@dxos/plugin-assistant 0.8.4-main.7ace549 → 0.8.4-main.8baae0fced

package/src/processor/processor.ts

@@ -3,126 +3,154 @@
 //
 
 import { Atom, Registry } from '@effect-atom/atom-react';
-import * as Cause from 'effect/Cause';
 import * as Effect from 'effect/Effect';
-import * as Exit from 'effect/Exit';
 import * as Fiber from 'effect/Fiber';
+import * as Layer from 'effect/Layer';
 import * as Option from 'effect/Option';
-import * as Runtime from 'effect/Runtime';
+import * as Stream from 'effect/Stream';
 
+import { type AiService, DEFAULT_EDGE_MODEL, type ModelName, type ModelRegistry, type OpaqueToolkit } from '@dxos/ai';
+import { Capabilities } from '@dxos/app-framework';
 import {
-  AiService,
-  DEFAULT_EDGE_MODEL,
-  type ModelName,
-  type ModelRegistry,
-  type ToolExecutionService,
-  type ToolResolverService,
-} from '@dxos/ai';
-import {
-  type AiConversation,
-  type AiConversationRunParams,
-  ArtifactDiffResolver,
-  GenerationObserver,
+  AiContext,
+  AiSession,
   createSystemPrompt,
+  formatSystemPrompt,
+  McpServerError,
+  PartialBlock,
+  ToolExecutionServices,
 } from '@dxos/assistant';
-import { type Blueprint } from '@dxos/blueprints';
-import { Obj } from '@dxos/echo';
-import { throwCause } from '@dxos/effect';
+import { type Chat } from '@dxos/assistant-toolkit';
 import {
-  type CredentialsService,
-  type DatabaseService,
-  type FunctionInvocationService,
-  type QueueService,
-  type TracingService,
-} from '@dxos/functions';
+  type Blueprint,
+  type Credential,
+  Operation,
+  type OperationRegistry,
+  type ServiceNotAvailableError,
+  Trace,
+} from '@dxos/compute';
+import { type Database, Feed, Obj, Ref } from '@dxos/echo';
+import { runAndForwardErrors, unwrapExit } from '@dxos/effect';
+import { type QueueService } from '@dxos/functions';
+import { AgentService } from '@dxos/functions-runtime';
 import { log } from '@dxos/log';
-import { type ContentBlock, Message } from '@dxos/types';
-
-import { type Assistant } from '../types';
+import { Message } from '@dxos/types';
 
-import { updateName } from './update-name';
+import { AssistantOperation } from '#types';
 
+/**
+ * @deprecated Services type for the old direct-conversation processor path.
+ * Retained for backward compatibility with CLI and update-name.
+ */
 export type AiChatServices =
-  | CredentialsService
-  | DatabaseService
+  | Credential.CredentialsService
+  | Database.Service
+  | QueueService
+  | AiService.AiService
+  | Trace.TraceService;
+
+/**
+ * Space-scoped services materialised by the layer passed into
+ * {@link AiChatProcessor}. Mirrors the tag list that
+ * {@link useChatProcessor} passes to {@link ServiceResolver.provide}.
+ */
+export type SpaceServices =
+  | Database.Service
   | QueueService
-  | FunctionInvocationService
+  | Feed.FeedService
+  | Credential.CredentialsService
   | AiService.AiService
-  | ToolExecutionService
-  | ToolResolverService
-  | TracingService;
+  | AgentService.AgentService
+  | OperationRegistry.Service
+  | OpaqueToolkit.OpaqueToolkitProvider;
 
 export type AiChatProcessorOptions = {
   model?: ModelName;
   modelRegistry?: ModelRegistry;
   blueprintRegistry?: Blueprint.Registry;
   observableRegistry?: Registry.Registry;
-  extensions?: ToolContextExtensions;
-} & Pick<AiConversationRunParams, 'system'>;
+  /**
+   * For tracing.
+   */
+  chat?: Ref.Ref<Chat.Chat>;
+  system?: string;
+  /**
+   * Probability of automatically updating chat name after each request.
+   * Chat name is always updated if it has no name.
+   * @default 0.1 (10%)
+   */
+  autoUpdateNameChance?: number;
+};
 
 const defaultOptions: Partial<AiChatProcessorOptions> = {
   model: DEFAULT_EDGE_MODEL,
+  autoUpdateNameChance: 0.1,
 };
 
