profclaw 2.0.0 → 2.4.0

package/dist/providers/ai-sdk.js

@@ -6,13 +6,14 @@
  *
  * @see https://sdk.vercel.ai/docs
  */
-import { generateText, streamText, tool as createTool, jsonSchema } from 'ai';
+import { generateText, streamText, stepCountIs, tool as createTool, jsonSchema } from 'ai';
 // Provider SDKs are lazily imported inside ensureProvider() to reduce startup memory
 // (each SDK is ~15MB; loading all 6 at startup costs ~100MB even if unused)
 import { zodToJsonSchema } from 'zod-to-json-schema';
 import { randomUUID } from 'node:crypto';
 import { logger } from '../utils/logger.js';
 import { normalizeToolSchema } from './schema-utils.js';
+import { aggregateStepUsage, applyAnthropicMessageCache, applyAnthropicToolCache, calculateCachedInputCost, getCacheConfigForModel, extractCacheUsage, } from './prompt-cache.js';
 // AI provider resilience settings (configurable via env)
 const AI_TIMEOUT_MS = parseInt(process.env['AI_PROVIDER_TIMEOUT_MS'] ?? '120000', 10);
 const AI_MAX_RETRIES = parseInt(process.env['AI_MAX_RETRIES'] ?? '2', 10);
@@ -26,8 +27,12 @@ async function withRetry(fn, maxRetries = AI_MAX_RETRIES) {
             return await fn();
         }
         catch (error) {
-            const isRetryable = error instanceof Error &&
-                (/429|503|rate.limit|too.many|service.unavailable|ECONNRESET|ETIMEDOUT|fetch.failed/i.test(error.message));
+            const msg = error instanceof Error ? error.message : '';
+            // Don't retry quota exhaustion — it won't resolve with backoff
+            const isQuotaError = /quota|exceeded your current|billing/i.test(msg);
+            const isRetryable = !isQuotaError &&
+                error instanceof Error &&
+                (/429|503|rate.limit|too.many|service.unavailable|ECONNRESET|ETIMEDOUT|fetch.failed/i.test(msg));
             if (!isRetryable || attempt === maxRetries)
                 throw error;
             const delay = Math.min(1000 * 2 ** attempt, 10000);
@@ -49,7 +54,7 @@ const PROVIDER_DEFAULT_MODEL = {
     anthropic: 'claude-sonnet-4-5-20250929',
     openai: 'gpt-4o',
     google: 'gemini-2.0-flash',
-    azure: 'gpt-4o',
+    azure: 'gpt-4o', // Resolved at runtime via deploymentName config
     groq: 'llama-3.3-70b-versatile',
     cerebras: 'llama-3.3-70b',
     fireworks: 'accounts/fireworks/models/llama-v3p1-70b-instruct',
@@ -172,17 +177,17 @@ class AIProviderManager {
                 enabled: true,
             });
         }
-        // Azure OpenAI (supports both standard and Foundry modes)
+        // Azure OpenAI (supports both base URL and resource name modes)
         if (process.env.AZURE_OPENAI_API_KEY) {
-            // Check if using Foundry (custom base URL) or standard (resource name)
             const baseUrl = process.env.AZURE_OPENAI_BASE_URL || process.env.AZURE_OPENAI_ENDPOINT;
-            const resourceName = process.env.AZURE_OPENAI_RESOURCE_NAME;
+            const resourceName = process.env.AZURE_OPENAI_RESOURCE_NAME || process.env.AZURE_RESOURCE_NAME;
             if (baseUrl || resourceName) {
                 this.configure('azure', {
                     type: 'azure',
                     apiKey: process.env.AZURE_OPENAI_API_KEY,
+                    // Prefer explicit base URL over resource name (resource name constructs URL automatically)
                     baseUrl: baseUrl,
-                    resourceName: resourceName,
+                    resourceName: baseUrl ? undefined : resourceName, // Don't use resourceName if baseUrl is set
                     deploymentName: process.env.AZURE_OPENAI_DEPLOYMENT_NAME,
