@aeye/ai 0.3.9 → 0.7.0

package/aeye-ai.md

@@ -1,219 +1,237 @@
-# @aeye/ai
-
-> Multi-provider AI library: a single typed instance that selects models, runs chat/image/speech/transcribe/embed requests, tracks cost, and binds tools/prompts/agents — built on `@aeye/core`.
-
-## When to use it
-
-Use `@aeye/ai` as the main entry point of an aeye application. You create one `AI` instance, register one or more provider packages (`@aeye/openai`, `@aeye/aws`, `@aeye/openrouter`, `@aeye/replicate`, or a custom `Provider`), and then call typed APIs. The instance handles automatic model selection by capability/cost/speed, per-call context + metadata injection, lifecycle hooks (budgets, logging), and cost estimation.
-
-- Use `@aeye/ai` for: provider-agnostic chat/embeddings/images/speech, model selection, cost tracking, and AI-bound `tool`/`prompt`/`agent` components.
-- Use `@aeye/core` directly for: the underlying `Tool`/`Prompt`/`Agent` classes, `Request`/`Response`/`Chunk`/`Message` types, and strict-mode descriptors (`@aeye/ai` re-exports the request/response types you need).
-
-## Installation & import
-
-```bash
-npm install @aeye/ai @aeye/core
-npm install @aeye/openai openai      # or @aeye/aws, @aeye/openrouter, @aeye/replicate
-npm install @aeye/models             # optional: curated model table + strict overrides
-```
-
-There is **one** export path. Everything is imported from the package root — there are no subpath exports.
-
-```typescript
-import {
-  AI, ModelRegistry, getProviderCapabilities, resolveStrictFormat,
-  isModelInfo, detectTier,
-  // types:
-  type Provider, type ModelInfo, type ModelOverride, type ModelSource,
-  type ModelHandler, type AIConfig, type AIHooks, type ModelCapability,
-} from '@aeye/ai';
-```
-
-## Core concepts
-
-- **AI instance** (`AI<T>`): central object. Holds `config`, `registry`, the API surfaces, `providers`, `components`, and `hooks`.
-- **Providers**: objects implementing the `Provider<TConfig>` interface. Each implements only the methods for the capabilities it supports (`createExecutor`/`createStreamer` for chat, `embed`, `speech`, `transcribe`, `generateImage`, etc.). Capabilities are auto-detected from which methods exist.
-- **Context** (`TContext`): your app data threaded through every call. Resolved by merging `defaultContext` → `providedContext(ctx)` → per-call required context → `{ ai }`. Fields satisfied by defaults/provided are not required at call time.
-- **Metadata** (`TMetadata` + `AIBaseMetadata`): controls model selection (`model`, `required`/`optional` capabilities, `requiredParameters`, `providers.allow/deny`, `pricing`/`contextWindow`/`outputTokens`/`metrics` constraints, `weights`, `weightProfile`, `tier`) plus your custom fields. Merged like context.
-- **Registry** (`ModelRegistry`): stores `ModelInfo` records keyed by both `id` and `provider/id`, applies overrides, scores and selects models.
-- **APIs**: `chat`, `image.{generate,edit,analyze}`, `speech`, `transcribe`, `embed`, `models`. Each request API has `get()` (await) and `stream()` (async iterable); **`embed` has only `get()`**. Selection/hooks/cost logic is shared via `BaseAPI`.
-- **Components**: `ai.tool()`, `ai.prompt()`, `ai.agent()` wrap the `@aeye/core` classes and auto-inject the executor/streamer/context.
-
-## Creating an instance
-
-`AI.with<TContext, TMetadata>().providers(providers).create(config)` — a 3-step fluent builder that infers the full type container `T`. Both generics default to `{}`.
-
-```typescript
-import { AI } from '@aeye/ai';
-import { OpenAIProvider } from '@aeye/openai';
-
-interface AppContext { userId: string; user?: User }
-
-const openai = new OpenAIProvider({ apiKey: process.env.OPENAI_API_KEY! });
-
-const ai = AI.with<AppContext>()
-  .providers({ openai })            // keys become provider identifiers ('openai')
-  .create({
-    providedContext: async (ctx) => ({ user: await getUser(ctx.userId) }),
-    defaultWeights: { cost: 0.5, speed: 0.3, accuracy: 0.2 },
-  });
-```
-
-Constructing `new AI(config)` directly is also valid but loses inference; prefer the builder.
-
-### Key `AIConfig` fields
-
-| Field | Purpose |
-|-------|---------|
-| `providers` | Required. `Record<string, Provider>`. |
-| `defaultContext` / `providedContext` | Static + async-resolved context. |
-| `defaultMetadata` / `providedMetadata` | Static + async-resolved metadata. |
-| `models` | `ModelInfo[]` registered up front (e.g. from `@aeye/models`). |
-| `modelOverrides` | `ModelOverride[]` patched onto matching models (by `provider`/`modelId`/`modelPattern`). |
-| `modelHandlers` | Per-model custom implementations (`ModelHandler`). |
-| `modelSources` | External `ModelSource[]` fetched on `models.refresh()`. |
-| `defaultWeights` / `weightProfiles` | Default and named scoring weights. |
-| `defaultCostPerMillionTokens` | Fallback price for unknown models (default `5.0`). |
-| `tokens` | Token-estimation divisors per content type. |
-
-## Using the request APIs
-
-All request APIs follow `api.get(request, ctx?)` / `api.stream(request, ctx?)`. The second arg is the **required context** (only the fields not satisfied by defaults/provided), and may carry `metadata` and `signal`.
-
-```typescript
-// Chat (await)
-const res = await ai.chat.get(
-  { messages: [{ role: 'user', content: 'What is TypeScript?' }], temperature: 0.7, maxTokens: 512 },
-  { userId: 'u1' }
-);
-console.log(res.content, res.usage?.cost, res.finishReason);
-
-// Chat (stream)
-for await (const chunk of ai.chat.stream({ messages: [{ role: 'user', content: 'Count to 5' }] }, { userId: 'u1' })) {
-  if (chunk.content) process.stdout.write(chunk.content);
-}
-```
-
-Vision/audio capabilities are auto-required from message content (an `image` part adds `'vision'`, an `audio` part adds `'hearing'`); `reason`, `responseFormat`, and `tools` add the matching capabilities/parameters during selection.
-
-```typescript
-// Embeddings (get only)
-const e = await ai.embed.get({ texts: ['hello', 'world'], dimensions: 1536 });
-e.embeddings.forEach(({ embedding, index }) => console.log(index, embedding.length));
-
-// Images
-await ai.image.generate.get({ prompt: 'a cat astronaut', n: 1, size: '1024x1024', quality: 'high' });
-await ai.image.edit.get({ prompt: 'add a sunset', image: buf, mask: maskBuf, size: '1024x1024' });
-await ai.image.analyze.get({ prompt: 'describe this', images: ['https://...jpg'] }); // routes through a vision chat model
-
-// Speech / Transcription
-const speech = await ai.speech.get({ text: 'Hello there', voice: 'alloy', speed: 1.0 });
-const text = await ai.transcribe.get({ audio: audioBuf, language: 'en' });
-```
-
-Request/response shapes are exported types: `Request`/`Response`/`Chunk`/`Message` (re-exported from core), plus `EmbeddingRequest`/`EmbeddingResponse`, `ImageGenerationRequest`/`ImageEditRequest`/`ImageGenerationResponse`/`ImageGenerationChunk`, `ImageAnalyzeRequest`, `SpeechRequest`/`SpeechResponse`, `TranscriptionRequest`/`TranscriptionResponse`/`TranscriptionChunk`.