-// TODO(burdon): Retry, timeout?
-export type AiRequestOptions = {};
+export type ProcessorRequestOptions = {};
 
-export type AiRequest = {
+export type ProcessorRequest = {
   message: string;
-  options?: AiRequestOptions;
+  options?: ProcessorRequestOptions;
 };
 
 /**
  * Handles interactions with the AI service.
- * Handles streaming responses from the conversation.
+ * Uses AgentService to spawn a process-backed agent and subscribes to ephemeral trace events for streaming.
  */
 export class AiChatProcessor {
-  private readonly _registry: Registry.Registry;
+  readonly #registry: Registry.Registry;
 
-  /** External observer. */
-  private readonly _observer: GenerationObserver;
+  /** Pending messages (finalized, non-streaming). */
+  readonly #pending = Atom.make<Message.Message[]>([]);
 
-  /** Currently active request fiber. */
-  private _fiber: Fiber.Fiber<void, any> | undefined;
+  /** Currently streaming messages (from ephemeral trace events). */
+  readonly #streaming = Atom.make<Message.Message[]>([]);
 
-  /** Last request (for retries). */
-  private _lastRequest: AiRequest | undefined;
+  /** Set of message IDs that have been finalized (non-pending delivered via ephemeral). */
+  readonly #finalizedIds = new Set<string>();
 
-  /** Pending messages (incl. the current user request). */
-  private readonly _pending = Atom.make<Message.Message[]>([]);
+  /** Currently active request fiber. */
+  #requestFiber: Fiber.RuntimeFiber<void, unknown> | undefined;
 
-  /** Currently streaming message (from the AI service). */
-  private readonly _streaming = Atom.make<Option.Option<Message.Message>>(Option.none());
+  /** Last request (for retries). */
+  #lastRequest: ProcessorRequest | undefined;
 
   /** Streaming state. */
-  public readonly streaming = Atom.make<boolean>((get) => Option.isSome(get(this._streaming)));
+  public readonly streaming = Atom.make<boolean>((get) => get(this.#streaming).length > 0);
 
   /** Active state. */
   public readonly active = Atom.make(false);
 
   /** Array of Messages (incl. the current message being streamed). */
-  public readonly messages = Atom.make<Message.Message[]>((get) =>
-    Option.match(get(this._streaming), {
-      onNone: () => get(this._pending),
-      onSome: (streaming) => [...get(this._pending), streaming],
-    }),
-  );
+  public readonly messages = Atom.make<Message.Message[]>((get) => [...get(this.#pending), ...get(this.#streaming)]);
 
   /** Last error. */
   public readonly error = Atom.make<Option.Option<Error>>(Option.none());
 
+  /**
+   * MCP server connection errors observed during the most recent request.
+   * Misconfigured/unreachable servers are dropped from the toolkit so the chat
+   * keeps working; the entries here let the UI display which servers failed.
+   */
+  public readonly mcpErrors = Atom.make<readonly Trace.PayloadType<typeof McpServerError>[]>([]);
+
   constructor(
-    private readonly _conversation: AiConversation,
-    // TODO(dmaretskyi): Replace this with effect's ManagedRuntime wrapping this layer.
-    private readonly _services: () => Promise<Runtime.Runtime<AiChatServices>>,
+    private readonly _conversation: AiSession.Session,
+    private readonly _runtime: Capabilities.ProcessManagerRuntime,
+    private readonly _feed: Feed.Feed,
+    /**
+     * Pre-built layer that materialises {@link SpaceServices}. Built via
+     * {@link ServiceResolver.provide} with the {@link ServiceResolver} already
+     * supplied (hence `RIn = never`); the {@link ServiceNotAvailableError}
+     * error channel surfaces when a tag is not available for the space.
+     * Provided to every effect run by the processor so the underlying
+     * {@link ProcessManagerRuntime} has access to space-affinity services.
+     */
+    private readonly _spaceLayer: Layer.Layer<SpaceServices, ServiceNotAvailableError, never>,
     private readonly _options: AiChatProcessorOptions = defaultOptions,
   ) {
-    // Initialize registries and defaults before using in other logic.
-    this._registry = this._options.observableRegistry ?? Registry.make();
-    this._observer = GenerationObserver.make({
-      onBlock: this._onBlock,
-      onMessage: this._onMessage,
-    });
+    this.#registry = this._options.observableRegistry ?? Registry.make();
     if (this._options.model && !this._options.system) {
       const capabilities = this._options.modelRegistry?.getCapabilities(this._options.model) ?? {};
       this._options.system = createSystemPrompt(capabilities);
@@ -141,54 +169,96 @@ export class AiChatProcessor {
     return this._options.blueprintRegistry;
   }
 