                     apiVersion: process.env.AZURE_OPENAI_API_VERSION,
                     enabled: true,
@@ -689,7 +694,7 @@ class AIProviderManager {
         }
     }
     resolveModel(modelOrAlias) {
-        // Check if it's a full provider/model path
+        // Check if it's a full provider/model path (e.g. "azure/gpt4o", "ollama/llama3.2")
         if (modelOrAlias.includes('/')) {
             const [provider, model] = modelOrAlias.split('/');
             return { provider: provider, model };
@@ -700,15 +705,35 @@ class AIProviderManager {
             // For Azure, use the configured deployment name instead of the default alias model
             if (alias.provider === 'azure') {
                 const azureConfig = this.configs.get('azure');
-                if (azureConfig?.defaultModel) {
-                    return { provider: 'azure', model: azureConfig.defaultModel };
+                const azureModel = azureConfig?.defaultModel
+                    || azureConfig?.deploymentName
+                    || process.env.AZURE_OPENAI_DEPLOYMENT_NAME;
+                if (azureModel) {
+                    return { provider: 'azure', model: azureModel };
                 }
             }
-            return alias;
+            // If the alias points to a configured provider, use it directly
+            if (this.isConfigured(alias.provider)) {
+                return alias;
+            }
+            // If the alias provider isn't configured, try Azure as a fallback
+            // (Azure can serve OpenAI models like gpt-4o via deployment names)
+            if (this.isConfigured('azure') && (alias.provider === 'openai')) {
+                return this.resolveToAzureDeployment(modelOrAlias);
+            }
         }
         // Try to find in catalog
         const catalogModel = MODEL_CATALOG.find((m) => m.id === modelOrAlias);
         if (catalogModel) {
+            // If the catalog provider is configured, use it
+            if (this.isConfigured(catalogModel.provider)) {
+                return { provider: catalogModel.provider, model: catalogModel.id };
+            }
+            // OpenAI models can be served by Azure via deployment names
+            if (catalogModel.provider === 'openai' && this.isConfigured('azure')) {
+                return this.resolveToAzureDeployment(modelOrAlias);
+            }
+            // Fallback: use catalog as-is (will fail at API call time with clear error)
             return { provider: catalogModel.provider, model: catalogModel.id };
         }
         // Default to the auto-selected default provider
@@ -717,6 +742,38 @@ class AIProviderManager {
         const defaultModel = defaultConfig?.defaultModel || modelOrAlias;
         return { provider: this.defaultProvider, model: defaultModel };
     }
+    /**
+     * Get the default model for a provider, respecting runtime config.
+     * For Azure, uses the deployment name instead of the static "gpt-4o".
+     */
+    getDefaultModelForProvider(provider) {
+        if (provider === 'azure') {
+            const config = this.configs.get('azure');
+            return config?.deploymentName
+                || config?.defaultModel
+                || process.env.AZURE_OPENAI_DEPLOYMENT_NAME
+                || PROVIDER_DEFAULT_MODEL.azure
+                || 'gpt-4o';
+        }
+        return PROVIDER_DEFAULT_MODEL[provider] || 'llama3.2';
+    }
+    /**
+     * Resolve a model name to an Azure deployment.
+     * Azure uses deployment names (e.g. "gpt4o") instead of model IDs (e.g. "gpt-4o").
+     * Checks: configured deployment name → env var → strip dashes from model name.
