@fgv/ts-extras 5.1.0-44 → 5.1.0-45

package/dist/packlets/ai-assist/apiClient.js

@@ -28,7 +28,7 @@
  */
 import { isJsonObject } from '@fgv/ts-json-base';
 import { fail, mapResults, succeed, Validators } from '@fgv/ts-utils';
-import { allModelCapabilities, isResponsesOnlyModel, resolveProviderModel } from './model';
+import { allModelCapabilities, DEFAULT_ANTHROPIC_MAX_TOKENS, isAdaptiveThinkingModel, isResponsesOnlyModel, resolveProviderModel, usesMaxCompletionTokensField } from './model';
 import { anthropicEffortToBudgetTokens, checkTemperatureConflict, mergeThinkingConfig, providerDiscriminatorForId } from './thinkingOptionsResolver';
 import { buildAnthropicMessages, buildGeminiContents, buildMessages, buildOpenAiChatUserContent, buildOpenAiResponsesUserContent, normalizeOutboundMessages, splitChatRequest } from './chatRequestBuilders';
 import { bearerAuthHeader, resolveEffectiveBaseUrl } from './endpoint';
@@ -221,14 +221,15 @@ const geminiResponse = Validators.object({
  * Works for xAI Grok, OpenAI, Groq, and Mistral.
  * @internal
  */
-async function callOpenAiCompletion(config, prompt, head, temperature, logger, signal, resolvedThinking) {
+async function callOpenAiCompletion(config, prompt, head, temperature, logger, signal, resolvedThinking, maxTokens, useMaxCompletionTokensField = false) {
     var _a;
     const url = `${config.baseUrl}/chat/completions`;
     const messages = buildMessages(prompt.system, buildOpenAiChatUserContent(prompt), {
         head
     });
     const effort = (_a = resolvedThinking === null || resolvedThinking === void 0 ? void 0 : resolvedThinking.openAiEffort) !== null && _a !== void 0 ? _a : resolvedThinking === null || resolvedThinking === void 0 ? void 0 : resolvedThinking.xaiEffort;
-    const body = Object.assign(Object.assign({ model: config.model, messages }, (temperature !== undefined ? { temperature } : {})), (effort !== undefined && config.model !== 'grok-4' ? { reasoning_effort: effort } : {}));
+    const maxTokensField = useMaxCompletionTokensField ? 'max_completion_tokens' : 'max_tokens';
+    const body = Object.assign(Object.assign(Object.assign({ model: config.model, messages }, (temperature !== undefined ? { temperature } : {})), (effort !== undefined && config.model !== 'grok-4' ? { reasoning_effort: effort } : {})), (maxTokens !== undefined ? { [maxTokensField]: maxTokens } : {}));
     if ((resolvedThinking === null || resolvedThinking === void 0 ? void 0 : resolvedThinking.otherParams) !== undefined) {
         Object.assign(body, resolvedThinking.otherParams);
     }
@@ -274,7 +275,7 @@ function extractResponsesApiText(output) {
  * Used when tools are configured for an openai-format provider.
  * @internal
  */
-async function callOpenAiResponsesCompletion(config, prompt, tools = [], head, temperature, logger, signal, resolvedThinking) {
+async function callOpenAiResponsesCompletion(config, prompt, tools = [], head, temperature, logger, signal, resolvedThinking, maxTokens) {
     var _a;
     const url = `${config.baseUrl}/responses`;
     const input = buildMessages(prompt.system, buildOpenAiResponsesUserContent(prompt), {
@@ -282,6 +283,10 @@ async function callOpenAiResponsesCompletion(config, prompt, tools = [], head, t
     });
     const effort = (_a = resolvedThinking === null || resolvedThinking === void 0 ? void 0 : resolvedThinking.openAiEffort) !== null && _a !== void 0 ? _a : resolvedThinking === null || resolvedThinking === void 0 ? void 0 : resolvedThinking.xaiEffort;
     const body = Object.assign(Object.assign(Object.assign({ model: config.model, input }, (tools.length > 0 ? { tools: toResponsesApiTools(tools) } : {})), (temperature !== undefined ? { temperature } : {})), (effort !== undefined && config.model !== 'grok-4' ? { reasoning: { effort } } : {}));
+    // Shared by OpenAI and xAI — both route through the Responses API with the same field name.
+    if (maxTokens !== undefined) {
+        body.max_output_tokens = maxTokens;
+    }
     if ((resolvedThinking === null || resolvedThinking === void 0 ? void 0 : resolvedThinking.otherParams) !== undefined) {
         Object.assign(body, resolvedThinking.otherParams);
     }
@@ -328,13 +333,24 @@ function extractAnthropicText(content) {
     return succeed(textParts.join(''));
 }
 /** Calls the Anthropic Messages API with optional tool support. @internal */
-async function callAnthropicCompletion(config, prompt, head, temperature, logger, tools, signal, resolvedThinking) {
+async function callAnthropicCompletion(config, prompt, head, temperature, logger, tools, signal, resolvedThinking, useAdaptiveThinking = false, maxTokens) {
     const url = `${config.baseUrl}/messages`;
     const messages = buildAnthropicMessages(prompt, { head });
-    const body = Object.assign({ model: config.model, system: prompt.system, messages, max_tokens: 4096 }, (temperature !== undefined ? { temperature } : {}));
+    const body = Object.assign({ model: config.model, system: prompt.system, messages, 
+        // Anthropic's Messages API requires max_tokens on every request — see
+        // AiAssist.DEFAULT_ANTHROPIC_MAX_TOKENS for why only this provider defaults it.
