@neon/ai-sdk-provider 0.7.2 → 0.8.0

package/dist/lib/neon-chat-options.js.map

@@ -1 +1 @@
-{"version":3,"file":"neon-chat-options.js","names":[],"sources":["../../src/lib/neon-chat-options.ts"],"sourcesContent":["// Models served by the Neon AI Gateway. Anthropic and OpenAI models are routed\n// to their provider-native endpoints; everything else uses the unified MLflow\n// (OpenAI-compatible) endpoint.\n//\n// Ids use the canonical Neon (unprefixed) form, matching the `neon` provider on\n// models.dev. The gateway also accepts the legacy `databricks-` prefixed form\n// (the `databricks` provider on models.dev), which still resolves via the\n// `(string & {})` fallback on `NeonChatModelId`.\n//\n// `NEON_MODELS_DEV_IDS` mirrors the models.dev `neon` catalog exactly — the\n// `neon-catalog-drift.test.ts` maintainer check fails if the two diverge.\n// `NEON_EXTRA_MODEL_IDS` are additional ids the gateway serves that models.dev\n// does not (yet) list; they are kept for autocomplete and verified to resolve\n// against the gateway. The authoritative, always-current catalog is shown in\n// the Neon Console under the branch's \"AI Gateway\" tab.\n\n/** The models.dev `neon` catalog (canonical, unprefixed ids). */\nexport const NEON_MODELS_DEV_IDS = [\n\t// Anthropic (native Messages API)\n\t\"claude-opus-4-7\",\n\t\"claude-opus-4-6\",\n\t\"claude-opus-4-5\",\n\t\"claude-opus-4-1\",\n\t\"claude-sonnet-4-6\",\n\t\"claude-sonnet-4-5\",\n\t\"claude-sonnet-4\",\n\t\"claude-haiku-4-5\",\n\t// OpenAI (native Responses API)\n\t\"gpt-5\",\n\t\"gpt-5-mini\",\n\t\"gpt-5-nano\",\n\t\"gpt-5-1\",\n\t\"gpt-5-2\",\n\t\"gpt-5-4\",\n\t\"gpt-5-4-mini\",\n\t\"gpt-5-4-nano\",\n\t\"gpt-5-5\",\n\t// OpenAI open-weight (unified MLflow endpoint)\n\t\"gpt-oss-120b\",\n\t\"gpt-oss-20b\",\n\t// Google (unified MLflow endpoint)\n\t\"gemini-3-pro\",\n\t\"gemini-3-flash\",\n\t\"gemini-3-1-pro\",\n\t\"gemini-3-1-flash-lite\",\n\t\"gemini-2-5-pro\",\n\t\"gemini-2-5-flash\",\n] as const;\n\n/**\n * Ids the gateway serves that the models.dev `neon` provider does not list.\n * Verified to resolve against the gateway; retained for autocomplete. If\n * models.dev later lists one of these, the drift check flags it for promotion\n * into `NEON_MODELS_DEV_IDS`.\n */\nexport const NEON_EXTRA_MODEL_IDS = [\n\t// Anthropic (native Messages API)\n\t\"claude-opus-4-8\",\n\t// OpenAI (native Responses API) — Codex is served only natively\n\t\"gpt-5-2-codex\",\n\t\"gpt-5-3-codex\",\n\t\"gpt-5-5-pro\",\n\t// Google (unified MLflow endpoint)\n\t\"gemini-3-5-flash\",\n\t\"gemma-3-12b\",\n\t// Meta (unified MLflow endpoint)\n\t\"llama-4-maverick\",\n\t\"meta-llama-3-3-70b-instruct\",\n\t\"meta-llama-3-1-8b-instruct\",\n\t// Alibaba (unified MLflow endpoint)\n\t\"qwen3-next-80b-a3b-instruct\",\n\t\"qwen35-122b-a10b\",\n] as const;\n\n/** A known Neon AI Gateway model id (models.dev catalog + gateway extras). */\nexport type NeonKnownModelId =\n\t| (typeof NEON_MODELS_DEV_IDS)[number]\n\t| (typeof NEON_EXTRA_MODEL_IDS)[number];\n\n/**\n * A Neon AI Gateway model id. Known ids are listed for autocomplete; any other\n * id (e.g. a brand-new model, or the legacy `databricks-` prefixed form) is\n * accepted via the `(string & {})` fallback.\n */\nexport type NeonChatModelId = NeonKnownModelId | (string & {});\n"],"mappings":";;AAiBA,MAAa,sBAAsB;CAElC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CAEA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CAEA;CACA;CAEA;CACA;CACA;CACA;CACA;CACA;AACD;;;;;;;AAQA,MAAa,uBAAuB;CAEnC;CAEA;CACA;CACA;CAEA;CACA;CAEA;CACA;CACA;CAEA;CACA;AACD"}
+{"version":3,"file":"neon-chat-options.js","names":[],"sources":["../../src/lib/neon-chat-options.ts"],"sourcesContent":["// Models served by the Neon AI Gateway. Anthropic and OpenAI models are routed\n// to their provider-native endpoints; everything else uses the unified MLflow\n// (OpenAI-compatible) endpoint.\n//\n// Ids use the canonical Neon (unprefixed) form, matching the `neon` provider on\n// models.dev. The gateway also accepts the legacy `databricks-` prefixed form\n// (the `databricks` provider on models.dev), which still resolves via the\n// `(string & {})` fallback on `NeonChatModelId`.\n//\n// `NEON_MODELS_DEV_IDS` mirrors the models.dev `neon` catalog exactly — the\n// `neon-catalog-drift.test.ts` maintainer check fails if the two diverge. The\n// authoritative, always-current catalog is shown in the Neon Console under the\n// branch's \"AI Gateway\" tab.\n//\n// There is deliberately no second list for ids the gateway serves ahead of\n// models.dev. Such an id already works — `NeonChatModelId` accepts any string —\n// so a hand-maintained list would buy autocomplete for a few days at the cost of\n// a surface that silently rots, which is what happened last time. The fix for a\n// gateway model missing here is to add it to models.dev; Orbit's catalog-parity\n// job reports that gap daily.\n\n/** The models.dev `neon` catalog (canonical, unprefixed ids). */\nexport const NEON_MODELS_DEV_IDS = [\n\t// Anthropic (native Messages API)\n\t\"claude-fable-5\",\n\t\"claude-haiku-4-5\",\n\t\"claude-opus-4-1\",\n\t\"claude-opus-4-5\",\n\t\"claude-opus-4-6\",\n\t\"claude-opus-4-7\",\n\t\"claude-opus-4-8\",\n\t\"claude-opus-5\",\n\t\"claude-sonnet-4-5\",\n\t\"claude-sonnet-4-6\",\n\t\"claude-sonnet-5\",\n\t// OpenAI (native Responses API)\n\t\"gpt-5\",\n\t\"gpt-5-1\",\n\t\"gpt-5-2\",\n\t\"gpt-5-3-codex\",\n\t\"gpt-5-4\",\n\t\"gpt-5-4-mini\",\n\t\"gpt-5-4-nano\",\n\t\"gpt-5-5\",\n\t\"gpt-5-5-pro\",\n\t\"gpt-5-6-luna\",\n\t\"gpt-5-6-sol\",\n\t\"gpt-5-6-terra\",\n\t\"gpt-5-mini\",\n\t\"gpt-5-nano\",\n\t// OpenAI open-weight (unified MLflow endpoint)\n\t\"gpt-oss-120b\",\n\t\"gpt-oss-20b\",\n\t// Google (unified MLflow endpoint)\n\t\"gemini-3-1-flash-lite\",\n\t\"gemini-3-1-pro\",\n\t\"gemini-3-5-flash\",\n\t\"gemini-3-flash\",\n\t\"gemma-3-12b\",\n\t// Meta (unified MLflow endpoint)\n\t\"llama-4-maverick\",\n\t\"meta-llama-3-1-8b-instruct\",\n\t\"meta-llama-3-3-70b-instruct\",\n\t// Alibaba (unified MLflow endpoint)\n\t\"qwen3-next-80b-a3b-instruct\",\n\t\"qwen35-122b-a10b\",\n\t// Zhipu (unified MLflow endpoint)\n\t\"glm-5-2\",\n\t// Thinking Machines (unified MLflow endpoint)\n\t\"inkling\",\n] as const;\n\n/** A known Neon AI Gateway model id. */\nexport type NeonKnownModelId = (typeof NEON_MODELS_DEV_IDS)[number];\n\n/**\n * A Neon AI Gateway model id. Known ids are listed for autocomplete; any other\n * id (e.g. a brand-new model, or the legacy `databricks-` prefixed form) is\n * accepted via the `(string & {})` fallback.\n */\nexport type NeonChatModelId = NeonKnownModelId | (string & {});\n"],"mappings":";;AAsBA,MAAa,sBAAsB;CAElC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CAEA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CAEA;CACA;CAEA;CACA;CACA;CACA;CACA;CAEA;CACA;CACA;CAEA;CACA;CAEA;CAEA;AACD"}