-
-## Model selection & the Models API
-
-When `request.model` / `metadata.model` is set, that model is used directly. Otherwise the registry scores **applicable** models (those passing required capabilities, provider allow/deny, and range constraints) using weighted `cost`/`speed`/`accuracy`/`contextWindow` (weights resolve as `metadata.weights` → `metadata.weightProfile` → config `defaultWeights` → `{ cost:0.5, speed:0.3, accuracy:0.2 }`). Optional capabilities/parameters act as preference multipliers, not filters.
-
-```typescript
-ai.models.list(providedOnly?)                 // all (or only provider-backed) ModelInfo[]
-ai.models.get('gpt-4o' | 'openai/gpt-4o')     // ModelInfo | undefined
-ai.models.search({ required: ['chat','structured'], weights: { cost: 0.6, speed: 0.4 } }) // ScoredModel[]
-ai.models.select(criteria)                     // SelectedModel | undefined
-await ai.models.refresh()                      // re-fetch from providers + modelSources
-```
-
-`ModelCapability` = `'chat' | 'tools' | 'vision' | 'json' | 'structured' | 'streaming' | 'reasoning' | 'image' | 'audio' | 'hearing' | 'embedding' | 'zdr' | 'toolsStrict'`. `ModelTier` = `'flagship' | 'efficient' | 'legacy' | 'experimental'`.
-
-## Tools, prompts & agents
-
-`ai.tool()`, `ai.prompt()`, `ai.agent()` return enhanced `@aeye/core` `Tool`/`Prompt`/`Agent` instances with the AI executor/streamer/context auto-wired (you don't pass `types`). Every created component is pushed to `ai.components`.
-
-```typescript
-import z from 'zod';
-
-const searchKnowledge = ai.tool({
-  name: 'searchKnowledge',
-  description: 'Search the knowledge base',
-  schema: z.object({ query: z.string(), limit: z.number().optional() }),
-  call: async ({ query, limit = 5 }, _refs, ctx) => ({ results: await search(query, limit) }),
-});
-
-const assistant = ai.prompt({
-  name: 'assistant',
-  description: 'Answers using the knowledge base',
-  content: 'Use searchKnowledge then answer.\n\nQuestion: {{question}}',
-  input: (i: { question: string }) => ({ question: i.question }),
-  tools: [searchKnowledge],
-  schema: z.object({ answer: z.string(), sources: z.array(z.string()) }),
-});
-
-const result = await assistant.get('result', { question: 'How does X work?' }, { userId: 'u1' });
-for await (const t of assistant.get('streamContent', { question: '...' })) process.stdout.write(t);
-
-// Or run any component with context injected:
-const out = await ai.run(assistant, { question: '...' }, { userId: 'u1' });
-```
-
-`ai.agent({ name, description, refs, call })` composes other components; `call(input, [..refs], ctx)` receives the resolved refs tuple and the full AI context.
-
-`ai.tool()` and `ai.prompt()` forward a trailing **`TDecoded`** generic to the underlying core `Tool`/`Prompt`, so a custom `parse` (which replaces Zod — see the [structured-output](../docs/guides/structured-output.md) / [tool-calling](../docs/guides/tool-calling.md) guides) has its **return type surface through the wrapper**: `call`/`validate` receive the decoded value, not the wire params/output. Absent a custom `parse`, `TDecoded` defaults to the wire type.
-
-## Provider & core integration
-
-- **Provider packages** export classes implementing the core `Provider` interface, e.g. `new OpenAIProvider({ apiKey })`, `new OpenRouterProvider({ apiKey })`, `new AWSProvider(...)`, `new ReplicateProvider(...)`. You map them under arbitrary keys in `.providers({ ... })`; the key is the provider identifier used in `metadata.providers.allow/deny`, model IDs (`provider/model`), and `ai.providers`.
-- **`@aeye/models`** provides a curated `models` array and a `strictSupport` `ModelOverride[]`. Splat both so selection knows real pricing/capabilities and which models truly support strict tool/output mode:
-
-  ```typescript
-  import { models, strictSupport } from '@aeye/models';
-  const ai = AI.with().providers({ openai }).create({ models, modelOverrides: [...strictSupport] });
-  ```
-- **`@aeye/core`** supplies the building blocks `@aeye/ai` re-exports (`Request`, `Response`, `Chunk`, `Message`, `Usage`, `Executor`, `Streamer`, `ToolInput`, `PromptInput`, `AgentInput`, `Component*`). Strict-mode descriptor mechanics live in core; `@aeye/ai` only wires the `strict` flag into selection (`true` → required `toolsStrict`, numeric/omitted → optional preference). Custom format families are registered in core via `registerDescriptor`.
-
-### Hooks
-
-Attach with `.withHooks(...)` (returns `this`) or via `config.hooks`-style assignment. Hooks run around every request:
-
-```typescript
-ai.withHooks({
-  beforeModelSelection: async (ctx, request, metadata) => metadata,         // adjust criteria
-  onModelSelected: async (ctx, request, selected) => {/* return to override */},
-  beforeRequest: async (ctx, request, selected, estimatedUsage, estimatedCost) => {
-    if (ctx.user && estimatedCost > ctx.user.budget) throw new Error('over budget'); // throw to cancel
-  },
-  afterRequest: async (ctx, request, response, complete, selected, usage, cost) => {/* track */},
-  onError: (type, message, error, ctx, request) => console.error(type, message),
-});
-```
-
-## Extending instances
-
-`ai.extend<TExtraContext, TExtraMetadata>(config?)` returns a new `AI` sharing providers/config, with merged defaults and additively-merged `models`/`modelOverrides`/`modelHandlers`. Useful for feature-scoped context.
-
-```typescript
-const chatAI = ai.extend<{ chatId: string }>({ defaultContext: {} });
-```
-
-## Other exports
-
-- `ModelRegistry` — the registry class (`getModel`, `registerModel(s)`, `listModels`, `providedModels`, `searchModels`, `selectModel`, `getProvider`, `getProviderCapabilities`, `getStrictFormat`, `refresh`). Accessible as `ai.registry`.
-- `getProviderCapabilities(provider)` → `Set<ModelCapability>` inferred from provider methods.
-- `resolveStrictFormat(model)` → strict dialect family or `undefined` (checks `strictFormat`, then `provider`, then `id` prefix).
-- `isModelInfo(input)` — type guard distinguishing a full `ModelInfo` from a model reference.
-- `detectTier(name)`, `detectCapabilitiesFromModality(modality, id)`, `detectZDRFromModeration(isModerated)` — heuristics used as fallbacks during refresh.
