@build-astron-co/nimbus 0.2.0 → 0.4.0

package/dist/src/llm/config-loader.js

@@ -0,0 +1,172 @@
+/**
+ * LLM Config Loader
+ *
+ * Reads ~/.nimbus/config.yaml and maps LLM-related settings
+ * to RouterConfig for the LLM Router.
+ */
+import * as fs from 'fs';
+import * as path from 'path';
+import * as os from 'os';
+import { logger } from '../utils';
+/**
+ * Load LLM configuration from ~/.nimbus/config.yaml
+ * Returns a Partial<RouterConfig> that can be passed to LLMRouter constructor
+ */
+export function loadLLMConfig() {
+    const configPath = process.env.NIMBUS_CONFIG_PATH || path.join(os.homedir(), '.nimbus', 'config.yaml');
+    if (!fs.existsSync(configPath)) {
+        logger.info(`No config file found at ${configPath}, using defaults`);
+        return {};
+    }
+    try {
+        const content = fs.readFileSync(configPath, 'utf-8');
+        const parsed = parseSimpleYaml(content);
+        const llmSection = parsed.llm || {};
+        const config = {};
+        // Map default provider
+        if (llmSection.default_provider) {
+            config.defaultProvider = llmSection.default_provider;
+        }
+        // Map default model
+        if (llmSection.defaultModel) {
+            config.defaultModel = llmSection.defaultModel;
+        }
+        // Map cost optimization settings
+        const costOpt = llmSection.cost_optimization;
+        if (costOpt) {
+            config.costOptimization = {
+                enabled: costOpt.enabled ?? false,
+                cheapModel: costOpt.cheap_model || '',
+                expensiveModel: costOpt.expensive_model || '',
+                cheapModelFor: costOpt.use_cheap_model_for || [],
+                expensiveModelFor: costOpt.use_expensive_model_for || [],
+            };
+        }
+        // Map fallback settings
+        const fallback = llmSection.fallback;
+        if (fallback) {
+            config.fallback = {
+                enabled: fallback.enabled ?? true,
+                providers: fallback.providers || [],
+            };
+        }
+        // Map token budget
+        if (llmSection.maxTokens) {
+            config.tokenBudget = {
+                maxTokensPerRequest: llmSection.maxTokens,
+            };
+        }
+        logger.info(`Loaded LLM config from ${configPath}`);
+        return config;
+    }
+    catch (error) {
+        logger.warn(`Failed to load LLM config from ${configPath}: ${error.message}`);
+        return {};
+    }
+}
+/**
+ * Minimal YAML parser for flat/nested key-value configs.
+ * Handles the subset of YAML used by Nimbus config files.
+ */
+function parseSimpleYaml(content) {
+    const result = {};
+    const lines = content.split('\n');
+    // Stack tracks: indent level, the parent object, and the last key set on that object
+    const stack = [
+        { indent: -1, obj: result },
+    ];
+    for (const rawLine of lines) {
+        // Skip comments and empty lines
+        const commentIdx = rawLine.indexOf('#');
+        const line = commentIdx >= 0 ? rawLine.slice(0, commentIdx) : rawLine;
+        if (line.trim() === '') {
+            continue;
+        }
+        const indent = line.length - line.trimStart().length;
+        const trimmed = line.trim();
+        // Handle list items (- value)
+        if (trimmed.startsWith('- ')) {
+            const listValue = trimmed.slice(2).trim();
+            // Pop stack to find the owner of the list key.
+            // If we're inside an empty object created by "key:" with no value,
+            // pop past it to find the frame that owns the key.
+            while (stack.length > 1 && stack[stack.length - 1].indent >= indent) {
+                stack.pop();
+            }
+            // If the current frame has no lastKey, it's a placeholder empty object --
+            // pop one more to reach the frame that holds the actual key reference.
+            let frame = stack[stack.length - 1];
+            if (!frame.lastKey && stack.length > 1) {
+                stack.pop();
+                frame = stack[stack.length - 1];
+            }
+            const parentObj = frame.obj;
+            const lastKey = frame.lastKey;
+            if (lastKey) {
+                if (!Array.isArray(parentObj[lastKey])) {
+                    parentObj[lastKey] = [];
+                }
+                parentObj[lastKey].push(parseYamlValue(listValue));
+            }
+            continue;
+        }
+        // Handle key: value pairs
+        const colonIdx = trimmed.indexOf(':');
+        if (colonIdx === -1) {
+            continue;
+        }
+        const key = trimmed.slice(0, colonIdx).trim();
+        const rawValue = trimmed.slice(colonIdx + 1).trim();
+        // Pop stack to find parent at correct indent level
+        while (stack.length > 1 && stack[stack.length - 1].indent >= indent) {
+            stack.pop();
+        }
+        const frame = stack[stack.length - 1];
+        const current = frame.obj;
+        if (rawValue === '' || rawValue === undefined) {
+            // Nested object (or potentially a list -- will be converted if list items follow)
+            current[key] = {};
+            frame.lastKey = key;
+            stack.push({ indent, obj: current[key] });
+        }
+        else {
+            current[key] = parseYamlValue(rawValue);
+            frame.lastKey = key;
+        }
+    }
+    return result;
+}
+/**
+ * Parse a YAML scalar value
+ */
+function parseYamlValue(value) {
+    // Remove surrounding quotes
+    if ((value.startsWith('"') && value.endsWith('"')) ||
+        (value.startsWith("'") && value.endsWith("'"))) {
+        return value.slice(1, -1);
+    }
+    // Boolean
+    if (value === 'true') {
+        return true;
+    }
+    if (value === 'false') {
+        return false;
+    }
+    // Null
+    if (value === 'null' || value === '~') {
+        return null;
+    }
+    // Number
+    const num = Number(value);
+    if (!isNaN(num) && value !== '') {
+        return num;
+    }
+    // Inline list [a, b, c]
+    if (value.startsWith('[') && value.endsWith(']')) {
+        return value
+            .slice(1, -1)
+            .split(',')
+            .map(v => parseYamlValue(v.trim()));
+    }
+    return value;
+}

