@absolutejs/ai 0.0.16 → 0.0.18

package/dist/ai/index.js

@@ -43,6 +43,111 @@ async function* tapStream(source, params, providerName) {
   }
 }
 
+// src/ai/providers/reasoning.ts
+var EFFORT_ORDER = [
+  "minimal",
+  "low",
+  "medium",
+  "high",
+  "max"
+];
+var ANTHROPIC_EFFORT = [
+  /opus-4-[5-8]/,
+  /sonnet-4-6/,
+  /fable-5/,
+  /mythos-5/
+];
+var ANTHROPIC_ADAPTIVE_ONLY = [];
+var ANTHROPIC_LEGACY_THINKING = [
+  /sonnet-4-5/,
+  /sonnet-4-0/,
+  /sonnet-4-2025/,
+  /opus-4-0/,
+  /opus-4-1/,
+  /opus-4-2025/,
+  /3-7-sonnet/
+];
+var ANTHROPIC_NO_SAMPLING = [/opus-4-[78]/, /fable-5/, /mythos-5/];
+var ANTHROPIC_MAX_EFFORT = [
+  /opus-4-[678]/,
+  /sonnet-4-6/,
+  /fable-5/,
+  /mythos-5/
+];
+var OPENAI_REASONING = [/(^|[^a-z])o[1345](-|$)/, /gpt-5/];
+var OPENAI_MINIMAL_EFFORT = [/gpt-5/];
+var matches = (model, patterns) => patterns.some((pattern) => pattern.test(model));
+var anthropicReasoningMode = (model) => {
+  if (matches(model, ANTHROPIC_EFFORT))
+    return "effort";
+  if (matches(model, ANTHROPIC_ADAPTIVE_ONLY))
+    return "adaptive";
+  if (matches(model, ANTHROPIC_LEGACY_THINKING))
+    return "legacy";
+  return "none";
+};
+var anthropicSupportsSampling = (model) => !matches(model, ANTHROPIC_NO_SAMPLING);
+var isOpenAIReasoningModel = (model) => matches(model, OPENAI_REASONING);
+var EFFORT_BUDGET = {
+  high: 16384,
+  low: 2048,
+  max: 32768,
+  medium: 8192,
+  minimal: 1024
+};
+var budgetToEffort = (budget) => {
+  if (budget <= 2048)
+    return "low";
+  if (budget <= 8192)
+    return "medium";
+  if (budget <= 16384)
+    return "high";
+  return "max";
+};
+var resolveEffort = (reasoning) => {
+  if (reasoning.effort)
+    return reasoning.effort;
+  if (typeof reasoning.budgetTokens === "number") {
+    return budgetToEffort(reasoning.budgetTokens);
+  }
+  return null;
+};
+var resolveBudgetTokens = (reasoning) => {
+  if (typeof reasoning.budgetTokens === "number")
+    return reasoning.budgetTokens;
+  if (reasoning.effort)
+    return EFFORT_BUDGET[reasoning.effort];
+  return null;
+};
+var clampEffort = (effort, allowed) => {
+  if (allowed.includes(effort))
+    return effort;
+  const idx = EFFORT_ORDER.indexOf(effort);
+  for (let lower = idx - 1;lower >= 0; lower -= 1) {
+    const candidate = EFFORT_ORDER[lower];
+    if (candidate && allowed.includes(candidate))
+      return candidate;
+  }
+  return allowed[0] ?? effort;
+};
+var anthropicEffortValue = (model, reasoning) => {
+  const effort = resolveEffort(reasoning);
+  if (!effort)
+    return null;
+  const allowed = matches(model, ANTHROPIC_MAX_EFFORT) ? ["low", "medium", "high", "max"] : ["low", "medium", "high"];
+  return clampEffort(effort, allowed);
+};
+var openaiEffortValue = (model, reasoning) => {
+  if (!isOpenAIReasoningModel(model))
+    return null;
+  const effort = resolveEffort(reasoning);
+  if (!effort)
+    return null;
+  const allowed = matches(model, OPENAI_MINIMAL_EFFORT) ? ["minimal", "low", "medium", "high"] : ["low", "medium", "high"];
+  const requested = effort === "max" ? "high" : effort;
+  return clampEffort(requested, allowed);
+};
+
 // src/ai/providers/openai.ts
 var h2IfHttps = (url) => url.startsWith("https://") ? { protocol: "http2" } : {};
 var DEFAULT_BASE_URL = "https://api.openai.com";
@@ -169,12 +274,24 @@ var buildRequestBody = (params) => {
       body.parallel_tool_calls = params.parallelToolCalls;
     }
   }
-  if (typeof params.temperature === "number")
-    body.temperature = params.temperature;