+     */
+    resolveToAzureDeployment(modelOrAlias) {
+        const azureConfig = this.configs.get('azure');
+        // Use configured deployment name if it looks like it matches the requested model
+        const deployment = azureConfig?.deploymentName
+            || process.env.AZURE_OPENAI_DEPLOYMENT_NAME;
+        if (deployment) {
+            return { provider: 'azure', model: deployment };
+        }
+        // Try stripping dashes (gpt-4o → gpt4o) as Azure often uses this convention
+        const stripped = modelOrAlias.replace(/-/g, '');
+        return { provider: 'azure', model: stripped };
+    }
     async getModel(provider, modelId) {
         await this.ensureProvider(provider);
         const key = this.providerKey(provider);
@@ -758,9 +815,12 @@ class AIProviderManager {
         // Azure uses .chat() method for deployment-based access
         if (provider === 'azure') {
             let deployment = modelId;
-            if (modelId === 'default') {
+            if (deployment === 'default' || !deployment) {
                 const azureConfig = this.configs.get('azure');
-                deployment = azureConfig?.defaultModel || modelId;
+                deployment = azureConfig?.defaultModel
+                    || azureConfig?.deploymentName
+                    || process.env.AZURE_OPENAI_DEPLOYMENT_NAME
+                    || 'default';
                 if (deployment === 'default') {
                     throw new Error('Azure deployment not configured. Set defaultModel in Azure config or AZURE_OPENAI_DEPLOYMENT_NAME.');
                 }
@@ -793,7 +853,7 @@ class AIProviderManager {
         // Resolve model
         const modelRef = request.model
             ? this.resolveModel(request.model)
-            : { provider: this.defaultProvider, model: PROVIDER_DEFAULT_MODEL[this.defaultProvider] || 'llama3.2' };
+            : { provider: this.defaultProvider, model: this.getDefaultModelForProvider(this.defaultProvider) };
         const model = await this.getModel(modelRef.provider, modelRef.model);
         // Convert messages to AI SDK format
         const messages = request.messages.map((msg) => ({
@@ -807,7 +867,7 @@ class AIProviderManager {
         try {
             const result = await withRetry(() => generateText({
                 model,
-                messages,
+                messages: applyAnthropicMessageCache(messages, modelRef.provider),
                 temperature: request.temperature ?? 0.7,
                 maxOutputTokens: request.maxTokens ?? 4096,
                 abortSignal: AbortSignal.timeout(AI_TIMEOUT_MS),
@@ -815,12 +875,9 @@ class AIProviderManager {
             const duration = Date.now() - startTime;
             const modelInfo = this.getModelInfo(modelRef.model);
             // Calculate cost - AI SDK v6 uses usage.totalTokens and similar
-            const promptTokens = result.usage?.promptTokens ??
-                result.usage?.inputTokens ?? 0;
-            const completionTokens = result.usage?.completionTokens ??
-                result.usage?.outputTokens ?? 0;
+            const { promptTokens, completionTokens, cacheReadTokens, cacheWriteTokens } = extractCacheUsage(result.usage);
             const cost = modelInfo
-                ? (promptTokens / 1_000_000) * modelInfo.costPer1MInput +
+                ? calculateCachedInputCost(promptTokens, cacheReadTokens, cacheWriteTokens, modelInfo.costPer1MInput, getCacheConfigForModel(modelInfo.id)) +
                     (completionTokens / 1_000_000) * modelInfo.costPer1MOutput
                 : 0;
             return {
@@ -834,6 +891,8 @@ class AIProviderManager {
                     completionTokens,
                     totalTokens: promptTokens + completionTokens,
                     cost,
+                    cacheReadTokens,
+                    cacheWriteTokens,
                 },
                 duration,
             };
@@ -851,7 +910,7 @@ class AIProviderManager {
         // Resolve model
         const modelRef = request.model
             ? this.resolveModel(request.model)
-            : { provider: this.defaultProvider, model: PROVIDER_DEFAULT_MODEL[this.defaultProvider] || 'llama3.2' };
+            : { provider: this.defaultProvider, model: this.getDefaultModelForProvider(this.defaultProvider) };
         const model = await this.getModel(modelRef.provider, modelRef.model);
         // Convert messages
         const messages = request.messages.map((msg) => ({
@@ -864,7 +923,7 @@ class AIProviderManager {
         try {
             const result = streamText({
                 model,
-                messages,
+                messages: applyAnthropicMessageCache(messages, modelRef.provider),
                 temperature: request.temperature ?? 0.7,
                 maxOutputTokens: request.maxTokens ?? 4096,
                 abortSignal: AbortSignal.timeout(AI_TIMEOUT_MS),
@@ -894,12 +953,9 @@ class AIProviderManager {
             const usage = await result.usage;
             const duration = Date.now() - startTime;
             const modelInfo = this.getModelInfo(modelRef.model);
-            const promptTokens = usage?.promptTokens ??
-                usage?.inputTokens ?? 0;
-            const completionTokens = usage?.completionTokens ??