package/dist/src/llm/cost-calculator.js

@@ -0,0 +1,137 @@
+/**
+ * Per-request cost calculation for LLM providers
+ *
+ * Pricing data is based on published rates from each provider.
+ * Prices are per 1K tokens (input / output).
+ * Ollama models are local and free.
+ * Unknown models return 0 cost with a warning log.
+ */
+import { logger } from '../utils';
+/**
+ * Pricing data by provider and model.
+ * Prices are in USD per 1,000 tokens.
+ */
+const PRICING = {
+    anthropic: {
+        // Claude 4 family
+        'claude-4-opus': [0.015, 0.075],
+        'claude-opus-4-20250514': [0.015, 0.075],
+        'claude-4-sonnet': [0.003, 0.015],
+        'claude-sonnet-4-20250514': [0.003, 0.015],
+        'claude-4-haiku': [0.0008, 0.004],
+        'claude-haiku-4-20250514': [0.0008, 0.004],
+        // Claude 3.5 family
+        'claude-3.5-sonnet': [0.003, 0.015],
+        'claude-3-5-sonnet-20241022': [0.003, 0.015],
+        'claude-3.5-haiku': [0.0008, 0.004],
+        'claude-3-5-haiku-20241022': [0.0008, 0.004],
+        // Claude 3 family
+        'claude-3-opus': [0.015, 0.075],
+        'claude-3-opus-20240229': [0.015, 0.075],
+        'claude-3-haiku': [0.00025, 0.00125],
+        'claude-3-haiku-20240307': [0.00025, 0.00125],
+    },
+    openai: {
+        'gpt-4': [0.03, 0.06],
+        'gpt-4-turbo': [0.01, 0.03],
+        'gpt-4-turbo-2024-04-09': [0.01, 0.03],
+        'gpt-4o': [0.005, 0.015],
+        'gpt-4o-2024-11-20': [0.005, 0.015],
+        'gpt-4o-mini': [0.00015, 0.0006],
+        'gpt-4o-mini-2024-07-18': [0.00015, 0.0006],
+        'gpt-3.5-turbo': [0.0005, 0.0015],
+    },
+    google: {
+        'gemini-pro': [0.00025, 0.0005],
+        'gemini-1.5-pro': [0.00125, 0.005],
+        'gemini-1.5-flash': [0.000075, 0.0003],
+        'gemini-2.0-flash-exp': [0.000075, 0.0003],
+    },
+    ollama: {
+    // All local models are free
+    },
+    groq: {
+        'llama-3.1-70b-versatile': [0.00059, 0.00079],
+        'llama-3.1-8b-instant': [0.00005, 0.00008],
+        'llama-3.3-70b-versatile': [0.00059, 0.00079],
+        'mixtral-8x7b-32768': [0.00024, 0.00024],
+        'gemma2-9b-it': [0.0002, 0.0002],
+    },
+    together: {
+        'meta-llama/Llama-3.1-70B-Instruct-Turbo': [0.00088, 0.00088],
+        'meta-llama/Llama-3.1-8B-Instruct-Turbo': [0.00018, 0.00018],
+        'meta-llama/Llama-3.3-70B-Instruct-Turbo': [0.00088, 0.00088],
+        'mistralai/Mixtral-8x22B-Instruct-v0.1': [0.0012, 0.0012],
+        'Qwen/Qwen2.5-72B-Instruct-Turbo': [0.0012, 0.0012],
+    },
+    deepseek: {
+        'deepseek-chat': [0.00014, 0.00028],
+        'deepseek-coder': [0.00014, 0.00028],
+        'deepseek-reasoner': [0.00055, 0.00219],
+    },
+    fireworks: {
+        'accounts/fireworks/models/llama-v3p1-70b-instruct': [0.0009, 0.0009],
