@strav/brain 1.0.0-alpha.9 → 1.0.2

package/src/translate/translator.ts

@@ -0,0 +1,271 @@
+/**
+ * `Translator` — LLM-backed translation primitive on top of
+ * `BrainManager`. Sonnet-uniform by default (`tier: 'balanced'`),
+ * which routes to `claude-sonnet-4-6` on the Anthropic driver — apps
+ * override with `options.model` or `options.provider` per call.
+ *
+ * Two entry points:
+ *
+ *   - `translate(text, { to: [...] })` — fan-out one string into
+ *     every target language in parallel. Returns
+ *     `{ [langCode]: translated }`.
+ *
+ *   - `translateBatch(fields, { to: [...] })` — translate a
+ *     fixed-shape object (`{ title, body }`) into every target
+ *     language. Each target language runs in parallel; within a
+ *     language, all fields land in one model call so the model
+ *     keeps shared context (a `title` and `body` translated
+ *     together stay tonally consistent).
+ *
+ * Cross-cutting:
+ *
+ *   - **Structured output.** Uses `brain.generate(input, schema)`
+ *     with a JSON Schema that locks the response to the expected
+ *     keys, so models never sneak in commentary or transliterations.
+ *
+ *   - **Prompt caching.** The system prompt is identical across
+ *     every call (per-language hints ride in the user message), so
+ *     Anthropic prompt caching kicks in once the cache window warms.
+ *     Set `cache: false` on the constructor to opt out.
+ *
+ *   - **In-memory cache.** Identical `(model, from, to, text)`
+ *     tuples are served from a process-local LRU (default 1000
+ *     entries) — see `TranslateCache`. Pass `cacheSize: 0` to
+ *     disable.
+ *
+ *   - **Source language auto-detect.** Omit `from` and the user
+ *     message tells the model to detect the source. Apps that know
+ *     the source pass it explicitly for marginal quality + token
+ *     savings.
+ */
+
+import type { BrainManager } from '../brain_manager.ts'
+import type { OutputSchema } from '../output_schema.ts'
+import type { ChatOptions, ModelTier } from '../types.ts'
+import { cacheKey, TranslateCache } from './translate_cache.ts'
+
+export interface TranslatorOptions {
+  brain: BrainManager
+  /** Brain provider name. Defaults to the configured `brain.default`. */
+  provider?: string
+  /** Brain tier sugar — overridden by `model`. Default `'balanced'` (Sonnet on Anthropic per ADR-0004-style routing). */
+  tier?: ModelTier
+  /** Explicit model id. Wins over `tier`. */
+  model?: string
+  /** Override the system prompt. Apps localising the prompt itself reach for this. */
+  systemPrompt?: string
+  /** LRU capacity for the translation cache. `0` disables. Default `1000`. */
+  cacheSize?: number
+  /** Enable Anthropic prompt caching on the system prompt. Default `true`. Non-Anthropic providers ignore. */
+  cache?: boolean
+}
+
+export interface TranslateOptions {
+  /** Target BCP-47 language codes (`'th'`, `'zh-Hant'`, `'ja'`). */
+  to: readonly string[]
+  /** Source BCP-47 code. Omit to ask the model to detect. */
+  from?: string
+  /** Per-call model override (wins over the constructor's tier/model). */
+  model?: string
+  /** Per-call provider override. */
+  provider?: string
+  /** Cancellation signal — forwarded to every parallel `brain.generate` call. */
+  signal?: AbortSignal
+}
+
+export type BatchTranslateOptions = TranslateOptions
+
+/**
+ * Default system prompt — kept stable across every call so prompt
+ * caching can warm. Per-call specifics (source/target language,
+ * text, field shape) ride in the user message.
+ */
+export const DEFAULT_SYSTEM_PROMPT = `You are a translation engine.
+
+The user supplies (a) a source-language code (or "auto"), (b) a target BCP-47 language code, and (c) the source text or a JSON object of named source fields. Translate the source into the target language and output ONLY the translation in the required JSON shape.
