@jaypie/llm 1.3.5 → 1.3.6

package/dist/cjs/index.cjs

@@ -101,7 +101,10 @@ const PROVIDER = {
     ANTHROPIC: {
         // https://docs.anthropic.com/en/docs/about-claude/models/overview
         MAX_TOKENS: {
-            DEFAULT: 4096,
+            // Non-streaming ceiling: responses above ~16K output tokens risk HTTP
+            // timeouts; streaming requests resolve to the model maximum instead
+            // (see util/maxOutputTokens.ts)
+            DEFAULT: 16384,
         },
         MODEL: {
             DEFAULT: FIRST_CLASS_PROVIDER.ANTHROPIC.DEFAULT,
@@ -450,14 +453,12 @@ class BaseProviderAdapter {
     /**
      * Default implementation returns false - override for providers with native structured output
      */
-    // eslint-disable-next-line @typescript-eslint/no-unused-vars
     hasStructuredOutput(_response) {
         return false;
     }
     /**
      * Default implementation returns undefined - override for providers with native structured output
      */
-    // eslint-disable-next-line @typescript-eslint/no-unused-vars
     extractStructuredOutput(_response) {
         return undefined;
     }
@@ -925,6 +926,52 @@ function jsonSchemaToOpenApi3(schema) {
 
 const getLogger$6 = () => log$1.log.lib({ lib: kit.JAYPIE.LIB.LLM });
 
+//
+//
+// Constants
+//
+// Non-streaming requests above ~16K output tokens risk HTTP timeouts
+// (Anthropic guidance is to stream anything larger), so non-streaming
+// defaults are capped here and only streaming requests resolve to the
+// full model maximum.
+const NON_STREAMING_MAX_TOKENS = PROVIDER.ANTHROPIC.MAX_TOKENS.DEFAULT;
+// Maximum output tokens by model; first match wins.
+// Providers without low output ceilings (OpenAI, xAI) are intentionally
+// absent — their requests leave the limit unset.
+const MODEL_MAX_OUTPUT_TOKENS = [
+    // Anthropic — https://platform.claude.com/docs/en/about-claude/models/overview
+    { pattern: /^claude-opus-4-(0$|1$|1-|2025)/, tokens: 32000 },
+    { pattern: /^claude-opus-4-5/, tokens: 64000 },
+    { pattern: /^claude-sonnet-4-(0$|5$|5-|2025)/, tokens: 64000 },
+    { pattern: /haiku/, tokens: 64000 },
+    { pattern: /claude|fable|mythos|opus|sonnet/, tokens: 128000 },
+    // Google — https://ai.google.dev/gemini-api/docs/models
+    { pattern: /gemini-(2\.5|3)/, tokens: 65536 },
+    { pattern: /gemini/, tokens: 8192 },
+];
+//
+//
+// Main
+//
+/**
+ * Maximum output tokens the model supports, or undefined when unknown.
+ */
+function maxOutputTokens(model) {
+    const match = MODEL_MAX_OUTPUT_TOKENS.find(({ pattern }) => pattern.test(model));
+    return match?.tokens;
+}
+/**
+ * Default output token limit for a request: the model maximum when
+ * streaming, capped at the non-streaming maximum otherwise. Returns
+ * undefined when the model's maximum is unknown.
+ */
+function resolveMaxOutputTokens(model, { stream = false } = {}) {
+    const max = maxOutputTokens(model);
+    if (max === undefined)
+        return undefined;
+    return stream ? max : Math.min(max, NON_STREAMING_MAX_TOKENS);
+}
+
 // Turn policy constants
 const MAX_TURNS_ABSOLUTE_LIMIT = 72;
 const MAX_TURNS_DEFAULT_LIMIT = 24;
@@ -1749,11 +1796,13 @@ class AnthropicAdapter extends BaseProviderAdapter {
                 lastMsg.content = lastMsg.content + "\n\n" + request.instructions;
             }
         }
+        const model = (request.model ||
+            this.defaultModel);
         const anthropicRequest = {
-            model: (request.model ||
-                this.defaultModel),
+            model,
             messages,
-            max_tokens: PROVIDER.ANTHROPIC.MAX_TOKENS.DEFAULT,
+            max_tokens: resolveMaxOutputTokens(model, { stream: request.stream }) ??
+                PROVIDER.ANTHROPIC.MAX_TOKENS.DEFAULT,
             stream: false,
         };
         if (request.system) {
@@ -1816,7 +1865,6 @@ class AnthropicAdapter extends BaseProviderAdapter {
     // outputSchema is part of the interface contract but Anthropic now uses
     // native `output_format` (set in buildRequest), so we no longer inject a
     // synthetic structured-output tool here.
-    // eslint-disable-next-line @typescript-eslint/no-unused-vars
     _outputSchema) {
         return toolkit.tools.map((tool) => ({
             name: tool.name,
@@ -3003,6 +3051,18 @@ class GoogleAdapter extends BaseProviderAdapter {
                     : structuredOutputInstruction,
             };
         }
+        // Default output ceiling: Google's own default (8,192) silently
+        // truncates long generations, so resolve to the model maximum (capped
+        // for non-streaming); providerOptions below override
+        const maxOutputTokens = resolveMaxOutputTokens(geminiRequest.model, {
+            stream: request.stream,
+        });
+        if (maxOutputTokens !== undefined) {
+            geminiRequest.config = {
+                ...geminiRequest.config,
+                maxOutputTokens,
+            };
+        }
         // Add provider-specific options
         if (request.providerOptions) {
             geminiRequest.config = {
@@ -3024,7 +3084,6 @@ class GoogleAdapter extends BaseProviderAdapter {
     // structured output via `responseJsonSchema`/`responseSchema` (or the
     // legacy fake-tool injected in buildRequest as a fallback). We no longer
     // inject a synthetic structured-output tool here.
-    // eslint-disable-next-line @typescript-eslint/no-unused-vars
     _outputSchema) {
         return toolkit.tools.map((tool) => ({
             name: tool.name,
@@ -4689,7 +4748,6 @@ class OpenRouterAdapter extends BaseProviderAdapter {
     // native `response_format` (set in buildRequest), so we no longer inject a
     // synthetic structured-output tool here. The legacy fake-tool injection
     // happens in buildRequest only as a runtime fallback.
-    // eslint-disable-next-line @typescript-eslint/no-unused-vars
     _outputSchema) {
         return toolkit.tools.map((tool) => ({
             name: tool.name,
@@ -7021,6 +7079,7 @@ class StreamLoop {
                     messages: state.currentInput,
                     model: options.model ?? this.adapter.defaultModel,
                     providerOptions: options.providerOptions,
+                    stream: true,
                     system: options.system,
                     temperature: options.temperature,
                     tools: state.formattedTools,
@@ -7089,6 +7148,7 @@ class StreamLoop {
             messages: state.currentInput,
             model: options.model ?? this.adapter.defaultModel,
             providerOptions: options.providerOptions,
+            stream: true,
             system: options.system,
             temperature: options.temperature,
             tools: state.formattedTools,