@independo/inderun-web 0.3.3 → 0.4.0-rc.1

package/src/core/provenance.ts

@@ -0,0 +1,90 @@
+import type { ExecutionProvenance, IndeRunErrorClass } from "@independo/inderun-contracts";
+import type { ProviderDescriptor } from "./provider.js";
+import type { SharedPlannerRoutePlan } from "./route-planner.js";
+import { AUTO_MODEL_PROFILE } from "./model-profiles.js";
+
+type ExecutionProvenanceAttempt = ExecutionProvenance["attempts"][number];
+type RouteRejection = ExecutionProvenance["routeRejections"][number];
+
+/**
+ * Accumulates what one execution did -- each attempt and how it ended -- and renders it as the
+ * normalized {@link ExecutionProvenance} carried on results, terminal outcomes, and errors.
+ *
+ * Everything recorded is either observed by the engine (attempts, their outcomes) or declared or
+ * reported by a provider (descriptor type and model, the model a response names). Nothing is
+ * inferred: an unknown is left absent or reported as `unknown`.
+ */
+export class ProvenanceRecorder {
+  private readonly attempts: ExecutionProvenanceAttempt[] = [];
+  private routeRejections: RouteRejection[] = [];
+
+  constructor(private readonly modelProfile: string = AUTO_MODEL_PROFILE) {}
+
+  /** Records the planner's rejections; codes only, never the provider-supplied messages. */
+  routed(routePlan: SharedPlannerRoutePlan): void {
+    this.routeRejections = routeRejectionsFrom(routePlan.rejectedProviders);
+  }
+
+  /** Records the rejections a routing failure carried out on its exception's details. */
+  routingFailed(details: Record<string, unknown> | undefined): void {
+    const rejected = details?.rejectedProviders;
+    if (Array.isArray(rejected)) {
+      this.routeRejections = routeRejectionsFrom(
+        rejected as SharedPlannerRoutePlan["rejectedProviders"]
+      );
+    }
+  }
+
+  /** Records how one provider attempt ended. */
+  attempted(
+    providerId: string,
+    outcome: ExecutionProvenanceAttempt["outcome"],
+    errorClass?: IndeRunErrorClass
+  ): void {
+    this.attempts.push({
+      providerId,
+      outcome,
+      ...(errorClass !== undefined ? { errorClass } : {})
+    });
+  }
+
+  /**
+   * Renders the record. `attributedTo` is the provider the outcome belongs to -- the one that
+   * produced the result, the one a stream was committed to, or the one running at cancellation --
+   * and is absent when no provider was reached or every attempt failed before a commit.
+   */
+  build(attributedTo?: ProviderDescriptor, reportedModel?: string): ExecutionProvenance {
+    return {
+      schemaVersion: "1.0",
+      modelProfile: this.modelProfile,
+      ...(attributedTo !== undefined ? { providerId: attributedTo.id } : {}),
+      locality: attributedTo !== undefined ? localityOf(attributedTo) : "unknown",
+      ...(attributedTo?.model !== undefined ? { configuredModel: attributedTo.model } : {}),
+      ...(attributedTo !== undefined && reportedModel !== undefined ? { reportedModel } : {}),
+      attemptCount: this.attempts.length,
+      fallbackOccurred: this.attempts.length > 1,
+      attempts: [...this.attempts],
+      routeRejections: [...this.routeRejections]
+    };
+  }
+}
+
+function localityOf(descriptor: ProviderDescriptor): ExecutionProvenance["locality"] {
+  switch (descriptor.type) {
+    case "local":
+      return "on_device";
+    case "edge":
+      return "edge";
+    case "cloud":
+      return "cloud";
+  }
+}
+
+function routeRejectionsFrom(
+  rejected: SharedPlannerRoutePlan["rejectedProviders"]
+): RouteRejection[] {
+  return rejected.map((provider) => ({
+    providerId: provider.providerId,
+    codes: provider.reasons.map((reason) => reason.code)
+  }));
+}