-- `ai.stats()` → `LibraryStats` (model counts, cumulative requests, average cost/latency).
-
-## Gotchas
-
-- **`embed` has no `stream()`** — only `embed.get()`. Image/speech "streaming" may fall back to a single chunk if the provider lacks a true streamer (the base API converts executor↔streamer automatically).
-- **Capabilities are per-provider AND per-model.** A required capability must be supported by both the model (`ModelInfo.capabilities`) and the provider (auto-detected from its methods), or the model is filtered out. Missing the right provider method → "No compatible model found".
-- **Models registered under two keys** (`id` and `provider/id`); colliding registrations are merged, not replaced. Use `provider/model` IDs to disambiguate across providers.
-- **`refresh()` clears and re-fetches** all models. Models passed via `config.models` are preserved because they are also added as an internal `'config'` model source.
-- **Pricing is per 1M tokens** (image output pricing is per-image by quality/size; `perRequest` is a flat add-on). `estimate*Usage` is heuristic (char-count divisors), not a real tokenizer — treat estimated cost as approximate.
-- **Selection without registered models fails** — with no `models`/`modelSources`/provider `listModels`, only an explicit `metadata.model` will resolve. Call `models.refresh()` or pass `config.models` first.
-- **`strict: true` is a hard filter**; numeric or omitted `strict` is only a scoring preference. Without `@aeye/models` `strictSupport` overrides (or explicit `strictFormat`), models default to lenient even if the API supports strict.
-- **Builder generics carry no runtime data** — `AI.with<Ctx, Meta>()` only types the instance; actual defaults come from `.create(config)`.
-</content>
-</invoke>
+# @aeye/ai
+
+> Multi-provider AI library: a single typed instance that selects models, runs chat/image/speech/transcribe/embed/rerank/decide requests and provider batches, tracks cost, and binds tools/prompts/agents — built on `@aeye/core`.
+
+## When to use it
+
+Use `@aeye/ai` as the main entry point of an aeye application. You create one `AI` instance, register one or more provider packages (`@aeye/openai`, `@aeye/aws`, `@aeye/openrouter`, `@aeye/replicate`, or a custom `Provider`), and then call typed APIs. The instance handles automatic model selection by capability/cost/speed, per-call context + metadata injection, lifecycle hooks (budgets, logging), and cost estimation.
+
+- Use `@aeye/ai` for: provider-agnostic chat/embeddings/images/speech, model selection, cost tracking, and AI-bound `tool`/`prompt`/`agent` components.
+- Use `@aeye/core` directly for: the underlying `Tool`/`Prompt`/`Agent` classes, `Request`/`Response`/`Chunk`/`Message` types, and strict-mode descriptors (`@aeye/ai` re-exports the request/response types you need).
+
+## Installation & import
+
+```bash
+npm install @aeye/ai @aeye/core
+npm install @aeye/openai openai      # or @aeye/aws, @aeye/openrouter, @aeye/replicate
+npm install @aeye/models             # optional: curated model table + strict overrides
+```
+
+There is **one** export path. Everything is imported from the package root — there are no subpath exports.
+
+```typescript
+import {
+  AI, ModelRegistry, getProviderCapabilities, resolveStrictFormat,
+  isModelInfo, detectTier,
+  // types:
+  type Provider, type ModelInfo, type ModelOverride, type ModelSource,
+  type ModelHandler, type AIConfig, type AIHooks, type ModelCapability,
+} from '@aeye/ai';
+```
+
+## Core concepts
+
+- **AI instance** (`AI<T>`): central object. Holds `config`, `registry`, the API surfaces, `providers`, `components`, and `hooks`.
+- **Providers**: objects implementing the `Provider<TConfig>` interface. Each implements only the methods for the capabilities it supports (`createExecutor`/`createStreamer` for chat, `embed`, `speech`, `transcribe`, `generateImage`, etc.). Capabilities are auto-detected from which methods exist.
+- **Context** (`TContext`): your app data threaded through every call. Resolved by merging `defaultContext` → `providedContext(ctx)` → per-call required context → `{ ai }`. Fields satisfied by defaults/provided are not required at call time.
+- **Metadata** (`TMetadata` + `AIBaseMetadata`): controls model selection (`model`, `required`/`optional` capabilities, `requiredParameters`, `providers.allow/deny`, `pricing`/`contextWindow`/`outputTokens`/`metrics` constraints, `weights`, `weightProfile`, `tier`) plus your custom fields. Merged like context.
+- **Registry** (`ModelRegistry`): stores `ModelInfo` records keyed by both `id` and `provider/id`, applies overrides, scores and selects models.
+- **APIs**: `chat`, `image.{generate,edit,analyze}`, `speech`, `transcribe`, `embed`, `rerank`, `decide`, `batch`, `models`. Each request API has `get()` (await) and `stream()` (async iterable); **`embed`, `rerank` and `decide` have only `get()`**, and `batch` has `submit()`, `get()` and `wait()`. Selection/hooks/cost logic is shared via `BaseAPI`.
+- **Components**: `ai.tool()`, `ai.prompt()`, `ai.agent()` wrap the `@aeye/core` classes and auto-inject the executor/streamer/context.
+
+## Creating an instance
+
+`AI.with<TContext, TMetadata>().providers(providers).create(config)` — a 3-step fluent builder that infers the full type container `T`. Both generics default to `{}`.
+
+```typescript
+import { AI } from '@aeye/ai';
+import { OpenAIProvider } from '@aeye/openai';
+
+interface AppContext { userId: string; user?: User }
+
+const openai = new OpenAIProvider({ apiKey: process.env.OPENAI_API_KEY! });
+
+const ai = AI.with<AppContext>()
+  .providers({ openai })            // keys become provider identifiers ('openai')
+  .create({
+    providedContext: async (ctx) => ({ user: await getUser(ctx.userId) }),
+    defaultWeights: { cost: 0.5, speed: 0.3, accuracy: 0.2 },
+  });
+```
+
+Constructing `new AI(config)` directly is also valid but loses inference; prefer the builder.
+
+### Key `AIConfig` fields
+
+| Field | Purpose |
+|-------|---------|
+| `providers` | Required. `Record<string, Provider>`. |
+| `defaultContext` / `providedContext` | Static + async-resolved context. |
+| `defaultMetadata` / `providedMetadata` | Static + async-resolved metadata. |
+| `models` | `ModelInfo[]` registered up front (e.g. from `@aeye/models`). |
+| `modelOverrides` | `ModelOverride[]` patched onto matching models (by `provider`/`modelId`/`modelPattern`). |
+| `modelHandlers` | Per-model custom implementations (`ModelHandler`). |
+| `modelSources` | External `ModelSource[]` fetched on `models.refresh()`. |
+| `defaultWeights` / `weightProfiles` | Default and named scoring weights. |
+| `defaultCostPerMillionTokens` | Fallback price for unknown models (default `5.0`). |
+| `tokens` | Token-estimation divisors per content type. |
+
+Hooks are not a config field: attach them with `ai.withHooks(...)`. Callbacks in the config (e.g. `providedContext`) are typed from `TContext`/`TMetadata`.
