@sriinnu/kosha-discovery 0.5.0

package/dist/discovery/openrouter.js

@@ -0,0 +1,184 @@
+/**
+ * kosha-discovery — OpenRouter aggregator discoverer.
+ *
+ * OpenRouter is a unified gateway that proxies requests to many upstream
+ * providers (OpenAI, Anthropic, Google, Meta, Mistral, etc.).  Its model
+ * list endpoint returns rich metadata including pricing, context lengths,
+ * and architecture info — making it one of the most informative sources.
+ * @module
+ */
+import { BaseDiscoverer } from "./base.js";
+/**
+ * Discovers models available through the OpenRouter aggregator.
+ *
+ * OpenRouter acts as a single gateway to dozens of providers.  Its model
+ * list is publicly accessible (no API key required), though an
+ * authenticated request gets higher rate limits.
+ */
+export class OpenRouterDiscoverer extends BaseDiscoverer {
+    providerId = "openrouter";
+    providerName = "OpenRouter";
+    baseUrl = "https://openrouter.ai";
+    /**
+     * Fetch the full model catalogue from OpenRouter.
+     *
+     * @param credential - API key is **optional** for this provider.
+     *                     The endpoint works without auth, just with stricter rate limits.
+     * @param options    - Optional timeout override (default 15 s — larger because the
+     *                     response contains hundreds of models).
+     */
+    async discover(credential, options) {
+        const timeoutMs = this.validateTimeout(options?.timeout, 15_000);
+        const headers = {};
+        // API key is optional — the /api/v1/models endpoint is publicly
+        // accessible; auth only provides higher rate limits
+        const apiKey = credential.apiKey ?? credential.accessToken;
+        if (apiKey) {
+            headers.Authorization = `Bearer ${apiKey}`;
+        }
+        const response = await this.fetchJSON(`${this.baseUrl}/api/v1/models`, Object.keys(headers).length > 0 ? headers : undefined, timeoutMs);
+        return response.data
+            .filter((model) => this.isAvailable(model))
+            .map((model) => this.toModelCard(model));
+    }
+    /**
+     * Filter out unavailable models. OpenRouter marks delisted models
+     * with `pricing.prompt === "-1"`.
+     */
+    isAvailable(model) {
+        return model.pricing?.prompt !== "-1";
+    }
+    /** Convert an OpenRouter model object into a normalized {@link ModelCard}. */
+    toModelCard(model) {
+        const mode = this.inferMode(model);
+        const capabilities = this.inferCapabilities(model);
+        const pricing = this.parsePricing(model.pricing);
+        const originProvider = this.extractOriginProvider(model.id);
+        return this.makeCard({
+            id: model.id,
+            name: model.name || model.id,
+            provider: this.providerId,
+            originProvider,
+            mode,
+            capabilities,
+            contextWindow: model.context_length ?? 0,
+            maxOutputTokens: model.top_provider?.max_completion_tokens ?? 0,
+            pricing,
+        });
+    }
+    /**
+     * Extract the canonical origin-provider slug from an OpenRouter compound
+     * model ID of the form `{vendor}/{model-name}`.
+     *
+     * OpenRouter uses vendor-namespaced IDs (e.g. `anthropic/claude-opus-4-6`,
+     * `meta-llama/llama-3-70b`).  We take the prefix before the first `/` and
+     * normalise it to our canonical slug so that downstream consumers can group
+     * models by original creator regardless of which gateway serves them.
+     *
+     * @param modelId - Raw OpenRouter model ID.
+     * @returns Canonical provider slug (e.g. `"anthropic"`, `"meta"`, `"openai"`).
+     */
+    extractOriginProvider(modelId) {
+        const slashIndex = modelId.indexOf("/");
+        if (slashIndex === -1) {
+            // No vendor prefix — treat the whole ID as the slug
+            return modelId;
+        }
+        const prefix = modelId.slice(0, slashIndex);
+        // Map well-known OpenRouter vendor prefixes to canonical slugs.
+        // Unknown prefixes fall through and are returned as-is.
+        const vendorMap = {
+            anthropic: "anthropic",
+            openai: "openai",
+            google: "google",
+            "meta-llama": "meta",
+            mistralai: "mistral",
+            cohere: "cohere",
+            deepseek: "deepseek",
+            qwen: "qwen",
+        };
+        return vendorMap[prefix] ?? prefix;
+    }
+    /** Infer the primary {@link ModelMode} from modality and naming patterns. */
+    inferMode(model) {
+        const modality = (model.architecture?.modality ?? "").toLowerCase();
+        const id = model.id.toLowerCase();
+        const name = (model.name ?? "").toLowerCase();
+        if (modality.includes("image") && !modality.includes("text")) {
+            return "image";
+        }
+        if (modality.includes("audio")) {
+            return "audio";
+        }
+        if (id.includes("embed") || name.includes("embedding")) {
+            return "embedding";
+        }
+        return "chat";
+    }
+    /**
+     * Infer capability flags from architecture modality and model ID.
+     */
+    inferCapabilities(model) {
+        const modality = (model.architecture?.modality ?? "").toLowerCase();
+        const id = model.id.toLowerCase();
+        const capabilities = [];
+        if (id.includes("embed")) {
+            return ["embedding"];
+        }
+        // Image-generation-only models should advertise image capability.
+        if (modality.includes("image") && !modality.includes("text")) {
+            return ["image_generation"];
+        }
+        capabilities.push("chat");
+        // Vision from modality: "text+image" means multimodal input
+        if (modality.includes("image") && modality.includes("text")) {
+            capabilities.push("vision");
+        }
+        // Heuristic: well-known modern chat models support tool use and code
+        if (this.isModernChatModel(id)) {
+            capabilities.push("function_calling", "code", "nlu");
+        }
+        return capabilities;
+    }
+    /**
+     * Heuristic to identify "modern chat models" that support structured
+     * tool use (function calling), code generation, and NLU.
+     *
+     * We match on known provider/model family prefixes. This is intentionally
+     * broad — false positives are benign (extra capability flags), while false
+     * negatives would hide useful features.
+     */
+    isModernChatModel(id) {
+        return (id.includes("gpt-4") ||
+            id.includes("claude-3") ||
+            id.includes("claude-sonnet") ||
+            id.includes("claude-opus") ||
+            id.includes("gemini") ||
+            id.includes("command-r") ||
+            id.includes("mistral-large") ||
+            id.includes("llama-3") ||
+            id.includes("deepseek"));
+    }
+    /**
+     * Parse OpenRouter's per-token pricing into our per-million format.
+     *
+     * OpenRouter returns pricing as **cost per single token** (stringified
+     * floats).  We multiply by 1,000,000 to convert to our standard
+     * per-million-token pricing representation.
+     */
+    parsePricing(pricing) {
+        if (!pricing)
+            return undefined;
+        const promptCost = Number.parseFloat(pricing.prompt);
+        const completionCost = Number.parseFloat(pricing.completion);
+        if (Number.isNaN(promptCost) || Number.isNaN(completionCost)) {
+            return undefined;
+        }
+        return {
+            // Convert per-token cost to per-million-token cost
+            inputPerMillion: promptCost * 1_000_000,
+            outputPerMillion: completionCost * 1_000_000,
+        };
+    }
+}
+//# sourceMappingURL=openrouter.js.map