+
+Rules:
+- Output ONLY the translated text in the requested JSON shape. Do not add explanations, notes, alternatives, or transliterations.
+- Preserve Markdown, HTML tags, links, mentions, hashtags, code spans, and emoji exactly as in the source.
+- Keep numbers, dates, currency symbols, and proper nouns recognisable in the target locale; do not invent translations for brand names.
+- If the source is already in the target language, output it unchanged.
+- For batch translations, every requested field must appear in the output — never drop a field.`
+
+export class Translator {
+  private readonly brain: BrainManager
+  private readonly provider: string | undefined
+  private readonly tier: ModelTier
+  private readonly explicitModel: string | undefined
+  private readonly systemPrompt: string
+  private readonly cache: TranslateCache
+  private readonly promptCache: boolean
+
+  constructor(options: TranslatorOptions) {
+    this.brain = options.brain
+    this.provider = options.provider
+    this.tier = options.tier ?? 'balanced'
+    this.explicitModel = options.model
+    this.systemPrompt = options.systemPrompt ?? DEFAULT_SYSTEM_PROMPT
+    this.cache = new TranslateCache(options.cacheSize ?? 1000)
+    this.promptCache = options.cache ?? true
+  }
+
+  /**
+   * Translate one string into every target language in parallel.
+   * Returns a `{ [lang]: translated }` map containing one entry per
+   * code in `options.to`. Calls fan out concurrently; a single
+   * thrown call rejects the whole `Promise.all`.
+   */
+  async translate(
+    text: string,
+    options: TranslateOptions,
+  ): Promise<Record<string, string>> {
+    if (options.to.length === 0) return {}
+
+    const results = await Promise.all(
+      options.to.map(async (lang) => {
+        const translated = await this.translateOne(text, lang, options)
+        return [lang, translated] as const
+      }),
+    )
+    return Object.fromEntries(results)
+  }
+
+  /**
+   * Translate a fixed-shape object of fields into every target
+   * language. Each target language runs in parallel; within a
+   * language, all fields are translated in one model call so context
+   * is shared.
+   *
+   * Returns `{ [lang]: { ...fields } }`. The shape of every per-
+   * language object matches the input keys exactly — missing keys
+   * are treated as a hard error (the model is instructed to never
+   * drop a field) and surface as a `BrainError` from `generate`'s
+   * schema parser.
+   */
+  async translateBatch<T extends Record<string, string>>(
+    fields: T,
+    options: BatchTranslateOptions,
+  ): Promise<Record<string, T>> {
+    if (options.to.length === 0) return {}
+    const fieldNames = Object.keys(fields) as Array<keyof T & string>
+    if (fieldNames.length === 0) return Object.fromEntries(options.to.map((l) => [l, {} as T]))
+
+    const results = await Promise.all(
+      options.to.map(async (lang) => {
+        const translated = await this.translateBatchOne(fields, fieldNames, lang, options)
+        return [lang, translated] as const
+      }),
+    )
+    return Object.fromEntries(results)
+  }
+
+  /** Drop the in-memory LRU. Useful in tests to keep cases isolated. */
+  clearCache(): void {
+    this.cache.clear()
+  }
+
+  // ─── internals ──────────────────────────────────────────────────────
+
+  private resolvedModel(per: TranslateOptions): string {
+    return per.model ?? this.explicitModel ?? this.tier
+  }
+
+  private buildChatOptions(per: TranslateOptions): ChatOptions {
+    const opts: ChatOptions = {
+      system: this.promptCache
+        ? { text: this.systemPrompt, cache: true }
+        : this.systemPrompt,
+    }
+    if (per.model) opts.model = per.model
+    else if (this.explicitModel) opts.model = this.explicitModel
+    else opts.tier = this.tier
+    if (per.provider ?? this.provider) opts.provider = (per.provider ?? this.provider)!
