@crossworks/voice-client 0.230.43

package/src/catalogs/huggingface.ts

@@ -0,0 +1,180 @@
+/**
+ * Hugging Face Inference Providers static catalog (chat models only).
+ *
+ * HF's API is fundamentally different from OpenAI's — it's a *router*
+ * that proxies to many sub-providers (Cerebras, Groq, Together,
+ * SambaNova, Fireworks, etc.) under one OpenAI-compatible endpoint
+ * at `router.huggingface.co/v1/chat/completions`. You can append
+ * routing-policy suffixes to the model id:
+ *
+ *   openai/gpt-oss-120b:fastest     ← default; HF picks lowest latency
+ *   openai/gpt-oss-120b:cheapest    ← lowest cost-per-output-token
+ *   openai/gpt-oss-120b:preferred   ← honour user's preference list
+ *   openai/gpt-oss-120b:groq        ← pin to a specific provider
+ *
+ * The :fastest / :cheapest / :preferred / :<provider> suffix is
+ * handled server-side — we just include or omit it on the model
+ * string we pass to /v1/chat/completions. The adapter exposes a
+ * `routing` knob on the worker params so operators can pick.
+ *
+ * The catalog below picks notable, broadly-available open-weight
+ * models. HF lists thousands; we curate the ones worth defaulting to.
+ * The `GET /v1/models` endpoint returns the live list with per-provider
+ * pricing/throughput, so discovery fills in the rest.
+ *
+ * Capabilities: HF Inference Providers cover chat, vision (VLM),
+ * embedding, text-to-image, text-to-video, and speech-to-text via
+ * specialty endpoints, but ONLY the chat endpoint is OpenAI-compatible.
+ * Non-chat tasks need custom call shapes — we ship chat now and add
+ * other tasks as adapters when we want them.
+ */
+
+import type { ChatModelInfo, ImageGenModelInfo } from '../adapters/types';
+
+export const HUGGINGFACE_BASE_URL = 'https://router.huggingface.co/v1';
+/** Image-gen inference uses a separate, repo-id-based endpoint. The
+ *  router doesn't proxy image tasks today — calls go straight to
+ *  api-inference for the per-model URL. */
+export const HUGGINGFACE_INFERENCE_BASE_URL = 'https://api-inference.huggingface.co';
+
+/** Suffixes the HF router accepts on the model id. The adapter
+ *  appends one of these to whatever model the user picks. */
+export const HUGGINGFACE_ROUTING_POLICIES = ['fastest', 'cheapest', 'preferred'] as const;
+export type HuggingfaceRoutingPolicy = (typeof HUGGINGFACE_ROUTING_POLICIES)[number];
+
+/**
+ * Curated list of notable open-weight chat models reliably available
+ * through HF's router. Live `/v1/models` returns more — these are
+ * useful defaults for the dropdown.