package/dist/discovery/openrouter.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"openrouter.js","sourceRoot":"","sources":["../../src/discovery/openrouter.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAGH,OAAO,EAAE,cAAc,EAAE,MAAM,WAAW,CAAC;AAwC3C;;;;;;GAMG;AACH,MAAM,OAAO,oBAAqB,SAAQ,cAAc;IAC9C,UAAU,GAAG,YAAY,CAAC;IAC1B,YAAY,GAAG,YAAY,CAAC;IAC5B,OAAO,GAAG,uBAAuB,CAAC;IAE3C;;;;;;;OAOG;IACH,KAAK,CAAC,QAAQ,CAAC,UAA4B,EAAE,OAA8B;QAC1E,MAAM,SAAS,GAAG,IAAI,CAAC,eAAe,CAAC,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,CAAC;QACjE,MAAM,OAAO,GAA2B,EAAE,CAAC;QAE3C,gEAAgE;QAChE,oDAAoD;QACpD,MAAM,MAAM,GAAG,UAAU,CAAC,MAAM,IAAI,UAAU,CAAC,WAAW,CAAC;QAC3D,IAAI,MAAM,EAAE,CAAC;YACZ,OAAO,CAAC,aAAa,GAAG,UAAU,MAAM,EAAE,CAAC;QAC5C,CAAC;QAED,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,SAAS,CACpC,GAAG,IAAI,CAAC,OAAO,gBAAgB,EAC/B,MAAM,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,SAAS,EACrD,SAAS,CACT,CAAC;QAEF,OAAO,QAAQ,CAAC,IAAI;aAClB,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC;aAC1C,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,IAAI,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC,CAAC;IAC3C,CAAC;IAED;;;OAGG;IACK,WAAW,CAAC,KAAsB;QACzC,OAAO,KAAK,CAAC,OAAO,EAAE,MAAM,KAAK,IAAI,CAAC;IACvC,CAAC;IAED,8EAA8E;IACtE,WAAW,CAAC,KAAsB;QACzC,MAAM,IAAI,GAAG,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;QACnC,MAAM,YAAY,GAAG,IAAI,CAAC,iBAAiB,CAAC,KAAK,CAAC,CAAC;QACnD,MAAM,OAAO,GAAG,IAAI,CAAC,YAAY,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;QACjD,MAAM,cAAc,GAAG,IAAI,CAAC,qBAAqB,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;QAE5D,OAAO,IAAI,CAAC,QAAQ,CAAC;YACpB,EAAE,EAAE,KAAK,CAAC,EAAE;YACZ,IAAI,EAAE,KAAK,CAAC,IAAI,IAAI,KAAK,CAAC,EAAE;YAC5B,QAAQ,EAAE,IAAI,CAAC,UAAU;YACzB,cAAc;YACd,IAAI;YACJ,YAAY;YACZ,aAAa,EAAE,KAAK,CAAC,cAAc,IAAI,CAAC;YACxC,eAAe,EAAE,KAAK,CAAC,YAAY,EAAE,qBAAqB,IAAI,CAAC;YAC/D,OAAO;SACP,CAAC,CAAC;IACJ,CAAC;IAED;;;;;;;;;;;OAWG;IACK,qBAAqB,CAAC,OAAe;QAC5C,MAAM,UAAU,GAAG,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;QACxC,IAAI,UAAU,KAAK,CAAC,CAAC,EAAE,CAAC;YACvB,oDAAoD;YACpD,OAAO,OAAO,CAAC;QAChB,CAAC;QAED,MAAM,MAAM,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC;QAE5C,gEAAgE;QAChE,wDAAwD;QACxD,MAAM,SAAS,GAA2B;YACzC,SAAS,EAAE,WAAW;YACtB,MAAM,EAAE,QAAQ;YAChB,MAAM,EAAE,QAAQ;YAChB,YAAY,EAAE,MAAM;YACpB,SAAS,EAAE,SAAS;YACpB,MAAM,EAAE,QAAQ;YAChB,QAAQ,EAAE,UAAU;YACpB,IAAI,EAAE,MAAM;SACZ,CAAC;QAEF,OAAO,SAAS,CAAC,MAAM,CAAC,IAAI,MAAM,CAAC;IACpC,CAAC;IAED,6EAA6E;IACrE,SAAS,CAAC,KAAsB;QACvC,MAAM,QAAQ,GAAG,CAAC,KAAK,CAAC,YAAY,EAAE,QAAQ,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;QACpE,MAAM,EAAE,GAAG,KAAK,CAAC,EAAE,CAAC,WAAW,EAAE,CAAC;QAClC,MAAM,IAAI,GAAG,CAAC,KAAK,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;QAE9C,IAAI,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;YAC9D,OAAO,OAAO,CAAC;QAChB,CAAC;QACD,IAAI,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,EAAE,CAAC;YAChC,OAAO,OAAO,CAAC;QAChB,CAAC;QACD,IAAI,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,IAAI,CAAC,QAAQ,CAAC,WAAW,CAAC,EAAE,CAAC;YACxD,OAAO,WAAW,CAAC;QACpB,CAAC;QAED,OAAO,MAAM,CAAC;IACf,CAAC;IAED;;OAEG;IACK,iBAAiB,CAAC,KAAsB;QAC/C,MAAM,QAAQ,GAAG,CAAC,KAAK,CAAC,YAAY,EAAE,QAAQ,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;QACpE,MAAM,EAAE,GAAG,KAAK,CAAC,EAAE,CAAC,WAAW,EAAE,CAAC;QAClC,MAAM,YAAY,GAAa,EAAE,CAAC;QAElC,IAAI,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,EAAE,CAAC;YAC1B,OAAO,CAAC,WAAW,CAAC,CAAC;QACtB,CAAC;QAED,kEAAkE;QAClE,IAAI,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;YAC9D,OAAO,CAAC,kBAAkB,CAAC,CAAC;QAC7B,CAAC;QAED,YAAY,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAE1B,4DAA4D;QAC5D,IAAI,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,QAAQ,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;YAC7D,YAAY,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QAC7B,CAAC;QAED,qEAAqE;QACrE,IAAI,IAAI,CAAC,iBAAiB,CAAC,EAAE,CAAC,EAAE,CAAC;YAChC,YAAY,CAAC,IAAI,CAAC,kBAAkB,EAAE,MAAM,EAAE,KAAK,CAAC,CAAC;QACtD,CAAC;QAED,OAAO,YAAY,CAAC;IACrB,CAAC;IAED;;;;;;;OAOG;IACK,iBAAiB,CAAC,EAAU;QACnC,OAAO,CACN,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC;YACpB,EAAE,CAAC,QAAQ,CAAC,UAAU,CAAC;YACvB,EAAE,CAAC,QAAQ,CAAC,eAAe,CAAC;YAC5B,EAAE,CAAC,QAAQ,CAAC,aAAa,CAAC;YAC1B,EAAE,CAAC,QAAQ,CAAC,QAAQ,CAAC;YACrB,EAAE,CAAC,QAAQ,CAAC,WAAW,CAAC;YACxB,EAAE,CAAC,QAAQ,CAAC,eAAe,CAAC;YAC5B,EAAE,CAAC,QAAQ,CAAC,SAAS,CAAC;YACtB,EAAE,CAAC,QAAQ,CAAC,UAAU,CAAC,CACvB,CAAC;IACH,CAAC;IAED;;;;;;OAMG;IACK,YAAY,CAAC,OAAsC;QAC1D,IAAI,CAAC,OAAO;YAAE,OAAO,SAAS,CAAC;QAE/B,MAAM,UAAU,GAAG,MAAM,CAAC,UAAU,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC;QACrD,MAAM,cAAc,GAAG,MAAM,CAAC,UAAU,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;QAE7D,IAAI,MAAM,CAAC,KAAK,CAAC,UAAU,CAAC,IAAI,MAAM,CAAC,KAAK,CAAC,cAAc,CAAC,EAAE,CAAC;YAC9D,OAAO,SAAS,CAAC;QAClB,CAAC;QAED,OAAO;YACN,mDAAmD;YACnD,eAAe,EAAE,UAAU,GAAG,SAAS;YACvC,gBAAgB,EAAE,cAAc,GAAG,SAAS;SAC5C,CAAC;IACH,CAAC;CACD"}

