@dxos/plugin-assistant 0.8.3 → 0.8.4-main.1c7ec43d41

package/src/processor/presets.ts

@@ -0,0 +1,68 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import * as Schema from 'effect/Schema';
+
+import { type ModelName } from '@dxos/ai';
+
+const ModelProviders = ['dxos-local', 'dxos-remote', 'lm-studio', 'ollama'] as const;
+
+const ModelProvider = Schema.Literal(...ModelProviders);
+type ModelProvider = Schema.Schema.Type<typeof ModelProvider>;
+
+export type AiServicePreset = {
+  id: string;
+  provider: ModelProvider;
+  model: ModelName;
+  label?: string;
+};
+
+const createModelLabel = (model: ModelName) => {
+  const parts = model.split('/');
+  return parts[parts.length - 1];
+};
+
+// TODO(burdon): Users should be able to create and edit presets.
+export const AiServicePresets: AiServicePreset[] = [
+  {
+    provider: 'dxos-remote' as const,
+    model: '@anthropic/claude-opus-4-6' as const,
+  },
+  {
+    provider: 'dxos-remote' as const,
+    model: '@anthropic/claude-sonnet-4-5' as const,
+  },
+  {
+    provider: 'dxos-remote' as const,
+    model: '@anthropic/claude-3-5-haiku-20241022' as const,
+  },
+  {
+    provider: 'ollama' as const,
+    model: 'gpt-oss:20b' as const,
+  },
+  {
+    provider: 'ollama' as const,
+    model: 'gemma4:latest' as const,
+  },
+  {
+    provider: 'lm-studio' as const,
+    model: '@google/gemma-3-27b' as const,
+  },
+  {
+    provider: 'lm-studio' as const,
+    model: '@meta/llama-3.2-3b-instruct' as const,
+  },
+  {
+    model: 'deepseek-r1:latest' as const,
+    provider: 'dxos-local' as const,
+  },
+].map(
+  ({ model, provider }, i) =>
+    ({
+      id: `preset-${i}`,
+      provider,
+      model,
+      label: createModelLabel(model),
+    }) satisfies AiServicePreset,
+);

package/src/processor/processor.test.ts

@@ -0,0 +1,40 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import { describe, it } from '@effect/vitest';
+import * as Effect from 'effect/Effect';
+import * as ManagedRuntime from 'effect/ManagedRuntime';
+
+import { AiSession } from '@dxos/assistant';
+import { Chat } from '@dxos/assistant-toolkit';
+import { Database, Feed } from '@dxos/echo';
+import { acquireReleaseResource } from '@dxos/effect';
+import { TestHelpers } from '@dxos/effect/testing';
+import { AssistantTestLayer } from '@dxos/functions-runtime/testing';
+
+import { AiChatProcessor } from './processor';
+
+const TestLayer = AssistantTestLayer({ tracing: 'noop', types: [Chat.Chat, Feed.Feed] });
+
+describe('Chat processor', () => {
+  it.scoped(
+    'basic',
+    Effect.fn(
+      function* ({ expect }) {
+        const feed = Feed.make();
+        yield* Database.add(feed);
+        const runtime = yield* Effect.runtime<Feed.FeedService>();
+        const session = yield* acquireReleaseResource(() => new AiSession({ feed, runtime }));
+        const managedRuntime = ManagedRuntime.make(
+          Effect.runSync(Effect.map(Effect.context<never>(), () => undefined as any)) as any,
+        );
+        const processor = new AiChatProcessor(session, managedRuntime as any, feed);
+        expect(processor).toBeDefined();
+        expect(processor.active).toBeDefined();
+      },
+      Effect.provide(TestLayer),
+      TestHelpers.provideTestContext,
+    ),
+  );
+});

