feat(ai): use zstd compression for codex sse transport

packages/ai/src/api/openai-codex-responses.ts

@@ -1,4 +1,5 @@
 import type * as NodeOs from "node:os";
+import type * as NodeZlib from "node:zlib";
 import type {
 	Tool as OpenAITool,
 	ResponseCreateParamsStreaming,
@@ -58,6 +59,9 @@ const DEFAULT_MAX_RETRIES = 0;
 const BASE_DELAY_MS = 1000;
 const DEFAULT_MAX_RETRY_DELAY_MS = 60_000;
 const DEFAULT_WEBSOCKET_CONNECT_TIMEOUT_MS = 15_000;
+// The Codex backend accepts zstd-compressed request bodies on the SSE responses
+// endpoint (the same endpoint the official Codex client compresses against).
+const REQUEST_COMPRESSION_ZSTD_LEVEL = 3;
 const CODEX_TOOL_CALL_PROVIDERS = new Set(["openai", "openai-codex", "opencode"]);
 const WEBSOCKET_MESSAGE_TOO_BIG_CLOSE_CODE = 1009;
 const WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE = "websocket_connection_limit_reached";
@@ -177,6 +181,39 @@ function normalizeTimeoutMs(value: number | undefined): number | undefined {
 	return Math.floor(value);
 }
 
+// ============================================================================
+// Request Compression
+// ============================================================================
+
+type ProcessWithBuiltinModule = typeof process & {
+	getBuiltinModule?: (id: "node:zlib") => typeof NodeZlib;
+};
+
+function loadNodeZlib(): typeof NodeZlib | null {
+	if (typeof process === "undefined" || !(process.versions?.node || process.versions?.bun)) {
+		return null;
+	}
+	return (process as ProcessWithBuiltinModule).getBuiltinModule?.("node:zlib") ?? null;
+}
+
+// Returns the zstd-compressed body bytes, or null when compression is
+// unavailable (browser/Vite builds). Callers fall back to sending the
+// uncompressed JSON when this returns null.
+function compressRequestBodyZstd(bodyJson: string): Uint8Array | null {
+	const zlib = loadNodeZlib();
+	if (!zlib || typeof zlib.zstdCompressSync !== "function") {
+		return null;
+	}
+	try {
+		const compressed = zlib.zstdCompressSync(bodyJson, {
+			params: { [zlib.constants.ZSTD_c_compressionLevel]: REQUEST_COMPRESSION_ZSTD_LEVEL },
+		});
+		return new Uint8Array(compressed.buffer, compressed.byteOffset, compressed.byteLength);
+	} catch {
+		return null;
+	}
+}
+
 // ============================================================================
 // Main Stream Function
 // ============================================================================
@@ -298,6 +335,15 @@ export const stream: StreamFunction<"openai-codex-responses", OpenAICodexRespons
 				}
 			}
 
+			// Compress the request body once for the SSE path. The Codex backend
+			// decodes Content-Encoding: zstd; the WebSocket transport above sends the
+			// uncompressed JSON frame, matching the official Codex client.
+			const compressedBody = compressRequestBodyZstd(bodyJson);
+			if (compressedBody) {
+				sseHeaders.set("content-encoding", "zstd");
+			}
+			const sseBody: Uint8Array | string = compressedBody ?? bodyJson;
+
 			// Fetch with retry logic for rate limits and transient errors
 			let response: Response | undefined;
 			let lastError: Error | undefined;
@@ -316,7 +362,7 @@ export const stream: StreamFunction<"openai-codex-responses", OpenAICodexRespons
 						response = await fetch(resolveCodexUrl(model.baseUrl), {
 							method: "POST",
 							headers: sseHeaders,
-							body: bodyJson,
+							body: sseBody,
 							signal: combinedSignal.signal,
 						});
 					} catch (error) {