package/src/core/provider.ts

@@ -19,6 +19,14 @@ export interface ProviderDescriptor {
    * Execution target type: 'local' (on-device), 'edge', or 'cloud'.
    */
   type: "local" | "edge" | "cloud";
+  /**
+   * The model this provider instance is bound to, when it can name one. A provider
+   * instance bound to one model is a *target* that model profiles list by `id`.
+   * Cloud providers report the model id they send; platform-managed models report
+   * an opaque platform alias rather than a guessed weight name. Absent means the
+   * provider cannot name its model.
+   */
+  model?: string;
   /**
    * Connection transport layer protocol.
    */
@@ -188,6 +196,11 @@ export type ProviderStreamEvent =
        */
       finishReason?: TaskResult["finishReason"];
       usage?: TaskResult["usage"];
+      /**
+       * The model identifier the provider's backend reported for this run, when it reports
+       * one. Carried into the terminal outcome's provenance; never inferred from configuration.
+       */
+      reportedModel?: string;
     }
   | { kind: "error"; error: unknown };
 

package/src/core/route-planner.ts

@@ -177,7 +177,8 @@ export function buildSharedPlannerInput(
   request: TaskRequest,
   snapshots: ProviderRuntimeSnapshot[],
   networkOnline: boolean,
-  interactionMode: InteractionMode = "run"
+  interactionMode: InteractionMode = "run",
+  modelProfile?: SharedPlannerInput["modelProfile"]
 ): SharedPlannerInput {
   const constraints = request.constraints ?? {};
   const preferences = request.preferences ?? {};
@@ -196,6 +197,7 @@ export function buildSharedPlannerInput(
     preferences: {
       optimizeFor: preferences.optimizeFor ?? "balanced"
     },
+    ...(modelProfile ? { modelProfile } : {}),
     providers: snapshots.map((snapshot) => ({
       descriptor: snapshot.descriptor,
       capabilities: snapshot.capabilities

package/src/core/router.ts

@@ -14,6 +14,7 @@ import {
   type InteractionMode,
   type ProviderRuntimeSnapshot,
   type RoutePlanner,
+  type SharedPlannerInput,
   type SharedPlannerRoutePlan,
   WasmRoutePlanner
 } from "./route-planner.js";
@@ -63,17 +64,25 @@ export class Router {
    *
    * The requested `interactionMode` is a planning input, not a post-filter: the
    * shared planner rejects providers that cannot satisfy it and says why, so the
-   * whole returned chain is mode-compatible.
+   * whole returned chain is mode-compatible. The resolved model profile is a
+   * planning input in the same way: the planner rejects targets outside it.
    */
   async selectRoute(
     request: TaskRequest,
     hostServices: HostServices,
-    interactionMode: InteractionMode = "run"
+    interactionMode: InteractionMode = "run",
+    modelProfile?: SharedPlannerInput["modelProfile"]
   ): Promise<RouteSelection> {
     const online = await hostServices.connectivity.isOnline();
     const snapshots = await collectProviderRuntimeSnapshots(this.registry.list(), hostServices);
 
-    const planInput = buildSharedPlannerInput(request, snapshots, online, interactionMode);
+    const planInput = buildSharedPlannerInput(
+      request,
+      snapshots,
+      online,
+      interactionMode,
+      modelProfile
+    );
     const outcome = await this.planner.planRoute(planInput);
 
