@corbat-tech/coco 2.30.0 → 2.31.0

package/dist/index.d.ts

@@ -503,16 +503,16 @@ declare const CocoConfigSchema: z.ZodObject<{
             "kimi-code": "kimi-code";
             lmstudio: "lmstudio";
             codex: "codex";
-            qwen: "qwen";
-            deepseek: "deepseek";
-            mistral: "mistral";
             copilot: "copilot";
-            vertex: "vertex";
-            ollama: "ollama";
             groq: "groq";
             openrouter: "openrouter";
+            mistral: "mistral";
+            deepseek: "deepseek";
             together: "together";
             huggingface: "huggingface";
+            qwen: "qwen";
+            vertex: "vertex";
+            ollama: "ollama";
         }>>;
         apiKey: z.ZodOptional<z.ZodString>;
         model: z.ZodDefault<z.ZodString>;
@@ -523,6 +523,15 @@ declare const CocoConfigSchema: z.ZodObject<{
         location: z.ZodOptional<z.ZodString>;
     }, z.core.$strip>>;
     providerModels: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
+    providerThinking: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodEnum<{
+        high: "high";
+        medium: "medium";
+        low: "low";
+        off: "off";
+        auto: "auto";
+    }>, z.ZodObject<{
+        budget: z.ZodNumber;
+    }, z.core.$strip>]>>>;
     quality: z.ZodDefault<z.ZodObject<{
         minScore: z.ZodDefault<z.ZodNumber>;
         minCoverage: z.ZodDefault<z.ZodNumber>;
@@ -926,9 +935,33 @@ interface InputAnalysis {
  */
 type ProjectType = "cli" | "api" | "web_app" | "library" | "service" | "full_stack" | "automation" | "unknown";
 
+/**
+ * Unified thinking/reasoning mode support for all LLM providers.
+ *
+ * Normalizes three distinct API surfaces:
+ *   - Anthropic: thinking.budget_tokens
+ *   - OpenAI Chat Completions: reasoning_effort
+ *   - OpenAI Responses API: reasoning.effort
+ *   - Gemini: thinkingConfig.thinkingBudget
+ *   - Kimi: thinking.type enabled/disabled
+ */
+/**
+ * Provider-agnostic thinking mode.
+ * "off"    — disable thinking entirely (or keep disabled for models like Kimi)
+ * "auto"   — provider default / dynamic budget
+ * "low"    — minimal reasoning
+ * "medium" — balanced reasoning
+ * "high"   — maximum reasoning
+ * { budget: N } — explicit token budget (Anthropic / Gemini only; rejected for effort-only providers)
+ */
+type ThinkingMode = "off" | "auto" | "low" | "medium" | "high" | {
+    budget: number;
+};
+
 /**
  * LLM Provider types for Corbat-Coco
  */
+
 /**
  * Message role
  */
@@ -1016,6 +1049,8 @@ interface ChatOptions {
     timeout?: number;
     /** Abort signal to cancel in-flight requests */
     signal?: AbortSignal;
+    /** Thinking/reasoning mode to pass to the model (if supported) */
+    thinking?: ThinkingMode;
 }
 /**
  * Chat response

package/dist/index.js

@@ -1032,6 +1032,7 @@ __export(schema_exports, {
   ShipConfigSchema: () => ShipConfigSchema,
   SkillsConfigSchema: () => SkillsConfigSchema,
   StackConfigSchema: () => StackConfigSchema,
+  ThinkingModeSchema: () => ThinkingModeSchema,
   ToolsConfigSchema: () => ToolsConfigSchema,
   createDefaultConfigObject: () => createDefaultConfigObject,
   validateConfig: () => validateConfig
@@ -1076,9 +1077,13 @@ function createDefaultConfigObject(projectName, language = "typescript") {
     }
   };
 }