-                usage?.outputTokens ?? 0;
+            const { promptTokens, completionTokens, cacheReadTokens, cacheWriteTokens } = extractCacheUsage(usage);
             const cost = modelInfo
-                ? (promptTokens / 1_000_000) * modelInfo.costPer1MInput +
+                ? calculateCachedInputCost(promptTokens, cacheReadTokens, cacheWriteTokens, modelInfo.costPer1MInput, getCacheConfigForModel(modelInfo.id)) +
                     (completionTokens / 1_000_000) * modelInfo.costPer1MOutput
                 : 0;
             return {
@@ -913,6 +969,8 @@ class AIProviderManager {
                     completionTokens,
                     totalTokens: promptTokens + completionTokens,
                     cost,
+                    cacheReadTokens,
+                    cacheWriteTokens,
                 },
                 duration,
             };
@@ -931,7 +989,7 @@ class AIProviderManager {
         // Resolve model
         const modelRef = request.model
             ? this.resolveModel(request.model)
-            : { provider: this.defaultProvider, model: PROVIDER_DEFAULT_MODEL[this.defaultProvider] || 'llama3.2' };
+            : { provider: this.defaultProvider, model: this.getDefaultModelForProvider(this.defaultProvider) };
         const model = await this.getModel(modelRef.provider, modelRef.model);
         // Check if model supports native tools
         const toolSupport = this.modelSupportsTools(modelRef.model || modelRef.provider + '/' + modelRef.model);
@@ -980,10 +1038,10 @@ class AIProviderManager {
         try {
             const result = await withRetry(() => generateText({
                 model,
-                messages,
+                messages: applyAnthropicMessageCache(messages, modelRef.provider),
                 temperature: request.temperature ?? 0.7,
                 maxOutputTokens: request.maxTokens ?? 4096,
-                tools: useNativeTools ? aiSdkTools : undefined,
+                tools: useNativeTools ? applyAnthropicToolCache(aiSdkTools, modelRef.provider) : undefined,
                 abortSignal: AbortSignal.timeout(AI_TIMEOUT_MS),
             }));
             let finalContent = result.text;
@@ -1042,12 +1100,9 @@ class AIProviderManager {
             }
             const duration = Date.now() - startTime;
             const modelInfo = this.getModelInfo(modelRef.model);
-            const promptTokens = result.usage?.promptTokens ??
-                result.usage?.inputTokens ?? 0;
-            const completionTokens = result.usage?.completionTokens ??
-                result.usage?.outputTokens ?? 0;
+            const { promptTokens, completionTokens, cacheReadTokens, cacheWriteTokens } = extractCacheUsage(result.usage);
             const cost = modelInfo
-                ? (promptTokens / 1_000_000) * modelInfo.costPer1MInput +
+                ? calculateCachedInputCost(promptTokens, cacheReadTokens, cacheWriteTokens, modelInfo.costPer1MInput, getCacheConfigForModel(modelInfo.id)) +
                     (completionTokens / 1_000_000) * modelInfo.costPer1MOutput
                 : 0;
             return {
@@ -1061,6 +1116,8 @@ class AIProviderManager {
                     completionTokens,
                     totalTokens: promptTokens + completionTokens,
                     cost,
+                    cacheReadTokens,
+                    cacheWriteTokens,
                 },
                 duration,
                 toolCalls: detectedToolCalls.length > 0 ? detectedToolCalls : undefined,
@@ -1078,7 +1135,7 @@ class AIProviderManager {
     modelSupportsTools(modelOrAlias) {
         const modelRef = modelOrAlias
             ? this.resolveModel(modelOrAlias)
-            : { provider: this.defaultProvider, model: PROVIDER_DEFAULT_MODEL[this.defaultProvider] || 'llama3.2' };
+            : { provider: this.defaultProvider, model: this.getDefaultModelForProvider(this.defaultProvider) };
         // Azure OpenAI: All GPT-4, GPT-4o, and GPT-3.5-turbo models support function calling
         // We check provider first since Azure deployment names won't be in MODEL_CATALOG
         if (modelRef.provider === 'azure') {
@@ -1123,7 +1180,7 @@ class AIProviderManager {
         // Resolve model
         const modelRef = request.model
             ? this.resolveModel(request.model)
-            : { provider: this.defaultProvider, model: PROVIDER_DEFAULT_MODEL[this.defaultProvider] || 'llama3.2' };
+            : { provider: this.defaultProvider, model: this.getDefaultModelForProvider(this.defaultProvider) };
         const model = await this.getModel(modelRef.provider, modelRef.model);
         // Check if model supports native tool calling
         const toolSupport = this.modelSupportsTools(request.model);
@@ -1234,22 +1291,19 @@ class AIProviderManager {
         try {
             const result = await withRetry(() => generateText({
                 model,
-                messages,
-                tools: Object.keys(tools).length > 0 ? tools : undefined,
+                messages: applyAnthropicMessageCache(messages, modelRef.provider),
+                tools: Object.keys(tools).length > 0 ? applyAnthropicToolCache(tools, modelRef.provider) : undefined,
                 temperature: request.temperature ?? 0.7,
                 maxOutputTokens: request.maxTokens ?? 4096,
-                maxSteps: request.maxToolRoundtrips ?? 5,
+                stopWhen: stepCountIs(request.maxToolRoundtrips ?? 5),
                 abortSignal: AbortSignal.timeout(AI_TIMEOUT_MS),
             }));
             const duration = Date.now() - startTime;
             const modelInfo = this.getModelInfo(modelRef.model);
-            // Calculate usage
-            const promptTokens = result.usage?.promptTokens ??