+    if (per.signal) opts.signal = per.signal
+    return opts
+  }
+
+  private async translateOne(
+    text: string,
+    lang: string,
+    per: TranslateOptions,
+  ): Promise<string> {
+    const model = this.resolvedModel(per)
+    const key = cacheKey({ model, from: per.from, to: lang, text })
+    const hit = this.cache.get(key)
+    if (hit !== undefined) return hit
+
+    const schema: OutputSchema<{ translation: string }> = {
+      name: 'translation',
+      description: `Translation of the source text into ${lang}.`,
+      jsonSchema: {
+        type: 'object',
+        properties: { translation: { type: 'string' } },
+        required: ['translation'],
+        additionalProperties: false,
+      },
+    }
+
+    const userMessage = `SOURCE_LANGUAGE: ${per.from ?? 'auto'}\nTARGET_LANGUAGE: ${lang}\nTEXT:\n${text}`
+
+    const result = await this.brain.generate(userMessage, schema, this.buildChatOptions(per))
+    const translated = result.value.translation
+    this.cache.set(key, translated)
+    return translated
+  }
+
+  private async translateBatchOne<T extends Record<string, string>>(
+    fields: T,
+    fieldNames: readonly (keyof T & string)[],
+    lang: string,
+    per: BatchTranslateOptions,
+  ): Promise<T> {
+    const model = this.resolvedModel(per)
+
+    // Per-field cache: check every field; only call the model when at
+    // least one field is missing. The single model call still covers
+    // all fields (we don't sub-call per missing field — the context
+    // gain from a single call outweighs the extra translation work).
+    const fromCache: Partial<Record<string, string>> = {}
+    let allHit = true
+    for (const name of fieldNames) {
+      const hit = this.cache.get(
+        cacheKey({ model, from: per.from, to: lang, text: fields[name]! }),
+      )
+      if (hit === undefined) {
+        allHit = false
+      } else {
+        fromCache[name] = hit
+      }
+    }
+    if (allHit) return fromCache as T
+
+    const properties: Record<string, unknown> = {}
+    for (const name of fieldNames) properties[name] = { type: 'string' }
+    const schema: OutputSchema<T> = {
+      name: 'batch_translation',
+      description: `Translation of every named field into ${lang}.`,
+      jsonSchema: {
+        type: 'object',
+        properties,
+        required: [...fieldNames],
+        additionalProperties: false,
+      },
+    }
+
+    const fieldsBlock = fieldNames
+      .map((n) => `- ${n}: ${JSON.stringify(fields[n]!)}`)
+      .join('\n')
+    const userMessage = `SOURCE_LANGUAGE: ${per.from ?? 'auto'}\nTARGET_LANGUAGE: ${lang}\nFIELDS:\n${fieldsBlock}\n\nOutput a JSON object with these exact keys: ${fieldNames.join(', ')}.`
+
+    const result = await this.brain.generate(userMessage, schema, this.buildChatOptions(per))
+    const translated = result.value
+    for (const name of fieldNames) {
+      this.cache.set(
+        cacheKey({ model, from: per.from, to: lang, text: fields[name]! }),
+        translated[name]!,
+      )
+    }
+    return translated
+  }
+}

package/src/types.ts

@@ -71,7 +71,147 @@ export interface ToolResultBlock {
   isError?: boolean
 }
 
-export type ContentBlock = TextBlock | ToolUseBlock | ToolResultBlock
+/**
+ * Provider-emitted MCP tool-use block. Read-only — apps don't construct
+ * these; they appear in `assistant`-role messages when the model calls
+ * a tool exposed by a configured MCP server. Anthropic's backend
+ * invokes the MCP server itself and inlines the result as an
+ * `MCPToolResultBlock` in the same response, so the framework's
+ * agentic loop doesn't need to handle the call.
+ *
+ * Apps render these for observability (showing users that the model
+ * consulted Linear / Notion / GitHub via MCP) and for audit trails.