package/dist/discovery/perplexity.d.ts

@@ -0,0 +1,66 @@
+/**
+ * kosha-discovery — Perplexity provider discoverer.
+ *
+ * Queries the Perplexity API `/v1/models` endpoint (OpenAI-compatible),
+ * and maps results into {@link ModelCard} objects.
+ *
+ * Perplexity is a single-vendor catalog — all models are Perplexity's own
+ * or their fine-tuned customisations (e.g. `r1-1776`, a DeepSeek R1 variant
+ * tuned and served by Perplexity). The origin provider is always "perplexity".
+ *
+ * Sonar models carry Perplexity's unique real-time web search / grounding
+ * feature, represented by the `web_search` capability flag.
+ *
+ * @module
+ */
+import type { ModelClassification, OpenAICompatibleModel } from "./openai-compatible.js";
+import { OpenAICompatibleDiscoverer } from "./openai-compatible.js";
+/**
+ * Discovers models available through the Perplexity API.
+ *
+ * All models in the Perplexity catalog are served by Perplexity, so
+ * `originProvider` is always `"perplexity"`. The catalog is clean
+ * (no reward models, guard models, or embedding entries), so
+ * {@link isRelevantModel} keeps every entry.
+ *
+ * Classification rules:
+ * - Mode: always `"chat"`
+ * - Base capabilities: `["chat", "function_calling"]`
+ * - `"web_search"` added for any model whose ID contains `"sonar"`
+ *   (Perplexity's grounding/search-augmented family)
+ * - `"nlu"` added for any model whose ID contains `"pro"`
+ *   (enhanced reasoning / language understanding tier)
+ */
+export declare class PerplexityDiscoverer extends OpenAICompatibleDiscoverer {
+    readonly providerId = "perplexity";
+    readonly providerName = "Perplexity";
+    readonly baseUrl = "https://api.perplexity.ai";
+    /**
+     * All models in the Perplexity catalog are relevant.
+     *
+     * The catalog is curated and does not include reward models,
+     * embedding models, or other non-inference entries.
+     */
+    protected isRelevantModel(_model: OpenAICompatibleModel): boolean;
+    /**
+     * Classify a Perplexity model.
+     *
+     * Origin is always "perplexity". Mode is always "chat".
+     * Capabilities are built from the model ID:
+     * - `"chat"` and `"function_calling"` for all models
+     * - `"web_search"` for sonar family models (real-time grounding)
+     * - `"nlu"` for pro-tier models (enhanced NLU / reasoning)
+     */
+    protected classifyModel(model: OpenAICompatibleModel): ModelClassification;
+    /**
+     * Build the capability list for a model based on its ID.
+     *
+     * Rules applied in order (all are additive — multiple can apply):
+     * 1. `"chat"` — always included
+     * 2. `"function_calling"` — always included (all Perplexity models support it)
+     * 3. `"web_search"` — added when the model ID contains `"sonar"`
+     * 4. `"nlu"` — added when the model ID contains `"pro"`
+     */
+    private buildCapabilities;
+}
+//# sourceMappingURL=perplexity.d.ts.map

package/dist/discovery/perplexity.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"perplexity.d.ts","sourceRoot":"","sources":["../../src/discovery/perplexity.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,KAAK,EAAE,mBAAmB,EAAE,qBAAqB,EAAE,MAAM,wBAAwB,CAAC;AACzF,OAAO,EAAE,0BAA0B,EAAE,MAAM,wBAAwB,CAAC;AAEpE;;;;;;;;;;;;;;;GAeG;AACH,qBAAa,oBAAqB,SAAQ,0BAA0B;IACnE,QAAQ,CAAC,UAAU,gBAAgB;IACnC,QAAQ,CAAC,YAAY,gBAAgB;IACrC,QAAQ,CAAC,OAAO,+BAA+B;IAE/C;;;;;OAKG;IACH,SAAS,CAAC,eAAe,CAAC,MAAM,EAAE,qBAAqB,GAAG,OAAO;IAIjE;;;;;;;;OAQG;IACH,SAAS,CAAC,aAAa,CAAC,KAAK,EAAE,qBAAqB,GAAG,mBAAmB;IAW1E;;;;;;;;OAQG;IACH,OAAO,CAAC,iBAAiB;CAezB"}