+        'accounts/fireworks/models/llama-v3p1-8b-instruct': [0.0002, 0.0002],
+        'accounts/fireworks/models/mixtral-8x22b-instruct': [0.0009, 0.0009],
+        'accounts/fireworks/models/qwen2p5-72b-instruct': [0.0009, 0.0009],
+    },
+    perplexity: {
+        'llama-3.1-sonar-large-128k-online': [0.001, 0.001],
+        'llama-3.1-sonar-small-128k-online': [0.0002, 0.0002],
+        'llama-3.1-sonar-huge-128k-online': [0.005, 0.005],
+    },
+    openrouter: {
+        // OpenRouter models use provider/model format; prices approximate the underlying provider
+        'anthropic/claude-sonnet-4-20250514': [0.003, 0.015],
+        'anthropic/claude-haiku-4-20250514': [0.0008, 0.004],
+        'anthropic/claude-opus-4-20250514': [0.015, 0.075],
+        'anthropic/claude-3.5-sonnet': [0.003, 0.015],
+        'openai/gpt-4o': [0.005, 0.015],
+        'openai/gpt-4o-mini': [0.00015, 0.0006],
+        'google/gemini-2.0-flash-exp': [0.000075, 0.0003],
+        'meta-llama/llama-3.1-405b-instruct': [0.003, 0.003],
+        'meta-llama/llama-3.1-70b-instruct': [0.0008, 0.0008],
+        'mistralai/mixtral-8x22b-instruct': [0.0009, 0.0009],
+    },
+};
+/**
+ * Calculate the cost for a single LLM request.
+ *
+ * @param provider  - Provider name (e.g. "anthropic", "openai", "google", "ollama", "openrouter")
+ * @param model     - Model identifier (e.g. "claude-sonnet-4-20250514", "gpt-4o")
+ * @param inputTokens  - Number of input/prompt tokens
+ * @param outputTokens - Number of output/completion tokens
+ * @returns Cost result with total USD cost and per-component breakdown
+ */
+export function calculateCost(provider, model, inputTokens, outputTokens) {
+    // Ollama is always free
+    if (provider === 'ollama') {
+        return { costUSD: 0, breakdown: { input: 0, output: 0 } };
+    }
+    const providerPricing = PRICING[provider];
+    if (!providerPricing) {
+        logger.debug(`No pricing data for provider "${provider}", returning zero cost`);
+        return { costUSD: 0, breakdown: { input: 0, output: 0 } };
+    }
+    const entry = providerPricing[model];
+    if (!entry) {
+        logger.debug(`No pricing data for model "${model}" on provider "${provider}", returning zero cost`);
+        return { costUSD: 0, breakdown: { input: 0, output: 0 } };
+    }
+    const [inputPricePer1K, outputPricePer1K] = entry;
+    const inputCost = (inputTokens / 1000) * inputPricePer1K;
+    const outputCost = (outputTokens / 1000) * outputPricePer1K;
+    return {
+        costUSD: inputCost + outputCost,
+        breakdown: {
+            input: inputCost,
+            output: outputCost,
+        },
+    };
+}
+/**
+ * Get available pricing data (useful for debugging / admin endpoints).
+ */
+export function getPricingData() {
+    return PRICING;
+}