package/src/processor/processor.ts

@@ -0,0 +1,352 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import { Atom, Registry } from '@effect-atom/atom-react';
+import * as Effect from 'effect/Effect';
+import * as Fiber from 'effect/Fiber';
+import * as Option from 'effect/Option';
+import * as Stream from 'effect/Stream';
+
+import { type AiService, DEFAULT_EDGE_MODEL, type ModelName, type ModelRegistry } from '@dxos/ai';
+import {
+  AiContextService,
+  type AiSession,
+  createSystemPrompt,
+  formatSystemPrompt,
+  McpServerError,
+  PartialBlock,
+  ToolExecutionServices,
+} from '@dxos/assistant';
+import { type Chat } from '@dxos/assistant-toolkit';
+import { type Blueprint, type Credential, Trace, Operation } 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 { AutomationCapabilities } from '@dxos/plugin-automation/types';
+import { Message } from '@dxos/types';
+
+import { UpdateChatName } from '../operations/definitions';
+
+/**
+ * @deprecated Services type for the old direct-conversation processor path.
+ * Retained for backward compatibility with CLI and update-name.
+ */
+export type AiChatServices =
+  | Credential.CredentialsService
+  | Database.Service
+  | QueueService
+  | AiService.AiService
+  | Trace.TraceService;
+
+export type AiChatProcessorOptions = {
+  model?: ModelName;
+  modelRegistry?: ModelRegistry;
+  blueprintRegistry?: Blueprint.Registry;
+  observableRegistry?: Registry.Registry;
+  /**
+   * 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,
+};
+
+export type AiRequestOptions = {};
+
+export type AiRequest = {
+  message: string;
+  options?: AiRequestOptions;
+};
+
+/**
+ * Handles interactions with the AI service.
+ * Uses AgentService to spawn a process-backed agent and subscribes to ephemeral trace events for streaming.
+ */
+export class AiChatProcessor {
+  readonly #registry: Registry.Registry;
+
+  /** Pending messages (finalized, non-streaming). */
+  readonly #pending = Atom.make<Message.Message[]>([]);
+
+  /** Currently streaming messages (from ephemeral trace events). */
+  readonly #streaming = Atom.make<Message.Message[]>([]);
+
+  /** Set of message IDs that have been finalized (non-pending delivered via ephemeral). */
+  readonly #finalizedIds = new Set<string>();
+
+  /** Currently active request fiber. */
+  #requestFiber: Fiber.RuntimeFiber<void, unknown> | undefined;
+
+  /** Last request (for retries). */
+  #lastRequest: AiRequest | undefined;
+
+  /** Streaming state. */
+  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) => [...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: AiSession,
+    private readonly _runtime: AutomationCapabilities.ComputeRuntime,
+    private readonly _feed: Feed.Feed,
+    private readonly _options: AiChatProcessorOptions = defaultOptions,
+  ) {
+    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);
+    }
+  }
+
+  get context() {
+    return this._conversation.context;
+  }
+
+  get conversation() {
+    return this._conversation;
+  }
+
+  get blueprintRegistry() {
+    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));
+  }
+
+  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(AiContextService, { binder: this.context }), Effect.orDie),
+    );
+  }
+
+  /**
+   * Initiates a new request via AgentService.
+   */
+  async request(requestProp: AiRequest): Promise<void> {
+    if (this.#requestFiber) {
+      await this.cancel();
+    }
+
+    try {
+      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,
+        );
+
+        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');
+
+        this.#flushStreaming();
+
+        yield* this.#maybeUpdateChatName();
+      });
+
+      this.#requestFiber = this._runtime.runFork(effect);
+
+      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.#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 {
+      log.info('setting active to false');
+      this.#registry.set(this.active, false);
+      this.#requestFiber = undefined;
+    }
+  }
+
+  /**
+   * Cancels the current request.
+   */
+  async cancel(): Promise<void> {
+    await runAndForwardErrors(
+      Effect.gen(this, function* () {
+        if (this.#requestFiber) {
+          yield* Fiber.interrupt(this.#requestFiber);
+        }
+      }),
+    );
+
+    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);
+    }
+  }
+
+  /**
+   * 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: Chat.Chat): Promise<void> {
+    unwrapExit(await this._runtime.runPromiseExit(Operation.invoke(UpdateChatName, { chat })));
+  }
+
+  /**
+   * 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],
+    });
+
+    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];
+      });
+    }
+  }
+
+  /**
+   * 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(UpdateChatName, { chat });
+  }
+}

package/src/queue-logger.ts

@@ -2,47 +2,57 @@
 // Copyright 2025 DXOS.org
 //
 
-import { type Blueprint, type BlueprintEvent, type BlueprintLogger } from '@dxos/assistant';
-import { getSpace, Ref, type Queue, type Space } from '@dxos/client/echo';
-import { DXN, Key, Obj } from '@dxos/echo';
-import {
-  InvocationOutcome,
-  InvocationTraceEndEvent,
-  InvocationTraceEventType,
-  InvocationTraceStartEvent,
-  TraceEvent,
-  type InvocationTraceEvent,
-} from '@dxos/functions';
+import { type Queue, Ref, type Space, getSpace } from '@dxos/client/echo';
+import { type Sequence, type SequenceEvent, type SequenceLogger } from '@dxos/conductor';
+import { DXN, Feed, Obj } from '@dxos/echo';
+import { InvocationTraceEndEvent, InvocationTraceEventType, InvocationTraceStartEvent } from '@dxos/functions-runtime';
+import { TraceEvent } from '@dxos/functions-runtime';
+import { InvocationOutcome } from '@dxos/functions-runtime';
+import { type InvocationTraceEvent } from '@dxos/functions-runtime';
 import { invariant } from '@dxos/invariant';
 import { QueueSubspaceTags } from '@dxos/keys';
 