     if (!outcome.routePlan) {

package/src/facade/text.ts

@@ -0,0 +1,349 @@
+import type {
+  ExecutionProvenance,
+  StreamEvent,
+  StreamTerminalOutcome,
+  TaskRequest,
+  TaskResult
+} from "@independo/inderun-contracts";
+import { IndeRunException } from "../core/errors.js";
+import type { IndeRunApi } from "../core/generated/inderun-api.js";
+import { AUTO_MODEL_PROFILE } from "../core/model-profiles.js";
+
+/**
+ * The primary text API: `generateText()` and `streamText()`.
+ *
+ * Both are translation only. Each call compiles to exactly one canonical `TaskRequest`
+ * ({@link toTaskRequest}) and calls `run()` or `stream()` on the `IndeRunApi` it is given, so
+ * routing, fallback, cancellation, credentials, telemetry, and error normalization are the
+ * engine's and identical to a direct call. Written against the generated interface rather than
+ * the engine class, so anything that implements `IndeRunApi` -- a bridge included -- gets the
+ * same facade. See `docs/architecture/api-layering.md`.
+ */
+
+type Message = NonNullable<TaskRequest["messages"]>[number];
+
+/**
+ * Compact execution policy. Reuses the canonical enum values, so moving a call between the
+ * facade and `run()` never changes what a constraint means.
+ */
+export interface TextExecutionOptions {
+  /** `constraints.privacy`: how strict placement is, e.g. `"local_required"`. */
+  privacy?: NonNullable<TaskRequest["constraints"]>["privacy"];
+  /** `constraints.cloud`: `"forbidden"`, `"allowed"`, or `"required"`. */
+  cloud?: NonNullable<TaskRequest["constraints"]>["cloud"];
+  /** `constraints.timeoutMs`. */
+  timeoutMs?: number;
+  /** `preferences.optimizeFor`. */
+  optimizeFor?: NonNullable<TaskRequest["preferences"]>["optimizeFor"];
+}
+
+interface TextGenerationCommonOptions {
+  /**
+   * Model profile to route within, as configured on the engine. Omitted or `"auto"` routes
+   * across every registered target. Never a vendor model id.
+   */
+  model?: string;
+  /** Instructions sent as a leading `system` message. */
+  system?: string;
+  maxOutputTokens?: number;
+  temperature?: number;
+  topP?: number;
+  stop?: string[];
+  seed?: number;
+  execution?: TextExecutionOptions;
+  /** Secure-storage slot for provider credentials. Never a raw secret. */
+  authContextRef?: string;
+  telemetry?: TaskRequest["telemetry"];
+  /** Fixes the run id, for correlation. */
+  requestId?: string;
+}
+
+/** Options for {@link generateText} and {@link streamText}: a prompt or a message list. */
+export type TextGenerationOptions = TextGenerationCommonOptions &
+  ({ prompt: string; messages?: never } | { messages: Message[]; prompt?: never });
+
+/** What {@link generateText} returns. */
+export interface GeneratedText {
+  text: string;
+  finishReason: TaskResult["finishReason"];
+  usage?: TaskResult["usage"];
+  runId: string;
+  /** How the run played out; see `ExecutionProvenance`. */
+  provenance?: ExecutionProvenance;
+  /** The canonical result, for anything the summary above leaves out. */
+  taskResult: TaskResult;
+}
+
+/** How a {@link streamText} run ended, when it did not fail. */
+export interface StreamedText {
+  /** The full text, or the partial text a cancelled run had delivered. */
+  text: string;
+  /** `"cancelled"` when the run was cancelled; otherwise as the provider reported it. */
+  finishReason?: TaskResult["finishReason"];
+  usage?: TaskResult["usage"];
+  runId: string;
+  provenance?: ExecutionProvenance;
+  /** The canonical terminal outcome. */
+  outcome: StreamTerminalOutcome;
+}
+
+/** What {@link streamText} returns. */
+export interface TextStream {
+  runId: string;
+  /**
+   * Text increments, in order. Ends when the run completes or is cancelled; throws the run's
+   * `IndeRunException` when it fails, and {@link TextStreamRetractedError} when the provider
+   * replaces delivered text (use `fullStream` to render retractions).
+   */
+  textStream: AsyncIterable<string>;
+  /** The canonical `StreamEvent` sequence, terminal included. */
+  fullStream: AsyncIterable<StreamEvent>;
+  /**
+   * Settles with the run's ending: resolves on completion or cancellation, rejects with the
+   * run's `IndeRunException` on failure.
+   */
+  result: Promise<StreamedText>;
+  /** Cancels the run; idempotent. */
+  cancel(reason?: string): void;
+}
+
+/**
+ * Thrown by `textStream` when the provider replaces text it already delivered with text that