+  get system(): string {
+    return this._options.system ?? '';
+  }
+
+  async getTools(): Promise<Record<string, any>> {
+    return this._runtime.runPromise(
+      Effect.provide(this._conversation.getTools(), ToolExecutionServices).pipe(Effect.provide(this._spaceLayer)),
+    );
+  }
+
+  async getSystemPrompt(): Promise<string> {
+    return this._runtime.runPromise(
+      Effect.gen(this, function* () {
+        const blueprints = this.context.getBlueprints();
+        const objects = this.context.getObjects();
+        return yield* formatSystemPrompt({ system: this._options.system, blueprints, objects });
+      }).pipe(
+        Effect.provideService(AiContext.Service, { binder: this.context }),
+        Effect.provide(this._spaceLayer),
+        Effect.orDie,
+      ),
+    );
+  }
+
   /**
-   * Initiates a new request.
+   * Initiates a new request via AgentService.
    */
-  async request(requestParam: AiRequest): Promise<void> {
-    if (this._fiber) {
+  async request(requestProp: ProcessorRequest): Promise<void> {
+    if (this.#requestFiber) {
       await this.cancel();
     }
 
     try {
-      this._lastRequest = requestParam;
-      this._registry.set(this.error, Option.none());
-      this._registry.set(this.active, true);
-
-      // Create request.
-      const request = this._conversation.createRequest({
-        system: this._options.system,
-        prompt: requestParam.message,
-        observer: this._observer,
-      });
+      this.#lastRequest = requestProp;
+      this.#registry.set(this.error, Option.none());
+      this.#registry.set(this.mcpErrors, []);
+      this.#registry.set(this.active, true);
+
+      const effect = Effect.gen(this, function* () {
+        // NOTE: Gets or creates a session for the feed.
+        log.info('init agent session', { feed: Obj.getDXN(this._feed).toString(), model: this._options.model });
+        const session = yield* AgentService.getSession(this._feed, { model: this._options.model });
+        const ephemeralStream = session.subscribeEphemeral();
+        yield* ephemeralStream.pipe(
+          Stream.runForEach((message) =>
+            Effect.sync(() => {
+              for (const event of message.events) {
+                if (Trace.isOfType(PartialBlock, event)) {
+                  this.#handleEphemeralMessage(event.data);
+                } else if (Trace.isOfType(McpServerError, event)) {
+                  this.#handleMcpError(event.data);
+                }
+              }
+            }),
+          ),
+          Effect.fork,
+        );
 
-      const runtime = await this._services();
+        log('chat processor submitting prompt', { length: requestProp.message.length });
+        yield* session.submitPrompt(requestProp.message);
+        log('chat processor submitPrompt returned, waiting for agent', {});
+        yield* session.waitForCompletion();
+        log.info('session complete');
 
-      // Create fiber.
-      this._fiber = request.pipe(
-        Effect.provide(AiService.model(this._options.model ?? DEFAULT_EDGE_MODEL)),
+        this.#flushStreaming();
 