package/dist/discovery/perplexity.js

@@ -0,0 +1,86 @@
+/**
+ * kosha-discovery — Perplexity provider discoverer.
+ *
+ * Queries the Perplexity API `/v1/models` endpoint (OpenAI-compatible),
+ * and maps results into {@link ModelCard} objects.
+ *
+ * Perplexity is a single-vendor catalog — all models are Perplexity's own
+ * or their fine-tuned customisations (e.g. `r1-1776`, a DeepSeek R1 variant
+ * tuned and served by Perplexity). The origin provider is always "perplexity".
+ *
+ * Sonar models carry Perplexity's unique real-time web search / grounding
+ * feature, represented by the `web_search` capability flag.
+ *
+ * @module
+ */
+import { OpenAICompatibleDiscoverer } from "./openai-compatible.js";
+/**
+ * Discovers models available through the Perplexity API.
+ *
+ * All models in the Perplexity catalog are served by Perplexity, so
+ * `originProvider` is always `"perplexity"`. The catalog is clean
+ * (no reward models, guard models, or embedding entries), so
+ * {@link isRelevantModel} keeps every entry.
+ *
+ * Classification rules:
+ * - Mode: always `"chat"`
+ * - Base capabilities: `["chat", "function_calling"]`
+ * - `"web_search"` added for any model whose ID contains `"sonar"`
+ *   (Perplexity's grounding/search-augmented family)
+ * - `"nlu"` added for any model whose ID contains `"pro"`
+ *   (enhanced reasoning / language understanding tier)
+ */
+export class PerplexityDiscoverer extends OpenAICompatibleDiscoverer {
+    providerId = "perplexity";
+    providerName = "Perplexity";
+    baseUrl = "https://api.perplexity.ai";
+    /**
+     * All models in the Perplexity catalog are relevant.
+     *
+     * The catalog is curated and does not include reward models,
+     * embedding models, or other non-inference entries.
+     */
+    isRelevantModel(_model) {
+        return true;
+    }
+    /**
+     * Classify a Perplexity model.
+     *
+     * Origin is always "perplexity". Mode is always "chat".
+     * Capabilities are built from the model ID:
+     * - `"chat"` and `"function_calling"` for all models
+     * - `"web_search"` for sonar family models (real-time grounding)
+     * - `"nlu"` for pro-tier models (enhanced NLU / reasoning)
+     */
+    classifyModel(model) {
+        const lower = model.id.toLowerCase();
+        const capabilities = this.buildCapabilities(lower);
+        return {
+            originProvider: "perplexity",
+            mode: "chat",
+            capabilities,
+        };
+    }
+    /**
+     * Build the capability list for a model based on its ID.
+     *
+     * Rules applied in order (all are additive — multiple can apply):
+     * 1. `"chat"` — always included
+     * 2. `"function_calling"` — always included (all Perplexity models support it)
+     * 3. `"web_search"` — added when the model ID contains `"sonar"`
+     * 4. `"nlu"` — added when the model ID contains `"pro"`
+     */
+    buildCapabilities(lower) {
+        const caps = ["chat", "function_calling"];
+        // Sonar models have built-in real-time web search / grounding
+        if (lower.includes("sonar")) {
+            caps.push("web_search");
+        }
+        // Pro-tier models have enhanced NLU / reasoning capabilities
+        if (lower.includes("pro")) {
+            caps.push("nlu");
+        }
+        return caps;
+    }
+}
+//# sourceMappingURL=perplexity.js.map

package/dist/discovery/perplexity.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"perplexity.js","sourceRoot":"","sources":["../../src/discovery/perplexity.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAGH,OAAO,EAAE,0BAA0B,EAAE,MAAM,wBAAwB,CAAC;AAEpE;;;;;;;;;;;;;;;GAeG;AACH,MAAM,OAAO,oBAAqB,SAAQ,0BAA0B;IAC1D,UAAU,GAAG,YAAY,CAAC;IAC1B,YAAY,GAAG,YAAY,CAAC;IAC5B,OAAO,GAAG,2BAA2B,CAAC;IAE/C;;;;;OAKG;IACO,eAAe,CAAC,MAA6B;QACtD,OAAO,IAAI,CAAC;IACb,CAAC;IAED;;;;;;;;OAQG;IACO,aAAa,CAAC,KAA4B;QACnD,MAAM,KAAK,GAAG,KAAK,CAAC,EAAE,CAAC,WAAW,EAAE,CAAC;QACrC,MAAM,YAAY,GAAG,IAAI,CAAC,iBAAiB,CAAC,KAAK,CAAC,CAAC;QAEnD,OAAO;YACN,cAAc,EAAE,YAAY;YAC5B,IAAI,EAAE,MAAM;YACZ,YAAY;SACZ,CAAC;IACH,CAAC;IAED;;;;;;;;OAQG;IACK,iBAAiB,CAAC,KAAa;QACtC,MAAM,IAAI,GAAa,CAAC,MAAM,EAAE,kBAAkB,CAAC,CAAC;QAEpD,8DAA8D;QAC9D,IAAI,KAAK,CAAC,QAAQ,CAAC,OAAO,CAAC,EAAE,CAAC;YAC7B,IAAI,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC;QACzB,CAAC;QAED,6DAA6D;QAC7D,IAAI,KAAK,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC;YAC3B,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;QAClB,CAAC;QAED,OAAO,IAAI,CAAC;IACb,CAAC;CACD"}

package/dist/discovery/together.d.ts

@@ -0,0 +1,41 @@
+/**
+ * kosha-discovery — Together AI provider discoverer.
+ *
+ * Queries the Together AI `/v1/models` endpoint (OpenAI-compatible),
+ * filters relevant models, and maps them into {@link ModelCard} objects.
+ *
+ * Together AI's catalog is multi-vendor, hosting models from meta-llama,
+ * mistralai, Qwen, google, deepseek-ai, microsoft, nvidia, databricks,
+ * togethercomputer, and others.
+ * @module
+ */
+import { OpenAICompatibleDiscoverer, type ModelClassification, type OpenAICompatibleModel } from "./openai-compatible.js";
+/**
+ * Discovers models available through the Together AI API (api.together.xyz).
+ *
+ * The Together AI catalog serves models from multiple vendors using an
+ * OpenAI-compatible API. Model IDs are namespaced (e.g.
+ * `meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo`,
+ * `mistralai/Mixtral-8x22B-Instruct-v0.1`), allowing origin provider
+ * extraction from the prefix.
+ */
+export declare class TogetherDiscoverer extends OpenAICompatibleDiscoverer {
+    readonly providerId = "together";
+    readonly providerName = "Together AI";
+    readonly baseUrl = "https://api.together.xyz";
+    /**
+     * Determine whether a model ID represents a model we want to track.
+     *
+     * We keep chat/instruct, embedding, code, and vision models and filter out:
+     * - Reward models (used for RLHF, not inference)
+     */
+    protected isRelevantModel(model: OpenAICompatibleModel): boolean;
+    /**
+     * Classify a Together AI model: determine origin provider, mode, and capabilities.
+     *
+     * Uses namespaced model IDs (vendor/model-name) to extract origin provider.
+     * Normalizes known prefixes via aliases map.
+     */
+    protected classifyModel(model: OpenAICompatibleModel): ModelClassification;
+}
+//# sourceMappingURL=together.d.ts.map