-var ProviderConfigSchema, QualityConfigSchema, PersistenceConfigSchema, StackConfigSchema, ProjectConfigSchema2, GitHubConfigSchema, IntegrationsConfigSchema, MCPServerConfigEntrySchema, MCPConfigSchema, ToolsConfigSchema, ShipConfigSchema, SkillsConfigSchema, CocoConfigSchema;
+var ThinkingModeSchema, ProviderConfigSchema, QualityConfigSchema, PersistenceConfigSchema, StackConfigSchema, ProjectConfigSchema2, GitHubConfigSchema, IntegrationsConfigSchema, MCPServerConfigEntrySchema, MCPConfigSchema, ToolsConfigSchema, ShipConfigSchema, SkillsConfigSchema, CocoConfigSchema;
 var init_schema = __esm({
   "src/config/schema.ts"() {
+    ThinkingModeSchema = z.union([
+      z.enum(["off", "auto", "low", "medium", "high"]),
+      z.object({ budget: z.number().int().min(0).max(2e5) })
+    ]);
     ProviderConfigSchema = z.object({
       type: z.enum([
         "anthropic",
@@ -1240,6 +1245,7 @@ var init_schema = __esm({
         timeout: 12e4
       }),
       providerModels: z.record(z.string(), z.string()).optional(),
+      providerThinking: z.record(z.string(), ThinkingModeSchema).optional(),
       quality: QualityConfigSchema.default({
         minScore: 85,
         minCoverage: 80,
@@ -13563,6 +13569,94 @@ async function withRetry(fn, config = {}) {
 
 // src/providers/anthropic.ts
 init_logger();
+
+// src/providers/thinking.ts
+var ANTHROPIC_BUDGET = {
+  low: 2048,
+  medium: 8e3,
+  high: 16e3
+};
+var GEMINI_BUDGET = {
+  low: 2048,
+  medium: 8e3,
+  high: 16e3
+};
+function isAnthropicThinkingModel(model) {
+  const m = model.toLowerCase();
+  if (m === "kimi-for-coding") return false;
+  return m.includes("claude-3-7") || m.includes("claude-opus-4") || m.includes("claude-sonnet-4") || m.includes("claude-haiku-4-5") || m.includes("claude-4");
+}
+function isOpenAIReasoningModel(model) {
+  const m = model.toLowerCase();
+  return m.startsWith("o1") || m.startsWith("o3") || m.startsWith("o4") || m.startsWith("gpt-5") || m.includes("codex");
+}
+function isGeminiThinkingModel(model) {
+  const m = model.toLowerCase();
+  return m.includes("gemini-2.5-pro") || m.includes("gemini-2.5-flash") || m.includes("gemini-3") && !m.includes("flash-lite") || m.includes("gemini-2.0-flash-thinking");
+}
+function isKimiThinkingModel(model) {
+  const m = model.toLowerCase();
+  return m.includes("kimi-k2") || m === "kimi-latest";
+}
+var ANTHROPIC_CAPABILITY = {
+  budgetRange: { min: 1024, max: 64e3, default: ANTHROPIC_BUDGET.medium }};
+var GEMINI_CAPABILITY = {
+  budgetRange: { min: 0, max: 32e3}};
+function mapToAnthropic(mode, model) {
+  if (!mode || mode === "off") return void 0;
+  if (!isAnthropicThinkingModel(model)) return void 0;
+  const cap = ANTHROPIC_CAPABILITY;
+  const { min, max } = cap.budgetRange;
+  if (typeof mode === "object") {
+    return { type: "enabled", budget_tokens: Math.min(Math.max(mode.budget, min), max) };
+  }
+  const budgetMap = {
+    auto: cap.budgetRange.default,
+    low: ANTHROPIC_BUDGET.low,
+    medium: ANTHROPIC_BUDGET.medium,
+    high: ANTHROPIC_BUDGET.high
+  };
+  const budget = budgetMap[mode];
+  if (budget === void 0) return void 0;
+  return { type: "enabled", budget_tokens: budget };
+}
+function mapToOpenAIEffort(mode, model) {
+  if (!mode || mode === "off") return void 0;
+  if (!isOpenAIReasoningModel(model)) return void 0;
+  if (typeof mode === "object") {
+    const { budget } = mode;
+    if (budget <= 2048) return "low";
+    if (budget <= 8e3) return "medium";
+    return "high";
+  }
+  if (mode === "auto") return "medium";