+ * does not extend it, which an append-only text view cannot represent.
+ */
+export class TextStreamRetractedError extends Error {
+  constructor(readonly runId: string) {
+    super(
+      `Run ${runId} replaced text it had already delivered; read fullStream to render retractions.`
+    );
+    this.name = "TextStreamRetractedError";
+  }
+}
+
+/** The canonical `TaskRequest` a facade call compiles to. */
+export function toTaskRequest(options: TextGenerationOptions): TaskRequest {
+  const request: TaskRequest = { schemaVersion: "1.0", task: { kind: "text_to_text" } };
+
+  if (options.system !== undefined) {
+    const conversation =
+      options.messages ??
+      (options.prompt !== undefined ? [{ role: "user" as const, content: options.prompt }] : []);
+    request.messages = [{ role: "system", content: options.system }, ...conversation];
+  } else if (options.messages !== undefined) {
+    // An empty list is the engine's to reject, with the same validation error as run().
+    request.messages = options.messages as NonNullable<TaskRequest["messages"]>;
+  } else if (options.prompt !== undefined) {
+    request.prompt = options.prompt;
+  }
+
+  if (options.model !== undefined && options.model !== AUTO_MODEL_PROFILE) {
+    request.modelProfile = options.model;
+  }
+
+  const generation: NonNullable<TaskRequest["generation"]> = {};
+  if (options.maxOutputTokens !== undefined) generation.maxOutputTokens = options.maxOutputTokens;
+  if (options.temperature !== undefined) generation.temperature = options.temperature;
+  if (options.topP !== undefined) generation.topP = options.topP;
+  if (options.stop !== undefined) generation.stop = options.stop;
+  if (options.seed !== undefined) generation.seed = options.seed;
+  if (Object.keys(generation).length > 0) request.generation = generation;
+
+  const execution = options.execution ?? {};
+  const constraints: NonNullable<TaskRequest["constraints"]> = {};
+  if (execution.privacy !== undefined) constraints.privacy = execution.privacy;
+  if (execution.cloud !== undefined) constraints.cloud = execution.cloud;
+  if (execution.timeoutMs !== undefined) constraints.timeoutMs = execution.timeoutMs;
+  if (Object.keys(constraints).length > 0) request.constraints = constraints;
+  if (execution.optimizeFor !== undefined) {
+    request.preferences = { optimizeFor: execution.optimizeFor };
+  }
+
+  if (options.authContextRef !== undefined) request.authContextRef = options.authContextRef;
+  if (options.telemetry !== undefined) request.telemetry = options.telemetry;
+  if (options.requestId !== undefined) request.requestId = options.requestId;
+
+  return request;
+}
+
+/**
+ * Generates text in one call: `run()` over the canonical request. Not cancellable, for the same
+ * reason `run()` is not; use {@link streamText} when cancellation matters.
+ * @throws {IndeRunException} Exactly as `run()` does.
+ */
+export async function generateText(
+  inderun: IndeRunApi,
+  options: TextGenerationOptions
+): Promise<GeneratedText> {
+  const taskResult = await inderun.run(toTaskRequest(options));
+  return {
+    text: taskResult.output.text,
+    finishReason: taskResult.finishReason,
+    ...(taskResult.usage !== undefined ? { usage: taskResult.usage } : {}),
+    runId: taskResult.runId,
+    ...(taskResult.provenance !== undefined ? { provenance: taskResult.provenance } : {}),
+    taskResult
+  };
+}
+
+/**
+ * Streams text: `stream()` over the canonical request. The run starts immediately and is read
+ * once, into a replay buffer, so `textStream`, `fullStream`, and `result` can each be consumed
+ * in any order (or not at all) without stalling the others.
+ * @throws {IndeRunException} Exactly as `stream()` does: validation and routing refusals reject
+ *   here, before there is a run.
+ */
+export async function streamText(
+  inderun: IndeRunApi,
+  options: TextGenerationOptions
+): Promise<TextStream> {
+  const run = await inderun.stream(toTaskRequest(options));
+  const runId = run.handle.runId;
+  const replay = new ReplayBuffer<StreamEvent>();
+
+  const result = (async (): Promise<StreamedText> => {
+    // The terminal event is the last one a run delivers, so it settles everything; a source
+    // that stays open after it (a bridge subscription, say) must not hold the result back.
+    // Releasing the source is not awaited for the same reason.
+    const events = run.events[Symbol.asyncIterator]();