-                result.usage?.inputTokens ?? 0;
-            const completionTokens = result.usage?.completionTokens ??
-                result.usage?.outputTokens ?? 0;
+            // Calculate usage, summed across all steps (result.usage is last-step only)
+            const { promptTokens, completionTokens, cacheReadTokens, cacheWriteTokens } = aggregateStepUsage(result.steps, result.usage);
             const cost = modelInfo
-                ? (promptTokens / 1_000_000) * modelInfo.costPer1MInput +
+                ? calculateCachedInputCost(promptTokens, cacheReadTokens, cacheWriteTokens, modelInfo.costPer1MInput, getCacheConfigForModel(modelInfo.id)) +
                     (completionTokens / 1_000_000) * modelInfo.costPer1MOutput
                 : 0;
             // Check if any tool calls require approval
@@ -1269,6 +1323,8 @@ class AIProviderManager {
                     completionTokens,
                     totalTokens: promptTokens + completionTokens,
                     cost,
+                    cacheReadTokens,
+                    cacheWriteTokens,
                 },
                 duration,
                 toolCalls: toolCallHistory.length > 0 ? toolCallHistory : undefined,
@@ -1297,7 +1353,7 @@ class AIProviderManager {
         // Resolve model
         const modelRef = request.model
             ? this.resolveModel(request.model)
-            : { provider: this.defaultProvider, model: PROVIDER_DEFAULT_MODEL[this.defaultProvider] || 'llama3.2' };
+            : { provider: this.defaultProvider, model: this.getDefaultModelForProvider(this.defaultProvider) };
         const model = await this.getModel(modelRef.provider, modelRef.model);
         // Convert messages
         const messages = toAiMessages(request.messages);
@@ -1379,11 +1435,11 @@ class AIProviderManager {
         try {
             const result = streamText({
                 model,
-                messages,
-                tools: Object.keys(tools).length > 0 ? tools : undefined,
+                messages: applyAnthropicMessageCache(messages, modelRef.provider),
+                tools: Object.keys(tools).length > 0 ? applyAnthropicToolCache(tools, modelRef.provider) : undefined,
                 temperature: request.temperature ?? 0.7,
                 maxOutputTokens: request.maxTokens ?? 4096,
-                maxSteps: request.maxToolRoundtrips ?? 5,
+                stopWhen: stepCountIs(request.maxToolRoundtrips ?? 5),
                 abortSignal: AbortSignal.timeout(AI_TIMEOUT_MS),
             });
             let fullText = '';
@@ -1394,12 +1450,9 @@ class AIProviderManager {
             const usage = await result.usage;
             const duration = Date.now() - startTime;
             const modelInfo = this.getModelInfo(modelRef.model);
-            const promptTokens = usage?.promptTokens ??
-                usage?.inputTokens ?? 0;
-            const completionTokens = usage?.completionTokens ??