+ */
+export const HUGGINGFACE_CHAT_MODELS: readonly ChatModelInfo[] = [
+  // ── OpenAI's open-weights releases ──────────────────────────────
+  {
+    id: 'openai/gpt-oss-120b',
+    label: 'gpt-oss 120B (OpenAI open)',
+    description: 'Large open-weights model from OpenAI. Solid general-purpose chat.',
+    contextTokens: 128_000,
+    capabilities: ['function_calling'],
+  },
+  {
+    id: 'openai/gpt-oss-20b',
+    label: 'gpt-oss 20B (OpenAI open)',
+    description: 'Lighter open-weights model. Fast, cheaper, weaker reasoning.',
+    contextTokens: 128_000,
+    capabilities: ['function_calling'],
+  },
+
+  // ── DeepSeek (strong reasoning at low cost) ─────────────────────
+  {
+    id: 'deepseek-ai/DeepSeek-R1',
+    label: 'DeepSeek R1',
+    description: 'Open reasoning model. Strong on math/code; verbose chain-of-thought.',
+    contextTokens: 64_000,
+    capabilities: ['reasoning'],
+  },
+  {
+    id: 'deepseek-ai/DeepSeek-V3',
+    label: 'DeepSeek V3',
+    description: 'General-purpose open chat. Cheap; solid coding ability.',
+    contextTokens: 64_000,
+    capabilities: ['function_calling'],
+  },
+
+  // ── Llama family (Meta open-weights) ────────────────────────────
+  {
+    id: 'meta-llama/Llama-3.3-70B-Instruct',
+    label: 'Llama 3.3 70B Instruct',
+    description: 'Latest open Llama instruction-tuned. Good balance of cost vs. quality.',
+    contextTokens: 128_000,
+    capabilities: ['function_calling'],
+  },
+  {
+    id: 'meta-llama/Llama-3.2-11B-Vision-Instruct',
+    label: 'Llama 3.2 11B Vision',
+    description: 'Llama with image understanding. Whiteboard / receipt OCR candidate.',
+    contextTokens: 128_000,
+    capabilities: ['vision', 'function_calling'],
+  },
+
+  // ── Mistral open weights ────────────────────────────────────────
+  {
+    id: 'mistralai/Mistral-Large-Instruct-2411',
+    label: 'Mistral Large',
+    description: 'Mistral flagship open chat model.',
+    contextTokens: 128_000,
+    capabilities: ['function_calling'],
+  },
+  {
+    id: 'mistralai/Mixtral-8x22B-Instruct-v0.1',
+    label: 'Mixtral 8×22B Instruct',
+    description: 'Mixture-of-experts. Strong multilingual; cheaper than dense large models.',
+    contextTokens: 64_000,
+    capabilities: ['function_calling'],
+  },
+
+  // ── Qwen (Alibaba — strong on code) ─────────────────────────────
+  {
+    id: 'Qwen/Qwen2.5-72B-Instruct',
+    label: 'Qwen 2.5 72B',
+    description: 'Strong general-purpose chat; particularly good at code.',
+    contextTokens: 128_000,
+    capabilities: ['function_calling'],
+  },
+  {
+    id: 'Qwen/Qwen2.5-Coder-32B-Instruct',
+    label: 'Qwen 2.5 Coder 32B',
+    description: 'Code-specialist. Pair-programming / refactoring duty.',
+    contextTokens: 128_000,
+    capabilities: ['function_calling'],
+  },
+  {
+    id: 'Qwen/Qwen2-VL-72B-Instruct',
+    label: 'Qwen 2 VL 72B (vision)',
+    description: 'Vision-language model. Image understanding plus chat.',
+    contextTokens: 32_000,
+    capabilities: ['vision'],
+  },
+];
+
+// ─── Hugging Face image generation ───────────────────────────────────
+//
+// Endpoint: POST {HUGGINGFACE_INFERENCE_BASE_URL}/models/<repo>
+// Auth:     Bearer
+// Body:     { inputs: <prompt>, parameters: { negative_prompt, seed,
+//             num_inference_steps, guidance_scale, width, height } }
+// Response: raw image bytes (Content-Type image/png or image/jpeg).
+//
+// HF's catalogue is huge; we surface only the proven defaults so the
+// dropdown isn't 200 items long. Operators can still type any repo id
+// into the worker's `model` field for custom selections (the adapter
+// passes it through verbatim).
