@strav/brain 1.0.0-alpha.15 → 1.0.0-alpha.17

package/src/tool.ts

@@ -23,6 +23,13 @@ export interface ToolContext {
   readonly callId: string
   /** Per-run free-form context bag passed by the caller. Optional. */
   readonly context: Readonly<Record<string, unknown>>
+  /**
+   * Cancellation signal forwarded from the run's `options.signal`.
+   * Tools that wrap network calls (HTTP fetches, MCP servers, child
+   * processes) should pass this through so cancellation actually
+   * unwinds in-flight work.
+   */
+  readonly signal?: AbortSignal
 }
 
 export interface Tool<TInput = unknown, TOutput = unknown> {

package/src/tool_runner.ts

@@ -0,0 +1,81 @@
+/**
+ * `runToolWithRecovery` — shared helper used by every provider's
+ * agentic loop to execute one tool call.
+ *
+ * Encapsulates two error paths and the optional `onToolError`
+ * recovery callback:
+ *
+ *   1. **Tool not registered** — the model called a name that
+ *      isn't in `toolMap`. Without recovery, throw
+ *      `ToolExecutionError`. With recovery, the callback's return
+ *      string becomes the `tool_result.content` (with `isError:
+ *      true`) and the loop continues — the model sees "unknown
+ *      tool" and adapts.
+ *
+ *   2. **`execute()` throws** — the tool's body raised. Same
+ *      pattern: either rethrow as `ToolExecutionError` or feed
+ *      back as an error result.
+ *
+ * The returned shape is the framework-agnostic `{ content, isError }`
+ * pair each provider then wraps into its own `tool_result` block
+ * shape (Anthropic `tool_result` with `is_error`; OpenAI tool-role
+ * message content; Gemini `functionResponse` with `{ error }`).
+ */
+
+import type { RunWithToolsOptions } from './provider.ts'
+import type { Tool, ToolContext } from './tool.ts'
+import { ToolExecutionError } from './tool_execution_error.ts'
+
+export interface ToolRunResult {
+  content: string
+  isError: boolean
+}
+
+export async function runToolWithRecovery(
+  tool: Tool | undefined,
+  toolName: string,
+  callId: string,
+  input: unknown,
+  options: RunWithToolsOptions,
+): Promise<ToolRunResult> {
+  if (!tool) {
+    return recoverOrThrow(
+      new ToolExecutionError(
+        toolName,
+        callId,
+        new Error(`Tool "${toolName}" is not registered.`),
+      ),
+      options,
+    )
+  }
+
+  const ctx: ToolContext = {
+    callId,
+    context: options.context ?? {},
+    ...(options.signal !== undefined ? { signal: options.signal } : {}),
+  }
+  let output: unknown
+  try {
+    output = await tool.execute(input, ctx)
+  } catch (cause) {
+    return recoverOrThrow(new ToolExecutionError(toolName, callId, cause), options)
+  }
+  return {
+    content: typeof output === 'string' ? output : JSON.stringify(output),
+    isError: false,
+  }
+}
+
+/**
+ * Resolve a `ToolExecutionError` through the `onToolError` callback
+ * (when set) or rethrow. Used by providers for failures that happen
+ * outside `tool.execute` — e.g., OpenAI's JSON-parse-arguments path.
+ */
+export function recoverOrThrow(
+  error: ToolExecutionError,
+  options: RunWithToolsOptions,
+): ToolRunResult {
+  const recovered = options.onToolError?.(error)
+  if (typeof recovered !== 'string') throw error
+  return { content: recovered, isError: true }
+}

package/src/types.ts

@@ -105,8 +105,79 @@ export interface MCPToolResultBlock {
   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 }
+}
+
 export type ContentBlock =
   | TextBlock
+  | ImageBlock
+  | DocumentBlock
+  | AudioBlock
   | ToolUseBlock
   | ToolResultBlock
   | MCPToolUseBlock
@@ -134,6 +205,55 @@ 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. */
+    }
+
 export interface ChatOptions {
   /** Override the configured default model. Wins over `tier`. */
   model?: string
@@ -168,6 +288,26 @@ 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[]
 }
 
 /** Token usage for a single call. Cache-hit fields are populated when caching is in play. */
@@ -201,6 +341,99 @@ 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