-  if (typeof params.topP === "number")
-    body.top_p = params.topP;
-  if (typeof params.maxTokens === "number")
-    body.max_tokens = params.maxTokens;
+  if (isOpenAIReasoningModel(params.model)) {
+    if (typeof params.maxTokens === "number") {
+      body.max_completion_tokens = params.maxTokens;
+    }
+    if (params.reasoning) {
+      const effort = openaiEffortValue(params.model, params.reasoning);
+      if (effort)
+        body.reasoning_effort = effort;
+    }
+  } else {
+    if (typeof params.temperature === "number") {
+      body.temperature = params.temperature;
+    }
+    if (typeof params.topP === "number")
+      body.top_p = params.topP;
+    if (typeof params.maxTokens === "number")
+      body.max_tokens = params.maxTokens;
+  }
   if (params.stopSequences && params.stopSequences.length > 0)
     body.stop = params.stopSequences;
   if (typeof params.seed === "number")
@@ -591,11 +708,14 @@ var buildRequestBody2 = (params, isImageModel) => {
       };
     }
   }
-  if (params.thinking) {
-    body.reasoning = {
-      effort: "high",
-      summary: "auto"
-    };
+  if (params.reasoning && isOpenAIReasoningModel(params.model)) {
+    const effort = openaiEffortValue(params.model, params.reasoning);
+    if (effort) {
+      body.reasoning = {
+        effort,
+        summary: "auto"
+      };
+    }
   }
   return body;
 };
@@ -1221,7 +1341,13 @@ var buildRequestBody4 = (params) => {
     stream: true
   };
   if (params.systemPrompt) {
-    body.system = params.systemPrompt;
+    body.system = params.cacheSystemPrompt ? [
+      {
+        cache_control: { type: "ephemeral" },
+        text: params.systemPrompt,
+        type: "text"
+      }
+    ] : params.systemPrompt;
   }
   if (params.tools && params.tools.length > 0) {
     body.tools = params.tools.map(mapToolDefinition2);
@@ -1233,18 +1359,33 @@ var buildRequestBody4 = (params) => {
       body.tool_choice = { name: params.toolChoice.name, type: "tool" };
     }
   }
-  if (typeof params.temperature === "number")
-    body.temperature = params.temperature;
-  if (typeof params.topP === "number")
-    body.top_p = params.topP;
   if (typeof params.maxTokens === "number")
     body.max_tokens = params.maxTokens;
   if (params.stopSequences && params.stopSequences.length > 0) {
     body.stop_sequences = params.stopSequences;
   }
-  if (params.thinking) {
-    body.thinking = params.thinking;
-    body.max_tokens = Math.max(typeof body.max_tokens === "number" ? body.max_tokens : MAX_TOKENS, params.thinking.budget_tokens + MAX_TOKENS);
+  const mode = params.reasoning ? anthropicReasoningMode(params.model) : "none";
+  const thinkingActive = mode !== "none";
+  if (!thinkingActive && anthropicSupportsSampling(params.model)) {
+    if (typeof params.temperature === "number") {
+      body.temperature = params.temperature;
+    }
+    if (typeof params.topP === "number")
+      body.top_p = params.topP;
+  }
+  if (mode === "effort" || mode === "adaptive") {
+    body.thinking = { type: "adaptive" };
+    if (mode === "effort" && params.reasoning) {
+      const effort = anthropicEffortValue(params.model, params.reasoning);
+      if (effort)
+        body.output_config = { effort };
+    }
+  } else if (mode === "legacy" && params.reasoning) {
+    const budget = resolveBudgetTokens(params.reasoning);
+    if (budget) {
+      body.thinking = { budget_tokens: budget, type: "enabled" };
+      body.max_tokens = Math.max(typeof body.max_tokens === "number" ? body.max_tokens : MAX_TOKENS, budget + MAX_TOKENS);
