@dxos/assistant 0.8.4-main.f9ba587 → 0.8.4-main.fcfe5033a5

package/src/session/request.test.ts

@@ -0,0 +1,195 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import * as Tool from '@effect/ai/Tool';
+import * as Toolkit from '@effect/ai/Toolkit';
+import { describe, expect, it } from '@effect/vitest';
+import * as Effect from 'effect/Effect';
+import * as Layer from 'effect/Layer';
+import * as Schema from 'effect/Schema';
+
+import { OpaqueToolkit } from '@dxos/ai';
+import { Obj, Type } from '@dxos/echo';
+import { TestHelpers } from '@dxos/effect/testing';
+import { log } from '@dxos/log';
+import { Message } from '@dxos/types';
+
+import { ToolExecutionServices } from '../functions';
+import { AssistantTestLayer } from '../testing';
+import { AiRequest } from './request';
+
+// Define a calendar event artifact schema.
+const CalendarEventSchema = Schema.Struct({
+  title: Schema.String,
+  startTime: Schema.String,
+  endTime: Schema.String,
+  description: Schema.String,
+}).pipe(
+  Type.object({
+    typename: 'com.example.type.calendar-event',
+    version: '0.1.0',
+  }),
+);
+
+type CalendarEvent = Schema.Schema.Type<typeof CalendarEventSchema>;
+
+const TestToolkit = Toolkit.make(
+  Tool.make('Calculator', {
+    description: 'Basic calculator tool',
+    parameters: {
+      input: Schema.String.annotations({
+        description: 'The calculation to perform.',
+      }),
+    },
+    success: Schema.Struct({
+      result: Schema.Number,
+    }),
+    failure: Schema.Never,
+  }),
+);
+
+// Tool handlers.
+const toolkitLayer = TestToolkit.toLayer({
+  Calculator: ({ input }) =>
+    Effect.gen(function* () {
+      const result = (() => {
+        // Restrict to basic arithmetic operations for safety.
+        const sanitizedInput = input.replace(/[^0-9+\-*/().\s]/g, '');
+        log.info('calculate', { sanitizedInput });
+
+        // eslint-disable-next-line @typescript-eslint/no-implied-eval
+        return Function(`"use strict"; return (${sanitizedInput})`)();
+      })();
+
+      return { result };
+    }),
+});
+
+const TestLayer = Layer.empty.pipe(
+  Layer.provideMerge(ToolExecutionServices),
+  Layer.provideMerge(
+    AssistantTestLayer({
+      types: [CalendarEventSchema],
+      tracing: 'pretty',
+    }),
+  ),
+  Layer.provideMerge(toolkitLayer),
+);
+
+describe('AiRequest', () => {
+  it.effect(
+    'no tools',
+    Effect.fnUntraced(
+      function* (_) {
+        const request = new AiRequest();
+        const response = yield* request.run({
+          prompt: 'Hello world!',
+          history: [],
+        });
+        log.info('response', { response });
+      },
+      Effect.provide(TestLayer),
+      TestHelpers.provideTestContext,
+    ),
+  );
+
+  it.effect(
+    'calculator',
+    Effect.fnUntraced(
+      function* (_) {
+        const request = new AiRequest();
+        const toolkit = yield* OpaqueToolkit.fromContext(TestToolkit);
+        const response = yield* request.run({
+          toolkit,
+          prompt: 'What is 10 + 30?',
+          history: [],
+        });
+        log.info('response', { response });
+      },
+      Effect.provide(TestLayer),
+      TestHelpers.provideTestContext,
+    ),
+  );
+
+  it.effect(
+    'tool schema error',
+    Effect.fnUntraced(
+      function* (_) {
+        const request = new AiRequest();
+        const toolkit = yield* OpaqueToolkit.fromContext(TestToolkit);
+        const response = yield* request.run({
+          toolkit,
+          prompt:
+            'I am testing error handling in tool paramter parsing. I want you to call the calculator tool but omit the input parameter to intentionally differ from the tool schema.',
+          history: [],
+        });
+        log.info('response', { response });
+      },
+      Effect.provide(TestLayer),
+      TestHelpers.provideTestContext,
+    ),
+  );
+
+  it.effect(
+    'summarization',
+    Effect.fnUntraced(
+      function* (_) {
+        const request = new AiRequest({ summarizationThreshold: 0 }); // Force summarization.