+  if (mode === "low" || mode === "medium" || mode === "high") return mode;
+  return void 0;
+}
+function mapToGeminiBudget(mode, model) {
+  if (!isGeminiThinkingModel(model)) return void 0;
+  if (!mode) return void 0;
+  if (mode === "off") return 0;
+  if (mode === "auto") return -1;
+  const { min, max } = GEMINI_CAPABILITY.budgetRange;
+  if (typeof mode === "object") {
+    return Math.min(Math.max(mode.budget, min), max);
+  }
+  const budgetMap = {
+    low: GEMINI_BUDGET.low,
+    medium: GEMINI_BUDGET.medium,
+    high: GEMINI_BUDGET.high
+  };
+  return budgetMap[mode];
+}
+function mapToKimiExtraBody(mode, model) {
+  if (!isKimiThinkingModel(model)) return void 0;
+  const effectiveMode = mode ?? "off";
+  const enabled = effectiveMode !== "off";
+  return { thinking: { type: enabled ? "enabled" : "disabled" } };
+}
+
+// src/providers/anthropic.ts
 var DEFAULT_MODEL = "claude-opus-4-6";
 var CONTEXT_WINDOWS = {
   // Kimi Code model (Anthropic-compatible endpoint)
@@ -13623,13 +13717,19 @@ var AnthropicProvider = class {
     this.ensureInitialized();
     return withRetry(async () => {
       try {
+        const model = options?.model ?? this.config.model ?? DEFAULT_MODEL;
+        const thinkingParam = mapToAnthropic(options?.thinking, model);
+        const baseMaxTokens = options?.maxTokens ?? this.config.maxTokens ?? 8192;
         const response = await this.client.messages.create({
-          model: options?.model ?? this.config.model ?? DEFAULT_MODEL,
-          max_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
-          temperature: options?.temperature ?? this.config.temperature ?? 0,
+          model,
+          // Anthropic requires max_tokens > budget_tokens
+          max_tokens: thinkingParam ? Math.max(baseMaxTokens, thinkingParam.budget_tokens + 1024) : baseMaxTokens,
+          // Anthropic requires temperature=1 when thinking is enabled
+          temperature: thinkingParam ? 1 : options?.temperature ?? this.config.temperature ?? 0,
           system: this.extractSystem(messages, options?.system),
           messages: this.convertMessages(messages),
-          stop_sequences: options?.stopSequences
+          stop_sequences: options?.stopSequences,
+          ...thinkingParam && { thinking: thinkingParam }
         });
         return {
           id: response.id,
@@ -13653,14 +13753,18 @@ var AnthropicProvider = class {
     this.ensureInitialized();
     return withRetry(async () => {
       try {
+        const model = options?.model ?? this.config.model ?? DEFAULT_MODEL;
+        const thinkingParam = mapToAnthropic(options?.thinking, model);
+        const baseMaxTokens = options?.maxTokens ?? this.config.maxTokens ?? 8192;
         const response = await this.client.messages.create({
-          model: options?.model ?? this.config.model ?? DEFAULT_MODEL,
-          max_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
-          temperature: options?.temperature ?? this.config.temperature ?? 0,
+          model,
+          max_tokens: thinkingParam ? Math.max(baseMaxTokens, thinkingParam.budget_tokens + 1024) : baseMaxTokens,
+          temperature: thinkingParam ? 1 : options?.temperature ?? this.config.temperature ?? 0,
           system: this.extractSystem(messages, options?.system),
           messages: this.convertMessages(messages),
           tools: this.convertTools(options.tools),
