INTEGER, stream INTEGER, prompt_tokens INTEGER, completion_tokens INTEGER,
|
|
cached_tokens INTEGER, cost REAL, fehler TEXT)`);
|
|
this.db.run(`CREATE INDEX IF NOT EXISTS requests_ts_key ON requests (ts, key)`);
|
|
}
|
|
schreiben(e: Eintrag): void {
|
|
appendFileSync(this.jsonl, JSON.stringify(e) + "\n");
|
|
this.db.run(
|
|
`INSERT INTO requests VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)`,
|
|
[e.ts, e.key, e.alias, e.angefragt, e.model, e.upstream, e.effort, e.pfad, e.status, e.ms, e.stream ? 1 : 0,
|
|
e.prompt_tokens, e.completion_tokens, e.cached_tokens, e.cost, e.fehler ?? null],
|
|
);
|
|
}
|
|
/** Je Tag und Key: Anfragen, Tokens, Kosten. */
|
|
statistik(tage: number): { tag: string; key: string; anfragen: number; prompt: number; completion: number; cost: number; modelle: string }[] {
|
|
return this.db.query(`
|
|
SELECT substr(ts,1,10) AS tag, key,
|
|
count(*) AS anfragen,
|
|
coalesce(sum(prompt_tokens),0) AS prompt,
|
|
coalesce(sum(completion_tokens),0) AS completion,
|
|
coalesce(sum(cost),0) AS cost,
|
|
group_concat(DISTINCT model) AS modelle
|
|
FROM requests WHERE ts >= datetime('now', ?)
|
|
GROUP BY tag, key ORDER BY tag DESC, cost DESC`).all(`-${tage} days`) as never;
|
|
}
|
|
}
|
|
|
|
/** Zieht Tokens und Kosten aus einer OpenAI-/OpenRouter-Usage. */
|
|
export function usageLesen(u: unknown): Pick<Eintrag, "prompt_tokens" | "completion_tokens" | "cached_tokens" | "cost"> {
|
|
// OpenAI-Form (prompt_tokens/completion_tokens) und Anthropic-Form (input_tokens/output_tokens)
|
|
const usage = (u ?? {}) as {
|
|
prompt_tokens?: number; completion_tokens?: number; cost?: number; prompt_tokens_details?: { cached_tokens?: number };
|
|
input_tokens?: number; output_tokens?: number; cache_read_input_tokens?: number;
|
|
};
|
|
return {
|
|
prompt_tokens: usage.prompt_tokens ?? usage.input_tokens ?? null,
|
|
completion_tokens: usage.completion_tokens ?? usage.output_tokens ?? null,
|
|
cached_tokens: usage.prompt_tokens_details?.cached_tokens ?? usage.cache_read_input_tokens ?? null,
|
|
cost: typeof usage.cost === "number" ? usage.cost : null,
|
|
};
|
|
}
|