+ */
+export interface MCPToolUseBlock {
+  type: 'mcp_tool_use'
+  id: string
+  /** MCP server identifier — matches `MCPServer.name`. */
+  serverName: string
+  /** Tool name as exposed by the MCP server. */
+  name: string
+  /** Parsed input the model passed to the MCP tool. */
+  input: unknown
+}
+
+/**
+ * Provider-emitted MCP tool result. Pairs with `MCPToolUseBlock` by
+ * `toolUseId`. `content` is either a string or text blocks; `isError`
+ * is `true` when the MCP server returned an error.
+ */
+export interface MCPToolResultBlock {
+  type: 'mcp_tool_result'
+  toolUseId: string
+  content: string | TextBlock[]
+  isError?: boolean
+}
+
+/**
+ * Image input — attaches a picture to a user message so vision-
+ * capable models can see it alongside the text. V1 covers images
+ * only; audio + video defer.
+ *
+ * `source` is a discriminated union:
+ *   - `{ type: 'base64', mediaType, data }` — inline bytes for
+ *     uploads, screenshots, attachments your app already holds in
+ *     memory. `mediaType` is the IANA MIME (`image/png`,
+ *     `image/jpeg`, `image/webp`, `image/gif`); `data` is the
+ *     base64-encoded image (no `data:` prefix — the provider
+ *     translation adds it where needed).
+ *   - `{ type: 'url', url }` — remote image URL. Anthropic, OpenAI,
+ *     and Gemini all accept HTTPS URLs; check the provider's
+ *     domain allowlist if calls 404 (Anthropic was historically
+ *     stricter). For Gemini, GCS URIs (`gs://...`) also work.
+ *
+ * Vision support is provider- AND model-dependent. Cloud picks:
+ * Anthropic Claude 4 family, OpenAI gpt-4o / gpt-5 family, Gemini
+ * 2.x. Local: `llama3.2-vision`, `llava`, `qwen2.5-vl` on Ollama.
+ * Models without vision either reject the call or ignore the image.
+ */
+export interface ImageBlock {
+  type: 'image'
+  source:
+    | { type: 'base64'; mediaType: string; data: string }
+    | { type: 'url'; url: string }
+}
+
+/**
+ * Document input — attaches a PDF (V1 only — the providers that
+ * support documents currently all gate on `application/pdf`) to a
+ * user message. Anthropic surfaces it as a native `document` block;
+ * Gemini accepts it via `inlineData` / `fileData` with
+ * `application/pdf` mime; OpenAI / Ollama / DeepSeek don't support
+ * PDF blocks at all (apps split the PDF to images and use
+ * `ImageBlock`s for those vendors).
+ *
+ * The optional `title` is shown to the model on Anthropic (helpful
+ * for multi-document calls — "the contract", "the invoice"); other
+ * providers ignore it.
+ */
+export interface DocumentBlock {
+  type: 'document'
+  source:
+    | { type: 'base64'; mediaType: string; data: string }
+    | { type: 'url'; url: string }
+  /** Optional title shown to the model (Anthropic uses it; others ignore). */
+  title?: string
+}
+
+/**
+ * Audio input — attaches a sound clip to a user message. V1
+ * coverage: Gemini supports audio natively via `inlineData` with
+ * audio MIMEs (`audio/mp3`, `audio/wav`, `audio/ogg`, `audio/flac`,
+ * `audio/webm`, `audio/aac`). Anthropic + OpenAI + Ollama don't
+ * accept audio in their chat APIs — OpenAI apps preprocess via
+ * Whisper; Anthropic apps wait for the audio block to land in the
+ * SDK; Ollama apps that need audio look at server-side
+ * transcription models.
+ */
+export interface AudioBlock {
+  type: 'audio'
+  source:
+    | { type: 'base64'; mediaType: string; data: string }
+    | { type: 'url'; url: string }
+}
+
+/**
+ * Server-side compaction block. Anthropic's `compact-2026-01-12`
+ * beta returns a `compaction` block when an auto-compaction trigger
+ * fires during a request. The framework surfaces it on
+ * `result.content` and Thread persists it on the assistant turn so
+ * subsequent requests echo it back verbatim — the model only sees
+ * the summary + opaque blob from then on, and the older raw turns
+ * stay out of context.