+
+## Using the request APIs
+
+All request APIs follow `api.get(request, ctx?)` / `api.stream(request, ctx?)`. The second arg is the **required context** (only the fields not satisfied by defaults/provided), and may carry `metadata` and `signal`.
+
+```typescript
+// Chat (await)
+const res = await ai.chat.get(
+  { messages: [{ role: 'user', content: 'What is TypeScript?' }], temperature: 0.7, maxTokens: 512 },
+  { userId: 'u1' }
+);
+console.log(res.content, res.usage?.cost, res.finishReason);
+
+// Chat (stream)
+for await (const chunk of ai.chat.stream({ messages: [{ role: 'user', content: 'Count to 5' }] }, { userId: 'u1' })) {
+  if (chunk.content) process.stdout.write(chunk.content);
+}
+```
+
+Vision/audio capabilities are auto-required from message content (an `image` part adds `'vision'`, an `audio` part adds `'hearing'`); `reason`, `responseFormat`, and `tools` add the matching capabilities/parameters during selection.
+
+```typescript
+// Embeddings (get only)
+const e = await ai.embed.get({ texts: ['hello', 'world'], dimensions: 1536 });
+e.embeddings.forEach(({ embedding, index }) => console.log(index, embedding.length));
+
+// Images
+await ai.image.generate.get({ prompt: 'a cat astronaut', n: 1, size: '1024x1024', quality: 'high' });
+await ai.image.edit.get({ prompt: 'add a sunset', image: buf, mask: maskBuf, size: '1024x1024' });
+await ai.image.analyze.get({ prompt: 'describe this', images: ['https://...jpg'] }); // routes through a vision chat model
+
+// Speech / Transcription
+const speech = await ai.speech.get({ text: 'Hello there', voice: 'alloy', speed: 1.0 });
+const text = await ai.transcribe.get({ audio: audioBuf, language: 'en' });
+
+// Rerank (OpenRouter) / Decisions (OpenAI, OpenRouter)
+const { results } = await ai.rerank.get({ query: 'capital of France', documents, topN: 1 });
+const { answers } = await ai.decide.get({ input: text, questions: { spam: { type: 'predicate', instructions: 'Is this spam?' } } });
+
+// Batches (OpenAI, OpenRouter): about half price, results within a day
+const job = await ai.batch.submit([{ messages }, { messages: other }], {});
+const done = await ai.batch.wait(job, {}); // done.results in request order
+```
+
+### Reasoning
+
+`reason: { effort }` is clamped to the model's `reasoningEfforts` with `clampReasoningEffort(effort, accepted)`: nearest accepted, lower on a tie, and `none` only when asked for or alone accepted (`low` on `[none, high, xhigh]` sends `high`). Keep `message.reasoning` when sending a conversation back: `Reasoning.provider` names who produced it, and a provider replays only its own or unnamed reasoning (`replaysTo` in core).
+
+Request/response shapes are exported types: `Request`/`Response`/`Chunk`/`Message` (re-exported from core), plus `EmbeddingRequest`/`EmbeddingResponse`, `ImageGenerationRequest`/`ImageEditRequest`/`ImageGenerationResponse`/`ImageGenerationChunk`, `ImageAnalyzeRequest`, `SpeechRequest`/`SpeechResponse`, `TranscriptionRequest`/`TranscriptionResponse`/`TranscriptionChunk`.
+
+## Model selection & the Models API
+
+When `request.model` / `metadata.model` is set, that model is used directly. Otherwise the registry scores **applicable** models (those passing required capabilities, provider allow/deny, and range constraints) using weighted `cost`/`speed`/`accuracy`/`contextWindow` (weights resolve as `metadata.weights` → `metadata.weightProfile` → config `defaultWeights` → `{ cost:0.5, speed:0.3, accuracy:0.2 }`). Optional capabilities/parameters act as preference multipliers, not filters.
+
+```typescript
+ai.models.list(providedOnly?)                 // all (or only provider-backed) ModelInfo[]
+ai.models.get('gpt-4o' | 'openai/gpt-4o')     // ModelInfo | undefined
+ai.models.search({ required: ['chat','structured'], weights: { cost: 0.6, speed: 0.4 } }) // ScoredModel[]
+ai.models.select(criteria)                     // SelectedModel | undefined
+await ai.models.refresh()                      // re-fetch from providers + modelSources
+```
+
+`ModelCapability` = `'chat' | 'tools' | 'vision' | 'files' | 'videoInput' | 'json' | 'structured' | 'streaming' | 'reasoning' | 'image' | 'audio' | 'hearing' | 'embedding' | 'rerank' | 'decision' | 'zdr' | 'toolsStrict'`. `ModelTier` = `'flagship' | 'efficient' | 'legacy' | 'experimental'`.
+
+## Tools, prompts & agents
+
+`ai.tool()`, `ai.prompt()`, `ai.agent()` return enhanced `@aeye/core` `Tool`/`Prompt`/`Agent` instances with the AI executor/streamer/context auto-wired (you don't pass `types`). Every created component is pushed to `ai.components`.
+
+```typescript
+import z from 'zod';
+
+const searchKnowledge = ai.tool({
+  name: 'searchKnowledge',
+  description: 'Search the knowledge base',
+  schema: z.object({ query: z.string(), limit: z.number().optional() }),
+  call: async ({ query, limit = 5 }, _refs, ctx) => ({ results: await search(query, limit) }),
+});
+
+const assistant = ai.prompt({
+  name: 'assistant',
+  description: 'Answers using the knowledge base',
+  content: 'Use searchKnowledge then answer.\n\nQuestion: {{question}}',
+  input: (i: { question: string }) => ({ question: i.question }),
+  tools: [searchKnowledge],
+  schema: z.object({ answer: z.string(), sources: z.array(z.string()) }),
+});
+
+const result = await assistant.get('result', { question: 'How does X work?' }, { userId: 'u1' });
+for await (const t of assistant.get('streamContent', { question: '...' })) process.stdout.write(t);
+
+// Or run any component with context injected:
+const out = await ai.run(assistant, { question: '...' }, { userId: 'u1' });
+```
+
+Every callback given to `ai.prompt` (`input`, `schema`, `config`, `reconfig`, `retool`, `parse`, `validate`, `review`, `onToolIntercept`, `applicable`, `metadataFn`) receives the full AI context. Prompts also take `review` / `reviewRetries` (check valid output, send feedback for another round), `schemaFallbacks` (steps taken when a provider rejects a schema; default `SCHEMA_FALLBACKS`, AI-wide via `defaultContext`), and `userTurn` (the user message added when run without a conversation, default `'Begin.'`). Errors carry `kind: 'model' | 'tool'` (`PromptError`, `ToolInputError`, `ToolExecutionError`, `OutputValidationError`, `ReviewRejectedError`, all from core).
+
+`ai.agent({ name, description, refs, call })` composes other components; `call(input, [..refs], ctx)` receives the resolved refs tuple and the full AI context.