-        // TODO(dmaretskyi): Move ArtifactDiffResolver upstream.
-        Effect.provideService(ArtifactDiffResolver, this._artifactDiffResolver),
+        yield* this.#maybeUpdateChatName();
+      });
 
-        Effect.asVoid,
-        Runtime.runFork(runtime), // Runs in the background.
-      );
+      this.#requestFiber = this._runtime.runFork(effect.pipe(Effect.provide(this._spaceLayer)));
 
-      // Execute request.
-      const response = await this._fiber.pipe(Fiber.join, Effect.runPromiseExit);
-      if (!Exit.isSuccess(response) && !Cause.isInterruptedOnly(response.cause)) {
-        throwCause(response.cause);
+      try {
+        await this._runtime.runPromise(Fiber.join(this.#requestFiber));
+      } catch (err: any) {
+        if (err._tag === 'InterruptedException' || err.message?.includes('interrupted')) {
+          return;
+        }
+        throw err;
       }
 
-      this._registry.set(this.error, Option.none());
-      this._lastRequest = undefined;
-      this._fiber = undefined;
-    } catch (error) {
-      log.error('request failed', { error });
-      this._registry.set(this.error, Option.some(new Error('AI service error', { cause: error })));
+      this.#registry.set(this.error, Option.none());
+      this.#lastRequest = undefined;
+      this.#requestFiber = undefined;
+    } catch (err) {
+      log.error('request failed', { error: err });
+      this.#registry.set(this.error, Option.some(new Error('AI service error', { cause: err })));
     } finally {
-      this._fiber = undefined;
-      this._registry.set(this.active, false);
+      log.info('setting active to false');
+      this.#registry.set(this.active, false);
+      this.#requestFiber = undefined;
     }
   }
 
@@ -196,82 +266,129 @@ export class AiChatProcessor {
    * Cancels the current request.
    */
   async cancel(): Promise<void> {
-    await Effect.runPromise(
+    await runAndForwardErrors(
       Effect.gen(this, function* () {
-        if (this._fiber) {
-          yield* this._fiber.pipe(Fiber.interrupt);
+        if (this.#requestFiber) {
+          yield* Fiber.interrupt(this.#requestFiber);
         }
       }),
     );
 
-    this._fiber = undefined;
-    this._registry.set(this.active, false);
+    this.#requestFiber = undefined;
+    this.#registry.set(this.active, false);
   }
 
   /**
    * Retry last failed request.
    */
   async retry(): Promise<void> {
-    if (this._lastRequest) {
-      return this.request(this._lastRequest);
+    if (this.#lastRequest) {
+      return this.request(this.#lastRequest);
     }
   }
 
+  /**
+   * Clears the recorded MCP server errors (e.g. after the user dismisses the warning banner).
+   */
+  dismissMcpErrors(): void {
+    this.#registry.set(this.mcpErrors, []);
+  }
+
   /**
    * Update the current chat's name.
    */
-  async updateName(chat: Assistant.Chat): Promise<void> {
-    const runtime = await this._services();
-    await updateName(runtime, this._conversation, chat, this._options.model);
+  async updateName(chat: Chat.Chat): Promise<void> {
+    unwrapExit(
+      await this._runtime.runPromiseExit(
+        Operation.invoke(AssistantOperation.UpdateChatName, { chat }).pipe(Effect.provide(this._spaceLayer)),
+      ),
+    );
   }
 