package/dist/lib/neon-gateway-error.d.ts

@@ -0,0 +1,56 @@
+import { FetchFunction } from "@ai-sdk/provider-utils";
+
+//#region src/lib/neon-gateway-error.d.ts
+
+/**
+ * Give every route of this provider the error envelope its model knows how to
+ * read, so a failed call surfaces the gateway's reason on `error.message`
+ * instead of a bare HTTP status line.
+ *
+ * The gateway emits several error shapes, and which one you get depends on
+ * which layer rejected the request rather than on which route you called:
+ *
+ * 1. The gateway's own rejection, in OpenAI shape:
+ *      { "error": { "message": "unknown model \"nope\"" } }
+ *
+ * 2. A Databricks rejection, flat and with its own code:
+ *      { "error_code": "BAD_REQUEST", "message": "service_tier='flex' is …" }
+ *
+ * 3. A Databricks rejection wrapping an upstream error as a JSON *string*:
+ *      { "error_code": "BAD_REQUEST", "message": "{\"error\":{\"message\":…}}" }
+ *
+ * Meanwhile each underlying model parses a different schema: the OpenAI and
+ * OpenAI-compatible models want `{ error: { message } }`, the Anthropic model
+ * wants `{ type: "error", error: { type, message } }`. Anything that does not
+ * match is dropped and the AI SDK falls back to the status text, which is how
+ * `AI_APICallError: Bad Request` reaches a caller with the real explanation
+ * stranded on `responseBody`.
+ *
+ * So the shape a body needs is a property of the route, not of the gateway:
+ * this reads any of the above into one reason and re-emits it in the dialect
+ * that route's model expects. Cross-dialect conversion is covered too, since
+ * nothing guarantees a given layer answers in the dialect of the route it was
+ * reached through. A body already valid for the target dialect is left alone,
+ * as is any successful response and anything unrecognised.
+ *
+ * The Anthropic direction is unverified against a live gateway: that endpoint
+ * currently answers every request with a plain-text 404, so no JSON envelope
+ * can be observed on it at all.
+ *
+ * None of the models expose an error hook (`OpenAIConfig` has none, and the
+ * Anthropic one takes a fixed handler), so this rides on `fetch` — the same
+ * lever `wrapFetchWithHarmonyNormalization` uses. The two are disjoint: that
+ * one returns early on a failed response, this one only acts on failures.
+ */
+/** The error envelope a route's model can parse. */
+type GatewayErrorDialect = "openai" | "anthropic";
+/**
+ * Map a gateway error body into `dialect`, or return null when it already
+ * parses there (or is not a shape we recognise).
+ */
+declare function normalizeGatewayErrorBody(body: unknown, dialect: GatewayErrorDialect): unknown | null;
+/** Wrap a fetch so failed responses reach the model in its own dialect. */
+declare function wrapFetchWithGatewayErrorNormalization(baseFetch: FetchFunction | undefined, dialect: GatewayErrorDialect): FetchFunction;
+//#endregion
+export { GatewayErrorDialect, normalizeGatewayErrorBody, wrapFetchWithGatewayErrorNormalization };
+//# sourceMappingURL=neon-gateway-error.d.ts.map

package/dist/lib/neon-gateway-error.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"neon-gateway-error.d.ts","names":[],"sources":["../../src/lib/neon-gateway-error.ts"],"mappings":";;;;;;AA4CA;AAqHA;AAiBA;AAAsD;AAC1C;AACF;AACP;AAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAzIJ,mBAAA;;;;;iBAqHI,yBAAA,yBAEN;;iBAeM,sCAAA,YACJ,oCACF,sBACP"}

package/dist/lib/neon-gateway-error.js

@@ -0,0 +1,102 @@
+//#region src/lib/neon-gateway-error.ts
+function isRecord(value) {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+function asOpenAIEnvelope(value) {
+	if (isRecord(value) && isRecord(value.error) && typeof value.error.message === "string") return value.error;
+	return null;
+}
+function asAnthropicEnvelope(value) {
+	if (isRecord(value) && value.type === "error" && isRecord(value.error) && typeof value.error.message === "string" && typeof value.error.type === "string") return value.error;
+	return null;
+}
+/** Read any envelope the gateway emits into a single reason. */
+function extractReason(body) {
+	const anthropic = asAnthropicEnvelope(body);
+	if (anthropic) return {
+		message: anthropic.message,
+		type: anthropic.type
+	};
+	const openai = asOpenAIEnvelope(body);
+	if (openai) return {
+		message: openai.message,
+		openaiFields: openai,
+		type: typeof openai.type === "string" ? openai.type : void 0
+	};
+	if (!isRecord(body) || typeof body.message !== "string") return null;
+	const code = typeof body.error_code === "string" ? body.error_code : void 0;
+	if (body.message.trimStart().startsWith("{")) try {
+		const nested = extractReason(JSON.parse(body.message));
+		if (nested) return {
+			...nested,
+			type: nested.type ?? code
+		};
+	} catch {}
+	return {
+		message: body.message,
+		type: code
+	};
+}
+/**
+* Overlay the target envelope on the original body rather than replacing it,
+* so whatever else the gateway sent — `error_code`, request ids, the encoded
+* original message — still reaches the caller on `APICallError.responseBody`.
+*/
+function emit(body, reason, dialect) {
+	const original = isRecord(body) ? body : {};
+	if (dialect === "anthropic") return {
+		...original,
+		type: "error",
+		error: {
+			...reason.openaiFields,
+			type: reason.type ?? "api_error",
+			message: reason.message
+		}
+	};
+	return {
+		...original,
+		error: {
+			...reason.openaiFields,
+			message: reason.message,
+			code: reason.openaiFields?.code ?? reason.type ?? null
+		}
+	};
+}
+/**
+* Map a gateway error body into `dialect`, or return null when it already
+* parses there (or is not a shape we recognise).
+*/
+function normalizeGatewayErrorBody(body, dialect) {
+	if (dialect === "anthropic" ? asAnthropicEnvelope(body) !== null : asOpenAIEnvelope(body) !== null) return null;
+	const reason = extractReason(body);
+	return reason === null ? null : emit(body, reason, dialect);
+}
+/** Wrap a fetch so failed responses reach the model in its own dialect. */
+function wrapFetchWithGatewayErrorNormalization(baseFetch, dialect) {
+	const inner = baseFetch ?? ((...args) => globalThis.fetch(...args));
+	return async (input, init) => {
+		const response = await inner(input, init);
+		if (response.ok) return response;
+		if (!(response.headers.get("content-type") ?? "").includes("application/json")) return response;
+		let parsed;
+		try {
+			parsed = JSON.parse(await response.clone().text());
+		} catch {
+			return response;
+		}
+		const normalized = normalizeGatewayErrorBody(parsed, dialect);
+		if (normalized === null) return response;
+		const headers = new Headers(response.headers);
+		headers.delete("content-length");
+		headers.delete("content-encoding");
+		return new Response(JSON.stringify(normalized), {
+			status: response.status,
+			statusText: response.statusText,
+			headers
+		});
+	};
+}
+//#endregion
+export { normalizeGatewayErrorBody, wrapFetchWithGatewayErrorNormalization };
+
+//# sourceMappingURL=neon-gateway-error.js.map