+
+`ai.tool()` and `ai.prompt()` forward a trailing **`TDecoded`** generic to the underlying core `Tool`/`Prompt`, so a custom `parse` (which replaces Zod — see the [structured-output](../docs/guides/structured-output.md) / [tool-calling](../docs/guides/tool-calling.md) guides) has its **return type surface through the wrapper**: `call`/`validate` receive the decoded value, not the wire params/output. Absent a custom `parse`, `TDecoded` defaults to the wire type.
+
+## Provider & core integration
+
+- **Provider packages** export classes implementing the core `Provider` interface, e.g. `new OpenAIProvider({ apiKey })`, `new OpenRouterProvider({ apiKey })`, `new AWSProvider(...)`, `new ReplicateProvider(...)`. You map them under arbitrary keys in `.providers({ ... })`; the key is the provider identifier used in `metadata.providers.allow/deny`, model IDs (`provider/model`), and `ai.providers`.
+- **`@aeye/models`** provides a curated `models` array and a `strictSupport` `ModelOverride[]`. Splat both so selection knows real pricing/capabilities and which models truly support strict tool/output mode:
+
+  ```typescript
+  import { models, strictSupport } from '@aeye/models';
+  const ai = AI.with().providers({ openai }).create({ models, modelOverrides: [...strictSupport] });
+  ```
+- **`@aeye/core`** supplies the building blocks `@aeye/ai` re-exports (`Request`, `Response`, `Chunk`, `Message`, `Usage`, `Executor`, `Streamer`, `ToolInput`, `PromptInput`, `AgentInput`, `Component*`). Strict-mode descriptor mechanics live in core; `@aeye/ai` only wires the `strict` flag into selection (`true` → required `toolsStrict`, numeric/omitted → optional preference). Custom format families are registered in core via `registerDescriptor`.
+
+### Hooks
+
+Attach with `.withHooks(...)` (returns `this`) or by assigning `ai.hooks`; `create(config)` has no hooks field. Hooks run around every request:
+
+```typescript
+ai.withHooks({
+  beforeModelSelection: async (ctx, request, metadata) => metadata,         // adjust criteria
+  onModelSelected: async (ctx, request, selected) => {/* return to override */},
+  beforeRequest: async (ctx, request, selected, estimatedUsage, estimatedCost) => {
+    if (ctx.user && estimatedCost > ctx.user.budget) throw new Error('over budget'); // throw to cancel
+  },
+  afterRequest: async (ctx, request, response, complete, selected, usage, cost) => {/* track */},
+  afterBatch: async (ctx, job, usage, cost) => {/* once per completed batch */},
+  onError: (type, message, error, ctx, request) => console.error(type, message),
+});
+```
+
+## Extending instances
+
+`ai.extend<TExtraContext, TExtraMetadata>(config?)` returns a new `AI` sharing providers/config, with merged defaults and additively-merged `models`/`modelOverrides`/`modelHandlers`. Useful for feature-scoped context.
+
+```typescript
+const chatAI = ai.extend<{ chatId: string }>({ defaultContext: {} });
+```
+
+## Other exports
+
+- `ModelRegistry` — the registry class (`getModel`, `registerModel(s)`, `listModels`, `providedModels`, `searchModels`, `selectModel`, `getProvider`, `getProviderCapabilities`, `getStrictFormat`, `refresh`). Accessible as `ai.registry`.
+- `getProviderCapabilities(provider)` → `Set<ModelCapability>` inferred from provider methods.
+- `resolveStrictFormat(model)` → strict dialect family or `undefined` (checks `strictFormat`, then `provider`, then `id` prefix).
+- `clampReasoningEffort(effort, accepted?)` — the accepted reasoning effort nearest the one asked for (unchanged when `accepted` is unknown).
+- `isModelInfo(input)` — type guard distinguishing a full `ModelInfo` from a model reference.
+- `detectTier(name)`, `detectCapabilitiesFromModality(modality, id)`, `detectZDRFromModeration(isModerated)` — heuristics used as fallbacks during refresh.
+- `ai.stats()` → `LibraryStats` (model counts, cumulative requests, average cost/latency).
+
+## Gotchas
+
+- **`embed` has no `stream()`** — only `embed.get()`. Image/speech "streaming" may fall back to a single chunk if the provider lacks a true streamer (the base API converts executor↔streamer automatically).
+- **Capabilities are per-provider AND per-model.** A required capability must be supported by both the model (`ModelInfo.capabilities`) and the provider (auto-detected from its methods), or the model is filtered out. Missing the right provider method → "No compatible model found".
+- **Models registered under two keys** (`id` and `provider/id`); colliding registrations are merged, not replaced. Use `provider/model` IDs to disambiguate across providers.
+- **`refresh()` clears and re-fetches** all models. Models passed via `config.models` are preserved because they are also added as an internal `'config'` model source.
+- **Pricing is per 1M tokens** (image output pricing is per-image by quality/size; `perRequest` is a flat add-on). `estimate*Usage` is heuristic (char-count divisors), not a real tokenizer — treat estimated cost as approximate.
+- **Selection without registered models fails** — with no `models`/`modelSources`/provider `listModels`, only an explicit `metadata.model` will resolve. Call `models.refresh()` or pass `config.models` first.
+- **`strict: true` is a hard filter**; numeric or omitted `strict` is only a scoring preference. Without `@aeye/models` `strictSupport` overrides (or explicit `strictFormat`), models default to lenient even if the API supports strict.
+- **Builder generics carry no runtime data** — `AI.with<Ctx, Meta>()` only types the instance; actual defaults come from `.create(config)`.
+</content>
+</invoke>

package/dist/ai.d.ts

@@ -5,8 +5,11 @@
  * Provides type-safe, scoped access to AI capabilities with hooks and context injection.