+        const response = yield* request.run({
+          prompt: 'What did we talk about?',
+          history: [
+            Obj.make(Message.Message, {
+              created: '2024-01-01T10:00:00Z',
+              sender: { role: 'user' },
+              blocks: [{ _tag: 'text', text: 'How many apples are in the basket?' }],
+            }),
+            Obj.make(Message.Message, {
+              created: '2024-01-01T11:00:00Z',
+              sender: { role: 'assistant' },
+              blocks: [{ _tag: 'text', text: 'There are 10 apples in the basket.' }],
+            }),
+          ],
+        });
+        expect(response).toBeDefined();
+      },
+      Effect.provide(TestLayer),
+      TestHelpers.provideTestContext,
+    ),
+  );
+});
+
+// Travel to rome, florence, livorno, siena, madrid for conferences
+
+const _CALENDAR_EVENTS: CalendarEvent[] = [
+  Obj.make(CalendarEventSchema, {
+    title: 'Exploring Ancient Ruins in Rome',
+    startTime: '2024-01-01T10:00:00Z',
+    endTime: '2024-01-01T11:00:00Z',
+    description: 'Tech conference at the historic Colosseum with networking opportunities',
+  }),
+  Obj.make(CalendarEventSchema, {
+    title: 'Renaissance Tech Summit in Florence',
+    startTime: '2024-01-01T11:00:00Z',
+    endTime: '2024-01-01T12:00:00Z',
+    description: 'Discussing AI innovations surrounded by Renaissance art',
+  }),
+  Obj.make(CalendarEventSchema, {
+    title: 'Travel to Livorno',
+    startTime: '2024-01-01T12:00:00Z',
+    endTime: '2024-01-01T13:00:00Z',
+    description: 'Travel to Livorno',
+  }),
+  Obj.make(CalendarEventSchema, {
+    title: 'Travel to Siena',
+    startTime: '2024-01-01T13:00:00Z',
+    endTime: '2024-01-01T14:00:00Z',
+    description: 'Travel to Siena',
+  }),
+  Obj.make(CalendarEventSchema, {
+    title: 'Travel to Madrid',
+    startTime: '2024-01-01T14:00:00Z',
+    endTime: '2024-01-01T15:00:00Z',
+    description: 'Travel to Madrid',
+  }),
+];

package/src/session/request.ts

@@ -0,0 +1,344 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import type * as AiError from '@effect/ai/AiError';
+import * as LanguageModel from '@effect/ai/LanguageModel';
+import * as Array from 'effect/Array';
+import * as Chunk from 'effect/Chunk';
+import * as Effect from 'effect/Effect';
+import { pipe } from 'effect/Function';
+import * as Option from 'effect/Option';
+import * as Stream from 'effect/Stream';
+
+import {
+  AiParser,
+  AiPreprocessor,
+  AiSummarizer,
+  type AiToolNotFoundError,
+  callTool,
+  type OpaqueToolkit,
+  type PromptPreprocessingError,
+  type ToolExecutionService,
+  type ToolResolverService,
+  withoutToolCallParising,
+} from '@dxos/ai';
+import { type Blueprint } from '@dxos/blueprints';
+import { Database, Obj } from '@dxos/echo';
+import { Trace } from '@dxos/functions';
+import { log } from '@dxos/log';
+import { Operation, OperationRegistry } from '@dxos/operation';
+import { ContentBlock, Message } from '@dxos/types';
+
+import { type AiAssistantError } from '../errors';
+import { CompleteBlock, PartialBlock } from '../tracing';
+import { formatSystemPrompt, formatUserPrompt } from './format';
+import { GenerationObserver } from './observer';
+
+export type AiRequestRunError = AiError.AiError | PromptPreprocessingError | AiToolNotFoundError | AiAssistantError;
+
+export type AiRequestRunRequirements =
+  | LanguageModel.LanguageModel
+  | ToolExecutionService
+  | ToolResolverService
+  | Database.Service
+  | Operation.Service
+  | OperationRegistry.Service
+  | Trace.TraceService;
+
+export type AiRequestOptions = {
+  /**
+   * Summarize before executing the prompt if the existing history exceeds this threshold.