+ *
+ * V1 produces these on Anthropic only. Other providers ignore the
+ * `compact` option silently, and never emit a `CompactionBlock`.
+ *
+ * Round-trip invariant: pass the block back unchanged. The
+ * `encryptedContent` blob is opaque metadata the server uses to
+ * stitch the compaction history together; the framework never
+ * mutates it.
+ *
+ * `content === null` means a compaction attempt failed (e.g.,
+ * malformed model output). The server treats these as no-ops on
+ * the next request, so apps don't need to special-case them.
+ */
+export interface CompactionBlock {
+  type: 'compaction'
+  /** Summary of compacted content. Null when compaction failed. */
+  content: string | null
+  /** Opaque metadata round-tripped verbatim on subsequent requests. */
+  encryptedContent: string | null
+}
+
+export type ContentBlock =
+  | TextBlock
+  | ImageBlock
+  | DocumentBlock
+  | AudioBlock
+  | ToolUseBlock
+  | ToolResultBlock
+  | MCPToolUseBlock
+  | MCPToolResultBlock
+  | CompactionBlock
 
 /** A single conversation turn. `content` can be a bare string or a typed block list. */
 export interface Message {
@@ -95,6 +235,85 @@ export type SystemPrompt =
  * escape hatch in `ChatResult` is what they reach for when they need
  * provider-specific fields.
  */
+/**
+ * Server-side tool — work the provider's backend runs on behalf
+ * of the model. Unlike framework-local tools (`Tool` /
+ * `defineTool`), the model's call doesn't round-trip through
+ * the app's process; the provider executes the tool and inlines
+ * the result in the response.
+ *
+ * V1 coverage:
+ *   - **Anthropic**: `web_search`, `code_execution`, `web_fetch`.
+ *   - **Gemini**: `web_search` (Google Search), `code_execution`,
+ *     `url_context`.
+ *   - **OpenAI / DeepSeek / Ollama**: throw — OpenAI's server tools
+ *     live on the Responses API (separate slice); the compat
+ *     providers don't expose them.
+ *
+ * Cross-provider portability:
+ *   - `web_search` + `code_execution` work on both Anthropic and
+ *     Gemini.
+ *   - `web_fetch` is Anthropic-only.
+ *   - `url_context` is Gemini-only.
+ *
+ * Server tools combine freely with framework-local `Tool[]` and
+ * MCP servers — the model sees all three sets in one tool list.
+ */
+export type ServerTool =
+  | {
+      type: 'web_search'
+      /** Max times the model can call this tool per turn (Anthropic; Gemini ignores). */
+      maxUses?: number
+      /** Domain allowlist (Anthropic; Gemini ignores). Mutually exclusive with `blockedDomains`. */
+      allowedDomains?: readonly string[]
+      /** Domain blocklist (Anthropic; Gemini ignores). */
+      blockedDomains?: readonly string[]
+    }
+  | { type: 'code_execution' }
+  | {
+      type: 'web_fetch'
+      /** Max URL fetches per turn (Anthropic). */
+      maxUses?: number
+      /** Domain allowlist. */
+      allowedDomains?: readonly string[]
+      /** Domain blocklist. */
+      blockedDomains?: readonly string[]
+    }
+  | {
+      type: 'url_context'
+      /** Gemini fetches the URL and surfaces grounded answers from it. */
+    }
+
+/**
+ * Per-call compaction configuration. Maps to Anthropic's
+ * `compact-2026-01-12` beta `edits[]` entry. All fields optional —
+ * omitting one falls back to the server's default (trigger:
+ * 150,000 input tokens; no extra instructions; no pause).
+ */
+export interface CompactConfig {
+  /**
+   * Trigger threshold in input tokens. Compaction fires once the
+   * conversation crosses this token count. Default 150,000 — same
+   * as the server-side default.
+   */
+  trigger?: number
+  /**
+   * Extra hint to the summarization model. Useful for biasing the
+   * compaction toward what your app actually cares to preserve
+   * ("keep all customer ids referenced", "preserve every diff
+   * hunk", ...).