package/dist/src/llm/index.js

@@ -0,0 +1,7 @@
+export { LLMRouter } from './router';
+export * from './types';
+export { calculateCost } from './cost-calculator';
+export { loadLLMConfig } from './config-loader';
+export { resolveModelAlias, stripProviderPrefix } from './model-aliases';
+export { detectProvider } from './provider-registry';
+export * from './auth-bridge';

package/dist/src/llm/model-aliases.js

@@ -0,0 +1,99 @@
+/**
+ * Model alias resolution
+ * Maps short names to full model identifiers
+ */
+const MODEL_ALIASES = {
+    // Anthropic shortcuts
+    sonnet: 'claude-sonnet-4-20250514',
+    haiku: 'claude-haiku-4-20250514',
+    opus: 'claude-opus-4-20250514',
+    claude: 'claude-sonnet-4-20250514',
+    // OpenAI shortcuts
+    gpt4: 'gpt-4o',
+    gpt4o: 'gpt-4o',
+    gpt4mini: 'gpt-4o-mini',
+    '4o': 'gpt-4o',
+    '4o-mini': 'gpt-4o-mini',
+    // Google shortcuts
+    gemini: 'gemini-2.0-flash-exp',
+    'gemini-flash': 'gemini-2.0-flash-exp',
+    'gemini-pro': 'gemini-1.5-pro',
+    // Groq shortcuts
+    groq: 'llama-3.1-70b-versatile',
+    'groq-fast': 'llama-3.1-8b-instant',
+    // Together shortcuts
+    together: 'meta-llama/Llama-3.1-70B-Instruct-Turbo',
+    'together-llama': 'meta-llama/Llama-3.1-70B-Instruct-Turbo',
+    // DeepSeek shortcuts
+    deepseek: 'deepseek-chat',
+    'deepseek-coder': 'deepseek-coder',
+    'deepseek-r1': 'deepseek-reasoner',
+    // Fireworks shortcuts
+    fireworks: 'accounts/fireworks/models/llama-v3p1-70b-instruct',
+    // Perplexity shortcuts
+    perplexity: 'llama-3.1-sonar-large-128k-online',
+    pplx: 'llama-3.1-sonar-large-128k-online',
+    sonar: 'llama-3.1-sonar-large-128k-online',
+    // Local shortcuts
+    llama: 'llama3.2',
+    mistral: 'mistral',
+    codellama: 'codellama',
+};
+/**
+ * Resolve a model alias to its full model identifier.
+ * Returns the original string if no alias match is found.
+ *
+ * @param modelOrAlias - Short alias or full model identifier
+ * @returns Resolved full model identifier
+ */
+export function resolveModelAlias(modelOrAlias) {
+    return MODEL_ALIASES[modelOrAlias.toLowerCase()] || modelOrAlias;
+}
+/**
+ * Strip a "provider/" prefix from a model string.
+ *
+ * Model strings can be specified as "provider/model" (e.g., "anthropic/claude-sonnet-4-20250514").
+ * The provider prefix is used for routing but must be stripped before sending to the
+ * provider's API, which expects just the model ID (e.g., "claude-sonnet-4-20250514").
+ *
+ * Preserves multi-segment model IDs used by OpenRouter (e.g., "meta-llama/llama-3.1-405b")
+ * and Fireworks (e.g., "accounts/fireworks/models/llama-v3p3-70b-instruct") by only
+ * stripping prefixes that match known provider names.
+ *
+ * @param model - Model string, optionally with provider prefix
+ * @returns Model string with provider prefix removed
+ */
+export function stripProviderPrefix(model) {
+    if (!model.includes('/')) {
+        return model;
+    }
+    const slashIndex = model.indexOf('/');
+    const prefix = model.slice(0, slashIndex).toLowerCase();
+    // Only strip if the prefix is a known provider name.
+    // This preserves OpenRouter-style "meta-llama/llama-3.1-405b" and
+    // Fireworks-style "accounts/fireworks/models/..." paths.
+    const KNOWN_PROVIDERS = new Set([
+        'anthropic',
+        'openai',
+        'google',
+        'ollama',
+        'groq',
+        'together',
+        'deepseek',
+        'fireworks',
+        'perplexity',
+        'bedrock',
+    ]);
+    if (KNOWN_PROVIDERS.has(prefix)) {
+        return model.slice(slashIndex + 1);
+    }
+    return model;
+}
+/**
+ * Get a copy of all registered model aliases.
+ *
+ * @returns Record mapping alias names to full model identifiers
+ */
+export function getAliases() {
+    return { ...MODEL_ALIASES };
+}