+    }
   }
   return body;
 };
@@ -1383,6 +1524,8 @@ var extractUsage4 = (usageRecord, existingUsage) => {
     return existingUsage;
   }
   return {
+    cacheReadInputTokens: getNumber(usageRecord, "cache_read_input_tokens") || existingUsage?.cacheReadInputTokens || 0,
+    cacheWriteInputTokens: getNumber(usageRecord, "cache_creation_input_tokens") || existingUsage?.cacheWriteInputTokens || 0,
     inputTokens: getNumber(usageRecord, "input_tokens") || existingUsage?.inputTokens || 0,
     outputTokens: getNumber(usageRecord, "output_tokens") || existingUsage?.outputTokens || 0
   };
@@ -1893,7 +2036,6 @@ var WS_OPEN = 1;
 var BACKPRESSURE_THRESHOLD = 1048576;
 var BACKPRESSURE_DELAY = 10;
 var DEFAULT_MAX_TURNS = 10;
-var DEFAULT_THINKING_BUDGET = 1e4;
 var INITIAL_TURN = 0;
 var delay = (milliseconds) => new Promise((resolve) => setTimeout(resolve, milliseconds));
 var checkBackpressure = async (socket) => {
@@ -2075,16 +2217,12 @@ var processToolTurn = async (socket, options, state, messageId, conversationId,
     role: "user"
   });
   const toolDefs = options.tools ? buildToolDefinitions(options.tools) : undefined;
-  const thinkingConfig = options.thinking ? {
-    budget_tokens: typeof options.thinking === "object" ? options.thinking.budgetTokens : DEFAULT_THINKING_BUDGET,
-    type: "enabled"
-  } : undefined;
   const stream = options.provider.stream({
     messages: state.currentMessages,
     model: options.model,
+    reasoning: options.reasoning,
     signal,
     systemPrompt: options.systemPrompt,
-    thinking: thinkingConfig,
     tools: toolDefs
   });
   await consumeToolStream(stream, state, options, socket, messageId, conversationId, signal);
@@ -2163,16 +2301,12 @@ var processStreamTextChunk = async (chunk, options, socket, messageId, conversat
 };
 var processStream = async (socket, options, messages, messageId, conversationId, signal, startTime) => {
   const toolDefs = options.tools ? buildToolDefinitions(options.tools) : undefined;
-  const thinkingConfig = options.thinking ? {
-    budget_tokens: typeof options.thinking === "object" ? options.thinking.budgetTokens : DEFAULT_THINKING_BUDGET,
-    type: "enabled"
-  } : undefined;
   const stream = options.provider.stream({
     messages,
     model: options.model,
+    reasoning: options.reasoning,
     signal,
     systemPrompt: options.systemPrompt,
-    thinking: thinkingConfig,
     tools: toolDefs
   });
   const result = await consumeStream(stream, options, socket, messages, messageId, conversationId, signal);
@@ -2282,7 +2416,6 @@ var handleStreamError = async (socket, err, messageId, conversationId, signal, s
 
 // src/ai/streamAIToSSE.ts
 var DEFAULT_MAX_TURNS2 = 10;
-var DEFAULT_THINKING_BUDGET2 = 1e4;
 var buildToolDefinitions2 = (tools) => Object.entries(tools).map(([name, def]) => ({
   description: def.description,
   input_schema: def.input,
@@ -2299,10 +2432,6 @@ var executeTool2 = async (options, toolName, toolInput) => {
     return `Error: ${err instanceof Error ? err.message : String(err)}`;
   }
 };
-var buildThinkingConfig = (options) => options.thinking ? {
-  budget_tokens: typeof options.thinking === "object" ? options.thinking.budgetTokens : DEFAULT_THINKING_BUDGET2,
-  type: "enabled"
-} : undefined;
 var serializeToolCall2 = (name, input) => `${name}:${JSON.stringify(input)}`;
 var streamAIToSSE = async function* (conversationId, messageId, options, renderers) {
   const signal = options.signal ?? new AbortController().signal;
