@code-yeongyu/senpi-ai 2026.8.18 → 2026.8.19

package/dist/api/openai-completions.js

@@ -5,6 +5,7 @@ import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import { shortHash } from "../utils/hash.js";
 import { headersToRecord } from "../utils/headers.js";
 import { parseStreamingJson } from "../utils/json-parse.js";
+import { forcePiUserAgent } from "../utils/pi-user-agent.js";
 import { getOpenAICompletionsCompat as getCompat, } from "../utils/prompt-cache-ttl.js";
 import { getProviderEnvValue } from "../utils/provider-env.js";
 import { retryProviderStreamRequest } from "../utils/provider-retry.js";
@@ -15,7 +16,7 @@ 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, clampReasoning, MIN_ANSWER_TOKENS, OPENAI_COMPLETIONS_RESERVED_BODY_KEYS, } from "./simple-options.js";
+import { applyExtraBody, buildBaseOptions, clampMaxForOpenAI, clampThinkingBudgetToAnswerRoom, OPENAI_COMPLETIONS_RESERVED_BODY_KEYS, thinkingBudgetForLevel, } 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 = {
@@ -647,7 +648,12 @@ export const stream = (model, context, options) => {
 };
 export const streamSimple = (model, context, options) => {
     resolveOpenAIClientAuth(model.provider, options?.apiKey, options?.headers);
-    const base = buildBaseOptions(model, context, options, options?.apiKey);
+    const base = {
+        ...buildBaseOptions(model, context, options, options?.apiKey),
+        // Adapter-native callers may pass the richer OpenAI tool_choice shape; the
+        // provider-neutral SimpleStreamOptions value is the fallback.
+        toolChoice: options?.toolChoice ?? options?.toolChoice,
+    };
     const compat = getCompat(model);
     const thinkingLevelMap = getThinkingLevelMap(model, compat);
     const thinkingModel = thinkingLevelMap === model.thinkingLevelMap ? model : { ...model, thinkingLevelMap };
@@ -657,11 +663,9 @@ export const streamSimple = (model, context, options) => {
         : clampedReasoning === "max" && supportsMax(thinkingModel)
             ? "max"
             : clampMaxForOpenAI(clampedReasoning, supportsXhigh(thinkingModel));
-    const toolChoice = options?.toolChoice;
     return stream(model, context, {
         ...base,
         reasoningEffort,
-        toolChoice,
         thinkingBudgets: options?.thinkingBudgets,
     });
 };
@@ -691,6 +695,9 @@ function createClient(model, context, apiKey, optionsHeaders, fetch, sessionId,
     if (optionsHeaders) {
         Object.assign(headers, optionsHeaders);
     }
+    if (model.provider === "xai") {
+        forcePiUserAgent(headers);
+    }
     return new OpenAI({
         apiKey,
         baseURL: model.baseUrl,
@@ -755,6 +762,8 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
     if (options?.toolChoice) {
         params.tool_choice = options.toolChoice;
     }
+    const thinkingTokenBudgetField = resolveThinkingTokenBudgetField(compat);
+    const thinkingBudget = resolveClampedThinkingBudget(model, options, params);
     if (compat.thinkingFormat === "zai" && model.reasoning) {
         const isGlm53 = /(?:^|[/:-])glm-5\.3(?:$|[/.:_-])/.test(model.id.toLowerCase());
         const zaiParams = params;
@@ -783,14 +792,14 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
         };
     }
     else if (compat.thinkingFormat === "chat-template" && model.reasoning) {
-        const chatTemplateKwargs = buildChatTemplateValues(model, options, compat, compat.chatTemplateKwargs);
+        const chatTemplateKwargs = buildChatTemplateValues(model, options, compat, compat.chatTemplateKwargs, thinkingBudget);
         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 ?? {});
+        const chatTemplateArgs = buildChatTemplateValues(model, options, compat, compat.chatTemplateArgs ?? {}, thinkingBudget);
         if (chatTemplateArgs) {
             basetenParams.chat_template_args = chatTemplateArgs;
         }
@@ -870,25 +879,12 @@ 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;
-        }
+    // Cap reasoning with a top-level budget field. 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 (thinkingTokenBudgetField && thinkingBudget !== undefined) {
+        Object.assign(params, { [thinkingTokenBudgetField]: thinkingBudget });
     }
     // OpenRouter provider routing preferences
     if (model.compat?.openRouterRouting) {
@@ -913,17 +909,31 @@ function buildParams(model, context, options, compat = getCompat(model), cacheRe
     }
     return params;
 }
-function buildChatTemplateValues(model, options, compat, values) {
+function resolveThinkingTokenBudgetField(compat) {
+    if (compat.thinkingTokenBudgetField)
+        return compat.thinkingTokenBudgetField;
+    if (compat.supportsThinkingTokenBudget)
+        return "thinking_token_budget";
+    return undefined;
+}
+function resolveClampedThinkingBudget(model, options, params) {
+    if (!options?.reasoningEffort || !model.reasoning)
+        return undefined;
+    const ceiling = params.max_tokens ?? params.max_completion_tokens ?? model.maxTokens;
+    const budget = clampThinkingBudgetToAnswerRoom(thinkingBudgetForLevel(options.reasoningEffort, options.thinkingBudgets), ceiling);
+    return budget > 0 ? budget : undefined;
+}
+function buildChatTemplateValues(model, options, compat, values, thinkingBudget) {
     const resolvedValues = {};
     for (const [key, value] of Object.entries(values)) {
-        const resolved = resolveChatTemplateKwargValue(model, options, compat, value);
+        const resolved = resolveChatTemplateKwargValue(model, options, compat, value, thinkingBudget);
         if (resolved !== undefined) {
             resolvedValues[key] = resolved;
         }
     }
     return Object.keys(resolvedValues).length > 0 ? resolvedValues : undefined;
 }
-function resolveChatTemplateKwargValue(model, options, compat, value) {
+function resolveChatTemplateKwargValue(model, options, compat, value, thinkingBudget) {
     if (typeof value !== "object" || value === null) {
         return value;
     }
@@ -934,6 +944,9 @@ function resolveChatTemplateKwargValue(model, options, compat, value) {
     if (value.$var === "thinking.enabled") {
         return !!reasoningEffort;
     }
+    if (value.$var === "thinking.budget") {
+        return thinkingBudget;
+    }
     const thinkingLevelMap = getThinkingLevelMap(model, compat);
     const mappedValue = reasoningEffort ? thinkingLevelMap?.[reasoningEffort] : thinkingLevelMap?.off;
     return mappedValue === undefined ? reasoningEffort : typeof mappedValue === "string" ? mappedValue : undefined;
@@ -1349,7 +1362,10 @@ function parseChunkUsage(rawUsage, model) {
     const cacheReadTokens = rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens ?? rawUsage.cached_tokens ?? 0;
     const cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;
     // Follow documented OpenAI/OpenRouter semantics: cached_tokens is cache-read
-    // tokens (hits). OpenAI does not document or emit cache_write_tokens, but
+    // tokens (hits). Providers disagree on placement: OpenAI/OpenRouter use
+    // prompt_tokens_details.cached_tokens, DeepSeek uses prompt_cache_hit_tokens,
+    // and Kimi documents top-level usage.cached_tokens on the final usage chunk.
+    // OpenAI does not document or emit cache_write_tokens, but
     // OpenRouter-compatible providers can include it as a separate write count.
     // OpenRouter's own provider/tests affirm the separate mapping:
     // https://github.com/OpenRouterTeam/ai-sdk-provider/pull/409