package/dist/lib/neon-gateway-error.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"neon-gateway-error.js","names":[],"sources":["../../src/lib/neon-gateway-error.ts"],"sourcesContent":["import type { FetchFunction } from \"@ai-sdk/provider-utils\";\n\n/**\n * Give every route of this provider the error envelope its model knows how to\n * read, so a failed call surfaces the gateway's reason on `error.message`\n * instead of a bare HTTP status line.\n *\n * The gateway emits several error shapes, and which one you get depends on\n * which layer rejected the request rather than on which route you called:\n *\n * 1. The gateway's own rejection, in OpenAI shape:\n *      { \"error\": { \"message\": \"unknown model \\\"nope\\\"\" } }\n *\n * 2. A Databricks rejection, flat and with its own code:\n *      { \"error_code\": \"BAD_REQUEST\", \"message\": \"service_tier='flex' is …\" }\n *\n * 3. A Databricks rejection wrapping an upstream error as a JSON *string*:\n *      { \"error_code\": \"BAD_REQUEST\", \"message\": \"{\\\"error\\\":{\\\"message\\\":…}}\" }\n *\n * Meanwhile each underlying model parses a different schema: the OpenAI and\n * OpenAI-compatible models want `{ error: { message } }`, the Anthropic model\n * wants `{ type: \"error\", error: { type, message } }`. Anything that does not\n * match is dropped and the AI SDK falls back to the status text, which is how\n * `AI_APICallError: Bad Request` reaches a caller with the real explanation\n * stranded on `responseBody`.\n *\n * So the shape a body needs is a property of the route, not of the gateway:\n * this reads any of the above into one reason and re-emits it in the dialect\n * that route's model expects. Cross-dialect conversion is covered too, since\n * nothing guarantees a given layer answers in the dialect of the route it was\n * reached through. A body already valid for the target dialect is left alone,\n * as is any successful response and anything unrecognised.\n *\n * The Anthropic direction is unverified against a live gateway: that endpoint\n * currently answers every request with a plain-text 404, so no JSON envelope\n * can be observed on it at all.\n *\n * None of the models expose an error hook (`OpenAIConfig` has none, and the\n * Anthropic one takes a fixed handler), so this rides on `fetch` — the same\n * lever `wrapFetchWithHarmonyNormalization` uses. The two are disjoint: that\n * one returns early on a failed response, this one only acts on failures.\n */\n\n/** The error envelope a route's model can parse. */\nexport type GatewayErrorDialect = \"openai\" | \"anthropic\";\n\ninterface Reason {\n\tmessage: string;\n\t/** Preserved verbatim when the source was already an OpenAI envelope. */\n\topenaiFields?: Record<string, unknown>;\n\t/** An Anthropic `error.type`, or a Databricks `error_code`. */\n\ttype?: string;\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n\treturn typeof value === \"object\" && value !== null && !Array.isArray(value);\n}\n\nfunction asOpenAIEnvelope(value: unknown): Record<string, unknown> | null {\n\tif (\n\t\tisRecord(value) &&\n\t\tisRecord(value.error) &&\n\t\ttypeof value.error.message === \"string\"\n\t) {\n\t\treturn value.error;\n\t}\n\treturn null;\n}\n\nfunction asAnthropicEnvelope(value: unknown): Record<string, unknown> | null {\n\tif (\n\t\tisRecord(value) &&\n\t\tvalue.type === \"error\" &&\n\t\tisRecord(value.error) &&\n\t\ttypeof value.error.message === \"string\" &&\n\t\ttypeof value.error.type === \"string\"\n\t) {\n\t\treturn value.error;\n\t}\n\treturn null;\n}\n\n/** Read any envelope the gateway emits into a single reason. */\nfunction extractReason(body: unknown): Reason | null {\n\tconst anthropic = asAnthropicEnvelope(body);\n\tif (anthropic) {\n\t\treturn {\n\t\t\tmessage: anthropic.message as string,\n\t\t\ttype: anthropic.type as string,\n\t\t};\n\t}\n\n\tconst openai = asOpenAIEnvelope(body);\n\tif (openai) {\n\t\treturn {\n\t\t\tmessage: openai.message as string,\n\t\t\topenaiFields: openai,\n\t\t\ttype: typeof openai.type === \"string\" ? openai.type : undefined,\n\t\t};\n\t}\n\n\tif (!isRecord(body) || typeof body.message !== \"string\") {\n\t\treturn null;\n\t}\n\tconst code =\n\t\ttypeof body.error_code === \"string\" ? body.error_code : undefined;\n\n\t// Shape 3: an upstream envelope encoded into `message`. Prefer it — it\n\t// names the offending parameter, which the outer wrapper does not.\n\tif (body.message.trimStart().startsWith(\"{\")) {\n\t\ttry {\n\t\t\tconst inner: unknown = JSON.parse(body.message);\n\t\t\tconst nested = extractReason(inner);\n\t\t\tif (nested) {\n\t\t\t\treturn { ...nested, type: nested.type ?? code };\n\t\t\t}\n\t\t} catch {\n\t\t\t// Not JSON after all; fall through to shape 2.\n\t\t}\n\t}\n\n\t// Shape 2: a flat Databricks rejection.\n\treturn { message: body.message, type: code };\n}\n\n/**\n * Overlay the target envelope on the original body rather than replacing it,\n * so whatever else the gateway sent — `error_code`, request ids, the encoded\n * original message — still reaches the caller on `APICallError.responseBody`.\n */\nfunction emit(\n\tbody: unknown,\n\treason: Reason,\n\tdialect: GatewayErrorDialect,\n): unknown {\n\tconst original = isRecord(body) ? body : {};\n\tif (dialect === \"anthropic\") {\n\t\treturn {\n\t\t\t...original,\n\t\t\ttype: \"error\",\n\t\t\terror: {\n\t\t\t\t...reason.openaiFields,\n\t\t\t\ttype: reason.type ?? \"api_error\",\n\t\t\t\tmessage: reason.message,\n\t\t\t},\n\t\t};\n\t}\n\treturn {\n\t\t...original,\n\t\terror: {\n\t\t\t...reason.openaiFields,\n\t\t\tmessage: reason.message,\n\t\t\tcode: reason.openaiFields?.code ?? reason.type ?? null,\n\t\t},\n\t};\n}\n\n/**\n * Map a gateway error body into `dialect`, or return null when it already\n * parses there (or is not a shape we recognise).\n */\nexport function normalizeGatewayErrorBody(\n\tbody: unknown,\n\tdialect: GatewayErrorDialect,\n): unknown | null {\n\tconst alreadyValid =\n\t\tdialect === \"anthropic\"\n\t\t\t? asAnthropicEnvelope(body) !== null\n\t\t\t: asOpenAIEnvelope(body) !== null;\n\tif (alreadyValid) {\n\t\treturn null;\n\t}\n\n\tconst reason = extractReason(body);\n\treturn reason === null ? null : emit(body, reason, dialect);\n}\n\n/** Wrap a fetch so failed responses reach the model in its own dialect. */\nexport function wrapFetchWithGatewayErrorNormalization(\n\tbaseFetch: FetchFunction | undefined,\n\tdialect: GatewayErrorDialect,\n): FetchFunction {\n\tconst inner: FetchFunction =\n\t\tbaseFetch ?? ((...args) => globalThis.fetch(...args));\n\n\treturn async (input, init) => {\n\t\tconst response = await inner(input, init);\n\t\tif (response.ok) {\n\t\t\treturn response;\n\t\t}\n\t\tconst contentType = response.headers.get(\"content-type\") ?? \"\";\n\t\tif (!contentType.includes(\"application/json\")) {\n\t\t\treturn response;\n\t\t}\n\n\t\tlet parsed: unknown;\n\t\ttry {\n\t\t\tparsed = JSON.parse(await response.clone().text());\n\t\t} catch {\n\t\t\treturn response;\n\t\t}\n\n\t\tconst normalized = normalizeGatewayErrorBody(parsed, dialect);\n\t\tif (normalized === null) {\n\t\t\treturn response;\n\t\t}\n\n\t\t// Rewriting the body invalidates both of these.\n\t\tconst headers = new Headers(response.headers);\n\t\theaders.delete(\"content-length\");\n\t\theaders.delete(\"content-encoding\");\n\n\t\treturn new Response(JSON.stringify(normalized), {\n\t\t\tstatus: response.status,\n\t\t\tstatusText: response.statusText,\n\t\t\theaders,\n\t\t});\n\t};\n}\n"],"mappings":";AAsDA,SAAS,SAAS,OAAkD;CACnE,OAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC3E;AAEA,SAAS,iBAAiB,OAAgD;CACzE,IACC,SAAS,KAAK,KACd,SAAS,MAAM,KAAK,KACpB,OAAO,MAAM,MAAM,YAAY,UAE/B,OAAO,MAAM;CAEd,OAAO;AACR;AAEA,SAAS,oBAAoB,OAAgD;CAC5E,IACC,SAAS,KAAK,KACd,MAAM,SAAS,WACf,SAAS,MAAM,KAAK,KACpB,OAAO,MAAM,MAAM,YAAY,YAC/B,OAAO,MAAM,MAAM,SAAS,UAE5B,OAAO,MAAM;CAEd,OAAO;AACR;;AAGA,SAAS,cAAc,MAA8B;CACpD,MAAM,YAAY,oBAAoB,IAAI;CAC1C,IAAI,WACH,OAAO;EACN,SAAS,UAAU;EACnB,MAAM,UAAU;CACjB;CAGD,MAAM,SAAS,iBAAiB,IAAI;CACpC,IAAI,QACH,OAAO;EACN,SAAS,OAAO;EAChB,cAAc;EACd,MAAM,OAAO,OAAO,SAAS,WAAW,OAAO,OAAO,KAAA;CACvD;CAGD,IAAI,CAAC,SAAS,IAAI,KAAK,OAAO,KAAK,YAAY,UAC9C,OAAO;CAER,MAAM,OACL,OAAO,KAAK,eAAe,WAAW,KAAK,aAAa,KAAA;CAIzD,IAAI,KAAK,QAAQ,UAAU,CAAC,CAAC,WAAW,GAAG,GAC1C,IAAI;EAEH,MAAM,SAAS,cADQ,KAAK,MAAM,KAAK,OACN,CAAC;EAClC,IAAI,QACH,OAAO;GAAE,GAAG;GAAQ,MAAM,OAAO,QAAQ;EAAK;CAEhD,QAAQ,CAER;CAID,OAAO;EAAE,SAAS,KAAK;EAAS,MAAM;CAAK;AAC5C;;;;;;AAOA,SAAS,KACR,MACA,QACA,SACU;CACV,MAAM,WAAW,SAAS,IAAI,IAAI,OAAO,CAAC;CAC1C,IAAI,YAAY,aACf,OAAO;EACN,GAAG;EACH,MAAM;EACN,OAAO;GACN,GAAG,OAAO;GACV,MAAM,OAAO,QAAQ;GACrB,SAAS,OAAO;EACjB;CACD;CAED,OAAO;EACN,GAAG;EACH,OAAO;GACN,GAAG,OAAO;GACV,SAAS,OAAO;GAChB,MAAM,OAAO,cAAc,QAAQ,OAAO,QAAQ;EACnD;CACD;AACD;;;;;AAMA,SAAgB,0BACf,MACA,SACiB;CAKjB,IAHC,YAAY,cACT,oBAAoB,IAAI,MAAM,OAC9B,iBAAiB,IAAI,MAAM,MAE9B,OAAO;CAGR,MAAM,SAAS,cAAc,IAAI;CACjC,OAAO,WAAW,OAAO,OAAO,KAAK,MAAM,QAAQ,OAAO;AAC3D;;AAGA,SAAgB,uCACf,WACA,SACgB;CAChB,MAAM,QACL,eAAe,GAAG,SAAS,WAAW,MAAM,GAAG,IAAI;CAEpD,OAAO,OAAO,OAAO,SAAS;EAC7B,MAAM,WAAW,MAAM,MAAM,OAAO,IAAI;EACxC,IAAI,SAAS,IACZ,OAAO;EAGR,IAAI,EADgB,SAAS,QAAQ,IAAI,cAAc,KAAK,GAAA,CAC3C,SAAS,kBAAkB,GAC3C,OAAO;EAGR,IAAI;EACJ,IAAI;GACH,SAAS,KAAK,MAAM,MAAM,SAAS,MAAM,CAAC,CAAC,KAAK,CAAC;EAClD,QAAQ;GACP,OAAO;EACR;EAEA,MAAM,aAAa,0BAA0B,QAAQ,OAAO;EAC5D,IAAI,eAAe,MAClB,OAAO;EAIR,MAAM,UAAU,IAAI,QAAQ,SAAS,OAAO;EAC5C,QAAQ,OAAO,gBAAgB;EAC/B,QAAQ,OAAO,kBAAkB;EAEjC,OAAO,IAAI,SAAS,KAAK,UAAU,UAAU,GAAG;GAC/C,QAAQ,SAAS;GACjB,YAAY,SAAS;GACrB;EACD,CAAC;CACF;AACD"}

package/dist/lib/neon-model-capabilities.d.ts

@@ -25,6 +25,13 @@ interface NeonModelCapabilities {
   supportsSeed: boolean;
   supportsStopSequences: boolean;
   supportsReasoningEffort: boolean;
+  /**
+   * True when a `claude-*` id could not be parsed into a version and was
+   * assumed to be new. The sampling restriction is then a precaution rather
+   * than something measured, and the warning says so instead of asserting a
+   * capability nobody checked.
+   */
+  claudeSamplingUnrecognized?: boolean;
 }
 /**
  * Heuristic, prefix-based capability detection for MLflow-routed Neon models.