package/dist/discovery/together.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"together.d.ts","sourceRoot":"","sources":["../../src/discovery/together.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAGH,OAAO,EAAE,0BAA0B,EAAE,KAAK,mBAAmB,EAAE,KAAK,qBAAqB,EAAE,MAAM,wBAAwB,CAAC;AAE1H;;;;;;;;GAQG;AACH,qBAAa,kBAAmB,SAAQ,0BAA0B;IACjE,QAAQ,CAAC,UAAU,cAAc;IACjC,QAAQ,CAAC,YAAY,iBAAiB;IACtC,QAAQ,CAAC,OAAO,8BAA8B;IAE9C;;;;;OAKG;IACH,SAAS,CAAC,eAAe,CAAC,KAAK,EAAE,qBAAqB,GAAG,OAAO;IAShE;;;;;OAKG;IACH,SAAS,CAAC,aAAa,CAAC,KAAK,EAAE,qBAAqB,GAAG,mBAAmB;CAiD1E"}

package/dist/discovery/together.js

@@ -0,0 +1,88 @@
+/**
+ * kosha-discovery — Together AI provider discoverer.
+ *
+ * Queries the Together AI `/v1/models` endpoint (OpenAI-compatible),
+ * filters relevant models, and maps them into {@link ModelCard} objects.
+ *
+ * Together AI's catalog is multi-vendor, hosting models from meta-llama,
+ * mistralai, Qwen, google, deepseek-ai, microsoft, nvidia, databricks,
+ * togethercomputer, and others.
+ * @module
+ */
+import { OpenAICompatibleDiscoverer } from "./openai-compatible.js";
+/**
+ * Discovers models available through the Together AI API (api.together.xyz).
+ *
+ * The Together AI catalog serves models from multiple vendors using an
+ * OpenAI-compatible API. Model IDs are namespaced (e.g.
+ * `meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo`,
+ * `mistralai/Mixtral-8x22B-Instruct-v0.1`), allowing origin provider
+ * extraction from the prefix.
+ */
+export class TogetherDiscoverer extends OpenAICompatibleDiscoverer {
+    providerId = "together";
+    providerName = "Together AI";
+    baseUrl = "https://api.together.xyz";
+    /**
+     * Determine whether a model ID represents a model we want to track.
+     *
+     * We keep chat/instruct, embedding, code, and vision models and filter out:
+     * - Reward models (used for RLHF, not inference)
+     */
+    isRelevantModel(model) {
+        const lower = model.id.toLowerCase();
+        // Skip reward models — used for RLHF, not direct inference
+        if (this.looksLikeReward(lower))
+            return false;
+        return true;
+    }
+    /**
+     * Classify a Together AI model: determine origin provider, mode, and capabilities.
+     *
+     * Uses namespaced model IDs (vendor/model-name) to extract origin provider.
+     * Normalizes known prefixes via aliases map.
+     */
+    classifyModel(model) {
+        const id = model.id;
+        const lower = id.toLowerCase();
+        // Extract origin from vendor/model-name format with alias normalization
+        const aliases = {
+            "meta-llama": "meta",
+            "mistralai": "mistral",
+            "deepseek-ai": "deepseek",
+            "togethercomputer": "together",
+        };
+        const originProvider = this.extractPrefixOrigin(id, "together", aliases);
+        // Detect mode: use model.type field for embedding, or infer from ID
+        let mode = "chat";
+        if (model.type === "embedding" || this.looksLikeEmbedding(lower)) {
+            mode = "embedding";
+        }
+        // Build capability flags
+        const capabilities = [];
+        if (mode === "embedding") {
+            capabilities.push("embedding");
+        }
+        else {
+            capabilities.push("chat");
+            // Add vision capability if detected
+            if (this.looksLikeVision(lower)) {
+                capabilities.push("vision");
+            }
+            // Add code capability if detected
+            if (this.looksLikeCode(lower)) {
+                capabilities.push("code");
+            }
+            // Add function_calling for instruct/chat models
+            if (lower.includes("instruct") || lower.includes("chat")) {
+                capabilities.push("function_calling");
+            }
+        }
+        return {
+            originProvider,
+            mode,
+            capabilities,
+        };
+    }
+}
+//# sourceMappingURL=together.js.map