+
+export const HUGGINGFACE_IMAGE_MODELS: readonly ImageGenModelInfo[] = [
+  {
+    id: 'black-forest-labs/FLUX.1-dev',
+    label: 'FLUX.1 dev',
+    description:
+      'Black Forest Labs FLUX.1 dev. Best open-source image quality today. Slower than Schnell.',
+    tier: 'quality',
+  },
+  {
+    id: 'black-forest-labs/FLUX.1-schnell',
+    label: 'FLUX.1 schnell',
+    description: 'Distilled FLUX. 4-step inference, 5-10x faster than dev. Slight quality drop.',
+    tier: 'fast',
+  },
+  {
+    id: 'stabilityai/stable-diffusion-xl-base-1.0',
+    label: 'SDXL Base 1.0',
+    description: 'Stability AI SDXL. Older baseline; widely supported by community LoRAs.',
+    tier: 'balanced',
+  },
+  {
+    id: 'stabilityai/stable-diffusion-3.5-large',
+    label: 'Stable Diffusion 3.5 Large',
+    description: 'Stability AI SD 3.5 Large. Higher quality than SDXL, more compute per image.',
+    tier: 'quality',
+  },
+];
+
+export const HUGGINGFACE_IMAGE_DEFAULT_MODEL = 'black-forest-labs/FLUX.1-schnell';

package/src/catalogs/openai-image.ts

@@ -0,0 +1,62 @@
+/**
+ * OpenAI image-generation catalog.
+ *
+ * Endpoint: POST https://api.openai.com/v1/images/generations
+ * Auth:     Bearer
+ *
+ * Models we surface in the worker form (May 2026 reality):
+ *
+ *   - `gpt-image-1`: the current default. Native multimodal model,
+ *     returns base64 by default (no `response_format` switch needed).
+ *     Better instruction-following than DALL-E for text-in-image.
+ *     ~$0.04-0.17 per image depending on size/quality.
+ *   - `dall-e-3`: still around but on the way out. Use when you
+ *     specifically need its style-steering ('vivid'/'natural').
+ *     Returns URLs by default; the adapter requests b64_json so we
+ *     get bytes uniformly.
+ *   - `dall-e-2`: legacy. Cheap (~$0.02 per image), lower quality.
+ *     Useful for high-volume placeholder generation.
+ *
+ * Why no live discovery: OpenAI's /v1/models endpoint returns ALL
+ * models the key can use (chat, embeddings, etc.) — not just image
+ * ones. We could filter by name prefix but the catalog stays simpler
+ * + clearer if we ship the curated list and skip discovery for image
+ * gen.
+ */
+
+import type { ImageGenModelInfo } from '../adapters/types';
+
+export const OPENAI_IMAGE_MODELS: readonly ImageGenModelInfo[] = [
+  {
+    id: 'gpt-image-1',
+    label: 'gpt-image-1',
+    description:
+      'Current OpenAI default. Best text-in-image rendering, strong instruction-following. Recommended.',
+    supportedSizes: ['1024x1024', '1024x1536', '1536x1024'],
+    supportedQualities: ['low', 'medium', 'high', 'auto'],
+    pricePerImage: 0.07,
+    tier: 'balanced',
+  },
+  {
+    id: 'dall-e-3',
+    label: 'DALL-E 3',
+    description:
+      'Steerable with style=vivid|natural. Good for stylised art; gpt-image-1 beats it for realism + text.',
+    supportedSizes: ['1024x1024', '1024x1792', '1792x1024'],
+    supportedStyles: ['vivid', 'natural'],
+    supportedQualities: ['standard', 'hd'],
+    pricePerImage: 0.08,
+    tier: 'quality',
+  },
+  {
+    id: 'dall-e-2',
+    label: 'DALL-E 2 (legacy)',
+    description: 'Older, cheaper. ~$0.02 per image. Use when quality matters less than throughput.',
+    supportedSizes: ['256x256', '512x512', '1024x1024'],
+    pricePerImage: 0.02,
+    tier: 'fast',
+  },
+];
+
+/** Default model for openai-image when no worker.model is set. */
+export const OPENAI_IMAGE_DEFAULT_MODEL = 'gpt-image-1';

package/src/catalogs/openai-vision.ts

@@ -0,0 +1,53 @@
+/**
+ * OpenAI vision model catalog.