-          tool_choice: options.toolChoice ? this.convertToolChoice(options.toolChoice) : void 0
+          tool_choice: options.toolChoice ? this.convertToolChoice(options.toolChoice) : void 0,
+          ...thinkingParam && { thinking: thinkingParam }
         });
         const toolCalls = this.extractToolCalls(response.content);
         return {
@@ -13686,13 +13790,17 @@ var AnthropicProvider = class {
     this.ensureInitialized();
     let timeoutTriggered = false;
     try {
+      const model = options?.model ?? this.config.model ?? DEFAULT_MODEL;
+      const thinkingParam = mapToAnthropic(options?.thinking, model);
+      const baseMaxTokens = options?.maxTokens ?? this.config.maxTokens ?? 8192;
       const stream = await this.client.messages.stream(
         {
-          model: options?.model ?? this.config.model ?? DEFAULT_MODEL,
-          max_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
-          temperature: options?.temperature ?? this.config.temperature ?? 0,
+          model,
+          max_tokens: thinkingParam ? Math.max(baseMaxTokens, thinkingParam.budget_tokens + 1024) : baseMaxTokens,
+          temperature: thinkingParam ? 1 : options?.temperature ?? this.config.temperature ?? 0,
           system: this.extractSystem(messages, options?.system),
-          messages: this.convertMessages(messages)
+          messages: this.convertMessages(messages),
+          ...thinkingParam && { thinking: thinkingParam }
         },
         { signal: options?.signal }
       );
@@ -13748,15 +13856,19 @@ var AnthropicProvider = class {
     this.ensureInitialized();
     let timeoutTriggered = false;
     try {
+      const model = options?.model ?? this.config.model ?? DEFAULT_MODEL;
+      const thinkingParam = mapToAnthropic(options?.thinking, model);
+      const baseMaxTokens = options?.maxTokens ?? this.config.maxTokens ?? 8192;
       const stream = await this.client.messages.stream(
         {
-          model: options?.model ?? this.config.model ?? DEFAULT_MODEL,
-          max_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
-          temperature: options?.temperature ?? this.config.temperature ?? 0,
+          model,
+          max_tokens: thinkingParam ? Math.max(baseMaxTokens, thinkingParam.budget_tokens + 1024) : baseMaxTokens,
+          temperature: thinkingParam ? 1 : options?.temperature ?? this.config.temperature ?? 0,
           system: this.extractSystem(messages, options?.system),
           messages: this.convertMessages(messages),
           tools: this.convertTools(options.tools),
-          tool_choice: options.toolChoice ? this.convertToolChoice(options.toolChoice) : void 0
+          tool_choice: options.toolChoice ? this.convertToolChoice(options.toolChoice) : void 0,
+          ...thinkingParam && { thinking: thinkingParam }
         },
         { signal: options?.signal }
       );
@@ -14421,26 +14533,15 @@ var OpenAIProvider = class {
     return !MODELS_WITHOUT_TEMPERATURE.some((m) => model.toLowerCase().includes(m.toLowerCase()));
   }
   /**
-   * Check if a model needs thinking mode disabled for tool use
-   * Kimi models have thinking mode enabled by default which requires
-   * reasoning_content in multi-turn conversations with tools
+   * Get extra body parameters for API calls.
+   * Honors the user's ThinkingMode for Kimi models; defaults to disabled
+   * (preserving existing behavior) when no mode is specified.