  */
 import { Agent, Extend, Fn, MessageContentType, Prompt, RetoolResult, Tool, ToolCompatible, Tuple } from '@aeye/core';
+import { BatchAPI } from './apis/batch';
 import { ChatAPI } from './apis/chat';
 import { EmbedAPI } from './apis/embed';
+import { RerankAPI } from './apis/rerank';
+import { DecideAPI } from './apis/decide';
 import { ImageAPI } from './apis/image';
 import { ModelsAPI } from './apis/models';
 import { SpeechAPI } from './apis/speech';
@@ -52,6 +55,9 @@ export declare class AI<T extends AIBaseTypes> {
     readonly speech: SpeechAPI<T>;
     readonly transcribe: TranscribeAPI<T>;
     readonly embed: EmbedAPI<T>;
+    readonly rerank: RerankAPI<T>;
+    readonly decide: DecideAPI<T>;
+    readonly batch: BatchAPI<T>;
     readonly models: ModelsAPI<AIProviders<T>>;
     readonly providers: AIProviders<T>;
     readonly components: ComponentFor<T>[];
@@ -70,7 +76,7 @@ export declare class AI<T extends AIBaseTypes> {
      */
     static with<TContext extends AIContextUser = {}, TMetadata extends AIMetadataUser = {}>(): {
         providers: <TProviders extends Providers>(providers: TProviders) => {
-            create: <TConfig extends Omit<AIConfig<TContext, TMetadata, TProviders>, "providers">>(config: TConfig) => AI<AITypesInfer<TContext, TMetadata, TProviders, TConfig>>;
+            create: <TConfig extends Omit<AIConfig<TContext, TMetadata, TProviders>, "providers">>(config?: TConfig) => AI<AITypesInfer<TContext, TMetadata, TProviders, TConfig>>;
         };
     };
     constructor(config: AIConfigOf<T>);

package/dist/ai.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"ai.d.ts","sourceRoot":"","sources":["../src/ai.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAmB,KAAK,EAAU,MAAM,EAAE,EAAE,EAAc,kBAAkB,EAAE,MAAM,EAAa,YAAY,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,YAAY,CAAC;AACtK,OAAO,EAAE,OAAO,EAAE,MAAM,aAAa,CAAC;AACtC,OAAO,EAAE,QAAQ,EAAE,MAAM,cAAc,CAAC;AACxC,OAAO,EAAE,QAAQ,EAAE,MAAM,cAAc,CAAC;AACxC,OAAO,EAAE,SAAS,EAAE,MAAM,eAAe,CAAC;AAC1C,OAAO,EAAE,SAAS,EAAE,MAAM,eAAe,CAAC;AAC1C,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAClD,OAAO,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAC3C,OAAO,KAAK,EACV,UAAU,EACV,cAAc,EACd,WAAW,EACX,QAAQ,EACR,UAAU,EACV,SAAS,EACT,iBAAiB,EACjB,iBAAiB,EACjB,aAAa,EACb,OAAO,EACP,UAAU,EACV,kBAAkB,EAClB,cAAc,EACd,eAAe,EACf,WAAW,EACX,YAAY,EACZ,YAAY,EACZ,cAAc,EACd,eAAe,EACf,OAAO,EAEP,YAAY,EACZ,OAAO,EACP,SAAS,EACT,aAAa,EACb,WAAW,EACX,SAAS,EACT,OAAO,EACP,gBAAgB,EAEhB,SAAS,EACT,KAAK,EACN,MAAM,SAAS,CAAC;AAIjB,KAAK,WAAW,CACd,CAAC,SAAS,WAAW,EACrB,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,OAAO,SAAS,MAAM,GAAG,EAAE,EAC3B,OAAO,GAAG,OAAO,EACjB,KAAK,SAAS,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EACzC,QAAQ,SAAS,OAAO,GAAG,OAAO,IAChC,IAAI,CAAC,SAAS,CAChB,SAAS,CAAC,CAAC,CAAC,EACZ,UAAU,CAAC,CAAC,CAAC,EACb,KAAK,EACL,OAAO,EACP,OAAO,EAEP,KAAK,EACL,QAAQ,CACT,EAAE,OAAO,CAAC,CAAC;AAEZ,KAAK,aAAa,CAChB,CAAC,SAAS,WAAW,EACrB,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,MAAM,SAAS,MAAM,GAAG,EAAE,EAC1B,OAAO,SAAS,MAAM,GAAG,MAAM,GAAG,MAAM,EACxC,MAAM,SAAS,KAAK,CAAC,cAAc,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EACtF,QAAQ,SAAS,OAAO,GAAG,OAAO,IAChC,IAAI,CAAC,WAAW,CAClB,SAAS,CAAC,CAAC,CAAC,EACZ,UAAU,CAAC,CAAC,CAAC,EACb,KAAK,EACL,MAAM,EACN,OAAO,EAEP,MAAM,EACN,QAAQ,CACT,EAAE,OAAO,GAAG,QAAQ,CAAC,GAAG;IACvB,MAAM,CAAC,EAAE,EAAE,CAAC,YAAY,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAAC,CAAC,EAAE,MAAM,CAAC,EAAE,CAAC,MAAM,GAAG,SAAS,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;CAClH,CAAC;AAEF,KAAK,YAAY,CACf,CAAC,SAAS,WAAW,EACrB,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,MAAM,SAAS,MAAM,GAAG,EAAE,EAC1B,OAAO,GAAG,OAAO,EACjB,KAAK,SAAS,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,IACvC,IAAI,CAAC,UAAU,CACjB,SAAS,CAAC,CAAC,CAAC,EACZ,UAAU,CAAC,CAAC,CAAC,EACb,KAAK,EACL,MAAM,EACN,OAAO,EAEP,KAAK,CACN,EAAE,OAAO,CAAC,CAAC;AAEZ;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AACH,qBAAa,EAAE,CAAC,CAAC,SAAS,WAAW;IAEnC,SAAgB,MAAM,EAAE,UAAU,CAAC,CAAC,CAAC,CAAC;IACtC,SAAgB,QAAQ,EAAE,aAAa,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;IAGxD,SAAgB,IAAI,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC;IACjC,SAAgB,KAAK,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC;IACnC,SAAgB,MAAM,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC;IACrC,SAAgB,UAAU,EAAE,aAAa,CAAC,CAAC,CAAC,CAAC;IAC7C,SAAgB,KAAK,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC;IACnC,SAAgB,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;IAGlD,SAAgB,SAAS,EAAE,WAAW,CAAC,CAAC,CAAC,CAAC;IAC1C,SAAgB,UAAU,EAAE,YAAY,CAAC,CAAC,CAAC,EAAE,CAAM;IAC5C,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC;IAGzB,OAAO,CAAC,cAAc,CAAa;IACnC,OAAO,CAAC,iBAAiB,CAAa;IACtC,OAAO,CAAC,sBAAsB,CAAa;IAG3C,SAAgB,MAAM,EAAE,MAAM,CAAC,kBAAkB,EAAE;QACjD,OAAO,EAAE,MAAM,CAAC;QAChB,aAAa,EAAE,MAAM,CAAC;QACtB,QAAQ,EAAE,MAAM,CAAC;QACjB,GAAG,CAAC,EAAE,MAAM,CAAC;KACd,CAAC,CAAC;IAEH;;OAEG;WACW,IAAI,CAAC,QAAQ,SAAS,aAAa,GAAG,EAAE,EAAE,SAAS,SAAS,cAAc,GAAG,EAAE;oBAE7E,UAAU,SAAS,SAAS,aAAa,UAAU;qBAElD,OAAO,SAAS,IAAI,CAAC,QAAQ,CAAC,QAAQ,EAAE,SAAS,EAAE,UAAU,CAAC,EAAE,WAAW,CAAC,UAC3E,OAAO;;;gBAYb,MAAM,EAAE,UAAU,CAAC,CAAC,CAAC;IA6EjC;;;;;OAKG;IACH,SAAS,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,GAAG,IAAI;IAKlC;;;;;;;OAOG;IACG,YAAY,CAAC,WAAW,EAAE,iBAAiB,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;IAgB5E;;;;;;;OAOG;IACG,gBAAgB,CAAC,WAAW,EAAE,iBAAiB,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;IAoB9E;;;;;;;OAOG;IACG,aAAa,CAAC,gBAAgB,EAAE,kBAAkB,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC;IAiBpF;;;;;;;;OAQG;IACH,aAAa,CAAC,GAAG,SAAS,EAAE,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC,CAAC,GAAG,SAAS,CAAC,CAAC,GAAG,UAAU,CAAC,CAAC,CAAC;IA8HrF;;;OAGG;IACH,WAAW,CAAC,QAAQ,EAAE,UAAU,CAAC,CAAC,CAAC,GAAG,gBAAgB,CAAC,CAAC,CAAC,GAAG,SAAS;IAIrE;;;;;OAKG;IACH,oBAAoB,CAAC,OAAO,EAAE,OAAO,GAAG,KAAK;IAuG7C;;;OAGG;IACH,oBAAoB,CAAC,OAAO,EAAE,OAAO,GAAG,KAAK;IAW7C;;;;OAIG;IACH,aAAa,CACX,KAAK,EAAE,SAAS,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,EACpC,KAAK,EAAE,KAAK,GACX,MAAM;IAoGT;;;;;;OAMG;IACH,eAAe,CAAC,KAAK,EAAE,SAAS,EAAE,QAAQ,EAAE,aAAa,GAAG,OAAO;IAmBnE;;;;;;OAMG;IACG,GAAG,CAAC,CAAC,SAAS,YAAY,CAAC,CAAC,CAAC,EACjC,SAAS,EAAE,CAAC,EACZ,KAAK,EAAE,cAAc,CAAC,CAAC,CAAC,EACxB,GAAG,CAAC,GAAG,CAAC,EAAE,iBAAiB,CAAC,CAAC,CAAC,GAC7B,OAAO,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC;IAS9B;;;;;;;;;OASG;IACH,MAAM,CACJ,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,MAAM,SAAS,MAAM,GAAG,EAAE,EAC1B,OAAO,SAAS,MAAM,GAAG,MAAM,GAAG,MAAM,EACxC,MAAM,SAAS,KAAK,CAAC,cAAc,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EACtF,QAAQ,SAAS,OAAO,GAAG,OAAO,EAElC,OAAO,EAAE,aAAa,CAAC,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,QAAQ,CAAC;IAwFrE;;OAEG;IACH,IAAI,CACF,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,OAAO,SAAS,MAAM,GAAG,EAAE,EAC3B,OAAO,GAAG,OAAO,EACjB,KAAK,SAAS,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EACzC,QAAQ,SAAS,OAAO,GAAG,OAAO,EAElC,OAAO,EAAE,WAAW,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,QAAQ,CAAC;IA0DnE;;OAEG;IACH,KAAK,CACH,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,MAAM,SAAS,MAAM,GAAG,EAAE,EAC1B,OAAO,GAAG,OAAO,EACjB,KAAK,SAAS,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EAEzC,OAAO,EAAE,YAAY,CAAC,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,CAAC;IAuCzD;;;;;;;;;;;;;;OAcG;IACH,MAAM,CACJ,gBAAgB,SAAS,MAAM,GAAG,EAAE,EACpC,iBAAiB,SAAS,MAAM,GAAG,EAAE,EAErC,MAAM,CAAC,EAAE,OAAO,CAAC,QAAQ,CACvB,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,EAAE,gBAAgB,CAAC,EACtC,MAAM,CAAC,CAAC,CAAC,UAAU,CAAC,EAAE,iBAAiB,CAAC,EACxC,WAAW,CAAC,CAAC,CAAC,CACf,CAAC;IA8DJ;;;;;;;OAOG;IACH,YAAY,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,IAAI;IAMjD;;OAEG;IACH,KAAK,IAAI,YAAY;CA+BtB"}
+{"version":3,"file":"ai.d.ts","sourceRoot":"","sources":["../src/ai.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAmB,KAAK,EAAU,MAAM,EAAE,EAAE,EAAc,kBAAkB,EAAE,MAAM,EAAa,YAAY,EAAE,IAAI,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,YAAY,CAAC;AACtK,OAAO,EAAE,QAAQ,EAAE,MAAM,cAAc,CAAC;AACxC,OAAO,EAAE,OAAO,EAAE,MAAM,aAAa,CAAC;AACtC,OAAO,EAAE,QAAQ,EAAE,MAAM,cAAc,CAAC;AACxC,OAAO,EAAE,SAAS,EAAE,MAAM,eAAe,CAAC;AAC1C,OAAO,EAAE,SAAS,EAAE,MAAM,eAAe,CAAC;AAC1C,OAAO,EAAE,QAAQ,EAAE,MAAM,cAAc,CAAC;AACxC,OAAO,EAAE,SAAS,EAAE,MAAM,eAAe,CAAC;AAC1C,OAAO,EAAE,SAAS,EAAE,MAAM,eAAe,CAAC;AAC1C,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAClD,OAAO,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAC3C,OAAO,KAAK,EACV,UAAU,EACV,cAAc,EACd,WAAW,EACX,QAAQ,EACR,UAAU,EACV,SAAS,EACT,iBAAiB,EACjB,iBAAiB,EACjB,aAAa,EACb,OAAO,EACP,UAAU,EACV,kBAAkB,EAClB,cAAc,EACd,eAAe,EACf,WAAW,EACX,YAAY,EACZ,YAAY,EACZ,cAAc,EACd,eAAe,EACf,OAAO,EAEP,YAAY,EACZ,OAAO,EACP,SAAS,EACT,aAAa,EACb,WAAW,EACX,SAAS,EACT,OAAO,EACP,gBAAgB,EAEhB,SAAS,EACT,KAAK,EACN,MAAM,SAAS,CAAC;AAIjB,KAAK,WAAW,CACd,CAAC,SAAS,WAAW,EACrB,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,OAAO,SAAS,MAAM,GAAG,EAAE,EAC3B,OAAO,GAAG,OAAO,EACjB,KAAK,SAAS,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EACzC,QAAQ,SAAS,OAAO,GAAG,OAAO,IAChC,IAAI,CAAC,SAAS,CAChB,SAAS,CAAC,CAAC,CAAC,EACZ,UAAU,CAAC,CAAC,CAAC,EACb,KAAK,EACL,OAAO,EACP,OAAO,EAEP,KAAK,EACL,QAAQ,CACT,EAAE,OAAO,CAAC,CAAC;AAEZ,KAAK,aAAa,CAChB,CAAC,SAAS,WAAW,EACrB,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,MAAM,SAAS,MAAM,GAAG,EAAE,EAC1B,OAAO,SAAS,MAAM,GAAG,MAAM,GAAG,MAAM,EACxC,MAAM,SAAS,KAAK,CAAC,cAAc,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EACtF,QAAQ,SAAS,OAAO,GAAG,OAAO,IAChC,IAAI,CAAC,WAAW,CAClB,SAAS,CAAC,CAAC,CAAC,EACZ,UAAU,CAAC,CAAC,CAAC,EACb,KAAK,EACL,MAAM,EACN,OAAO,EAEP,MAAM,EACN,QAAQ,CACT,EAAE,OAAO,GAAG,QAAQ,CAAC,GAAG;IACvB,MAAM,CAAC,EAAE,EAAE,CAAC,YAAY,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAAC,CAAC,EAAE,MAAM,CAAC,EAAE,CAAC,MAAM,GAAG,SAAS,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;CAClH,CAAC;AAEF,KAAK,YAAY,CACf,CAAC,SAAS,WAAW,EACrB,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,MAAM,SAAS,MAAM,GAAG,EAAE,EAC1B,OAAO,GAAG,OAAO,EACjB,KAAK,SAAS,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,IACvC,IAAI,CAAC,UAAU,CACjB,SAAS,CAAC,CAAC,CAAC,EACZ,UAAU,CAAC,CAAC,CAAC,EACb,KAAK,EACL,MAAM,EACN,OAAO,EAEP,KAAK,CACN,EAAE,OAAO,CAAC,CAAC;AAEZ;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AACH,qBAAa,EAAE,CAAC,CAAC,SAAS,WAAW;IAEnC,SAAgB,MAAM,EAAE,UAAU,CAAC,CAAC,CAAC,CAAC;IACtC,SAAgB,QAAQ,EAAE,aAAa,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;IAGxD,SAAgB,IAAI,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC;IACjC,SAAgB,KAAK,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC;IACnC,SAAgB,MAAM,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC;IACrC,SAAgB,UAAU,EAAE,aAAa,CAAC,CAAC,CAAC,CAAC;IAC7C,SAAgB,KAAK,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC;IACnC,SAAgB,MAAM,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC;IACrC,SAAgB,MAAM,EAAE,SAAS,CAAC,CAAC,CAAC,CAAC;IACrC,SAAgB,KAAK,EAAE,QAAQ,CAAC,CAAC,CAAC,CAAC;IACnC,SAAgB,MAAM,EAAE,SAAS,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,CAAC;IAGlD,SAAgB,SAAS,EAAE,WAAW,CAAC,CAAC,CAAC,CAAC;IAC1C,SAAgB,UAAU,EAAE,YAAY,CAAC,CAAC,CAAC,EAAE,CAAM;IAC5C,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC;IAGzB,OAAO,CAAC,cAAc,CAAa;IACnC,OAAO,CAAC,iBAAiB,CAAa;IACtC,OAAO,CAAC,sBAAsB,CAAa;IAG3C,SAAgB,MAAM,EAAE,MAAM,CAAC,kBAAkB,EAAE;QACjD,OAAO,EAAE,MAAM,CAAC;QAChB,aAAa,EAAE,MAAM,CAAC;QACtB,QAAQ,EAAE,MAAM,CAAC;QACjB,GAAG,CAAC,EAAE,MAAM,CAAC;KACd,CAAC,CAAC;IAEH;;OAEG;WACW,IAAI,CAAC,QAAQ,SAAS,aAAa,GAAG,EAAE,EAAE,SAAS,SAAS,cAAc,GAAG,EAAE;oBAE7E,UAAU,SAAS,SAAS,aAAa,UAAU;qBAElD,OAAO,SAAS,IAAI,CAAC,QAAQ,CAAC,QAAQ,EAAE,SAAS,EAAE,UAAU,CAAC,EAAE,WAAW,CAAC,WAC1E,OAAO;;;gBAYd,MAAM,EAAE,UAAU,CAAC,CAAC,CAAC;IAsFjC;;;;;OAKG;IACH,SAAS,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,GAAG,IAAI;IAKlC;;;;;;;OAOG;IACG,YAAY,CAAC,WAAW,EAAE,iBAAiB,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;IAgB5E;;;;;;;OAOG;IACG,gBAAgB,CAAC,WAAW,EAAE,iBAAiB,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;IAoB9E;;;;;;;OAOG;IACG,aAAa,CAAC,gBAAgB,EAAE,kBAAkB,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC;IAiBpF;;;;;;;;OAQG;IACH,aAAa,CAAC,GAAG,SAAS,EAAE,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC,CAAC,GAAG,SAAS,CAAC,CAAC,GAAG,UAAU,CAAC,CAAC,CAAC;IA8HrF;;;OAGG;IACH,WAAW,CAAC,QAAQ,EAAE,UAAU,CAAC,CAAC,CAAC,GAAG,gBAAgB,CAAC,CAAC,CAAC,GAAG,SAAS;IAIrE;;;;;OAKG;IACH,oBAAoB,CAAC,OAAO,EAAE,OAAO,GAAG,KAAK;IAuG7C;;;OAGG;IACH,oBAAoB,CAAC,OAAO,EAAE,OAAO,GAAG,KAAK;IAW7C;;;;OAIG;IACH,aAAa,CACX,KAAK,EAAE,SAAS,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC,EACpC,KAAK,EAAE,KAAK,GACX,MAAM;IAoDT;;;;;;OAMG;IACH,eAAe,CAAC,KAAK,EAAE,SAAS,EAAE,QAAQ,EAAE,aAAa,GAAG,OAAO;IAmBnE;;;;;;OAMG;IACG,GAAG,CAAC,CAAC,SAAS,YAAY,CAAC,CAAC,CAAC,EACjC,SAAS,EAAE,CAAC,EACZ,KAAK,EAAE,cAAc,CAAC,CAAC,CAAC,EACxB,GAAG,CAAC,GAAG,CAAC,EAAE,iBAAiB,CAAC,CAAC,CAAC,GAC7B,OAAO,CAAC,eAAe,CAAC,CAAC,CAAC,CAAC;IAS9B;;;;;;;;;OASG;IACH,MAAM,CACJ,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,MAAM,SAAS,MAAM,GAAG,EAAE,EAC1B,OAAO,SAAS,MAAM,GAAG,MAAM,GAAG,MAAM,EACxC,MAAM,SAAS,KAAK,CAAC,cAAc,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,kBAAkB,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EACtF,QAAQ,SAAS,OAAO,GAAG,OAAO,EAElC,OAAO,EAAE,aAAa,CAAC,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,QAAQ,CAAC;IA0FrE;;OAEG;IACH,IAAI,CACF,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,OAAO,SAAS,MAAM,GAAG,EAAE,EAC3B,OAAO,GAAG,OAAO,EACjB,KAAK,SAAS,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EACzC,QAAQ,SAAS,OAAO,GAAG,OAAO,EAElC,OAAO,EAAE,WAAW,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,QAAQ,CAAC;IA0DnE;;OAEG;IACH,KAAK,CACH,KAAK,SAAS,MAAM,GAAG,MAAM,EAC7B,MAAM,SAAS,MAAM,GAAG,EAAE,EAC1B,OAAO,GAAG,OAAO,EACjB,KAAK,SAAS,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE,EAEzC,OAAO,EAAE,YAAY,CAAC,CAAC,EAAE,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,KAAK,CAAC;IAuCzD;;;;;;;;;;;;;;OAcG;IACH,MAAM,CACJ,gBAAgB,SAAS,MAAM,GAAG,EAAE,EACpC,iBAAiB,SAAS,MAAM,GAAG,EAAE,EAErC,MAAM,CAAC,EAAE,OAAO,CAAC,QAAQ,CACvB,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,EAAE,gBAAgB,CAAC,EACtC,MAAM,CAAC,CAAC,CAAC,UAAU,CAAC,EAAE,iBAAiB,CAAC,EACxC,WAAW,CAAC,CAAC,CAAC,CACf,CAAC;IA8DJ;;;;;;;OAOG;IACH,YAAY,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,IAAI;IAMjD;;OAEG;IACH,KAAK,IAAI,YAAY;CA+BtB"}