@jambonz/llm 0.4.0 → 0.4.1

package/dist/index.js

@@ -81,7 +81,10 @@ async function* streamFromOpenAI(client, req, options) {
   }
   if (req.temperature !== void 0) body.temperature = req.temperature;
   const effectiveMaxTokens = req.maxTokens ?? options.defaultMaxTokens;
-  if (effectiveMaxTokens !== void 0) body.max_tokens = effectiveMaxTokens;
+  if (effectiveMaxTokens !== void 0) {
+    const param = options.tokensParam ?? (needsMaxCompletionTokens(req.model) ? "max_completion_tokens" : "max_tokens");
+    body[param] = effectiveMaxTokens;
+  }
   let stream;
   try {
     stream = await client.chat.completions.create(body, { signal: req.signal });
@@ -264,6 +267,9 @@ function capabilitiesFor(modelId, knownModels) {
   if (known) return known.capabilities;
   return defaultCapabilitiesForUnknown(modelId);
 }
+function needsMaxCompletionTokens(modelId) {
+  return /^(o\d|gpt-5)/i.test(modelId);
+}
 function defaultCapabilitiesForUnknown(modelId) {
   const isReasoning = /^o[13]([-/]|$)/i.test(modelId);
   if (isReasoning) {
@@ -1641,7 +1647,12 @@ var VertexOpenAIAdapter = class {
   stream(req) {
     return streamFromOpenAI(this.ensureClient(), req, {
       knownModels: vertexOpenAIManifest.knownModels,
-      defaultMaxTokens: 4096
+      defaultMaxTokens: 4096,
+      // Pin legacy `max_tokens`: Vertex's OpenAI-compat partner-model
+      // endpoint predates `max_completion_tokens` and may not recognize it.
+      // The model-id heuristic wouldn't catch a partner model named in the
+      // OpenAI style (unlikely today, but defensive).
+      tokensParam: "max_tokens"
     });
   }
   appendAssistantToolCall(history, toolCalls) {
@@ -2389,7 +2400,8 @@ var AzureOpenAIAdapter = class {
   }
   stream(req) {
     return streamFromOpenAI(this.ensureClient(), req, {
-      knownModels: azureOpenAIManifest.knownModels
+      knownModels: azureOpenAIManifest.knownModels,
+      tokensParam: "max_completion_tokens"
     });
   }
   appendAssistantToolCall(history, toolCalls) {
@@ -2408,7 +2420,12 @@ var AzureOpenAIAdapter = class {
       // still requires it as a parameter.
       model: "probe",
       messages: [{ role: "user", content: "ping" }],
-      max_tokens: 1,
+      // Use `max_completion_tokens` (not legacy `max_tokens`): gpt-5 family
+      // and o-series reasoning models return 400 on the legacy parameter.
+      // The newer name is a superset — also accepted by gpt-4o, gpt-4-turbo,
+      // and gpt-3.5-turbo deployments — so it's the safe choice for a probe
+      // that has to work regardless of the underlying deployment model.
+      max_completion_tokens: 1,
       stream: false
     });
   }