car-runtime 0.16.1 → 0.18.0

package/index.d.ts

@@ -242,10 +242,19 @@ export class CarRuntime {
    *
    * **Note:** intent is not exposed on the tracked path until the
    * positional argument list is converted to an options object —
-   * this method already takes 8 positional parameters and adding a
-   * 9th would push call sites past readability. For new code, use
-   * {@link inferTrackedWithRequest} which takes a JSON-stringified
-   * `GenerateRequest` and exposes every field including `intent`.
+   * this method already takes 9 positional parameters and adding
+   * intent would push call sites past readability. For new code,
+   * use {@link inferTrackedWithRequest} which takes a JSON-
+   * stringified `GenerateRequest` and exposes every field
+   * including `intent`.
+   *
+   * `imagesJson` is a JSON-encoded array of `ContentBlock` image
+   * variants — either
+   * `{ "type": "image_base64", "data": "<b64>", "media_type": "image/png" }`
+   * or `{ "type": "image_url", "url": "https://…", "detail": "auto" }`.
+   * Vision-capable hosted models (Claude 3.5+/4.x, GPT-4o, Gemini)
+   * accept these directly; non-vision providers reject the request
+   * via a structured error from the daemon. See #230.
    */
   inferTracked(
     prompt: string,
@@ -256,6 +265,7 @@ export class CarRuntime {
     messagesJson?: string | null,
     toolChoice?: string | null,
     parallelToolCalls?: boolean | null,
+    imagesJson?: string | null,
   ): Promise<string>;
 
   /**
@@ -895,6 +905,26 @@ export function transcribeStreamPush(
 
 export function listVoiceSessions(rt: CarRuntime): string;
 
+/**
+ * Start a streaming TTS synthesis.
+ *
+ * Stub: not exposed in the FFI bindings. Connect to the daemon's
+ * WebSocket and use `voice.tts_stream.start`; chunks arrive as
+ * `voice.event` notifications with `type = "tts_chunk"`.
+ */
+export function ttsStreamStart(
+  rt: CarRuntime,
+  streamId: string,
+  text: string,
+  optionsJson?: string | null,
+): Promise<string>;
+
+/** Cancel an in-flight TTS stream. Idempotent. */
+export function ttsStreamCancel(rt: CarRuntime, streamId: string): Promise<string>;
+
+/** List the ids of all in-flight TTS streams. */
+export function listTtsStreams(rt: CarRuntime): string;
+
 // --- Voice turn dispatch (two-track sidecar pattern) ---
 
 export interface DispatchVoiceTurnRequest {
@@ -1174,11 +1204,23 @@ export function a2aServerStatus(): string;
 
 // --- Verification (stateless) ---
 
+/**
+ * Statically verify a proposal.
+ *
+ * `toolNames` checks tool existence only. To also validate each
+ * `tool_call`'s parameters against the tool's JSON Schema — catching
+ * type mismatches (`{path: 42}` for a `string`) and missing required
+ * fields — pass `toolSchemasJson`: a JSON array of tool schemas, e.g.
+ * `JSON.stringify([{ name: "echo", parameters: { type: "object",
+ * properties: { msg: { type: "string" } }, required: ["msg"] } }])`.
+ * When both are given, `toolSchemasJson` takes precedence.
+ */
 export function verify(
   proposalJson: string,
   initialStateJson?: string | null,
   toolNames?: string[] | null,
   maxActions?: number | null,
+  toolSchemasJson?: string | null,
 ): string;
 
 export function simulate(
@@ -1202,43 +1244,15 @@ export function registerAgentRunner(
   agentFn: (specJson: string, taskJson: string) => Promise<string>,
 ): Promise<void>;
 
-/**
- * Register the inference runner callback. Closes
- * Parslee-ai/car-releases#24 — when a model schema declares
- * `source: { type: "delegated", ... }`, CAR routes the request through
- * this callback. The host owns the wire format (Anthropic, OpenAI,
- * Vercel AI SDK, GitHub Models, etc.); CAR observes events and stays
- * in the policy / replay path.
- *
- * The runner emits stream events for every chunk it receives from its
- * provider via `inferenceRunnerEmitEvent(callId, eventJson)`, then
- * resolves the promise with the final aggregated result JSON
- * (`{text: string, tool_calls: ToolCall[]}`).
- *
- * Idempotent — re-calling overwrites the previous runner. Only one
- * runner can be registered per process (matches `registerAgentRunner`).
- *
- * Event JSON shapes match `inferStream`'s event taxonomy:
- *   `{type: "text", data: string}`
- *   `{type: "tool_start", name: string, index: number, id?: string}`
- *   `{type: "tool_delta", index: number, data: string}`
- *   `{type: "usage", input_tokens: number, output_tokens: number}`
- *   `{type: "done", text: string, tool_calls: ToolCall[]}`
- */
-export function registerInferenceRunner(
-  runnerFn: (requestJson: string, callId: string) => Promise<string>,
-): void;
-
-/**
- * Emit a stream event from inside the inference runner callback.
- * `callId` is the second argument the runner received; `eventJson`
- * is one of the shapes documented on `registerInferenceRunner`.
- *
- * Silently no-ops if `callId` is unknown (call already completed or
- * the runner emitted after returning) — racing the normal completion
- * path shouldn't error.
- */
-export function inferenceRunnerEmitEvent(callId: string, eventJson: string): void;
+// --- Inference runner (delegated inference, closes car-releases#24) ---
+//
+// In-process registration is not exposed in the FFI bindings (car-releases#55).
+// The v0.8+ daemon-only architecture moves delegated inference to the
+// WebSocket protocol: a runner client connects to car-server and calls
+// `inference.register_runner`; the daemon then sends
+// `inference.runner.invoke` notifications for every delegated call.
+// See docs/websocket-protocol.md §"Inference runner" for the wire shape.
+// `car-server` is shipped as a binary in this npm package (`bin/car-server`).
 
 /** Run a Swarm pattern. `mode` is "parallel", "sequential", or "debate". */
 export function runSwarm(

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "car-runtime",
-  "version": "0.16.1",
+  "version": "0.18.0",
   "description": "Common Agent Runtime — a deterministic execution layer for AI agents",
   "main": "index.js",
   "types": "index.d.ts",