package/dist/src/llm/provider-registry.js

@@ -0,0 +1,57 @@
+/**
+ * Provider Registry
+ * Auto-detects the provider from a model string
+ */
+/**
+ * Map of known provider prefixes used in "provider/model" format
+ */
+const PROVIDER_MAP = {
+    anthropic: 'anthropic',
+    openai: 'openai',
+    google: 'google',
+    'meta-llama': 'openrouter',
+    mistralai: 'openrouter',
+    groq: 'groq',
+    together: 'together',
+    deepseek: 'deepseek',
+    fireworks: 'fireworks',
+    perplexity: 'perplexity',
+};
+/**
+ * Detect the provider for a given model string.
+ *
+ * Detection strategy:
+ * 1. Explicit provider prefix (e.g., "groq/llama-3.1-70b") -- split on "/" and look up prefix
+ * 2. Model name pattern matching (e.g., "claude-*" -> anthropic, "gpt-*" -> openai)
+ * 3. Default to "anthropic" if no pattern matches
+ *
+ * @param model - Model identifier, optionally prefixed with "provider/"
+ * @returns Detected provider name
+ */
+export function detectProvider(model) {
+    // Explicit provider prefix (e.g., "groq/llama-3.1-70b")
+    if (model.includes('/')) {
+        const prefix = model.split('/')[0].toLowerCase();
+        return PROVIDER_MAP[prefix] || 'openrouter';
+    }
+    // Model name pattern matching
+    if (model.startsWith('claude')) {
+        return 'anthropic';
+    }
+    if (model.startsWith('gpt')) {
+        return 'openai';
+    }
+    if (model.startsWith('gemini')) {
+        return 'google';
+    }
+    if (model.startsWith('llama') ||
+        model.startsWith('mistral') ||
+        model.startsWith('codellama') ||
+        model.startsWith('phi')) {
+        return 'ollama';
+    }
+    if (model.startsWith('deepseek')) {
+        return 'deepseek';
+    }
+    return 'anthropic'; // default
+}