    */
-  needsThinkingDisabled(model) {
-    return MODELS_WITH_THINKING_MODE.some((m) => model.toLowerCase().includes(m.toLowerCase()));
-  }
-  /**
-   * Get extra body parameters for API calls
-   * Used to disable thinking mode for Kimi models
-   * See: https://huggingface.co/moonshotai/Kimi-K2.5
-   *
-   * For Official Moonshot API: {'thinking': {'type': 'disabled'}}
-   * For vLLM/SGLang: {'chat_template_kwargs': {"thinking": False}}
-   */
-  getExtraBody(model) {
-    if (this.needsThinkingDisabled(model)) {
-      return {
-        thinking: { type: "disabled" }
-      };
+  getExtraBody(model, thinking) {
+    const kimiBody = mapToKimiExtraBody(thinking, model);
+    if (kimiBody) return kimiBody;
+    if (MODELS_WITH_THINKING_MODE.some((m) => model.toLowerCase().includes(m.toLowerCase()))) {
+      return { thinking: { type: "disabled" } };
     }
     return void 0;
   }
@@ -14457,6 +14558,7 @@ var OpenAIProvider = class {
       try {
         const supportsTemp = this.supportsTemperature(model);
         const maxTokens = options?.maxTokens ?? this.config.maxTokens ?? 8192;
+        const reasoningEffort = mapToOpenAIEffort(options?.thinking, model);
         const response = await this.client.chat.completions.create({
           model,
           ...buildMaxTokensParam(model, maxTokens),
@@ -14464,7 +14566,8 @@ var OpenAIProvider = class {
           stop: options?.stopSequences,
           ...supportsTemp && {
             temperature: options?.temperature ?? this.config.temperature ?? 0
-          }
+          },
+          ...reasoningEffort && { reasoning_effort: reasoningEffort }
         });
         const choice = response.choices[0];
         return {
@@ -14494,7 +14597,8 @@ var OpenAIProvider = class {
     return withRetry(async () => {
       try {
         const supportsTemp = this.supportsTemperature(model);
-        const extraBody = this.getExtraBody(model);
+        const extraBody = this.getExtraBody(model, options?.thinking);
+        const reasoningEffort = mapToOpenAIEffort(options?.thinking, model);
         const maxTokens = options?.maxTokens ?? this.config.maxTokens ?? 8192;
         const requestParams = {
           model,
@@ -14506,6 +14610,9 @@ var OpenAIProvider = class {
         if (supportsTemp) {
           requestParams.temperature = options?.temperature ?? this.config.temperature ?? 0;
         }
+        if (reasoningEffort) {
+          requestParams.reasoning_effort = reasoningEffort;
+        }
         if (extraBody) {
           Object.assign(requestParams, extraBody);
         }
@@ -14543,12 +14650,14 @@ var OpenAIProvider = class {
     try {
       const supportsTemp = this.supportsTemperature(model);
       const maxTokens = options?.maxTokens ?? this.config.maxTokens ?? 8192;
+      const reasoningEffort = mapToOpenAIEffort(options?.thinking, model);
       const stream = await this.client.chat.completions.create({
         model,
         ...buildMaxTokensParam(model, maxTokens),
         messages: this.convertMessages(messages, options?.system),
         stream: true,
-        ...supportsTemp && { temperature: options?.temperature ?? this.config.temperature ?? 0 }
+        ...supportsTemp && { temperature: options?.temperature ?? this.config.temperature ?? 0 },
+        ...reasoningEffort && { reasoning_effort: reasoningEffort }
       });
       let streamStopReason;
       for await (const chunk of stream) {
@@ -14579,7 +14688,8 @@ var OpenAIProvider = class {
     let timeoutTriggered = false;
     try {
       const supportsTemp = this.supportsTemperature(model);
-      const extraBody = this.getExtraBody(model);
+      const extraBody = this.getExtraBody(model, options?.thinking);
+      const reasoningEffort = mapToOpenAIEffort(options?.thinking, model);
       const maxTokens = options?.maxTokens ?? this.config.maxTokens ?? 8192;
       const requestParams = {
         model,
@@ -14592,6 +14702,9 @@ var OpenAIProvider = class {
       if (supportsTemp) {
         requestParams.temperature = options?.temperature ?? this.config.temperature ?? 0;
       }
+      if (reasoningEffort) {
+        requestParams.reasoning_effort = reasoningEffort;