package/dist/lib/neon-model-capabilities.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"neon-model-capabilities.d.ts","names":[],"sources":["../../src/lib/neon-model-capabilities.ts"],"mappings":";KAAY,eAAA;AAAZ;AAiBA;AAEA;AAiBA;AAkCA;;;;;;KArDY,cAAA;iBAEI,iBAAA,mBAAoC;UAiBnC,qBAAA;UACR;;;;;;;;;;;;;;;;;;;;;;iBAiCO,wBAAA,mBAEb"}
+{"version":3,"file":"neon-model-capabilities.d.ts","names":[],"sources":["../../src/lib/neon-model-capabilities.ts"],"mappings":";KAAY,eAAA;AAAZ;AAiBA;AAEA;AAiBA;AA8DA;;;;;;KAjFY,cAAA;iBAEI,iBAAA,mBAAoC;UAiBnC,qBAAA;UACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA6DO,wBAAA,mBAEb"}

package/dist/lib/neon-model-capabilities.js

@@ -15,6 +15,20 @@ const PERMISSIVE = {
 	supportsReasoningEffort: true
 };
 /**
+* Whether a Claude id predates the 4.7 cutoff where the gateway stopped
+* accepting `temperature` and `top_p`. Ids look like `claude-<tier>-<major>` or
+* `claude-<tier>-<major>-<minor>`, optionally behind the legacy `databricks-`
+* prefix. An id we cannot parse is treated as new, which is the safe direction.
+*/
+function claudeVersion(id) {
+	const match = /^(?:databricks-)?claude-[a-z]+-(\d+)(?:-(\d+))?$/.exec(id);
+	if (!match) return null;
+	return {
+		major: Number(match[1]),
+		minor: match[2] === void 0 ? 0 : Number(match[2])
+	};
+}
+/**
 * Heuristic, prefix-based capability detection for MLflow-routed Neon models.
 *
 * The unified endpoint proxies to heterogeneous upstream providers, each of
@@ -25,16 +39,21 @@ const PERMISSIVE = {
 */
 function getNeonModelCapabilities(modelId) {
 	const id = modelId.toLowerCase();
-	if (id.includes("claude")) return {
-		family: "anthropic",
-		supportsTemperature: true,
-		supportsTopP: true,
-		temperatureTopPMutuallyExclusive: true,
-		supportsPenalties: false,
-		supportsSeed: false,
-		supportsStopSequences: true,
-		supportsReasoningEffort: false
-	};
+	if (id.includes("claude")) {
+		const version = claudeVersion(id);
+		const sampling = version !== null && (version.major < 4 || version.major === 4 && version.minor <= 6);
+		return {
+			family: "anthropic",
+			supportsTemperature: sampling,
+			supportsTopP: sampling,
+			temperatureTopPMutuallyExclusive: true,
+			supportsPenalties: false,
+			supportsSeed: false,
+			supportsStopSequences: true,
+			supportsReasoningEffort: false,
+			...version === null ? { claudeSamplingUnrecognized: true } : {}
+		};
+	}
 	if (id.includes("gemini")) return {
 		family: "google",
 		...PERMISSIVE,
@@ -63,6 +82,24 @@ function getNeonModelCapabilities(modelId) {
 		supportsStopSequences: true,
 		supportsReasoningEffort: true
 	};
+	if (id.includes("gpt-oss")) return {
+		family: "other",
+		...PERMISSIVE,
+		supportsPenalties: false,
+		supportsSeed: false,
+		supportsStopSequences: false
+	};
+	if (id.includes("qwen") || id.includes("gemma")) return {
+		family: "other",
+		...PERMISSIVE,
+		supportsPenalties: false,
+		supportsSeed: false
+	};
+	if (id.includes("glm") || id.includes("inkling")) return {
+		family: "other",
+		...PERMISSIVE,
+		supportsPenalties: false
+	};
 	return {
 		family: "other",
 		...PERMISSIVE