+    try {
+      for (let next = await events.next(); next.done !== true; next = await events.next()) {
+        replay.push(next.value);
+        if (next.value.type === "terminal") {
+          void events.return?.()?.catch(() => undefined);
+          break;
+        }
+      }
+      replay.close();
+    } catch (error) {
+      replay.fail(error);
+      throw error;
+    }
+    const terminal = replay.items.find((event) => event.type === "terminal");
+    if (terminal === undefined) {
+      throw new IndeRunException({
+        errorClass: "Internal",
+        message: `Run ${runId} ended without a terminal event.`,
+        runId
+      });
+    }
+    return settle(terminal.payload as unknown as StreamTerminalOutcome);
+  })();
+  // Consumers may read only one view; an unread rejection must not surface as unhandled.
+  result.catch(() => undefined);
+
+  return {
+    runId,
+    fullStream: { [Symbol.asyncIterator]: () => replay.iterate() },
+    textStream: { [Symbol.asyncIterator]: () => textIncrements(replay.iterate(), runId) },
+    result,
+    cancel: (reason?: string) => run.cancel(reason)
+  };
+}
+
+function settle(outcome: StreamTerminalOutcome): StreamedText {
+  const provenance = outcome.provenance !== undefined ? { provenance: outcome.provenance } : {};
+  switch (outcome.outcome) {
+    case "completed":
+      return {
+        text: outcome.finalText ?? "",
+        ...(outcome.finishReason !== undefined ? { finishReason: outcome.finishReason } : {}),
+        ...(outcome.usage !== undefined ? { usage: outcome.usage } : {}),
+        runId: outcome.runId,
+        ...provenance,
+        outcome
+      };
+    case "cancelled":
+      return {
+        text: outcome.partialText ?? "",
+        finishReason: "cancelled",
+        runId: outcome.runId,
+        ...provenance,
+        outcome
+      };
+    case "error": {
+      const error = outcome.error!;
+      throw new IndeRunException({
+        errorClass: error.errorClass,
+        message: error.message,
+        runId: outcome.runId,
+        ...(error.providerId !== undefined ? { providerId: error.providerId } : {}),
+        ...(error.retryable !== undefined ? { retryable: error.retryable } : {}),
+        ...(error.retryAfterMs !== undefined ? { retryAfterMs: error.retryAfterMs } : {}),
+        ...(error.details !== undefined ? { details: error.details } : {}),
+        ...(outcome.provenance !== undefined ? { provenance: outcome.provenance } : {})
+      });
+    }
+  }
+}
+
+async function* textIncrements(
+  events: AsyncIterable<StreamEvent>,
+  runId: string
+): AsyncGenerator<string> {
+  let text = "";
+  for await (const event of events) {
+    if (event.type === "content_delta") {
+      const delta = (event.payload as { text: string }).text;
+      text += delta;
+      yield delta;
+    } else if (event.type === "content_snapshot") {
+      const snapshot = (event.payload as { text: string }).text;
+      if (!snapshot.startsWith(text)) {
+        throw new TextStreamRetractedError(runId);
+      }
+      const suffix = snapshot.slice(text.length);
+      text = snapshot;
+      if (suffix.length > 0) yield suffix;
+    } else if (event.type === "terminal") {
+      settle(event.payload as unknown as StreamTerminalOutcome);
+      return;
+    }
+  }
+}
+
+/**
+ * Holds every item read from a single-use source so any number of readers can iterate it from
+ * the start, each at its own pace. Readers that catch up wait for the next item.
+ */
+class ReplayBuffer<T> {
+  readonly items: T[] = [];
+  private done = false;
+  private failure: { error: unknown } | undefined;
+  private waiters: (() => void)[] = [];
+
+  push(item: T): void {
+    this.items.push(item);
+    this.wake();
+  }
+
+  close(): void {
+    this.done = true;
+    this.wake();
+  }
+
+  fail(error: unknown): void {
+    this.failure = { error };
+    this.close();
+  }
+
+  async *iterate(): AsyncGenerator<T> {
+    let index = 0;
+    while (true) {
+      if (index < this.items.length) {
+        yield this.items[index++]!;
+        continue;
+      }
+      if (this.done) {
+        if (this.failure !== undefined) throw this.failure.error;
+        return;
+      }
+      await new Promise<void>((resolve) => this.waiters.push(resolve));
+    }
+  }
+
+  private wake(): void {
+    const waiters = this.waiters;
+    this.waiters = [];
+    for (const wake of waiters) wake();
+  }
+}