+ *
+ * Every gpt-4o family model accepts image_url content parts; older
+ * gpt-4-vision-preview is still around but deprecated. We list the
+ * three operators are most likely to pick:
+ *
+ *   - gpt-4o-mini: ~10x cheaper than gpt-4o. Strong on printed text
+ *     and clean handwriting. Good default for the photo-of-notes
+ *     case the user mentioned.
+ *   - gpt-4o: full quality. Better on cursive, faint pencil, and
+ *     multi-column layouts. Use when mini misses too much.
+ *   - gpt-4-vision-preview: legacy. Don't pick this unless you have
+ *     a specific reason — discoverModels still surfaces it if the
+ *     account has access.
+ *
+ * Endpoint is shared with chat: POST /v1/chat/completions. The image
+ * goes in messages[].content as `{type: 'image_url', image_url: {url:
+ * 'data:image/jpeg;base64,...'}}`. Adapter handles the encoding.
+ */
+
+import type { VisionModelInfo } from '../adapters/types';
+
+export const OPENAI_VISION_MODELS: readonly VisionModelInfo[] = [
+  {
+    id: 'gpt-4o-mini',
+    label: 'gpt-4o-mini',
+    description:
+      'Cheap, fast vision. Recommended default for high-volume OCR of notes, receipts, screenshots.',
+    contextTokens: 128_000,
+    inputPricePer1M: 0.15,
+    outputPricePer1M: 0.6,
+    tier: 'fast',
+  },
+  {
+    id: 'gpt-4o',
+    label: 'gpt-4o',
+    description:
+      'Full-quality vision. Use when gpt-4o-mini misreads cursive or you need diagram/chart comprehension alongside text.',
+    contextTokens: 128_000,
+    inputPricePer1M: 2.5,
+    outputPricePer1M: 10,
+    tier: 'balanced',
+  },
+  {
+    id: 'gpt-4-vision-preview',
+    label: 'gpt-4-vision-preview (legacy)',
+    description:
+      'Deprecated. Kept here only so existing workers configured for it still appear in the dropdown.',
+    contextTokens: 128_000,
+    tier: 'quality',
+  },
+];

package/src/catalogs/openrouter.ts

@@ -0,0 +1,221 @@
+/**
+ * Curated OpenRouter chat catalog — Mantle's headline picks.
+ *
+ * OpenRouter serves 300+ chat models across every major provider. We
+ * don't try to mirror that — the full live catalog is fetched at
+ * runtime by `discoverModels` against `https://openrouter.ai/api/v1/models`
+ * (keyless). This static list is what the worker form's model dropdown
+ * renders BEFORE discovery completes, and the fallback when discovery
+ * fails (rare; OR's /models endpoint is among the most reliable in
+ * the catalogue).
+ *
+ * Pick criteria — what makes it onto this list:
+ *   - The model the project's default workers ship pointing at, OR
+ *   - A current top-tier headline from each major lab (Anthropic,
+ *     OpenAI, Google, xAI), OR
+ *   - A notable open-weights model that's frequently asked for
+ *     (DeepSeek's reasoning lines, Llama 4).
+ *
+ * Capabilities + pricing here mirror the source-of-truth pricing
+ * table in [packages/tracing/src/pricing.ts](../../tracing/src/pricing.ts).
+ * Keep them aligned when models shift — the chat-adapter UI reads
+ * these for the dropdown hints, and tracing reads pricing.ts for the
+ * fallback cost calculation. Diverging would let a model show one
+ * price in the dropdown and bill at a different one.
+ */
+
+import type { ChatModelInfo, VisionModelInfo } from '../adapters/types';
+
+/** OpenRouter API base. The SDK constructs this internally too; we
+ *  expose it for the catalogue + discovery so both paths agree. */
+export const OPENROUTER_BASE_URL = 'https://openrouter.ai/api/v1';
+
+/** A small curated set of OpenRouter chat models we vouch for as
+ *  defaults. The full live catalog comes from discovery; this is the
+ *  "before the network call returns" picture. */
+export const OPENROUTER_CHAT_MODELS: readonly ChatModelInfo[] = [
+  // Anthropic — the responder's default lives here.