-                usage?.outputTokens ?? 0;
+            const { promptTokens, completionTokens, cacheReadTokens, cacheWriteTokens } = extractCacheUsage(usage);
             const cost = modelInfo
-                ? (promptTokens / 1_000_000) * modelInfo.costPer1MInput +
+                ? calculateCachedInputCost(promptTokens, cacheReadTokens, cacheWriteTokens, modelInfo.costPer1MInput, getCacheConfigForModel(modelInfo.id)) +
                     (completionTokens / 1_000_000) * modelInfo.costPer1MOutput
                 : 0;
             const pendingApprovals = toolResultHistory.filter(r => r.approvalRequired);
@@ -1418,6 +1471,8 @@ class AIProviderManager {
                     completionTokens,
                     totalTokens: promptTokens + completionTokens,
                     cost,
+                    cacheReadTokens,
+                    cacheWriteTokens,
                 },
                 duration,
                 toolCalls: toolCallHistory.length > 0 ? toolCallHistory : undefined,

package/dist/providers/core/models.js

@@ -7,8 +7,11 @@ import { ProviderType } from './types.js';
 // Model Aliases
 export const MODEL_ALIASES = {
     // Anthropic
-    opus: { provider: 'anthropic', model: 'claude-opus-4-6' },
-    sonnet: { provider: 'anthropic', model: 'claude-sonnet-4-5-20250929' },
+    opus: { provider: 'anthropic', model: 'claude-opus-5-5' },
+    sonnet: { provider: 'anthropic', model: 'claude-sonnet-5' },
+    fable: { provider: 'anthropic', model: 'claude-fable-5-1' },
+    'opus-legacy': { provider: 'anthropic', model: 'claude-opus-4-6' },
+    'sonnet-legacy': { provider: 'anthropic', model: 'claude-sonnet-4-5-20250929' },
     haiku: { provider: 'anthropic', model: 'claude-haiku-4-5-20251001' },
     // OpenAI
     gpt: { provider: 'openai', model: 'gpt-4o' },
@@ -32,8 +35,11 @@ export const MODEL_ALIASES = {
     local: { provider: 'ollama', model: 'llama3.2' },
     llama: { provider: 'ollama', model: 'llama3.2' },
     'deepseek-local': { provider: 'ollama', model: 'deepseek-r1:7b' },
-    qwen: { provider: 'ollama', model: 'qwen2.5:14b' },
+    qwen: { provider: 'ollama', model: 'qwen3:4b' },
+    'qwen-14b': { provider: 'ollama', model: 'qwen2.5:14b' },
     'mistral-local': { provider: 'ollama', model: 'mistral:7b' },
+    gemma4: { provider: 'ollama', model: 'gemma4:e4b' },
+    'gemma4-26b': { provider: 'ollama', model: 'gemma4:26b' },
     // xAI (Grok)
     grok: { provider: 'xai', model: 'grok-2' },
     'grok-3': { provider: 'xai', model: 'grok-3' },
@@ -110,6 +116,45 @@ export const MODEL_ALIASES = {
 // Model Catalog
 export const MODEL_CATALOG = [
     // Anthropic - All support native function calling
+    // Current Anthropic models. ids, context, max output and prices verified against
+    // https://platform.claude.com/docs/en/about-claude/models/overview (2026-09-26).
+    {
+        id: 'claude-fable-5-1',
+        name: 'Claude Fable 5.1',
+        provider: 'anthropic',
+        contextWindow: 1000000,
+        maxOutput: 128000,
+        supportsVision: true,
+        supportsStreaming: true,
+        supportsTools: true,
+        costPer1MInput: 10,
+        costPer1MOutput: 50,
+    },
+    {
+        id: 'claude-opus-5-5',
+        name: 'Claude Opus 5.5',
+        provider: 'anthropic',
+        contextWindow: 1000000,
+        maxOutput: 128000,
+        supportsVision: true,
+        supportsStreaming: true,
+        supportsTools: true,
+        costPer1MInput: 4,
+        costPer1MOutput: 20,
+    },
+    {
+        id: 'claude-sonnet-5',
+        name: 'Claude Sonnet 5',
+        provider: 'anthropic',
+        contextWindow: 1000000,
+        maxOutput: 128000,
+        supportsVision: true,
+        supportsStreaming: true,
+        supportsTools: true,
+        costPer1MInput: 2,
+        costPer1MOutput: 10,
+    },
+    // Legacy entries kept because other code and tests reference them.