+   */
+  instructions?: string
+  /**
+   * When `true`, the server returns the compaction block in-line
+   * but does NOT continue generation — the next assistant turn
+   * waits for an explicit re-prompt. Apps that want to inspect or
+   * gate compaction set this; default `false` (compaction is
+   * transparent).
+   */
+  pauseAfterCompaction?: boolean
+}
+
 export interface ChatOptions {
   /** Override the configured default model. Wins over `tier`. */
   model?: string
@@ -129,6 +348,56 @@ export interface ChatOptions {
    * provider by config; this is the override for that.
    */
   provider?: string
+  /**
+   * Cancel the in-flight operation. Aborting between iterations of
+   * a tool loop bails before the next model call; aborting mid-call
+   * propagates the SDK's native abort error (typically a `DOMException`
+   * with `name: 'AbortError'`). Streaming iterators reject on the
+   * next `for await` step.
+   */
+  signal?: AbortSignal
+  /**
+   * Server-side tools — work the provider's backend runs (web
+   * search, code execution, URL fetching). The model's calls
+   * don't round-trip through the framework's tool loop; results
+   * land inline in the response. Combines freely with
+   * framework-local `Tool[]` and MCP servers.
+   *
+   * V1 supports Anthropic + Gemini; OpenAI / DeepSeek / Ollama
+   * throw `BrainError` (use the Responses API for OpenAI, or
+   * route to Anthropic / Gemini).
+   */
+  serverTools?: readonly ServerTool[]
+  /**
+   * Server-side conversation compaction. When set, the provider
+   * auto-summarizes the older part of the message history once the
+   * `trigger` token threshold is reached; the summary lives on the
+   * response as a `CompactionBlock` that apps round-trip on
+   * subsequent requests (Thread does this automatically). Saves
+   * tokens on long threads without lossy client-side pruning.
+   *
+   * Only honored by `AnthropicBrainDriver` (driver `'anthropic'`),
+   * via the `compact-2026-01-12` beta. Silently ignored by every
+   * other provider so apps targeting multiple providers with the
+   * same options object don't have to special-case.
+   */
+  compact?: CompactConfig
+  /**
+   * Stateful conversation pointer — OpenAI Responses API. When set,
+   * the provider sends only the new turn(s); the server picks up
+   * from the prior `Response` identified by this id and replays
+   * the conversation server-side. Saves tokens on long threads.
+   *
+   * Only honored by `OpenAIResponsesBrainDriver` (driver
+   * `'openai-responses'`); silently ignored by every other provider
+   * — apps that target multiple providers with the same options
+   * object don't have to special-case.
+   *
+   * Pair with `ChatResult.responseId` (returned by every call) to
+   * thread the conversation forward. `Thread` does this
+   * automatically when its underlying provider supports it.
+   */
+  previousResponseId?: string
 }
 
 /** Token usage for a single call. Cache-hit fields are populated when caching is in play. */
@@ -151,6 +420,24 @@ export interface ChatResult<Raw = unknown> {
   stopReason: string | null
   usage: ChatUsage
   raw: Raw
+  /**
+   * Structured assistant content blocks — populated when the model
+   * emitted more than plain text on this turn (compaction blocks
+   * today; reasoning blocks once those surface). Apps that
+   * persist the conversation (`Thread`, custom stores) push this
+   * onto the message history when present so round-trippable
+   * blocks survive subsequent requests. Undefined when the turn
+   * was plain text only.
+   */
+  content?: ContentBlock[]
+  /**
+   * Provider response id when the provider exposes stateful
+   * conversations (currently OpenAI Responses API). Apps thread
+   * this forward via `ChatOptions.previousResponseId` so the
+   * server replays prior turns without re-sending them.
+   * Undefined for providers that don't support the pattern.
