@code-yeongyu/senpi-ai 2026.9.4 → 2026.9.5

package/dist/api/openai-completions.js

@@ -1,5 +1,5 @@
 import OpenAI from "openai";
-import { calculateCost, clampThinkingLevel, supportsMax, supportsXhigh } from "../models.js";
+import { calculateCost, clampThinkingLevel, inferOpenAIThinkingLevelMap, supportsMax, supportsXhigh, } from "../models.js";
 import { formatProviderError, normalizeProviderError } from "../utils/error-body.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import { shortHash } from "../utils/hash.js";
@@ -62,9 +62,9 @@ const OLLAMA_THINKING_LEVEL_MAP = {
     max: "high",
 };
 function getThinkingLevelMap(model, compat) {
-    if (model.thinkingLevelMap !== undefined) {
-        return model.thinkingLevelMap;
-    }
+    const inferred = inferOpenAIThinkingLevelMap(model);
+    if (inferred !== undefined)
+        return inferred;
     const id = model.id.toLowerCase();
     const isKimiK3 = id === "k3" || id.startsWith("k3-") || /(?:^|[/:-])kimi-k3(?:$|[/.:_-])/.test(id);
     const isDeepSeek = id.includes("deepseek");
@@ -262,6 +262,14 @@ export const stream = (model, context, options) => {
             stopReason: "pending",
             timestamp: Date.now(),
         };
+        // `reasoning_details` are replay metadata, not user-visible stream deltas.
+        // Keep them in memory during streaming and serialize once when the block is finalized.
+        let streamedReasoningDetails;
+        const applyStreamedReasoningDetails = (block) => {
+            if (streamedReasoningDetails !== undefined) {
+                block.thinkingSignature = JSON.stringify(streamedReasoningDetails);
+            }
+        };
         try {
             const clientAuth = resolveOpenAIClientAuth(model.provider, options?.apiKey, options?.headers);
             const compat = getCompat(model);
@@ -341,6 +349,7 @@ export const stream = (model, context, options) => {
                     });
                 }
                 else if (block.type === "thinking") {
+                    applyStreamedReasoningDetails(block);
                     stream.push({
                         type: "thinking_end",
                         contentIndex,
@@ -671,6 +680,9 @@ export const stream = (model, context, options) => {
         }
         catch (error) {
             for (const block of output.content) {
+                if (block.type === "thinking") {
+                    applyStreamedReasoningDetails(block);
+                }
                 delete block.index;
                 // Streaming scratch buffers are only used during parsing; never persist them.
                 delete block.partialArgs;
@@ -704,12 +716,17 @@ export const streamSimple = (model, context, options) => {
     const compat = getCompat(model);
     const thinkingLevelMap = getThinkingLevelMap(model, compat);
     const thinkingModel = thinkingLevelMap === model.thinkingLevelMap ? model : { ...model, thinkingLevelMap };
-    const clampedReasoning = options?.reasoning ? clampThinkingLevel(thinkingModel, options.reasoning) : undefined;
-    const reasoningEffort = clampedReasoning === "off"
+    const clampedReasoning = options?.reasoning
+        ? clampThinkingLevel(thinkingModel, options.reasoning)
+        : model.id.includes("gpt-6-astra")
+            ? "off"
+            : undefined;
+    const normalizedReasoning = clampedReasoning === "off" && model.id.includes("gpt-6-astra") ? "low" : clampedReasoning;
+    const reasoningEffort = normalizedReasoning === "off"
         ? undefined
-        : clampedReasoning === "max" && supportsMax(thinkingModel)
+        : normalizedReasoning === "max" && supportsMax(thinkingModel)
             ? "max"
-            : clampMaxForOpenAI(clampedReasoning, supportsXhigh(thinkingModel));
+            : clampMaxForOpenAI(normalizedReasoning, supportsXhigh(thinkingModel));
     return stream(model, context, {
         ...base,
         reasoningEffort,
@@ -803,9 +820,12 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
     if (cacheControl) {
         applyAnthropicCacheControl(messages, params.tools, cacheControl);
     }
-    if (options?.toolChoice && params.tools?.length) {
+    if (options?.toolChoice) {
         params.tool_choice = options.toolChoice;
     }
+    if (compat.vllmPriority !== undefined) {
+        params.priority = compat.vllmPriority;
+    }
     const thinkingTokenBudgetField = resolveThinkingTokenBudgetField(compat);
     const thinkingBudget = resolveClampedThinkingBudget(model, options, params);
     if (compat.thinkingFormat === "zai" && model.reasoning) {
@@ -849,7 +869,7 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
         }
         if (compat.supportsReasoningEffort) {
             const requestedEffort = options?.reasoningEffort;
-            const mappedEffort = requestedEffort ? model.thinkingLevelMap?.[requestedEffort] : model.thinkingLevelMap?.off;
+            const mappedEffort = requestedEffort ? thinkingLevelMap?.[requestedEffort] : thinkingLevelMap?.off;
             const effort = mappedEffort === undefined ? requestedEffort : mappedEffort;
             if (typeof effort === "string") {
                 basetenParams.reasoning_effort = effort;