-export class QueueLogger implements BlueprintLogger {
+export class QueueLogger implements SequenceLogger {
   private _space: Space;
   private _invocationTraceQueue: Queue<InvocationTraceEvent>;
 
-  constructor(private readonly blueprint: Blueprint) {
-    const space = getSpace(blueprint);
+  constructor(private readonly sequence: Sequence) {
+    const space = getSpace(sequence);
     invariant(space, 'Space not found');
     this._space = space;
-    let dxn = this._space.properties.invocationTraceQueue?.dxn;
-    if (!dxn) {
-      dxn = DXN.fromQueue(QueueSubspaceTags.TRACE, this._space.id, Key.ObjectId.random());
-      this._space.properties.invocationTraceQueue = Ref.fromDXN(dxn);
+
+    const existingFeedRef = this._space.properties.invocationTraceFeed;
+    let queueDxn: DXN | undefined;
+
+    if (existingFeedRef) {
+      // A feed reference exists; resolve its queue DXN. If the target isn't loaded yet,
+      // fail loudly rather than silently creating a new feed and orphaning existing traces.
+      invariant(existingFeedRef.target, 'invocationTraceFeed reference is not yet loaded');
+      queueDxn = Feed.getQueueDxn(existingFeedRef.target) ?? undefined;
+      invariant(queueDxn, 'invocationTraceFeed has no queue DXN');
+    } else {
+      const feed = space.db.add(Feed.make({ namespace: 'trace' }));
+      queueDxn = Feed.getQueueDxn(feed) ?? undefined;
+      invariant(queueDxn, 'New invocationTraceFeed has no queue DXN');
+      Obj.update(this._space.properties, (obj) => {
+        obj.invocationTraceFeed = Ref.make(feed);
+      });
     }
-    this._invocationTraceQueue = this._space.queues.get(dxn);
+
+    this._invocationTraceQueue = this._space.queues.get(queueDxn);
   }
 
-  log(event: BlueprintEvent) {
+  log(event: SequenceEvent) {
     switch (event.type) {
       case 'begin':
         void this._invocationTraceQueue.append([
           Obj.make(InvocationTraceStartEvent, {
             type: InvocationTraceEventType.START,
             invocationId: event.invocationId,
-            timestampMs: Date.now(),
+            timestamp: Date.now(),
             input: {},
             invocationTraceQueue: Ref.fromDXN(this._getTraceQueueDxn(event.invocationId)),
-            invocationTarget: Ref.make(this.blueprint),
+            invocationTarget: Ref.make(this.sequence),
           }),
         ]);
         break;
@@ -51,7 +61,7 @@ export class QueueLogger implements BlueprintLogger {
           Obj.make(InvocationTraceEndEvent, {
             type: InvocationTraceEventType.END,
             invocationId: event.invocationId,
-            timestampMs: Date.now(),
+            timestamp: Date.now(),
             outcome: InvocationOutcome.SUCCESS,
           }),
         ]);
@@ -62,10 +72,10 @@ export class QueueLogger implements BlueprintLogger {
           Obj.make(TraceEvent, {
             outcome: event.type,
             truncated: false,
-            ingestionTimestampMs: Date.now(),
+            ingestionTimestamp: Date.now(),
             logs: [
               {
-                timestampMs: Date.now(),
+                timestamp: Date.now(),
                 level: 'info',
                 message: event.type,
                 context: { step: event.step },
@@ -80,10 +90,10 @@ export class QueueLogger implements BlueprintLogger {
           Obj.make(TraceEvent, {
             outcome: event.type,
             truncated: false,
-            ingestionTimestampMs: Date.now(),
+            ingestionTimestamp: Date.now(),
             logs: [
               {
-                timestampMs: Date.now(),
+                timestamp: Date.now(),
                 level: 'info',
                 message: event.type,
                 context: { message: event.message },
@@ -98,10 +108,10 @@ export class QueueLogger implements BlueprintLogger {
           Obj.make(TraceEvent, {
             outcome: event.type,
             truncated: false,
-            ingestionTimestampMs: Date.now(),
+            ingestionTimestamp: Date.now(),
             logs: [
               {
-                timestampMs: Date.now(),
+                timestamp: Date.now(),
                 level: 'info',
                 message: event.type,
                 context: { block: event.block },