package/dist/lib/neon-model-capabilities.js.map

@@ -1 +1 @@
-{"version":3,"file":"neon-model-capabilities.js","names":[],"sources":["../../src/lib/neon-model-capabilities.ts"],"sourcesContent":["export type NeonModelFamily =\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"openai\"\n\t| \"meta\"\n\t| \"other\";\n\n/**\n * Which gateway endpoint a model should be routed to.\n *\n * - `anthropic`: native Messages API — streaming structured output + native reasoning.\n * - `openai`: native Responses API — models served only natively (e.g. Codex),\n *   native reasoning, and the image-generation tool.\n * - `mlflow`: the unified, OpenAI-compatible endpoint (the fallback). Gemini is\n *   routed here because the gateway's native Gemini endpoint does not support\n *   streaming (`streamGenerateContent` is rejected).\n */\nexport type NeonModelRoute = \"anthropic\" | \"openai\" | \"mlflow\";\n\nexport function getNeonModelRoute(modelId: string): NeonModelRoute {\n\tconst id = modelId.toLowerCase();\n\tif (id.includes(\"claude\")) {\n\t\treturn \"anthropic\";\n\t}\n\t// OpenAI proprietary models (gpt-4/gpt-5 families, Codex) are served only via\n\t// the native Responses API. `gpt-oss` is open-weight and served on the\n\t// unified chat endpoint, so it is intentionally excluded here.\n\tif (\n\t\t(id.includes(\"gpt-\") && !id.includes(\"gpt-oss\")) ||\n\t\tid.includes(\"codex\")\n\t) {\n\t\treturn \"openai\";\n\t}\n\treturn \"mlflow\";\n}\n\nexport interface NeonModelCapabilities {\n\tfamily: NeonModelFamily;\n\tsupportsTemperature: boolean;\n\tsupportsTopP: boolean;\n\t/**\n\t * Whether `temperature` and `topP` may be sent together. Anthropic models\n\t * accept only one of the two.\n\t */\n\ttemperatureTopPMutuallyExclusive: boolean;\n\tsupportsPenalties: boolean;\n\tsupportsSeed: boolean;\n\tsupportsStopSequences: boolean;\n\tsupportsReasoningEffort: boolean;\n}\n\nconst PERMISSIVE: Omit<NeonModelCapabilities, \"family\"> = {\n\tsupportsTemperature: true,\n\tsupportsTopP: true,\n\ttemperatureTopPMutuallyExclusive: false,\n\tsupportsPenalties: true,\n\tsupportsSeed: true,\n\tsupportsStopSequences: true,\n\tsupportsReasoningEffort: true,\n};\n\n/**\n * Heuristic, prefix-based capability detection for MLflow-routed Neon models.\n *\n * The unified endpoint proxies to heterogeneous upstream providers, each of\n * which accepts a different subset of the OpenAI-style parameters the AI SDK\n * emits. Sending a parameter an upstream rejects results in a hard `400`, so we\n * strip the parameters a family is known to reject and surface a warning\n * instead. Unknown/untested models stay permissive (passed through unchanged).\n */\nexport function getNeonModelCapabilities(\n\tmodelId: string,\n): NeonModelCapabilities {\n\tconst id = modelId.toLowerCase();\n\n\t// Anthropic (Claude): rejects penalties and seed, accepts only one of\n\t// temperature/topP, and rejects the OpenAI `reasoning_effort` field.\n\tif (id.includes(\"claude\")) {\n\t\treturn {\n\t\t\tfamily: \"anthropic\",\n\t\t\tsupportsTemperature: true,\n\t\t\tsupportsTopP: true,\n\t\t\ttemperatureTopPMutuallyExclusive: true,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: false,\n\t\t};\n\t}\n\n\t// Google (Gemini): accepts standard sampling params but rejects the OpenAI\n\t// `reasoning_effort` field (not part of Gemini's generation config).\n\tif (id.includes(\"gemini\")) {\n\t\treturn {\n\t\t\tfamily: \"google\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsReasoningEffort: false,\n\t\t};\n\t}\n\n\t// OpenAI GPT-5 reasoning family (routed to the native Responses API). The\n\t// Responses model strips parameters the Responses API doesn't accept\n\t// (penalties, seed, stop), but its reasoning-model detection matches the bare\n\t// model id (`gpt-5`), which the gateway's optional `databricks-` prefix can\n\t// defeat. So we only handle the temperature/topP restriction here: the\n\t// original gpt-5 / gpt-5-mini / gpt-5-nano require the default temperature and\n\t// reject topP, while gpt-5.1+ (a minor version digit follows) accept them\n\t// again. The regex matches both prefixed and unprefixed ids.\n\tif (/gpt-5/.test(id)) {\n\t\tconst hasMinorVersion = /gpt-5[.-]\\d/.test(id);\n\t\treturn {\n\t\t\tfamily: \"openai\",\n\t\t\tsupportsTemperature: hasMinorVersion,\n\t\t\tsupportsTopP: hasMinorVersion,\n\t\t\ttemperatureTopPMutuallyExclusive: false,\n\t\t\tsupportsPenalties: true,\n\t\t\tsupportsSeed: true,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: true,\n\t\t};\n\t}\n\n\t// Meta (Llama): rejects penalties and seed; accepts sampling params and stop.\n\tif (id.includes(\"llama\")) {\n\t\treturn {\n\t\t\tfamily: \"meta\",\n\t\t\tsupportsTemperature: true,\n\t\t\tsupportsTopP: true,\n\t\t\ttemperatureTopPMutuallyExclusive: false,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: true,\n\t\t};\n\t}\n\n\treturn { family: \"other\", ...PERMISSIVE };\n}\n"],"mappings":";AAmBA,SAAgB,kBAAkB,SAAiC;CAClE,MAAM,KAAK,QAAQ,YAAY;CAC/B,IAAI,GAAG,SAAS,QAAQ,GACvB,OAAO;CAKR,IACE,GAAG,SAAS,MAAM,KAAK,CAAC,GAAG,SAAS,SAAS,KAC9C,GAAG,SAAS,OAAO,GAEnB,OAAO;CAER,OAAO;AACR;AAiBA,MAAM,aAAoD;CACzD,qBAAqB;CACrB,cAAc;CACd,kCAAkC;CAClC,mBAAmB;CACnB,cAAc;CACd,uBAAuB;CACvB,yBAAyB;AAC1B;;;;;;;;;;AAWA,SAAgB,yBACf,SACwB;CACxB,MAAM,KAAK,QAAQ,YAAY;CAI/B,IAAI,GAAG,SAAS,QAAQ,GACvB,OAAO;EACN,QAAQ;EACR,qBAAqB;EACrB,cAAc;EACd,kCAAkC;EAClC,mBAAmB;EACnB,cAAc;EACd,uBAAuB;EACvB,yBAAyB;CAC1B;CAKD,IAAI,GAAG,SAAS,QAAQ,GACvB,OAAO;EACN,QAAQ;EACR,GAAG;EACH,yBAAyB;CAC1B;CAWD,IAAI,QAAQ,KAAK,EAAE,GAAG;EACrB,MAAM,kBAAkB,cAAc,KAAK,EAAE;EAC7C,OAAO;GACN,QAAQ;GACR,qBAAqB;GACrB,cAAc;GACd,kCAAkC;GAClC,mBAAmB;GACnB,cAAc;GACd,uBAAuB;GACvB,yBAAyB;EAC1B;CACD;CAGA,IAAI,GAAG,SAAS,OAAO,GACtB,OAAO;EACN,QAAQ;EACR,qBAAqB;EACrB,cAAc;EACd,kCAAkC;EAClC,mBAAmB;EACnB,cAAc;EACd,uBAAuB;EACvB,yBAAyB;CAC1B;CAGD,OAAO;EAAE,QAAQ;EAAS,GAAG;CAAW;AACzC"}
+{"version":3,"file":"neon-model-capabilities.js","names":[],"sources":["../../src/lib/neon-model-capabilities.ts"],"sourcesContent":["export type NeonModelFamily =\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"openai\"\n\t| \"meta\"\n\t| \"other\";\n\n/**\n * Which gateway endpoint a model should be routed to.\n *\n * - `anthropic`: native Messages API — streaming structured output + native reasoning.\n * - `openai`: native Responses API — models served only natively (e.g. Codex),\n *   native reasoning, and the image-generation tool.\n * - `mlflow`: the unified, OpenAI-compatible endpoint (the fallback). Gemini is\n *   routed here because the gateway's native Gemini endpoint does not support\n *   streaming (`streamGenerateContent` is rejected).\n */\nexport type NeonModelRoute = \"anthropic\" | \"openai\" | \"mlflow\";\n\nexport function getNeonModelRoute(modelId: string): NeonModelRoute {\n\tconst id = modelId.toLowerCase();\n\tif (id.includes(\"claude\")) {\n\t\treturn \"anthropic\";\n\t}\n\t// OpenAI proprietary models (gpt-4/gpt-5 families, Codex) are served only via\n\t// the native Responses API. `gpt-oss` is open-weight and served on the\n\t// unified chat endpoint, so it is intentionally excluded here.\n\tif (\n\t\t(id.includes(\"gpt-\") && !id.includes(\"gpt-oss\")) ||\n\t\tid.includes(\"codex\")\n\t) {\n\t\treturn \"openai\";\n\t}\n\treturn \"mlflow\";\n}\n\nexport interface NeonModelCapabilities {\n\tfamily: NeonModelFamily;\n\tsupportsTemperature: boolean;\n\tsupportsTopP: boolean;\n\t/**\n\t * Whether `temperature` and `topP` may be sent together. Anthropic models\n\t * accept only one of the two.\n\t */\n\ttemperatureTopPMutuallyExclusive: boolean;\n\tsupportsPenalties: boolean;\n\tsupportsSeed: boolean;\n\tsupportsStopSequences: boolean;\n\tsupportsReasoningEffort: boolean;\n\t/**\n\t * True when a `claude-*` id could not be parsed into a version and was\n\t * assumed to be new. The sampling restriction is then a precaution rather\n\t * than something measured, and the warning says so instead of asserting a\n\t * capability nobody checked.\n\t */\n\tclaudeSamplingUnrecognized?: boolean;\n}\n\nconst PERMISSIVE: Omit<NeonModelCapabilities, \"family\"> = {\n\tsupportsTemperature: true,\n\tsupportsTopP: true,\n\ttemperatureTopPMutuallyExclusive: false,\n\tsupportsPenalties: true,\n\tsupportsSeed: true,\n\tsupportsStopSequences: true,\n\tsupportsReasoningEffort: true,\n};\n\n/**\n * Whether a Claude id predates the 4.7 cutoff where the gateway stopped\n * accepting `temperature` and `top_p`. Ids look like `claude-<tier>-<major>` or\n * `claude-<tier>-<major>-<minor>`, optionally behind the legacy `databricks-`\n * prefix. An id we cannot parse is treated as new, which is the safe direction.\n */\nfunction claudeVersion(id: string): { major: number; minor: number } | null {\n\t// Anchored: `claude-opus-4-beta` and `claude-opus-4.7` must not match their\n\t// leading digits and be read as old models. A dated id such as\n\t// `claude-3-5-sonnet-20241022` does not parse either, and is reported as\n\t// unrecognised rather than silently classified.\n\tconst match = /^(?:databricks-)?claude-[a-z]+-(\\d+)(?:-(\\d+))?$/.exec(id);\n\tif (!match) {\n\t\treturn null;\n\t}\n\treturn {\n\t\tmajor: Number(match[1]),\n\t\tminor: match[2] === undefined ? 0 : Number(match[2]),\n\t};\n}\n\n/**\n * Heuristic, prefix-based capability detection for MLflow-routed Neon models.\n *\n * The unified endpoint proxies to heterogeneous upstream providers, each of\n * which accepts a different subset of the OpenAI-style parameters the AI SDK\n * emits. Sending a parameter an upstream rejects results in a hard `400`, so we\n * strip the parameters a family is known to reject and surface a warning\n * instead. Unknown/untested models stay permissive (passed through unchanged).\n */\nexport function getNeonModelCapabilities(\n\tmodelId: string,\n): NeonModelCapabilities {\n\tconst id = modelId.toLowerCase();\n\n\t// Anthropic (Claude): rejects penalties and seed, accepts only one of\n\t// temperature/topP, and rejects the OpenAI `reasoning_effort` field.\n\t//\n\t// From Claude 4.7 onward the gateway also rejects non-default sampling\n\t// entirely — `does not support the temperature parameter` /\n\t// `does not support sampling parameters: top_p` — because those models are\n\t// steered with `output_config.effort` instead. Measured: 4.1/4.5/4.6 accept\n\t// both, 4.7/4.8/5 reject both, and every id accepts the default\n\t// `temperature: 1.0`. A version comparison rather than a list of ids, so a\n\t// future Claude inherits the restriction: dropping a parameter the model\n\t// would have accepted costs a warning, claiming one it rejects costs a 400.\n\tif (id.includes(\"claude\")) {\n\t\tconst version = claudeVersion(id);\n\t\tconst sampling =\n\t\t\tversion !== null &&\n\t\t\t(version.major < 4 || (version.major === 4 && version.minor <= 6));\n\t\treturn {\n\t\t\tfamily: \"anthropic\",\n\t\t\tsupportsTemperature: sampling,\n\t\t\tsupportsTopP: sampling,\n\t\t\ttemperatureTopPMutuallyExclusive: true,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: false,\n\t\t\t...(version === null ? { claudeSamplingUnrecognized: true } : {}),\n\t\t};\n\t}\n\n\t// Google (Gemini): accepts standard sampling params but rejects the OpenAI\n\t// `reasoning_effort` field (not part of Gemini's generation config).\n\tif (id.includes(\"gemini\")) {\n\t\treturn {\n\t\t\tfamily: \"google\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsReasoningEffort: false,\n\t\t};\n\t}\n\n\t// OpenAI GPT-5 reasoning family (routed to the native Responses API). The\n\t// Responses model strips parameters the Responses API doesn't accept\n\t// (penalties, seed, stop), but its reasoning-model detection matches the bare\n\t// model id (`gpt-5`), which the gateway's optional `databricks-` prefix can\n\t// defeat. So we only handle the temperature/topP restriction here: the\n\t// original gpt-5 / gpt-5-mini / gpt-5-nano require the default temperature and\n\t// reject topP, while gpt-5.1+ (a minor version digit follows) accept them\n\t// again. The regex matches both prefixed and unprefixed ids.\n\tif (/gpt-5/.test(id)) {\n\t\tconst hasMinorVersion = /gpt-5[.-]\\d/.test(id);\n\t\treturn {\n\t\t\tfamily: \"openai\",\n\t\t\tsupportsTemperature: hasMinorVersion,\n\t\t\tsupportsTopP: hasMinorVersion,\n\t\t\ttemperatureTopPMutuallyExclusive: false,\n\t\t\tsupportsPenalties: true,\n\t\t\tsupportsSeed: true,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: true,\n\t\t};\n\t}\n\n\t// Meta (Llama): rejects penalties and seed; accepts sampling params and stop.\n\tif (id.includes(\"llama\")) {\n\t\treturn {\n\t\t\tfamily: \"meta\",\n\t\t\tsupportsTemperature: true,\n\t\t\tsupportsTopP: true,\n\t\t\ttemperatureTopPMutuallyExclusive: false,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: true,\n\t\t\tsupportsReasoningEffort: true,\n\t\t};\n\t}\n\n\t// Everything else on the unified endpoint rejects penalties with\n\t// `parameter \"frequency_penalty\" must be equal to 0`. Gemini, handled above,\n\t// is the only MLflow-routed family that accepts them. Seed and stop vary, so\n\t// each family carries what was measured rather than inheriting a guess.\n\tif (id.includes(\"gpt-oss\")) {\n\t\treturn {\n\t\t\tfamily: \"other\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t\tsupportsStopSequences: false,\n\t\t};\n\t}\n\n\tif (id.includes(\"qwen\") || id.includes(\"gemma\")) {\n\t\treturn {\n\t\t\tfamily: \"other\",\n\t\t\t...PERMISSIVE,\n\t\t\tsupportsPenalties: false,\n\t\t\tsupportsSeed: false,\n\t\t};\n\t}\n\n\tif (id.includes(\"glm\") || id.includes(\"inkling\")) {\n\t\treturn { family: \"other\", ...PERMISSIVE, supportsPenalties: false };\n\t}\n\n\treturn { family: \"other\", ...PERMISSIVE };\n}\n"],"mappings":";AAmBA,SAAgB,kBAAkB,SAAiC;CAClE,MAAM,KAAK,QAAQ,YAAY;CAC/B,IAAI,GAAG,SAAS,QAAQ,GACvB,OAAO;CAKR,IACE,GAAG,SAAS,MAAM,KAAK,CAAC,GAAG,SAAS,SAAS,KAC9C,GAAG,SAAS,OAAO,GAEnB,OAAO;CAER,OAAO;AACR;AAwBA,MAAM,aAAoD;CACzD,qBAAqB;CACrB,cAAc;CACd,kCAAkC;CAClC,mBAAmB;CACnB,cAAc;CACd,uBAAuB;CACvB,yBAAyB;AAC1B;;;;;;;AAQA,SAAS,cAAc,IAAqD;CAK3E,MAAM,QAAQ,mDAAmD,KAAK,EAAE;CACxE,IAAI,CAAC,OACJ,OAAO;CAER,OAAO;EACN,OAAO,OAAO,MAAM,EAAE;EACtB,OAAO,MAAM,OAAO,KAAA,IAAY,IAAI,OAAO,MAAM,EAAE;CACpD;AACD;;;;;;;;;;AAWA,SAAgB,yBACf,SACwB;CACxB,MAAM,KAAK,QAAQ,YAAY;CAa/B,IAAI,GAAG,SAAS,QAAQ,GAAG;EAC1B,MAAM,UAAU,cAAc,EAAE;EAChC,MAAM,WACL,YAAY,SACX,QAAQ,QAAQ,KAAM,QAAQ,UAAU,KAAK,QAAQ,SAAS;EAChE,OAAO;GACN,QAAQ;GACR,qBAAqB;GACrB,cAAc;GACd,kCAAkC;GAClC,mBAAmB;GACnB,cAAc;GACd,uBAAuB;GACvB,yBAAyB;GACzB,GAAI,YAAY,OAAO,EAAE,4BAA4B,KAAK,IAAI,CAAC;EAChE;CACD;CAIA,IAAI,GAAG,SAAS,QAAQ,GACvB,OAAO;EACN,QAAQ;EACR,GAAG;EACH,yBAAyB;CAC1B;CAWD,IAAI,QAAQ,KAAK,EAAE,GAAG;EACrB,MAAM,kBAAkB,cAAc,KAAK,EAAE;EAC7C,OAAO;GACN,QAAQ;GACR,qBAAqB;GACrB,cAAc;GACd,kCAAkC;GAClC,mBAAmB;GACnB,cAAc;GACd,uBAAuB;GACvB,yBAAyB;EAC1B;CACD;CAGA,IAAI,GAAG,SAAS,OAAO,GACtB,OAAO;EACN,QAAQ;EACR,qBAAqB;EACrB,cAAc;EACd,kCAAkC;EAClC,mBAAmB;EACnB,cAAc;EACd,uBAAuB;EACvB,yBAAyB;CAC1B;CAOD,IAAI,GAAG,SAAS,SAAS,GACxB,OAAO;EACN,QAAQ;EACR,GAAG;EACH,mBAAmB;EACnB,cAAc;EACd,uBAAuB;CACxB;CAGD,IAAI,GAAG,SAAS,MAAM,KAAK,GAAG,SAAS,OAAO,GAC7C,OAAO;EACN,QAAQ;EACR,GAAG;EACH,mBAAmB;EACnB,cAAc;CACf;CAGD,IAAI,GAAG,SAAS,KAAK,KAAK,GAAG,SAAS,SAAS,GAC9C,OAAO;EAAE,QAAQ;EAAS,GAAG;EAAY,mBAAmB;CAAM;CAGnE,OAAO;EAAE,QAAQ;EAAS,GAAG;CAAW;AACzC"}