+        max_tokens: maxTokens !== null && maxTokens !== void 0 ? maxTokens : DEFAULT_ANTHROPIC_MAX_TOKENS }, (temperature !== undefined ? { temperature } : {}));
     const effort = resolvedThinking === null || resolvedThinking === void 0 ? void 0 : resolvedThinking.anthropicEffort;
     if (effort !== undefined) {
-        body.thinking = { type: 'enabled', budget_tokens: anthropicEffortToBudgetTokens(effort) };
+        if (useAdaptiveThinking) {
+            // Claude 5 family: adaptive thinking — no budget_tokens; effort moves to the
+            // top-level output_config block. See AiAssist.isAdaptiveThinkingModel.
+            body.thinking = { type: 'adaptive' };
+            body.output_config = { effort };
+        }
+        else {
+            body.thinking = { type: 'enabled', budget_tokens: anthropicEffortToBudgetTokens(effort) };
+        }
     }
     if ((resolvedThinking === null || resolvedThinking === void 0 ? void 0 : resolvedThinking.otherParams) !== undefined) {
         Object.assign(body, resolvedThinking.otherParams);
@@ -378,7 +394,7 @@ async function callAnthropicCompletion(config, prompt, head, temperature, logger
  * When tools are configured, includes Google Search grounding.
  * @internal
  */
-async function callGeminiCompletion(config, prompt, head, temperature, logger, tools, signal, resolvedThinking) {
+async function callGeminiCompletion(config, prompt, head, temperature, logger, tools, signal, resolvedThinking, maxTokens) {
     const url = `${config.baseUrl}/models/${config.model}:generateContent`;
     const contents = buildGeminiContents(prompt, { head });
     // Temperature is sent only when explicitly provided; otherwise Gemini's default applies.
@@ -386,6 +402,9 @@ async function callGeminiCompletion(config, prompt, head, temperature, logger, t
     if (temperature !== undefined) {
         generationConfig.temperature = temperature;
     }
+    if (maxTokens !== undefined) {
+        generationConfig.maxOutputTokens = maxTokens;
+    }
     if ((resolvedThinking === null || resolvedThinking === void 0 ? void 0 : resolvedThinking.geminiThinkingBudget) !== undefined) {
         generationConfig.thinkingConfig = { thinkingBudget: resolvedThinking.geminiThinkingBudget };
     }
@@ -434,7 +453,7 @@ async function callGeminiCompletion(config, prompt, head, temperature, logger, t
  * @public
  */
 export async function callProviderCompletion(params) {
-    const { descriptor, apiKey, system, messages, temperature, modelOverride, tier, logger, tools, signal, endpoint, thinking } = params;
+    const { descriptor, apiKey, system, messages, temperature, modelOverride, tier, logger, tools, signal, endpoint, thinking, maxTokens } = params;
     const splitResult = splitChatRequest(system, messages);
     if (splitResult.isFailure()) {
         return fail(splitResult.message);
@@ -489,13 +508,13 @@ export async function callProviderCompletion(params) {
             // Responses-API-only models (e.g. gpt-5.5-pro) 400 on /chat/completions, so they route
             // to the Responses path even with no tools requested — same path the tools case uses.
             if (hasTools || isResponsesOnlyModel(descriptor, config.model)) {
-                return callOpenAiResponsesCompletion(config, prompt, tools, head, temperature, logger, signal, resolvedThinking);
+                return callOpenAiResponsesCompletion(config, prompt, tools, head, temperature, logger, signal, resolvedThinking, maxTokens);
             }
-            return callOpenAiCompletion(config, prompt, head, temperature, logger, signal, resolvedThinking);
+            return callOpenAiCompletion(config, prompt, head, temperature, logger, signal, resolvedThinking, maxTokens, usesMaxCompletionTokensField(descriptor));
         case 'anthropic':
-            return callAnthropicCompletion(config, prompt, head, temperature, logger, tools, signal, resolvedThinking);
+            return callAnthropicCompletion(config, prompt, head, temperature, logger, tools, signal, resolvedThinking, isAdaptiveThinkingModel(descriptor, config.model), maxTokens);
         case 'gemini':
-            return callGeminiCompletion(config, prompt, head, temperature, logger, tools, signal, resolvedThinking);
+            return callGeminiCompletion(config, prompt, head, temperature, logger, tools, signal, resolvedThinking, maxTokens);
         /* c8 ignore next 4 - defensive coding: exhaustive switch guaranteed by TypeScript */
         default: {
             const _exhaustive = descriptor.apiFormat;
@@ -1120,7 +1139,7 @@ export async function callProxiedListModels(proxyUrl, params) {
  * @public
  */
 export async function callProxiedCompletion(proxyUrl, params) {
-    const { descriptor, apiKey, system, messages, temperature, modelOverride, logger, tools, signal, thinking } = params;
+    const { descriptor, apiKey, system, messages, temperature, modelOverride, logger, tools, signal, thinking, maxTokens } = params;
     const splitResult = splitChatRequest(system, messages);
     if (splitResult.isFailure()) {
         return fail(splitResult.message);
@@ -1150,6 +1169,11 @@ export async function callProxiedCompletion(proxyUrl, params) {
     if (thinking !== undefined) {
         body.thinking = thinking;
     }
+    // Forwarded only when explicitly provided; the proxy is responsible for mapping it to the
+    // correct upstream provider field (see AiAssist.usesMaxCompletionTokensField).
+    if (maxTokens !== undefined) {
+        body.maxTokens = maxTokens;
+    }
     /* c8 ignore next 1 - optional logger */
     logger === null || logger === void 0 ? void 0 : logger.info(`AI proxy request: provider=${descriptor.id}, proxy=${proxyUrl}`);
     const url = `${proxyUrl}/api/ai/completion`;