@adaptic/utils 0.0.1017 → 0.0.1019

package/dist/index.mjs

@@ -64280,28 +64280,28 @@ var aliases = {
 			{
 				role: "secondary",
 				provider: "deepinfra",
-				model_id: "zai-org/GLM-5.3-Flash",
+				model_id: "openai/gpt-oss-20b",
 				model_id_status: "confirmed",
 				model_id_source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account",
-				model_family: "GLM-5.3 Flash",
-				lumic_model: "zai-org/GLM-5.3-Flash",
+				model_family: "gpt-oss-20B",
+				lumic_model: "openai/gpt-oss-20b",
 				params: {
 					temperature: null,
 					max_output_tokens: null,
 					supports_temperature: true,
 					supports_json_schema: true,
 					supports_tools: true,
-					supports_cache_control: true,
-					supports_vision: true,
-					context_window: 1048576
+					supports_cache_control: false,
+					supports_vision: false,
+					context_window: 131072
 				},
 				price_per_mtok: {
-					input: 0.15,
-					output: 0.5,
+					input: 0.03,
+					output: 0.14,
 					as_of: "2026-09-11",
 					source: "DeepInfra /v1/openai/models catalogue queried against the live account 2026-09-11; tool-calling and latency verified by direct call on the same account"
 				},
-				notes: "A different lineage from the primary, with a 1M context and JSON mode. Chosen over the faster Nemotron 3 Nano, which measured the quickest median of any candidate but rejects response_format outright with HTTP 405 — and llm.fast is a structured-output alias, so a leg that cannot express the contract is not a fallback at all, only a slower way to fail."
+				notes: "A different lineage from the primary, so one family's regression cannot take both serving legs, and fast enough to be a real fallback: median 5.5s against this alias's 30s hot-path budget on a production-scale prompt. It replaces GLM-5.3-Flash, which was chosen for the same lineage reason on a TRIVIAL prompt and measured 13-23s once put against production-sized input — close enough to the budget that the fallback was timing out and returning 408 instead of covering the primary. A fallback with no latency headroom is not a fallback; it is a slower way to fail, which is the exact test this alias's previous secondary was chosen to satisfy and then failed itself."
 			},
 			{
 				role: "closed_incumbent",
@@ -65293,7 +65293,16 @@ function gatewayFor(chain) {
         gatewayTransport = createGatewayTransport({
             baseUrl,
             apiKeyEnv: config.gatewayApiKeyEnv ?? DEFAULT_GATEWAY_KEY_ENV,
-            modelNameFor: (request) => gatewayModelNameFor(request.route, chain),
+            // The chain is re-resolved from the request's own route rather than
+            // captured from the caller above. The transport is cached for the life of
+            // the process, so a captured chain would bind every later alias to
+            // whichever alias happened to dispatch first: its head leg would never
+            // match, would be addressed as "<alias>.fallback.primary" — a name the
+            // gateway does not register — and the 400 would be absorbed by the
+            // fallback chain. The call still succeeds, from the SECONDARY leg, which
+            // is why this reads as healthy traffic while every alias but one quietly
+            // stops using the model it was chosen for.
+            modelNameFor: (request) => gatewayModelNameFor(request.route, resolveChain(request.route.alias, { isolated: request.route.isolated })),
         });
     }
     return gatewayTransport;
@@ -65381,7 +65390,7 @@ async function callLLMByAlias(content, responseFormat = "text", options) {
             cost: 0,
         });
     }
-    const gateway = gatewayFor(chain);
+    const gateway = gatewayFor();
     const attemptLog = [];
     /**
      * Run the chain, falling back from the gateway to the degraded direct path