+  {
+    id: 'anthropic/claude-sonnet-5',
+    label: 'Claude Sonnet 5',
+    description:
+      'Anthropic flagship. Strong reasoning + tool use, 1M context. Default for the responder and Saskia. Supports prompt caching (~10× cheaper on cache hits).',
+    contextTokens: 1_000_000,
+    capabilities: ['reasoning', 'function_calling', 'vision'],
+    inputPricePer1M: 2,
+    outputPricePer1M: 10,
+  },
+  {
+    id: 'anthropic/claude-sonnet-4.6',
+    label: 'Claude Sonnet 4.6',
+    description: 'Previous-generation Sonnet. 1M context; supports prompt caching.',
+    contextTokens: 1_000_000,
+    capabilities: ['reasoning', 'function_calling', 'vision'],
+    inputPricePer1M: 3,
+    outputPricePer1M: 15,
+  },
+  {
+    id: 'anthropic/claude-haiku-4.5',
+    label: 'Claude Haiku 4.5',
+    description:
+      'Anthropic small/fast. Cheap, fast, 200K context. Default for the extractor + summarizer. Supports prompt caching.',
+    contextTokens: 200_000,
+    capabilities: ['function_calling', 'vision'],
+    inputPricePer1M: 0.8,
+    outputPricePer1M: 4,
+  },
+  {
+    id: 'anthropic/claude-opus-4.7',
+    label: 'Claude Opus 4.7',
+    description:
+      'Anthropic top-tier. Highest capability, highest price. 1M context. Use when the responder needs to chew on something hard.',
+    contextTokens: 1_000_000,
+    capabilities: ['reasoning', 'function_calling', 'vision'],
+    inputPricePer1M: 15,
+    outputPricePer1M: 75,
+  },
+  // OpenAI — headline picks.
+  {
+    id: 'openai/gpt-5',
+    label: 'GPT-5',
+    description: 'OpenAI flagship. Strong all-rounder, native vision + tool use. 256K context.',
+    contextTokens: 256_000,
+    capabilities: ['reasoning', 'function_calling', 'vision'],
+    inputPricePer1M: 5,
+    outputPricePer1M: 15,
+  },
+  {
+    id: 'openai/gpt-4o',
+    label: 'GPT-4o',
+    description: 'OpenAI multimodal workhorse. Still widely used for cost reasons. 128K context.',
+    contextTokens: 128_000,
+    capabilities: ['function_calling', 'vision'],
+    inputPricePer1M: 2.5,
+    outputPricePer1M: 10,
+  },
+  {
+    id: 'openai/gpt-4o-mini',
+    label: 'GPT-4o Mini',
+    description:
+      'OpenAI cheap fast model. Good extractor candidate when keeping costs low matters more than precision.',
+    contextTokens: 128_000,
+    capabilities: ['function_calling', 'vision'],
+    inputPricePer1M: 0.15,
+    outputPricePer1M: 0.6,
+  },
+  // Google — Gemini 2.5+.
+  {
+    id: 'google/gemini-2.5-pro',
+    label: 'Gemini 2.5 Pro',
+    description:
+      'Google top-tier with 2M context. Strong long-document handling; implicit prompt caching on Gemini 2.5+.',
+    contextTokens: 2_000_000,
+    capabilities: ['reasoning', 'function_calling', 'vision'],
+    inputPricePer1M: 1.25,
+    outputPricePer1M: 10,
+  },
+  {
+    id: 'google/gemini-2.5-flash',
+    label: 'Gemini 2.5 Flash',
+    description:
+      'Google fast/cheap. 1M context, implicit caching. Good fit for extractor + summarizer when not using Anthropic.',
+    contextTokens: 1_000_000,
+    capabilities: ['function_calling', 'vision'],
+    inputPricePer1M: 0.3,
+    outputPricePer1M: 2.5,
+  },
+  // xAI.
