Function proxyRequest

src/proxy.ts:3498–6231

* Proxy a single request through x402 payment flow to BlockRun API. * * Optimizations applied in order: * 1. Dedup check — if same request body seen within 30s, replay cached response * 2. Streaming heartbeat — for stream:true, send 200 + heartbeats immediately * 3. Smart routing — when m

(
  req: IncomingMessage,
  res: ServerResponse,
  apiBase: string,
  payFetch: (input: RequestInfo | URL, init?: RequestInit) => Promise<Response>,
  options: ProxyOptions,
  routerOpts: RouterOptions,
  deduplicator: RequestDeduplicator,
  balanceMonitor: AnyBalanceMonitor,
  sessionStore: SessionStore,
  responseCache: ResponseCache,
  sessionJournal: SessionJournal,
)

 *   4. Fallback chain — on provider errors, try next model in tier's fallback list
 */
async function proxyRequest(
  req: IncomingMessage,
  res: ServerResponse,
  apiBase: string,
  payFetch: (input: RequestInfo | URL, init?: RequestInit) => Promise<Response>,
  options: ProxyOptions,
  routerOpts: RouterOptions,
  deduplicator: RequestDeduplicator,
  balanceMonitor: AnyBalanceMonitor,
  sessionStore: SessionStore,
  responseCache: ResponseCache,
  sessionJournal: SessionJournal,
): Promise<void> {
  const startTime = Date.now();

  // Build upstream URL: /v1/chat/completions → https://blockrun.ai/api/v1/chat/completions
  const upstreamUrl = `${apiBase}${req.url}`;

  // Collect request body
  const bodyChunks: Buffer[] = [];
  for await (const chunk of req) {
    bodyChunks.push(Buffer.isBuffer(chunk) ? chunk : Buffer.from(chunk));
  }
  let body = Buffer.concat(bodyChunks);

  // Track original context size for response headers
  const originalContextSizeKB = Math.ceil(body.length / 1024);

  // Routing debug info is on by default; disable per-request with
  // x-clawrouter-debug: false, or globally with CLAWROUTER_DEBUG_HEADERS=off
  const debugMode = debugHeadersEnabledFromEnv() && req.headers["x-clawrouter-debug"] !== "false";

  // --- Smart routing ---
  let routingDecision: RoutingDecision | undefined;
  let hasTools = false; // true when request includes a tools schema
  let hasVision = false; // true when request includes image_url content parts
  let isStreaming = false;
  let modelId = "";
  let maxTokens = 4096;
  let routingProfile: "eco" | "auto" | "premium" | null = null;
  let stickyExplicitModel: string | undefined;
  let balanceFallbackNotice: string | undefined;
  let budgetDowngradeNotice: string | undefined;
  let budgetDowngradeHeaderMode: "downgraded" | undefined;
  let accumulatedContent = ""; // For session journal event extraction
  let responseInputTokens: number | undefined;
  let responseOutputTokens: number | undefined;
  let requestHadError = false; // Set to true when all models fail → used in logUsage
  let requestSummaryForStore = ""; // Truncated user prompt — captured for response-store entry
  const isChatCompletion = req.url?.includes("/chat/completions");

  // Extract session ID early for journal operations (header-only at this point)
  const sessionId = getSessionId(req.headers as Record<string, string | string[] | undefined>);
  // Full session ID (header + content-derived) — populated once messages are parsed
  let effectiveSessionId: string | undefined = sessionId;

  if (isChatCompletion && body.length > 0) {
    try {