package/src/index.ts

@@ -87,7 +87,13 @@ export {
 } from "./core/errors.js";
 
 /** The engine itself, and the app-facing API surface it implements. */
-export { IndeRun, type StreamHandleResult } from "./core/engine.js";
+export { IndeRun, type IndeRunOptions, type StreamHandleResult } from "./core/engine.js";
+
+/**
+ * Model profiles: application-defined names for ordered sets of targets that a
+ * request can route within. `"auto"` is built in.
+ */
+export { AUTO_MODEL_PROFILE, type ModelProfiles } from "./core/model-profiles.js";
 export type { IndeRunApi } from "./core/generated/inderun-api.js";
 
 /**
@@ -140,8 +146,28 @@ export {
   type FetchHttpClientOptions
 } from "./core/browser-host.js";
 
+/**
+ * The primary text API: one canonical request per call, over any `IndeRunApi`.
+ * See `docs/architecture/api-layering.md`.
+ */
+export {
+  generateText,
+  streamText,
+  toTaskRequest,
+  TextStreamRetractedError,
+  type GeneratedText,
+  type StreamedText,
+  type TextExecutionOptions,
+  type TextGenerationOptions,
+  type TextStream
+} from "./facade/text.js";
+
 /**
  * The recommended way in: builds host services, registers the providers named in
  * the options, and returns a ready engine.
  */
-export { createIndeRunWeb, type CreateIndeRunWebOptions } from "./web-sdk.js";
+export {
+  createIndeRunWeb,
+  type CreateIndeRunWebOptions,
+  type OpenAIFactoryOptions
+} from "./web-sdk.js";

package/src/providers/onnx/provider.ts

@@ -101,6 +101,7 @@ export class OnnxRuntimeWebProvider implements ProviderAdapter {
     return {
       id: this.id,
       type: "local",
+      model: this.modelPackage.id,
       transport: "in_process",
       supports: {
         run: true,

package/src/providers/openai/provider.ts

@@ -90,6 +90,7 @@ export interface OpenAIProviderOptions {
 }
 
 type OpenAIResponseBody = {
+  model?: unknown;
   output_text?: unknown;
   output?: unknown;
   status?: unknown;
@@ -146,6 +147,7 @@ export class OpenAIResponsesProvider implements ProviderAdapter {
     return {
       id: this.id,
       type: "cloud",
+      model: this.options.model,
       transport: "http",
       streamingStyle: "tokens",
       supports: {
@@ -324,6 +326,11 @@ export class OpenAIResponsesProvider implements ProviderAdapter {
       }
     };
 
+    const reportedModel = getReportedModel(responseBody);
+    if (reportedModel !== undefined) {
+      result.telemetry.reportedModel = reportedModel;
+    }
+
     const usage = getUsage(responseBody);
     if (usage) {
       result.usage = usage;
@@ -440,6 +447,10 @@ export class OpenAIResponsesProvider implements ProviderAdapter {
         if (usage) {
           done.usage = usage;
         }
+        const reportedModel = getReportedModel(body);
+        if (reportedModel !== undefined) {
+          done.reportedModel = reportedModel;
+        }
         yield done;
         return;
       }
@@ -686,6 +697,16 @@ function getFinishReason(response: OpenAIResponseBody): TaskResult["finishReason
   return "stop";
 }
 
+/**
+ * The model id the Responses API echoes on the response object. It names the model that
+ * served the request, which can be a dated snapshot of the configured alias.
+ */
+function getReportedModel(responseBody: OpenAIResponseBody): string | undefined {
+  return typeof responseBody.model === "string" && responseBody.model.length > 0
+    ? responseBody.model
+    : undefined;
+}
+
 function getUsage(response: OpenAIResponseBody): TaskResult["usage"] | undefined {
   const usage: TaskResult["usage"] = {};