-  // TODO(burdon): Fix/factor out.
-  private _artifactDiffResolver: ArtifactDiffResolver.Service = {
-    resolve: async (_artifacts) => {
-      const versions = new Map();
-      // await Promise.all(
-      //   artifacts.map(async (artifact) => {
-      //     const {
-      //       objects: [object],
-      //     } = await space.db.query(Filter.ids(artifact.id)).run();
-      //     if (!object) {
-      //       return;
-      //     }
-
-      //     versions.set(artifact.id, {
-      //       version: getVersion(object),
-      //       diff: `Current state: ${JSON.stringify(object)}`,
-      //     });
-      //   }),
-      // );
-
-      return versions;
-    },
-  };
-
-  private _onMessage = Effect.fn(
-    function* (this: AiChatProcessor, message: Message.Message) {
-      this._registry.set(this._streaming, Option.none());
-      this._registry.update(this._pending, (pending) => [...pending, message]);
-    }.bind(this),
-  );
-
-  private _onBlock = Effect.fn(
-    function* (this: AiChatProcessor, block: ContentBlock.Any) {
-      this._registry.update(this._streaming, (streaming) => {
-        const blocks = streaming.pipe(
-          Option.map((streaming) => streaming.blocks.filter((b: ContentBlock.Any) => !b.pending)),
-          Option.getOrElse(() => []),
-        );
+  /**
+   * Handles an ephemeral message from the agent process.
+   * Both pending and completed blocks arrive here. Completed blocks are deduped
+   * against messages already written to the feed queue to handle the race between
+   * ephemeral delivery and feed replication.
+   */
+  #handleEphemeralMessage(event: Trace.PayloadType<typeof PartialBlock>) {
+    const isPending = event.block.pending;
+    const message = Obj.make(Message.Message, {
+      id: event.messageId,
+      created: new Date().toISOString(),
+      sender: { role: event.role },
+      blocks: [event.block],
+    });
 
-        return Option.some(
-          Obj.make(Message.Message, {
-            created: new Date().toISOString(),
-            sender: { role: 'assistant' },
-            blocks: [...blocks, block],
-          }),
-        );
+    if (isPending) {
+      if (this.#finalizedIds.has(event.messageId)) {
+        return;
+      }
+      this.#registry.update(this.#streaming, (streaming) => {
+        const idx = streaming.findIndex((existing) => existing.id === event.messageId);
+        if (idx >= 0) {
+          const updated = [...streaming];
+          updated[idx] = message;
+          return updated;
+        }
+        return [...streaming, message];
+      });
+    } else {
+      this.#finalizedIds.add(event.messageId);
+      this.#registry.update(this.#streaming, (streaming) => streaming.filter((existing) => existing.id !== message.id));
+      this.#registry.update(this.#pending, (pending) => {
+        if (pending.some((existing) => existing.id === message.id)) {
+          return pending;
+        }
+        return [...pending, message];
       });
-    }.bind(this),
-  );
+    }
+  }
+
+  /**
+   * Records a per-server MCP failure, deduped by url+protocol so repeat misconfigurations
+   * across turns do not spam the UI.
+   */
+  #handleMcpError(event: Trace.PayloadType<typeof McpServerError>) {
+    log.warn('MCP server error', event);
+    this.#registry.update(this.mcpErrors, (errors) => {
+      if (errors.some((existing) => existing.url === event.url && existing.protocol === event.protocol)) {
+        return errors;
+      }
+      return [...errors, event];
+    });
+  }
+
+  /**
+   * Move remaining streaming messages to pending (called when agent completes).
+   */
+  #flushStreaming() {
+    const remaining = this.#registry.get(this.#streaming);
+    if (remaining.length > 0) {
+      this.#registry.update(this.#pending, (pending) => [...pending, ...remaining]);
+      this.#registry.set(this.#streaming, []);
+    }
+    this.#finalizedIds.clear();
+  }
+
+  /**
+   * Conditionally schedule chat name update in detached fork mode.
+   * Updates if chat has no name OR based on random chance (default 10%).
+   */
+  #maybeUpdateChatName(): Effect.Effect<void, never, Operation.Service> {
+    const chat = this._options.chat?.target;
+    if (!chat) {
+      return Effect.void;
+    }
+
+    const chance = this._options.autoUpdateNameChance ?? defaultOptions.autoUpdateNameChance ?? 0.1;
+    const shouldUpdate = !chat.name || Math.random() < chance;
+    if (!shouldUpdate) {
+      return Effect.void;
+    }
+
+    // TODO(dmaretskyi): Operation.schedule didn't work.
+    log.info('scheduling chat name update', { hasName: !!chat.name, chance });
+    return Operation.schedule(AssistantOperation.UpdateChatName, { chat });
+  }
 }