+      }
       if (extraBody) {
         Object.assign(requestParams, extraBody);
       }
@@ -15059,6 +15172,7 @@ var OpenAIProvider = class {
         const model = options?.model ?? this.config.model ?? DEFAULT_MODEL2;
         const { input, instructions } = this.convertToResponsesInput(messages, options?.system);
         const supportsTemp = this.supportsTemperature(model);
+        const reasoningEffort = mapToOpenAIEffort(options?.thinking, model);
         const response = await this.client.responses.create({
           model,
           input,
@@ -15067,6 +15181,8 @@ var OpenAIProvider = class {
           ...supportsTemp && {
             temperature: options?.temperature ?? this.config.temperature ?? 0
           },
+          // Responses API uses nested reasoning.effort (not top-level reasoning_effort)
+          ...reasoningEffort && { reasoning: { effort: reasoningEffort } },
           store: false
         });
         return {
@@ -15095,6 +15211,7 @@ var OpenAIProvider = class {
         const { input, instructions } = this.convertToResponsesInput(messages, options?.system);
         const tools = this.convertToolsForResponses(options.tools);
         const supportsTemp = this.supportsTemperature(model);
+        const reasoningEffort = mapToOpenAIEffort(options?.thinking, model);
         const response = await this.client.responses.create({
           model,
           input,
@@ -15104,6 +15221,7 @@ var OpenAIProvider = class {
           ...supportsTemp && {
             temperature: options?.temperature ?? this.config.temperature ?? 0
           },
+          ...reasoningEffort && { reasoning: { effort: reasoningEffort } },
           store: false
         });
         let content = "";
@@ -15149,12 +15267,14 @@ var OpenAIProvider = class {
       const model = options?.model ?? this.config.model ?? DEFAULT_MODEL2;
       const { input, instructions } = this.convertToResponsesInput(messages, options?.system);
       const supportsTemp = this.supportsTemperature(model);
+      const reasoningEffort = mapToOpenAIEffort(options?.thinking, model);
       const stream = await this.client.responses.create({
         model,
         input,
         instructions: instructions ?? void 0,
         max_output_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
         ...supportsTemp && { temperature: options?.temperature ?? this.config.temperature ?? 0 },
+        ...reasoningEffort && { reasoning: { effort: reasoningEffort } },
         store: false,
         stream: true
       });
@@ -15212,12 +15332,14 @@ var OpenAIProvider = class {
       const { input, instructions } = this.convertToResponsesInput(messages, options?.system);
       const tools = options.tools.length > 0 ? this.convertToolsForResponses(options.tools) : void 0;
       const supportsTemp = this.supportsTemperature(model);
+      const reasoningEffort = mapToOpenAIEffort(options?.thinking, model);
       const requestParams = {
         model,
         input,
         instructions: instructions ?? void 0,
         max_output_tokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
         ...supportsTemp && { temperature: options?.temperature ?? this.config.temperature ?? 0 },
+        ...reasoningEffort && { reasoning: { effort: reasoningEffort } },
         store: false,
         stream: true
       };
@@ -16396,12 +16518,17 @@ var GeminiProvider = class {
     return model ?? this.config.model ?? DEFAULT_MODEL5;
   }
   buildConfig(messages, options, tools, toolChoice) {
+    const model = this.getModel(options?.model);
+    const thinkingBudget = mapToGeminiBudget(options?.thinking, model);
     const config = {
       maxOutputTokens: options?.maxTokens ?? this.config.maxTokens ?? 8192,
       temperature: options?.temperature ?? this.config.temperature ?? 0,
       stopSequences: options?.stopSequences,
       systemInstruction: this.extractSystem(messages, options?.system)
     };
+    if (thinkingBudget !== void 0) {
+      config.thinkingConfig = { thinkingBudget };
+    }
     if (tools && tools.length > 0) {
       config.tools = [{ functionDeclarations: this.convertTools(tools) }];
       config.toolConfig = {