package/dist/discovery/together.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"together.js","sourceRoot":"","sources":["../../src/discovery/together.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAGH,OAAO,EAAE,0BAA0B,EAAwD,MAAM,wBAAwB,CAAC;AAE1H;;;;;;;;GAQG;AACH,MAAM,OAAO,kBAAmB,SAAQ,0BAA0B;IACxD,UAAU,GAAG,UAAU,CAAC;IACxB,YAAY,GAAG,aAAa,CAAC;IAC7B,OAAO,GAAG,0BAA0B,CAAC;IAE9C;;;;;OAKG;IACO,eAAe,CAAC,KAA4B;QACrD,MAAM,KAAK,GAAG,KAAK,CAAC,EAAE,CAAC,WAAW,EAAE,CAAC;QAErC,2DAA2D;QAC3D,IAAI,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC;YAAE,OAAO,KAAK,CAAC;QAE9C,OAAO,IAAI,CAAC;IACb,CAAC;IAED;;;;;OAKG;IACO,aAAa,CAAC,KAA4B;QACnD,MAAM,EAAE,GAAG,KAAK,CAAC,EAAE,CAAC;QACpB,MAAM,KAAK,GAAG,EAAE,CAAC,WAAW,EAAE,CAAC;QAE/B,wEAAwE;QACxE,MAAM,OAAO,GAA2B;YACvC,YAAY,EAAE,MAAM;YACpB,WAAW,EAAE,SAAS;YACtB,aAAa,EAAE,UAAU;YACzB,kBAAkB,EAAE,UAAU;SAC9B,CAAC;QACF,MAAM,cAAc,GAAG,IAAI,CAAC,mBAAmB,CAAC,EAAE,EAAE,UAAU,EAAE,OAAO,CAAC,CAAC;QAEzE,oEAAoE;QACpE,IAAI,IAAI,GAAc,MAAM,CAAC;QAC7B,IAAI,KAAK,CAAC,IAAI,KAAK,WAAW,IAAI,IAAI,CAAC,kBAAkB,CAAC,KAAK,CAAC,EAAE,CAAC;YAClE,IAAI,GAAG,WAAW,CAAC;QACpB,CAAC;QAED,yBAAyB;QACzB,MAAM,YAAY,GAAa,EAAE,CAAC;QAElC,IAAI,IAAI,KAAK,WAAW,EAAE,CAAC;YAC1B,YAAY,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC;QAChC,CAAC;aAAM,CAAC;YACP,YAAY,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;YAE1B,oCAAoC;YACpC,IAAI,IAAI,CAAC,eAAe,CAAC,KAAK,CAAC,EAAE,CAAC;gBACjC,YAAY,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;YAC7B,CAAC;YAED,kCAAkC;YAClC,IAAI,IAAI,CAAC,aAAa,CAAC,KAAK,CAAC,EAAE,CAAC;gBAC/B,YAAY,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;YAC3B,CAAC;YAED,gDAAgD;YAChD,IAAI,KAAK,CAAC,QAAQ,CAAC,UAAU,CAAC,IAAI,KAAK,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;gBAC1D,YAAY,CAAC,IAAI,CAAC,kBAAkB,CAAC,CAAC;YACvC,CAAC;QACF,CAAC;QAED,OAAO;YACN,cAAc;YACd,IAAI;YACJ,YAAY;SACZ,CAAC;IACH,CAAC;CACD"}