     {
         id: 'claude-opus-4-6',
         name: 'Claude Opus 4.6',
@@ -139,12 +184,12 @@ export const MODEL_CATALOG = [
         name: 'Claude Haiku 4.5',
         provider: 'anthropic',
         contextWindow: 200000,
-        maxOutput: 8192,
+        maxOutput: 64000,
         supportsVision: true,
         supportsStreaming: true,
         supportsTools: true,
-        costPer1MInput: 0.25,
-        costPer1MOutput: 1.25,
+        costPer1MInput: 1,
+        costPer1MOutput: 5,
     },
     // OpenAI - GPT-4+ supports function calling
     {
@@ -378,6 +423,44 @@ export const MODEL_CATALOG = [
         costPer1MInput: 30.0,
         costPer1MOutput: 60.0,
     },
+    // Gemma 4 (local): supports native tool calling + multimodal
+    {
+        id: 'gemma4:e4b',
+        name: 'Gemma 4 E4B (Local)',
+        provider: 'ollama',
+        contextWindow: 128000,
+        maxOutput: 8192,
+        supportsVision: true,
+        supportsStreaming: true,
+        supportsTools: true,
+        costPer1MInput: 0,
+        costPer1MOutput: 0,
+    },
+    {
+        id: 'gemma4:26b',
+        name: 'Gemma 4 26B (Local)',
+        provider: 'ollama',
+        contextWindow: 128000,
+        maxOutput: 8192,
+        supportsVision: true,
+        supportsStreaming: true,
+        supportsTools: true,
+        costPer1MInput: 0,
+        costPer1MOutput: 0,
+    },
+    // Qwen 3 (local): best small model for tool calling
+    {
+        id: 'qwen3:4b',
+        name: 'Qwen 3 4B (Local)',
+        provider: 'ollama',
+        contextWindow: 32768,
+        maxOutput: 8192,
+        supportsVision: false,
+        supportsStreaming: true,
+        supportsTools: true,
+        costPer1MInput: 0,
+        costPer1MOutput: 0,
+    },
     // Ollama (local) - Most models don't reliably support native tool calling
     {
         id: 'llama3.2',
@@ -884,6 +967,31 @@ export const MODEL_CATALOG = [
         costPer1MInput: 0.40,
         costPer1MOutput: 1.20,
     },
+    // Cerebras (fast inference)
+    {
+        id: 'llama3.1-8b',
+        name: 'Llama 3.1 8B (Cerebras)',
+        provider: 'cerebras',
+        contextWindow: 8192,
+        maxOutput: 8192,
+        supportsVision: false,
+        supportsStreaming: true,
+        supportsTools: false,
+        costPer1MInput: 0.10,
+        costPer1MOutput: 0.10,
+    },
+    {
+        id: 'llama-3.3-70b',
+        name: 'Llama 3.3 70B (Cerebras)',
+        provider: 'cerebras',
+        contextWindow: 8192,
+        maxOutput: 8192,
+        supportsVision: false,
+        supportsStreaming: true,
+        supportsTools: true,
+        costPer1MInput: 0.60,
+        costPer1MOutput: 0.60,
+    },
     // IBM Watsonx
     {
         id: 'watsonx/ibm/granite-13b-chat-v2',

package/dist/providers/core/types.d.ts

@@ -55,13 +55,20 @@ export interface ChatResponse {
         completionTokens: number;
         totalTokens: number;
         cost: number;
+        /** Prompt tokens served from the provider cache (subset of promptTokens) */
+        cacheReadTokens?: number;
+        /** Prompt tokens written to the provider cache (subset of promptTokens) */
+        cacheWriteTokens?: number;
     };
     duration: number;
 }
+export type ToolTier = 'essential' | 'standard' | 'full';
+export type ModelCapabilityLevel = 'basic' | 'instruction' | 'reasoning';
 export interface ToolDefinition {
     name: string;
     description: string;
     parameters?: Record<string, unknown>;
+    tier?: ToolTier;
 }
 export interface ToolCall {
     id: string;

package/dist/providers/index.d.ts

@@ -4,4 +4,5 @@
  * Multi-provider AI chat using Vercel AI SDK.
  */
 export { aiProvider, chat, chatStream, ProviderType, MODEL_ALIASES, MODEL_CATALOG, type ProviderConfig, type ModelInfo, type ChatMessage, type ChatRequest, type ChatResponse, type NativeToolCall, type NativeToolResult, type NativeToolDefinition, type ToolExecutionResult, } from './ai-sdk.js';
+export { routeQuery, classifyComplexity, selectModel, getRoutingStats, recordRoutingDecision, configureSmartRouter, isSmartRouterEnabled, getCostComparison, getCostOptimizationAdvice, resolveDefaultModel, type ComplexityTier, type ComplexityResult, type RoutingDecision, type SmartRouterConfig, type ProviderRecommendation, } from './smart-router.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/providers/index.js

@@ -4,4 +4,5 @@
  * Multi-provider AI chat using Vercel AI SDK.