package/dist/lib/neon-responses-language-model.d.ts

@@ -5,20 +5,14 @@ import { LanguageModelV3, LanguageModelV3CallOptions, LanguageModelV3GenerateRes
 
 /**
  * Language model for OpenAI models served via the Neon AI Gateway's native
- * Responses API (`/openai/v1/responses`). This route is required for
- * models that are only served natively (e.g. Codex) and unlocks native
- * reasoning and the image-generation tool.
+ * Responses API (`/openai/v1/responses`). This route is required for models
+ * that are only served natively (e.g. Codex) and unlocks native reasoning and
+ * the image-generation tool.
  *
- * The shared OpenAI Responses model shapes a request based on whether the model
- * is a reasoning model, but its detection is brittle here: the gateway also
- * accepts the legacy `databricks-` prefix, which defeats the upstream bare-id
- * (`gpt-5`) match. For the GPT-5 reasoning family we set the model's own
- * `forceReasoning` provider option so it applies the correct reasoning behavior
- * for both id forms. Users can still override it.
+ * The request shaping lives in `neon-responses-options.ts`, which documents
+ * what this route requires and why.
  */
 declare class NeonResponsesLanguageModel extends OpenAIResponsesLanguageModel implements LanguageModelV3 {
-  private get isReasoningFamily();
-  private withForcedReasoning;
   doGenerate(options: LanguageModelV3CallOptions): Promise<LanguageModelV3GenerateResult>;
   doStream(options: LanguageModelV3CallOptions): Promise<LanguageModelV3StreamResult>;
 }

package/dist/lib/neon-responses-language-model.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"neon-responses-language-model.d.ts","names":[],"sources":["../../src/lib/neon-responses-language-model.ts"],"mappings":";;;;;;;AAqBA;AACC;AA4BU;AACC;AAAR;AAKO;AACC;AAAR;AAnCK;AACG;AAAe;cAFd,0BAAA,SACJ,4BAAA,YACG;;;sBA2BD,6BACP,QAAQ;oBAKD,6BACP,QAAQ"}
+{"version":3,"file":"neon-responses-language-model.d.ts","names":[],"sources":["../../src/lib/neon-responses-language-model.ts"],"mappings":";;;;;;;AAkBA;AACC;AAIU;AACC;AAAR;AAOO;AACC;AAAR,cAdS,0BAAA,SACJ,4BAAA,YACG,eAYR,CAAA;EAbK,UAAA,CAAA,OAAA,EAIE,0BAJF,CAAA,EAKL,OALK,CAKG,6BALH,CAAA;EACG,QAAA,CAAA,OAAA,EAWD,0BAXC,CAAA,EAYR,OAZQ,CAYA,2BAZA,CAAA;AAAe"}

package/dist/lib/neon-responses-language-model.js

@@ -1,42 +1,21 @@
+import { withResponsesGatewayDefaults } from "./neon-responses-options.js";
 import { OpenAIResponsesLanguageModel } from "@ai-sdk/openai/internal";
 //#region src/lib/neon-responses-language-model.ts
 /**
 * Language model for OpenAI models served via the Neon AI Gateway's native
-* Responses API (`/openai/v1/responses`). This route is required for
-* models that are only served natively (e.g. Codex) and unlocks native
-* reasoning and the image-generation tool.
+* Responses API (`/openai/v1/responses`). This route is required for models
+* that are only served natively (e.g. Codex) and unlocks native reasoning and
+* the image-generation tool.
 *
-* The shared OpenAI Responses model shapes a request based on whether the model
-* is a reasoning model, but its detection is brittle here: the gateway also
-* accepts the legacy `databricks-` prefix, which defeats the upstream bare-id
-* (`gpt-5`) match. For the GPT-5 reasoning family we set the model's own
-* `forceReasoning` provider option so it applies the correct reasoning behavior
-* for both id forms. Users can still override it.
+* The request shaping lives in `neon-responses-options.ts`, which documents
+* what this route requires and why.
 */
 var NeonResponsesLanguageModel = class extends OpenAIResponsesLanguageModel {
-	get isReasoningFamily() {
-		return /gpt-5/.test(this.modelId.toLowerCase());
-	}
-	withForcedReasoning(options) {
-		if (!this.isReasoningFamily) return options;
-		const openai = options.providerOptions?.openai;
-		if (openai != null && "forceReasoning" in openai) return options;
-		return {
-			...options,
-			providerOptions: {
-				...options.providerOptions,
-				openai: {
-					...openai,
-					forceReasoning: true
-				}
-			}
-		};
-	}
 	doGenerate(options) {
-		return super.doGenerate(this.withForcedReasoning(options));
+		return super.doGenerate(withResponsesGatewayDefaults(this.modelId, options));
 	}
 	doStream(options) {
-		return super.doStream(this.withForcedReasoning(options));
+		return super.doStream(withResponsesGatewayDefaults(this.modelId, options));
 	}
 };
 //#endregion