package/dist/discovery/vertex.d.ts

@@ -0,0 +1,146 @@
+/**
+ * kosha-discovery — Google Vertex AI provider discoverer.
+ *
+ * Resolution strategy (in order):
+ * 1. REST API with Application Default Credentials (ADC)
+ * 2. gcloud CLI fallback (`gcloud ai models list`)
+ * 3. Static fallback list of known Vertex AI models
+ *
+ * Credentials: Uses the Google ADC chain —
+ *   ~/.config/gcloud/application_default_credentials.json,
+ *   GOOGLE_APPLICATION_CREDENTIALS env var, or the GCP metadata server.
+ * @module
+ */
+import type { CredentialResult, ModelCard } from "../types.js";
+import { BaseDiscoverer } from "./base.js";
+/**
+ * Discovers models available through Google Vertex AI.
+ *
+ * Works without an explicit API key: access tokens are resolved via the
+ * standard ADC chain, so it behaves correctly in local dev environments
+ * (after `gcloud auth application-default login`) and on GCP VMs (metadata
+ * server).  When the REST API is unavailable it falls back to the gcloud CLI,
+ * and when that is missing it returns a curated static list so callers always
+ * receive useful results.
+ */
+export declare class VertexDiscoverer extends BaseDiscoverer {
+    readonly providerId = "vertex";
+    readonly providerName = "Google Vertex AI";
+    /** Template string — region is substituted per-request. */
+    readonly baseUrl = "https://{region}-aiplatform.googleapis.com";
+    /**
+     * Discover models from Vertex AI using a three-tier fallback strategy:
+     *   1. Vertex AI REST API (requires an ADC access token + project ID)
+     *   2. `gcloud ai models list` CLI
+     *   3. Curated static list
+     *
+     * @param credential - Credential bag; `accessToken` is used when present.
+     *   `metadata.projectId` and `metadata.region` take precedence over env vars.
+     * @param options    - Optional timeout in ms (default 10 000).
+     */
+    discover(credential: CredentialResult, options?: {
+        timeout?: number;
+    }): Promise<ModelCard[]>;
+    /**
+     * Fetch models from the Vertex AI Publisher Models REST endpoint.
+     *
+     * URL: `https://{region}-aiplatform.googleapis.com/v1/projects/{project}/
+     *       locations/{region}/publishers/google/models`
+     *
+     * @param credential - Used to obtain a Bearer token via {@link getAccessToken}.
+     * @param projectId  - GCP project ID (required for the REST path).
+     * @param region     - GCP region (e.g. "us-central1").
+     * @param options    - Optional timeout override.
+     */
+    private discoverViaApi;
+    /**
+     * Resolve a Google OAuth2 access token using the ADC chain:
+     *   1. `credential.accessToken` — already resolved upstream
+     *   2. ADC JSON file containing `access_token` (service account / workload identity)
+     *   3. ADC JSON file with `refresh_token` — exchange via Google OAuth endpoint
+     *   4. `gcloud auth print-access-token` CLI fallback
+     *
+     * Returns `undefined` when no token can be obtained.
+     *
+     * @param credential - The credential bag from the discovery pipeline.
+     */
+    getAccessToken(credential: CredentialResult): Promise<string | undefined>;
+    /**
+     * Exchange a refresh token for a short-lived access token via Google OAuth2.
+     *
+     * @param clientId     - OAuth2 client ID from the ADC file.
+     * @param clientSecret - OAuth2 client secret from the ADC file.
+     * @param refreshToken - Refresh token from the ADC file.
+     */
+    private exchangeRefreshToken;
+    /**
+     * Obtain an access token by calling `gcloud auth print-access-token`.
+     * Returns `undefined` when the CLI is not installed or the command fails.
