@rulvar/openai 1.27.0 → 1.29.0

package/dist/index.d.ts

@@ -221,9 +221,17 @@ declare function normalizeOpenAiUsage(raw: Record<string, unknown> | undefined):
 * output array, never the output_text aggregate. Raw output items ride
 * finish.providerMetadata.openai.outputItems so the runtime can retain
 * reasoning items as provider-raw parts.
+*
+* A stream that drains without any response terminal event
+* (`response.completed`, `response.incomplete`, `response.failed`, or
+* `error`) is a truncated wire read: the mapper fails closed with one
+* retryable transport error instead of ending silently, unless
+* `options.signal` shows the caller requested the abort (the documented
+* exception that ends a stream without a terminal event).
 */
 declare function mapResponsesStream(stream: AsyncIterable<ResponsesStreamEvent>, ids: OpenAiIdMap, options?: {
   effortDownmapped?: boolean;
+  signal?: AbortSignal;
 }): AsyncGenerator<ChatEvent, void>;
 /** Projects SDK/API errors into the retryable WireError vocabulary. */
 declare function openAiErrorToWire(error: unknown): WireError;
@@ -239,7 +247,17 @@ declare function buildChatCompletionsParams(req: ChatRequest, ids: OpenAiIdMap):
 * Delta-patched chunk assembly for the degraded path; yields each
 * canonical event as its chunk is consumed (same live-streaming contract
 * as mapResponsesStream).
+*
+* The chat dialect has no explicit terminal frame at this layer: the
+* only completion signal is a `finish_reason` on the last choice chunk.
+* A stream that drains without one is a truncated wire read, so the
+* mapper fails closed with one retryable transport error (after
+* forwarding any usage the provider did report, which was still paid
+* for) instead of synthesizing a `stop` finish, unless `options.signal`
+* shows the caller requested the abort.
 */
-declare function mapChatCompletionsStream(stream: AsyncIterable<Record<string, unknown>>, ids: OpenAiIdMap): AsyncGenerator<ChatEvent, void>;
+declare function mapChatCompletionsStream(stream: AsyncIterable<Record<string, unknown>>, ids: OpenAiIdMap, options?: {
+  signal?: AbortSignal;
+}): AsyncGenerator<ChatEvent, void>;
 //#endregion
 export { CONSERVATIVE_COMPATIBLE_CAPS, OPENAI_MODELS, OPENAI_PRICING, type OpenAiAdapterOptions, type OpenAiClientLike, type OpenAiCompatibleConfig, OpenAiIdMap, type OpenAiModelInfo, type OpenAiSdkOptions, type ResponsesStreamEvent, type V1190CacheAudit, auditV1190CacheJournal, buildChatCompletionsParams, buildResponsesParams, mapChatCompletionsStream, mapOpenAiEffort, mapResponsesStream, normalizeOpenAiUsage, openAiErrorToWire, openAiModelInfo, openai, openaiCompatible, undoV1190CacheDoubleCount };