package/dist/lib/neon-responses-language-model.js.map

@@ -1 +1 @@
-{"version":3,"file":"neon-responses-language-model.js","names":[],"sources":["../../src/lib/neon-responses-language-model.ts"],"sourcesContent":["import { OpenAIResponsesLanguageModel } from \"@ai-sdk/openai/internal\";\nimport type {\n\tLanguageModelV3,\n\tLanguageModelV3CallOptions,\n\tLanguageModelV3GenerateResult,\n\tLanguageModelV3StreamResult,\n} from \"@ai-sdk/provider\";\n\n/**\n * Language model for OpenAI models served via the Neon AI Gateway's native\n * Responses API (`/openai/v1/responses`). This route is required for\n * models that are only served natively (e.g. Codex) and unlocks native\n * reasoning and the image-generation tool.\n *\n * The shared OpenAI Responses model shapes a request based on whether the model\n * is a reasoning model, but its detection is brittle here: the gateway also\n * accepts the legacy `databricks-` prefix, which defeats the upstream bare-id\n * (`gpt-5`) match. For the GPT-5 reasoning family we set the model's own\n * `forceReasoning` provider option so it applies the correct reasoning behavior\n * for both id forms. Users can still override it.\n */\nexport class NeonResponsesLanguageModel\n\textends OpenAIResponsesLanguageModel\n\timplements LanguageModelV3\n{\n\tprivate get isReasoningFamily(): boolean {\n\t\treturn /gpt-5/.test(this.modelId.toLowerCase());\n\t}\n\n\tprivate withForcedReasoning(\n\t\toptions: LanguageModelV3CallOptions,\n\t): LanguageModelV3CallOptions {\n\t\tif (!this.isReasoningFamily) {\n\t\t\treturn options;\n\t\t}\n\t\tconst openai = options.providerOptions?.openai;\n\t\t// Respect an explicit user setting.\n\t\tif (openai != null && \"forceReasoning\" in openai) {\n\t\t\treturn options;\n\t\t}\n\t\treturn {\n\t\t\t...options,\n\t\t\tproviderOptions: {\n\t\t\t\t...options.providerOptions,\n\t\t\t\topenai: { ...openai, forceReasoning: true },\n\t\t\t},\n\t\t};\n\t}\n\n\toverride doGenerate(\n\t\toptions: LanguageModelV3CallOptions,\n\t): Promise<LanguageModelV3GenerateResult> {\n\t\treturn super.doGenerate(this.withForcedReasoning(options));\n\t}\n\n\toverride doStream(\n\t\toptions: LanguageModelV3CallOptions,\n\t): Promise<LanguageModelV3StreamResult> {\n\t\treturn super.doStream(this.withForcedReasoning(options));\n\t}\n}\n"],"mappings":";;;;;;;;;;;;;;;AAqBA,IAAa,6BAAb,cACS,6BAET;CACC,IAAY,oBAA6B;EACxC,OAAO,QAAQ,KAAK,KAAK,QAAQ,YAAY,CAAC;CAC/C;CAEA,oBACC,SAC6B;EAC7B,IAAI,CAAC,KAAK,mBACT,OAAO;EAER,MAAM,SAAS,QAAQ,iBAAiB;EAExC,IAAI,UAAU,QAAQ,oBAAoB,QACzC,OAAO;EAER,OAAO;GACN,GAAG;GACH,iBAAiB;IAChB,GAAG,QAAQ;IACX,QAAQ;KAAE,GAAG;KAAQ,gBAAgB;IAAK;GAC3C;EACD;CACD;CAEA,WACC,SACyC;EACzC,OAAO,MAAM,WAAW,KAAK,oBAAoB,OAAO,CAAC;CAC1D;CAEA,SACC,SACuC;EACvC,OAAO,MAAM,SAAS,KAAK,oBAAoB,OAAO,CAAC;CACxD;AACD"}
+{"version":3,"file":"neon-responses-language-model.js","names":[],"sources":["../../src/lib/neon-responses-language-model.ts"],"sourcesContent":["import { OpenAIResponsesLanguageModel } from \"@ai-sdk/openai/internal\";\nimport type {\n\tLanguageModelV3,\n\tLanguageModelV3CallOptions,\n\tLanguageModelV3GenerateResult,\n\tLanguageModelV3StreamResult,\n} from \"@ai-sdk/provider\";\nimport { withResponsesGatewayDefaults } from \"./neon-responses-options.js\";\n\n/**\n * Language model for OpenAI models served via the Neon AI Gateway's native\n * Responses API (`/openai/v1/responses`). This route is required for models\n * that are only served natively (e.g. Codex) and unlocks native reasoning and\n * the image-generation tool.\n *\n * The request shaping lives in `neon-responses-options.ts`, which documents\n * what this route requires and why.\n */\nexport class NeonResponsesLanguageModel\n\textends OpenAIResponsesLanguageModel\n\timplements LanguageModelV3\n{\n\toverride doGenerate(\n\t\toptions: LanguageModelV3CallOptions,\n\t): Promise<LanguageModelV3GenerateResult> {\n\t\treturn super.doGenerate(\n\t\t\twithResponsesGatewayDefaults(this.modelId, options),\n\t\t);\n\t}\n\n\toverride doStream(\n\t\toptions: LanguageModelV3CallOptions,\n\t): Promise<LanguageModelV3StreamResult> {\n\t\treturn super.doStream(\n\t\t\twithResponsesGatewayDefaults(this.modelId, options),\n\t\t);\n\t}\n}\n"],"mappings":";;;;;;;;;;;;AAkBA,IAAa,6BAAb,cACS,6BAET;CACC,WACC,SACyC;EACzC,OAAO,MAAM,WACZ,6BAA6B,KAAK,SAAS,OAAO,CACnD;CACD;CAEA,SACC,SACuC;EACvC,OAAO,MAAM,SACZ,6BAA6B,KAAK,SAAS,OAAO,CACnD;CACD;AACD"}

package/dist/lib/neon-responses-options.d.ts

@@ -0,0 +1,14 @@
+import { LanguageModelV3CallOptions } from "@ai-sdk/provider";
+
+//#region src/lib/neon-responses-options.d.ts
+/** Models the gateway serves as reasoning models on this route. */
+declare function isReasoningModelId(modelId: string): boolean;
+/**
+ * Apply the route's defaults to one call, or throw when a caller asked for
+ * something the gateway cannot do. Returns the options unchanged when there is
+ * nothing to add.
+ */
+declare function withResponsesGatewayDefaults(modelId: string, options: LanguageModelV3CallOptions): LanguageModelV3CallOptions;
+//#endregion
+export { isReasoningModelId, withResponsesGatewayDefaults };
+//# sourceMappingURL=neon-responses-options.d.ts.map

package/dist/lib/neon-responses-options.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"neon-responses-options.d.ts","names":[],"sources":["../../src/lib/neon-responses-options.ts"],"mappings":";;;;iBAmCgB,kBAAA;AAAhB;AASA;AAA4C;AAElC;AACP;AAA0B,iBAHb,4BAAA,CAGa,OAAA,EAAA,MAAA,EAAA,OAAA,EADnB,0BACmB,CAAA,EAA1B,0BAA0B"}

package/dist/lib/neon-responses-options.js

@@ -0,0 +1,66 @@
+import { UnsupportedFunctionalityError } from "@ai-sdk/provider";
+//#region src/lib/neon-responses-options.ts
+/**
+* Request defaults for the Neon AI Gateway's Responses route.
+*
+* `store: false` is the one value this gateway accepts, not a preference. It
+* serves the Responses API statelessly and keeps no items — a response comes
+* back with `"store": false` even when the request never set it. Left alone the
+* shared OpenAI model assumes stored-item semantics and replays earlier
+* reasoning as `{ type: "item_reference", id }`, which nothing upstream can
+* resolve; the gateway answers `502 INTERNAL_ERROR` on the second step of a
+* tool loop, after the first has already succeeded. Sending `false` makes the
+* model inline the encrypted reasoning instead.
+*
+* Since no other value can work, an explicit `store` that is not `false` is
+* refused rather than sent: the request would come back `400
+* INVALID_PARAMETER_VALUE` after a round trip, and quietly forcing it to
+* `false` would be the wrong kind of quiet for a data-retention flag. A
+* type-level constraint is not available — `providerOptions` is
+* `Record<string, JSONObject>`, so a provider cannot narrow it at the call
+* site.
+*
+* `forceReasoning: true` covers a gap in the shared model's own detection,
+* which matches a bare `gpt-5` prefix. The gateway also accepts the legacy
+* `databricks-` form, and `databricks-gpt-5-2` fails that match while still
+* being a reasoning model.
+*/
+const STATELESS_HELP = "See https://github.com/neondatabase/neon-pkgs/tree/main/packages/ai-sdk-provider#errors";
+/** Models the gateway serves as reasoning models on this route. */
+function isReasoningModelId(modelId) {
+	return /gpt-5/.test(modelId.toLowerCase());
+}
+/**
+* Apply the route's defaults to one call, or throw when a caller asked for
+* something the gateway cannot do. Returns the options unchanged when there is
+* nothing to add.
+*/
+function withResponsesGatewayDefaults(modelId, options) {
+	const openai = options.providerOptions?.openai;
+	const defaults = {};
+	if (openai?.store === void 0) defaults.store = false;
+	else if (openai.store === true || openai.store === null) throw new UnsupportedFunctionalityError({
+		functionality: "storing responses (providerOptions.openai.store)",
+		message: `\`store: ${String(openai.store)}\` is not available on the Neon AI Gateway, which serves the Responses API statelessly and never persists response items. Remove \`providerOptions.openai.store\`, or set it to \`false\`.` + (openai.store === null ? " `null` is not the same as omitting it — the AI SDK reads a null `store` as `true`." : "") + ` ${STATELESS_HELP}`
+	});
+	for (const option of ["previousResponseId", "conversation"]) if (openai?.[option] != null) throw new UnsupportedFunctionalityError({
+		functionality: `server-side conversation state (providerOptions.openai.${option})`,
+		message: `\`${option}\` is not available on the Neon AI Gateway, which serves the Responses API statelessly and never persists response items, so there is nothing for \`${option}\` to refer to. Send the full message history instead. ${STATELESS_HELP}`
+	});
+	if (isReasoningModelId(modelId) && openai?.forceReasoning === void 0) defaults.forceReasoning = true;
+	if (Object.keys(defaults).length === 0) return options;
+	return {
+		...options,
+		providerOptions: {
+			...options.providerOptions,
+			openai: {
+				...openai,
+				...defaults
+			}
+		}
+	};
+}
+//#endregion
+export { isReasoningModelId, withResponsesGatewayDefaults };
+
+//# sourceMappingURL=neon-responses-options.js.map