+   */
+  responseId?: string
 }
 
 /**
@@ -161,3 +448,113 @@ export interface ChatResult<Raw = unknown> {
 export type StreamEvent =
   | { type: 'text'; delta: string }
   | { type: 'stop'; stopReason: string | null; usage: ChatUsage }
+
+/**
+ * Per-call options for `brain.embed(...)`. Only the embed-relevant
+ * subset of `ChatOptions` — chat-specific knobs (system prompt,
+ * thinking, cache, tools) don't apply.
+ */
+export interface EmbedOptions {
+  /** Override the configured default embedding model. */
+  model?: string
+  /**
+   * Override the default provider. Must name a provider that
+   * implements `embed` (V1: OpenAI, Gemini, Ollama; Anthropic +
+   * DeepSeek throw with a clear "route to a different provider"
+   * message).
+   */
+  provider?: string
+  /**
+   * Optional dimensionality hint. OpenAI passes through as
+   * `dimensions`; Gemini as `outputDimensionality`. Providers
+   * that ignore it silently drop the field.
+   */
+  dimensions?: number
+  /** Cancellation signal — same shape as `ChatOptions.signal`. */
+  signal?: AbortSignal
+}
+
+/**
+ * Per-call options for `brain.transcribe(...)`.
+ */
+export interface TranscribeOptions {
+  /** Override the configured default transcription model. */
+  model?: string
+  /**
+   * Override the default provider. Must name a provider that
+   * implements `transcribe` (V1: OpenAI / Gemini / Ollama;
+   * Anthropic + DeepSeek throw).
+   */
+  provider?: string
+  /**
+   * Optional BCP-47 language hint (`en`, `fr`, `ja`). Improves
+   * accuracy when known; models without hint support ignore.
+   */
+  language?: string
+  /**
+   * Optional bias prompt to steer vocabulary / style / formatting.
+   * OpenAI calls this `prompt`; Gemini-via-chat threads it into
+   * the system message; others ignore.
+   */
+  prompt?: string
+  /** Cancellation signal — same shape as `ChatOptions.signal`. */
+  signal?: AbortSignal
+}
+
+/**
+ * Audio source — same discriminated union as
+ * `AudioBlock.source`, named separately for `transcribe(...)`
+ * which takes it directly (no wrapping `AudioBlock` shell).
+ */
+export type AudioSource =
+  | { type: 'base64'; mediaType: string; data: string }
+  | { type: 'url'; url: string }
+
+/**
+ * Result of one `transcribe` call. `text` is the transcribed
+ * audio; `language` / `duration` are surfaced when the provider
+ * returns them (OpenAI does on the `verbose_json` response
+ * format; Gemini's chat-wrap path doesn't). `raw` is the
+ * provider's full native response for fields the framework
+ * doesn't surface.
+ */
+export interface TranscribeResult<Raw = unknown> {
+  text: string
+  model: string
+  /** BCP-47 detected (or echoed) language. Optional. */
+  language?: string
+  /** Audio duration in seconds. Optional. */
+  duration?: number
+  raw: Raw
+}
+
+/**
+ * Result of one `embed` call. `embeddings[i]` is the vector for
+ * the i-th input text. `model` is the model the provider used
+ * (echoed back for logging). `usage.inputTokens` is the total
+ * tokens consumed across all inputs.
+ */
+export interface EmbedResult<Raw = unknown> {
+  embeddings: number[][]
+  model: string
+  usage: { inputTokens: number }
+  /** Provider's full native response — escape hatch for fields the framework doesn't surface. */
+  raw: Raw
+}
+
+/**
+ * Result of a structured-output call. `value` is the parsed JSON
+ * shaped to the `OutputSchema<T>` passed in. `text` is the raw JSON
+ * string the model produced (useful for logging / debugging when
+ * `parse` rejects). `raw` is the provider's full native response.
+ */
+export interface GenerateResult<T = unknown, Raw = unknown> {
+  value: T
+  text: string
+  model: string
+  stopReason: string | null
+  usage: ChatUsage
+  raw: Raw
+  /** See `ChatResult.responseId`. */
+  responseId?: string
+}