+  {
+    id: 'x-ai/grok-4',
+    label: 'Grok 4',
+    description:
+      'xAI flagship. 256K context, vision-capable. Real-time web access on some variants. Automatic prompt caching server-side.',
+    contextTokens: 256_000,
+    capabilities: ['reasoning', 'function_calling', 'vision'],
+    inputPricePer1M: 3,
+    outputPricePer1M: 15,
+  },
+  // DeepSeek — open-weights reasoning.
+  {
+    id: 'deepseek/deepseek-v3.1',
+    label: 'DeepSeek V3.1',
+    description:
+      'Open-weights MoE. Strong reasoning at low price; popular for cost-sensitive workloads.',
+    contextTokens: 128_000,
+    capabilities: ['function_calling'],
+    inputPricePer1M: 0.27,
+    outputPricePer1M: 1.1,
+  },
+  // Meta — Llama 4 if/when added; placeholder pricing.
+  {
+    id: 'meta-llama/llama-4-maverick',
+    label: 'Llama 4 Maverick',
+    description:
+      'Open-weights from Meta. Reasonable cost, broad capability. Pricing varies by sub-provider OR routes to.',
+    contextTokens: 256_000,
+    capabilities: ['function_calling', 'vision'],
+    inputPricePer1M: 0.5,
+    outputPricePer1M: 1.5,
+  },
+];
+
+/** Curated vision-capable OpenRouter routes for the Vision + Document worker
+ *  dropdowns (before discovery returns the live image-input list). The
+ *  Anthropic + Gemini routes also support native PDF, so they back the
+ *  Document worker's `extractDocument` path; gpt-4o is image-only here. */
+export const OPENROUTER_VISION_MODELS: readonly VisionModelInfo[] = [
+  {
+    id: 'anthropic/claude-sonnet-5',
+    label: 'Claude Sonnet 5',
+    description: 'Strong document + table reading; native PDF. Great default for invoices.',
+    contextTokens: 1_000_000,
+    inputPricePer1M: 2,
+    outputPricePer1M: 10,
+    tier: 'balanced',
+  },
+  {
+    id: 'anthropic/claude-sonnet-4.6',
+    label: 'Claude Sonnet 4.6',
+    description: 'Previous-generation Sonnet vision; native PDF.',
+    contextTokens: 1_000_000,
+    inputPricePer1M: 3,
+    outputPricePer1M: 15,
+    tier: 'balanced',
+  },
+  {
+    id: 'anthropic/claude-opus-4.7',
+    label: 'Claude Opus 4.7',
+    description: 'Top-tier vision/document fidelity; native PDF. Highest cost.',
+    contextTokens: 1_000_000,
+    inputPricePer1M: 15,
+    outputPricePer1M: 75,
+    tier: 'quality',
+  },
+  {
+    id: 'anthropic/claude-haiku-4.5',
+    label: 'Claude Haiku 4.5',
+    description: 'Cheap + fast vision; native PDF. Good for high-volume describe/OCR.',
+    contextTokens: 200_000,
+    inputPricePer1M: 0.8,
+    outputPricePer1M: 4,
+    tier: 'fast',
+  },
+  {
+    id: 'google/gemini-2.5-pro',
+    label: 'Gemini 2.5 Pro',
+    description: 'Google multimodal; native PDF. Strong on dense layouts.',
+    contextTokens: 1_000_000,
+    inputPricePer1M: 1.25,
+    outputPricePer1M: 10,
+    tier: 'balanced',
+  },
+  {
+    id: 'openai/gpt-4o',
+    label: 'GPT-4o',
+    description: 'OpenAI multimodal (images). PDFs fall back to page OCR via OpenRouter.',
+    contextTokens: 128_000,
+    inputPricePer1M: 2.5,
+    outputPricePer1M: 10,
+    tier: 'balanced',
+  },
+];