  */
 export { aiProvider, chat, chatStream, ProviderType, MODEL_ALIASES, MODEL_CATALOG, } from './ai-sdk.js';
+export { routeQuery, classifyComplexity, selectModel, getRoutingStats, recordRoutingDecision, configureSmartRouter, isSmartRouterEnabled, getCostComparison, getCostOptimizationAdvice, resolveDefaultModel, } from './smart-router.js';
 //# sourceMappingURL=index.js.map

package/dist/providers/prompt-cache.d.ts

@@ -0,0 +1,68 @@
+/**
+ * Prompt caching helpers
+ *
+ * Anthropic requires explicit cache breakpoints (providerOptions.anthropic.cacheControl).
+ * The cached prefix order is: tools, then system, then messages. A breakpoint on the
+ * last system message therefore caches tool schemas and the system prompt together;
+ * an extra breakpoint on the last tool lets the tool block hit even if the system
+ * prompt changes. OpenAI caches automatically and only reports cached token counts.
+ *
+ * Cache multipliers are relative to the base input price and are env-configurable.
+ */
+import type { ModelMessage, ToolSet } from 'ai';
+export interface PromptCacheConfig {
+    enabled: boolean;
+    ttl: '5m' | '1h';
+    /** Cost of a cache read relative to base input price */
+    readMultiplier: number;
+    /** Cost of a cache write relative to base input price */
+    writeMultiplier: number;
+}
+export interface CacheUsage {
+    /** Total prompt tokens, including cached reads and writes */
+    promptTokens: number;
+    completionTokens: number;
+    cacheReadTokens: number;
+    cacheWriteTokens: number;
+}
+/** Read cache settings from env on each call so tests and runtime changes apply. */
+export declare function getPromptCacheConfig(env?: NodeJS.ProcessEnv): PromptCacheConfig;
+/**
+ * Cache config for one model. An explicit PROFCLAW_CACHE_READ_MULTIPLIER always wins,
+ * otherwise the model's documented read multiplier applies.
+ */
+export declare function getCacheConfigForModel(model: string, env?: NodeJS.ProcessEnv): PromptCacheConfig;
+/**
+ * Add cache breakpoints to the leading system block: one on the first system message
+ * (the stable prompt) and one on the last leading system message. Callers put volatile
+ * per-request context in a later system message so the first breakpoint still hits
+ * across requests. History after the system block stays outside the cached prefix.
+ * Returns a new array; inputs are not mutated.
+ */
+export declare function applyAnthropicMessageCache<T extends ModelMessage>(messages: T[], provider: string, config?: PromptCacheConfig): T[];
+/**
+ * Return tools in a deterministic (name-sorted) order with a cache breakpoint on the
+ * last one. Object key order in the request body affects the prefix bytes, so sorting
+ * keeps the tool block byte-stable across steps and sessions.
+ */
+export declare function applyAnthropicToolCache<T extends ToolSet>(tools: T, provider: string, config?: PromptCacheConfig): T;
+/**
+ * Normalize AI SDK usage (v6 shape, legacy promptTokens shape, or raw provider
+ * counters) into total prompt tokens plus cache read/write counts.
+ * In AI SDK v6, inputTokens already includes cached reads and writes.
+ */
+export declare function extractCacheUsage(usage: unknown): CacheUsage;
+/**
+ * Sum usage across multi-step runs (tokens plus cache read/write). Falls back to
+ * `fallback` (the run-level usage) when there are no per-step usages, because
+ * result.usage only reflects the last step in AI SDK v6.
+ */
+export declare function aggregateStepUsage(steps: unknown, fallback: unknown): CacheUsage;
+/**
+ * Input cost in USD for a prompt that may include cached tokens.
+ * Uncached tokens bill at the base rate, reads and writes at the multiplied rate.
+ */
+export declare function calculateCachedInputCost(promptTokens: number, cacheReadTokens: number, cacheWriteTokens: number, inputRatePer1M: number, config?: PromptCacheConfig): number;
+/** Fraction of prompt tokens served from cache (0 when there were no prompt tokens). */
+export declare function cacheHitRate(cacheReadTokens: number, promptTokens: number): number;
+//# sourceMappingURL=prompt-cache.d.ts.map