+     */
+    private tokenFromGcloudCli;
+    /**
+     * Discover models by executing `gcloud ai models list` and parsing JSON output.
+     *
+     * @param projectId - GCP project ID passed as `--project`.
+     * @param region    - GCP region passed as `--region`.
+     */
+    private discoverViaCli;
+    /**
+     * Resolve a GCP project ID from the active `gcloud` configuration.
+     * Returns `undefined` when the CLI is unavailable or unconfigured.
+     */
+    private resolveProjectFromCli;
+    /**
+     * Return a curated list of well-known Vertex AI models (Feb 2026).
+     *
+     * Used when neither the REST API nor the gcloud CLI is accessible.
+     * All entries carry `source: "manual"` so consumers can distinguish them
+     * from live API data.
+     *
+     * @param region    - Propagated to each card's `region` field.
+     * @param projectId - Propagated to each card's `projectId` field (may be undefined).
+     */
+    private staticFallback;
+    /**
+     * Convert a Vertex AI REST API model object to a {@link ModelCard}.
+     *
+     * The `name` field is a full resource path such as
+     * `publishers/google/models/gemini-2.5-pro`; the prefix is stripped to
+     * obtain the bare model ID used in API calls.
+     *
+     * @param model     - Raw Vertex API model entry.
+     * @param region    - Region the model was fetched from.
+     * @param projectId - GCP project ID.
+     */
+    private toModelCard;
+    /**
+     * Convert a `gcloud ai models list` JSON entry to a {@link ModelCard}.
+     *
+     * gcloud returns full resource names such as
+     * `projects/{p}/locations/{r}/models/{id}` — we take only the last segment.
+     *
+     * @param model     - Raw gcloud JSON model entry.
+     * @param region    - Region used for the gcloud query.
+     * @param projectId - GCP project ID.
+     */
+    private gcloudModelToCard;
+    /**
+     * Determine the primary {@link ModelMode} from a model ID and its
+     * supported actions list.
+     *
+     * @param id      - Bare model ID.
+     * @param actions - `supportedActions` array from the API response.
+     */
+    private inferMode;
+    /**
+     * Infer capability flags from a model ID and its supported actions.
+     *
+     * Gemini pro/flash/ultra variants receive vision + function_calling because
+     * all modern Gemini editions support multimodal input and tool use.
+     *
+     * @param id      - Bare model ID.
+     * @param actions - `supportedActions` from the API response.
+     */
+    private inferCapabilities;
+}
+//# sourceMappingURL=vertex.d.ts.map

package/dist/discovery/vertex.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"vertex.d.ts","sourceRoot":"","sources":["../../src/discovery/vertex.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAQH,OAAO,KAAK,EAAE,gBAAgB,EAAE,SAAS,EAAa,MAAM,aAAa,CAAC;AAC1E,OAAO,EAAE,cAAc,EAAE,MAAM,WAAW,CAAC;AAiD3C;;;;;;;;;GASG;AACH,qBAAa,gBAAiB,SAAQ,cAAc;IACnD,QAAQ,CAAC,UAAU,YAAY;IAC/B,QAAQ,CAAC,YAAY,sBAAsB;IAC3C,2DAA2D;IAC3D,QAAQ,CAAC,OAAO,gDAAgD;IAEhE;;;;;;;;;OASG;IACG,QAAQ,CAAC,UAAU,EAAE,gBAAgB,EAAE,OAAO,CAAC,EAAE;QAAE,OAAO,CAAC,EAAE,MAAM,CAAA;KAAE,GAAG,OAAO,CAAC,SAAS,EAAE,CAAC;IAwClG;;;;;;;;;;OAUG;YACW,cAAc;IA8B5B;;;;;;;;;;OAUG;IACG,cAAc,CAAC,UAAU,EAAE,gBAAgB,GAAG,OAAO,CAAC,MAAM,GAAG,SAAS,CAAC;IA8B/E;;;;;;OAMG;YACW,oBAAoB;IA0BlC;;;OAGG;IACH,OAAO,CAAC,kBAAkB;IAkB1B;;;;;OAKG;YACW,cAAc;IAuB5B;;;OAGG;IACH,OAAO,CAAC,qBAAqB;IAmB7B;;;;;;;;;OASG;IACH,OAAO,CAAC,cAAc;IAkBtB;;;;;;;;;;OAUG;IACH,OAAO,CAAC,WAAW;IAkBnB;;;;;;;;;OASG;IACH,OAAO,CAAC,iBAAiB;IAqBzB;;;;;;OAMG;IACH,OAAO,CAAC,SAAS;IAOjB;;;;;;;;OAQG;IACH,OAAO,CAAC,iBAAiB;CAgBzB"}