package/dist/index.js

@@ -370,6 +370,13 @@ function normalizeOpenAiUsage(raw) {
 * output array, never the output_text aggregate. Raw output items ride
 * finish.providerMetadata.openai.outputItems so the runtime can retain
 * reasoning items as provider-raw parts.
+*
+* A stream that drains without any response terminal event
+* (`response.completed`, `response.incomplete`, `response.failed`, or
+* `error`) is a truncated wire read: the mapper fails closed with one
+* retryable transport error instead of ending silently, unless
+* `options.signal` shows the caller requested the abort (the documented
+* exception that ends a stream without a terminal event).
 */
 async function* mapResponsesStream(stream, ids, options) {
 	const callIdByItemId = /* @__PURE__ */ new Map();
@@ -487,6 +494,16 @@ async function* mapResponsesStream(stream, ids, options) {
 			return;
 		default: break;
 	}
+	if (options?.signal?.aborted === true) return;
+	yield {
+		type: "error",
+		error: {
+			code: "agent",
+			message: "Responses stream ended without a response terminal event (response.completed, response.incomplete, response.failed, or error); the read was truncated",
+			retryable: true,
+			data: { kind: "transport" }
+		}
+	};
 }
 /**
 * Classifies a terminal stream failure (`response.failed` /
@@ -518,6 +535,24 @@ function failedResponseError(code, message) {
 		}
 	};
 }
+/**
+* Largest delay a Node timer represents exactly; a bigger value would
+* overflow into an almost immediate (storm inducing) retry.
+*/
+const MAX_RETRY_AFTER_MS = 2147483647;
+/**
+* Parses a Retry-After header into milliseconds. Only the delta
+* seconds form is honored: the HTTP date form and any other
+* unparsable value return undefined so the engine falls back to its
+* computed policy backoff instead of receiving NaN, and a huge but
+* finite value is clamped to the Node timer maximum (v1.28.0 review
+* P2).
+*/
+function retryAfterMsFrom(header) {
+	const seconds = Number(header);
+	if (!Number.isFinite(seconds) || seconds < 0) return;
+	return Math.min(Math.round(seconds * 1e3), MAX_RETRY_AFTER_MS);
+}
 /** Projects SDK/API errors into the retryable WireError vocabulary. */
 function openAiErrorToWire(error) {
 	const record = error;
@@ -528,7 +563,7 @@ function openAiErrorToWire(error) {
 		const headers = record.headers;
 		if (headers !== void 0 && headers !== null) {
 			const value = typeof headers.get === "function" ? headers.get("retry-after") ?? void 0 : headers["retry-after"];
-			if (value !== void 0) retryAfterMs = Number(value) * 1e3;
+			if (value !== void 0) retryAfterMs = retryAfterMsFrom(value);
 		}
 		return {
 			code: "agent",
@@ -635,10 +670,19 @@ function buildChatCompletionsParams(req, ids) {
 * Delta-patched chunk assembly for the degraded path; yields each
 * canonical event as its chunk is consumed (same live-streaming contract
 * as mapResponsesStream).
+*
+* The chat dialect has no explicit terminal frame at this layer: the
+* only completion signal is a `finish_reason` on the last choice chunk.
+* A stream that drains without one is a truncated wire read, so the
+* mapper fails closed with one retryable transport error (after
+* forwarding any usage the provider did report, which was still paid
+* for) instead of synthesizing a `stop` finish, unless `options.signal`
+* shows the caller requested the abort.
 */
-async function* mapChatCompletionsStream(stream, ids) {
+async function* mapChatCompletionsStream(stream, ids, options) {
 	const pendingCalls = /* @__PURE__ */ new Map();
 	let finishReason;
+	let sawUsageChunk = false;
 	let usage = {
 		inputTokens: 0,
 		outputTokens: 0,
@@ -683,6 +727,7 @@ async function* mapChatCompletionsStream(stream, ids) {
 		if (typeof choice?.finish_reason === "string") finishReason = choice.finish_reason;
 		const chunkUsage = chunk.usage;
 		if (chunkUsage !== void 0 && chunkUsage !== null) {
+			sawUsageChunk = true;
 			const promptDetails = chunkUsage.prompt_tokens_details;
 			const promptTokens = typeof chunkUsage.prompt_tokens === "number" ? chunkUsage.prompt_tokens : 0;
 			const clamped = clampCacheSubsets(promptTokens, typeof promptDetails?.cached_tokens === "number" ? promptDetails.cached_tokens : 0, typeof promptDetails?.cache_write_tokens === "number" ? promptDetails.cache_write_tokens : 0);
@@ -694,6 +739,23 @@ async function* mapChatCompletionsStream(stream, ids) {
 			};
 		}
 	}
+	if (finishReason === void 0) {
+		if (options?.signal?.aborted === true) return;
+		if (sawUsageChunk) yield {
+			type: "usage",
+			usage
+		};
+		yield {
+			type: "error",
+			error: {
+				code: "agent",
+				message: "chat completions stream ended without a finish_reason; the read was truncated",
+				retryable: true,
+				data: { kind: "transport" }
+			}
+		};
+		return;
+	}
 	for (const [, pending] of pendingCalls) {
 		let args = {};
 		try {
@@ -768,14 +830,17 @@ function openai(options = {}) {
 					yield* mapResponsesStream(await client.responses.create({
 						...params,
 						stream: true
-					}, signal === void 0 ? void 0 : { signal }), ids, { effortDownmapped });
+					}, signal === void 0 ? void 0 : { signal }), ids, {
+						effortDownmapped,
+						...signal === void 0 ? {} : { signal }
+					});
 				} else {
 					const params = buildChatCompletionsParams(req, ids);
 					yield* mapChatCompletionsStream(await client.chat.completions.create({
 						...params,
 						stream: true,
 						stream_options: { include_usage: true }
-					}, signal === void 0 ? void 0 : { signal }), ids);
+					}, signal === void 0 ? void 0 : { signal }), ids, signal === void 0 ? void 0 : { signal });
 				}
 			} catch (thrown) {
 				if (signal?.aborted !== true) yield {
@@ -842,7 +907,7 @@ function openaiCompatible(cfg) {
 					...params,
 					stream: true,
 					stream_options: { include_usage: true }
-				}, signal === void 0 ? void 0 : { signal }), ids);
+				}, signal === void 0 ? void 0 : { signal }), ids, signal === void 0 ? void 0 : { signal });
 			} catch (thrown) {
 				if (signal?.aborted !== true) yield {
 					type: "error",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rulvar/openai",
-  "version": "1.27.0",
+  "version": "1.29.0",
   "description": "Rulvar first-class provider adapter for the OpenAI Responses API, plus the openaiCompatible factory.",
   "type": "module",
   "license": "Apache-2.0",
@@ -23,13 +23,13 @@
   },
   "dependencies": {
     "openai": "^6.45.0",
-    "@rulvar/core": "1.27.0"
+    "@rulvar/core": "1.29.0"
   },
   "devDependencies": {
     "@types/node": "^22.20.0",
     "tsdown": "^0.22.3",
     "typescript": "~6.0.3",
-    "@rulvar/testing": "1.27.0"
+    "@rulvar/testing": "1.29.0"
   },
   "repository": {
     "type": "git",