package/dist/lib/neon-responses-options.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"neon-responses-options.js","names":[],"sources":["../../src/lib/neon-responses-options.ts"],"sourcesContent":["import {\n\ttype LanguageModelV3CallOptions,\n\tUnsupportedFunctionalityError,\n} from \"@ai-sdk/provider\";\n\n/**\n * Request defaults for the Neon AI Gateway's Responses route.\n *\n * `store: false` is the one value this gateway accepts, not a preference. It\n * serves the Responses API statelessly and keeps no items — a response comes\n * back with `\"store\": false` even when the request never set it. Left alone the\n * shared OpenAI model assumes stored-item semantics and replays earlier\n * reasoning as `{ type: \"item_reference\", id }`, which nothing upstream can\n * resolve; the gateway answers `502 INTERNAL_ERROR` on the second step of a\n * tool loop, after the first has already succeeded. Sending `false` makes the\n * model inline the encrypted reasoning instead.\n *\n * Since no other value can work, an explicit `store` that is not `false` is\n * refused rather than sent: the request would come back `400\n * INVALID_PARAMETER_VALUE` after a round trip, and quietly forcing it to\n * `false` would be the wrong kind of quiet for a data-retention flag. A\n * type-level constraint is not available — `providerOptions` is\n * `Record<string, JSONObject>`, so a provider cannot narrow it at the call\n * site.\n *\n * `forceReasoning: true` covers a gap in the shared model's own detection,\n * which matches a bare `gpt-5` prefix. The gateway also accepts the legacy\n * `databricks-` form, and `databricks-gpt-5-2` fails that match while still\n * being a reasoning model.\n */\n\nconst STATELESS_HELP =\n\t\"See https://github.com/neondatabase/neon-pkgs/tree/main/packages/ai-sdk-provider#errors\";\n\n/** Models the gateway serves as reasoning models on this route. */\nexport function isReasoningModelId(modelId: string): boolean {\n\treturn /gpt-5/.test(modelId.toLowerCase());\n}\n\n/**\n * Apply the route's defaults to one call, or throw when a caller asked for\n * something the gateway cannot do. Returns the options unchanged when there is\n * nothing to add.\n */\nexport function withResponsesGatewayDefaults(\n\tmodelId: string,\n\toptions: LanguageModelV3CallOptions,\n): LanguageModelV3CallOptions {\n\tconst openai = options.providerOptions?.openai;\n\tconst defaults: Record<string, boolean> = {};\n\n\t// `undefined` counts as unset rather than as a choice: it serializes away,\n\t// so `{ store: undefined }` and `{}` are the same request on the wire.\n\tif (openai?.store === undefined) {\n\t\tdefaults.store = false;\n\t} else if (openai.store === true || openai.store === null) {\n\t\t// `null` reads as \"omitted\" but is not: the shared model resolves it\n\t\t// with `store ?? true`, so it asks for storage just as `true` does.\n\t\tthrow new UnsupportedFunctionalityError({\n\t\t\tfunctionality: \"storing responses (providerOptions.openai.store)\",\n\t\t\tmessage:\n\t\t\t\t`\\`store: ${String(openai.store)}\\` is not available on the Neon AI ` +\n\t\t\t\t\"Gateway, which serves the Responses API statelessly and never \" +\n\t\t\t\t\"persists response items. Remove `providerOptions.openai.store`, or \" +\n\t\t\t\t\"set it to `false`.\" +\n\t\t\t\t(openai.store === null\n\t\t\t\t\t? \" `null` is not the same as omitting it — the AI SDK reads a null\" +\n\t\t\t\t\t\t\" `store` as `true`.\"\n\t\t\t\t\t: \"\") +\n\t\t\t\t` ${STATELESS_HELP}`,\n\t\t});\n\t}\n\n\t// Any other type is left for the shared model's own provider-option schema\n\t// (`store: z.boolean().nullish()`), which rejects it locally with a type\n\t// error. Only `true` and `null` pass that schema and are refused by the\n\t// gateway, so only they need saying something about here.\n\n\t// The same statelessness rules these out, and they belong to the same set:\n\t// refusing one and forwarding the others would read as an oversight.\n\tfor (const option of [\"previousResponseId\", \"conversation\"] as const) {\n\t\tif (openai?.[option] != null) {\n\t\t\tthrow new UnsupportedFunctionalityError({\n\t\t\t\tfunctionality: `server-side conversation state (providerOptions.openai.${option})`,\n\t\t\t\tmessage:\n\t\t\t\t\t`\\`${option}\\` is not available on the Neon AI Gateway, which ` +\n\t\t\t\t\t\"serves the Responses API statelessly and never persists response \" +\n\t\t\t\t\t`items, so there is nothing for \\`${option}\\` to refer to. Send ` +\n\t\t\t\t\t`the full message history instead. ${STATELESS_HELP}`,\n\t\t\t});\n\t\t}\n\t}\n\n\tif (isReasoningModelId(modelId) && openai?.forceReasoning === undefined) {\n\t\tdefaults.forceReasoning = true;\n\t}\n\n\tif (Object.keys(defaults).length === 0) {\n\t\treturn options;\n\t}\n\treturn {\n\t\t...options,\n\t\tproviderOptions: {\n\t\t\t...options.providerOptions,\n\t\t\topenai: { ...openai, ...defaults },\n\t\t},\n\t};\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AA+BA,MAAM,iBACL;;AAGD,SAAgB,mBAAmB,SAA0B;CAC5D,OAAO,QAAQ,KAAK,QAAQ,YAAY,CAAC;AAC1C;;;;;;AAOA,SAAgB,6BACf,SACA,SAC6B;CAC7B,MAAM,SAAS,QAAQ,iBAAiB;CACxC,MAAM,WAAoC,CAAC;CAI3C,IAAI,QAAQ,UAAU,KAAA,GACrB,SAAS,QAAQ;MACX,IAAI,OAAO,UAAU,QAAQ,OAAO,UAAU,MAGpD,MAAM,IAAI,8BAA8B;EACvC,eAAe;EACf,SACC,YAAY,OAAO,OAAO,KAAK,EAAE,+LAIhC,OAAO,UAAU,OACf,wFAEA,MACH,IAAI;CACN,CAAC;CAUF,KAAK,MAAM,UAAU,CAAC,sBAAsB,cAAc,GACzD,IAAI,SAAS,WAAW,MACvB,MAAM,IAAI,8BAA8B;EACvC,eAAe,0DAA0D,OAAO;EAChF,SACC,KAAK,OAAO,sJAEwB,OAAO,yDACN;CACvC,CAAC;CAIH,IAAI,mBAAmB,OAAO,KAAK,QAAQ,mBAAmB,KAAA,GAC7D,SAAS,iBAAiB;CAG3B,IAAI,OAAO,KAAK,QAAQ,CAAC,CAAC,WAAW,GACpC,OAAO;CAER,OAAO;EACN,GAAG;EACH,iBAAiB;GAChB,GAAG,QAAQ;GACX,QAAQ;IAAE,GAAG;IAAQ,GAAG;GAAS;EAClC;CACD;AACD"}

package/dist/lib/provider.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"provider.d.ts","names":[],"sources":["../../src/lib/provider.ts"],"mappings":";;;;;;;cAoCM,iBAAe,CAAA,CAAA;;IAAf,OAAA,aAOJ;IAAA,IAAA,eAAA,cAAA,YAAA,CAAA,CAAA;;;;GAAA,CAAA,CAAA,IAAA,CAAA;KAEU,aAAA,GAAgB,CAAA,CAAE,aAAa;UA4C1B,oBAAA;;;;;;;EA9Cf;AAPmB;AAAA;AASrB;EAAyB,MAAA,CAAA,EAAA,MAAA;EAAkB;EAAf,OAAE,CAAA,EA2DnB,MA3DmB,CAAA,MAAA,EAAA,MAAA,CAAA;EAAK;EA4ClB,KAAA,CAAA,EAkBR,aAlBQ;AAAoB;AAe1B,UAMM,YAAA,SAAqB,UAN3B,CAAA;EAGF;EAAa,CAAA,OAAA,EAKX,eALW,CAAA,EAKO,eALP;EAGL;EAAa,aAAA,CAAA,OAAA,EAKN,eALM,CAAA,EAKY,eALZ;EAEnB;EAAkB,IAAA,CAAA,OAAA,EAMd,eANc,CAAA,EAMI,eANJ;EAGL;EAAkB,KAAA,EAAA,OAM3B,eAN2B;EAG3B;EAAkB,kBAAA,CAAA,OAAA,EAAA,MAAA,CAAA,EAAA,KAAA;AAGlB;AAXuB,iBAiBtB,UAAA,CAjBsB,OAAA,CAAA,EAiBF,oBAjBE,CAAA,EAiB0B,YAjB1B;AAAU;AAiBhC,cA+FH,IA/Fa,EA+FT,YA/FS"}
+{"version":3,"file":"provider.d.ts","names":[],"sources":["../../src/lib/provider.ts"],"mappings":";;;;;;;cAqCM,iBAAe,CAAA,CAAA;;IAAf,OAAA,aAOJ;IAAA,IAAA,eAAA,cAAA,YAAA,CAAA,CAAA;;;;GAAA,CAAA,CAAA,IAAA,CAAA;KAEU,aAAA,GAAgB,CAAA,CAAE,aAAa;UA4C1B,oBAAA;;;;;;;EA9Cf;AAPmB;AAAA;AASrB;EAAyB,MAAA,CAAA,EAAA,MAAA;EAAkB;EAAf,OAAE,CAAA,EA2DnB,MA3DmB,CAAA,MAAA,EAAA,MAAA,CAAA;EAAK;EA4ClB,KAAA,CAAA,EAkBR,aAlBQ;AAAoB;AAe1B,UAMM,YAAA,SAAqB,UAN3B,CAAA;EAGF;EAAa,CAAA,OAAA,EAKX,eALW,CAAA,EAKO,eALP;EAGL;EAAa,aAAA,CAAA,OAAA,EAKN,eALM,CAAA,EAKY,eALZ;EAEnB;EAAkB,IAAA,CAAA,OAAA,EAMd,eANc,CAAA,EAMI,eANJ;EAGL;EAAkB,KAAA,EAAA,OAM3B,eAN2B;EAG3B;EAAkB,kBAAA,CAAA,OAAA,EAAA,MAAA,CAAA,EAAA,KAAA;AAGlB;AAXuB,iBAiBtB,UAAA,CAjBsB,OAAA,CAAA,EAiBF,oBAjBE,CAAA,EAiB0B,YAjB1B;AAAU;AAiBhC,cAiHH,IAjHa,EAiHT,YAjHS"}

package/dist/lib/provider.js

@@ -1,7 +1,8 @@
-import { NeonAnthropicLanguageModel } from "./neon-anthropic-language-model.js";
 import { getNeonModelRoute } from "./neon-model-capabilities.js";
+import { NeonAnthropicLanguageModel } from "./neon-anthropic-language-model.js";
 import { NeonChatLanguageModel } from "./neon-chat-language-model.js";
 import { NeonResponsesLanguageModel } from "./neon-responses-language-model.js";
+import { wrapFetchWithGatewayErrorNormalization } from "./neon-gateway-error.js";
 import { wrapFetchWithHarmonyNormalization } from "./neon-harmony-normalize.js";
 import { neonOpenAITools } from "./neon-openai-tools.js";
 import { VERSION } from "./version.js";
@@ -61,21 +62,21 @@ function createNeon(options = {}) {
 		provider: "neon.anthropic",
 		baseURL: `${getHost()}/anthropic/v1`,
 		headers: () => getHeaders({ "anthropic-version": "2023-06-01" }),
-		fetch: options.fetch,
+		fetch: wrapFetchWithGatewayErrorNormalization(options.fetch, "anthropic"),
 		generateId
 	});
 	const createOpenAIModel = (modelId) => new NeonResponsesLanguageModel(modelId, {
 		provider: "neon.openai.responses",
 		url: ({ path }) => `${getHost()}/openai/v1${path}`,
 		headers: getHeaders,
-		fetch: options.fetch,
+		fetch: wrapFetchWithGatewayErrorNormalization(options.fetch, "openai"),
 		fileIdPrefixes: ["file-"]
 	});
 	const createChatModel = (modelId) => new NeonChatLanguageModel(modelId, {
 		provider: "neon.chat",
 		url: ({ path }) => `${getHost()}/v1${path}`,
 		headers: getHeaders,
-		fetch: wrapFetchWithHarmonyNormalization(options.fetch),
+		fetch: wrapFetchWithHarmonyNormalization(wrapFetchWithGatewayErrorNormalization(options.fetch, "openai")),
 		errorStructure: neonErrorStructure,
 		transformRequestBody: transformNeonRequestBody,
 		supportsStructuredOutputs: true