@@ -2463,7 +2592,6 @@ var streamTurns = async function* (options, renderers, messages, signal, startTi
     turn: 0
   };
   const toolDefs = options.tools ? buildToolDefinitions2(options.tools) : undefined;
-  const thinkingConfig = buildThinkingConfig(options);
   for (;turnState.turn <= maxTurns && !signal.aborted; turnState.turn++) {
     const chunkState = {
       contentBlocks: [],
@@ -2474,9 +2602,9 @@ var streamTurns = async function* (options, renderers, messages, signal, startTi
     const stream = options.provider.stream({
       messages: turnState.currentMessages,
       model: options.model,
+      reasoning: options.reasoning,
       signal,
       systemPrompt: options.systemPrompt,
-      thinking: thinkingConfig,
       tools: toolDefs
     });
     yield* consumeStream2(stream, chunkState, renderers, options, turnState, signal);
@@ -2632,7 +2760,7 @@ var resolveModel = (config, parsed) => {
   return parsed.providerName;
 };
 var resolveTools = (config, providerName, model) => typeof config.tools === "function" ? config.tools(providerName, model) : config.tools;
-var resolveThinking = (config, providerName, model) => typeof config.thinking === "function" ? config.thinking(providerName, model) : config.thinking;
+var resolveReasoning = (config, providerName, model) => typeof config.reasoning === "function" ? config.reasoning(providerName, model) : config.reasoning;
 var aiChat = (config) => {
   const path = config.path ?? DEFAULT_PATH;
   const store = config.store ?? createMemoryStore();
@@ -2683,7 +2811,7 @@ var aiChat = (config) => {
       provider: config.provider(providerName),
       signal: controller.signal,
       systemPrompt: config.systemPrompt,
-      thinking: resolveThinking(config, providerName, model),
+      reasoning: resolveReasoning(config, providerName, model),
       tools: resolveTools(config, providerName, model),
       onComplete: async (fullResponse, usage) => {
         const conv = await store.get(conversationId);
@@ -2757,7 +2885,7 @@ var aiChat = (config) => {
         provider: config.provider(providerName),
         signal: controller.signal,
         systemPrompt: config.systemPrompt,
-        thinking: resolveThinking(config, providerName, model),
+        reasoning: resolveReasoning(config, providerName, model),
         tools: resolveTools(config, providerName, model)
       }, renderers);
       for await (const event of sseStream) {
@@ -2821,9 +2949,11 @@ var aiChat = (config) => {
 var DEFAULT_OBJECT_TOOL_NAME = "respond";
 var generateAI = async (options) => {
   const stream = options.provider.stream({
+    cacheSystemPrompt: options.cacheSystemPrompt,
     maxTokens: options.maxTokens,
     messages: options.messages,
     model: options.model,
+    reasoning: options.reasoning,
     responseFormat: options.responseFormat,
     signal: options.signal,
     stopSequences: options.stopSequences,
@@ -2855,10 +2985,12 @@ var generateObjectAI = async (options) => {
     name: toolName
   };
   const { toolCalls, usage } = await generateAI({
+    cacheSystemPrompt: options.cacheSystemPrompt,
     maxTokens: options.maxTokens,
     messages: options.messages,
     model: options.model,
     provider: options.provider,
+    reasoning: options.reasoning,
     signal: options.signal,
     systemPrompt: options.systemPrompt,
     temperature: options.temperature,
@@ -3725,5 +3857,5 @@ export {
   aiChat
 };
 
-//# debugId=A6D6E7B52873D16664756E2164756E21
+//# debugId=F448AA17322C183164756E2164756E21
 //# sourceMappingURL=index.js.map