@code-yeongyu/senpi-ai 2026.8.12 → 2026.8.13

package/dist/api/openai-completions.js

@@ -15,7 +15,7 @@ import { appendGrammarToolInputJsonDelta, createGrammarToolInputProperties, getG
 import { buildCopilotDynamicHeaders, hasCopilotVisionInput } from "./github-copilot-headers.js";
 import { resolveOpenAIClientAuth } from "./openai-client-auth.js";
 import { clampOpenAIPromptCacheKey } from "./openai-prompt-cache.js";
-import { applyExtraBody, buildBaseOptions, clampMaxForOpenAI, OPENAI_COMPLETIONS_RESERVED_BODY_KEYS, } from "./simple-options.js";
+import { applyExtraBody, buildBaseOptions, clampMaxForOpenAI, clampReasoning, MIN_ANSWER_TOKENS, OPENAI_COMPLETIONS_RESERVED_BODY_KEYS, } from "./simple-options.js";
 import { transformMessages } from "./transform-messages.js";
 export { getOpenAICompletionsCompat as getCompat, } from "../utils/prompt-cache-ttl.js";
 const KIMI_K3_THINKING_LEVEL_MAP = {
@@ -662,6 +662,7 @@ export const streamSimple = (model, context, options) => {
         ...base,
         reasoningEffort,
         toolChoice,
+        thinkingBudgets: options?.thinkingBudgets,
     });
 };
 function createClient(model, context, apiKey, optionsHeaders, fetch, sessionId, compat = getCompat(model)) {
@@ -780,11 +781,26 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
         };
     }
     else if (compat.thinkingFormat === "chat-template" && model.reasoning) {
-        const chatTemplateKwargs = buildChatTemplateKwargs(model, options, compat);
+        const chatTemplateKwargs = buildChatTemplateValues(model, options, compat, compat.chatTemplateKwargs);
         if (chatTemplateKwargs) {
             params.chat_template_kwargs = chatTemplateKwargs;
         }
     }
+    else if (compat.thinkingFormat === "baseten" && model.reasoning) {
+        const basetenParams = params;
+        const chatTemplateArgs = buildChatTemplateValues(model, options, compat, compat.chatTemplateArgs ?? {});
+        if (chatTemplateArgs) {
+            basetenParams.chat_template_args = chatTemplateArgs;
+        }
+        if (compat.supportsReasoningEffort) {
+            const requestedEffort = options?.reasoningEffort;
+            const mappedEffort = requestedEffort ? model.thinkingLevelMap?.[requestedEffort] : model.thinkingLevelMap?.off;
+            const effort = mappedEffort === undefined ? requestedEffort : mappedEffort;
+            if (typeof effort === "string") {
+                basetenParams.reasoning_effort = effort;
+            }
+        }
+    }
     else if (compat.thinkingFormat === "deepseek" && model.reasoning) {
         if (options?.reasoningEffort) {
             params.thinking = { type: "enabled" };
@@ -852,6 +868,26 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
             params.reasoning_effort = offValue;
         }
     }
+    // vLLM caps reasoning with a top-level thinking_token_budget. Independent of
+    // thinkingFormat: the same server can serve zai, qwen or chat-template models.
+    // Reasoning and the answer share max_tokens here, so an uncapped reasoning
+    // phase can consume the whole response and leave no answer and no tool call.
+    if (compat.supportsThinkingTokenBudget && options?.reasoningEffort && model.reasoning) {
+        const level = clampReasoning(options.reasoningEffort);
+        const budgets = {
+            minimal: 1024,
+            low: 2048,
+            medium: 8192,
+            high: 16384,
+            ...options.thinkingBudgets,
+        };
+        const ceiling = params.max_tokens ?? params.max_completion_tokens ?? model.maxTokens;
+        // Always leave room for the answer, otherwise the budget recreates the bug it prevents.
+        const budget = Math.min(budgets[level], Math.max(0, ceiling - MIN_ANSWER_TOKENS));
+        if (budget > 0) {
+            params.thinking_token_budget = budget;
+        }
+    }
     // OpenRouter provider routing preferences
     if (model.compat?.openRouterRouting) {
         params.provider = model.compat.openRouterRouting;
@@ -869,17 +905,21 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
         }
     }
     applyExtraBody(params, options?.extraBody, OPENAI_COMPLETIONS_RESERVED_BODY_KEYS);
+    // Last so custom keys override the named request fields.
+    if (options?.samplingParams) {
+        Object.assign(params, options.samplingParams);
+    }
     return params;
 }
-function buildChatTemplateKwargs(model, options, compat) {
-    const kwargs = {};
-    for (const [key, value] of Object.entries(compat.chatTemplateKwargs)) {
+function buildChatTemplateValues(model, options, compat, values) {
+    const resolvedValues = {};
+    for (const [key, value] of Object.entries(values)) {
         const resolved = resolveChatTemplateKwargValue(model, options, compat, value);
         if (resolved !== undefined) {
-            kwargs[key] = resolved;
+            resolvedValues[key] = resolved;
         }
     }
-    return Object.keys(kwargs).length > 0 ? kwargs : undefined;
+    return Object.keys(resolvedValues).length > 0 ? resolvedValues : undefined;
 }
 function resolveChatTemplateKwargValue(model, options, compat, value) {
     if (typeof value !== "object" || value === null) {