+   */
+  summarizationThreshold?: number;
+
+  // TODO(dmaretskyi): Plan to phase out in favor of TracingService and the return type being a stream.
+  observer?: GenerationObserver;
+  /**
+   * Callback for when a message is received from the user, model, or tool.
+   * This is useful for streaming the output to a queue.
+   */
+  onOutput?: (message: Message.Message) => Effect.Effect<void, never, never>;
+};
+
+export type AiRequestRunProps<R = never> = {
+  prompt: string;
+  // TODO(wittjosiah): Rename to systemPrompt.
+  system?: string;
+  history?: Message.Message[];
+  objects?: Obj.Unknown[];
+  blueprints?: readonly Blueprint.Blueprint[];
+  toolkit?: OpaqueToolkit.OpaqueToolkit<R>;
+};
+
+export type AiRequestBeginProps = {
+  prompt: string;
+  system?: string;
+  history?: Message.Message[];
+  objects?: Obj.Unknown[];
+  blueprints?: readonly Blueprint.Blueprint[];
+};
+
+export type AiRequestTurnProps<R = never> = {
+  system: string;
+  toolkit?: OpaqueToolkit.OpaqueToolkit<R>;
+};
+
+export type AiRequestTurnResult = {
+  messages: Message.Message[];
+  done: boolean;
+};
+
+/**
+ * Contains message history, tools, current context.
+ * Current context means the state of the app, time of day, and other contextual information.
+ * It makes requests to the model, its a state machine.
+ * It keeps track of the current goal.
+ * It manages the context window.
+ * Tracks the success criteria of reaching the goal, exposing metrics (stretch).
+ * Could be run locally in the app or remotely.
+ * Could be personal or shared.
+ */
+export class AiRequest {
+  /** Prevents concurrent execution of session. */
+  private readonly _semaphore = Effect.runSync(Effect.makeSemaphore(1));
+
+  private readonly _observer: GenerationObserver;
+  private readonly _onOutput: (message: Message.Message) => Effect.Effect<void, never, never>;
+
+  /** Message history from queue. */
+  // TODO(burdon): Evolve into supporting a git-like graph of messages.
+  private _history: Message.Message[] = [];
+
+  /** Pending messages for this session (incl. the current prompt). */
+  private _pending: Message.Message[] = [];
+
+  private _started = 0;
+  private _ended = 0;
+  private _toolCalls = 0;
+
+  constructor(private readonly _options: AiRequestOptions = {}) {
+    this._observer = _options.observer ?? GenerationObserver.noop();
+    this._onOutput = _options.onOutput ?? (() => Effect.void);
+  }
+
+  get duration(): number {
+    return this._ended - this._started;
+  }
+
+  get toolCalls(): number {
+    return this._toolCalls;
+  }
+
+  get pending(): readonly Message.Message[] {
+    return this._pending;
+  }
+
+  private _submitMessage = (message: Message.Message): Effect.Effect<Message.Message, never, Trace.TraceService> =>
+    Effect.gen(this, function* () {
+      this._pending.push(message);
+      yield* this._observer.onMessage(message);
+      for (const block of message.blocks) {
+        yield* Trace.write(CompleteBlock, {
+          messageId: message.id,
+          role: message.sender.role!,
+          block,
+        });
+      }
+      yield* this._onOutput(message);
+      return message;
+    });
+
+  getToolCalls = () =>
+    pipe(
+      [...this._history, ...this._pending],
+      Array.reverse,
+      Array.takeWhile((_) => _.sender.role === 'assistant'),
+      Array.flatMap((_) => _.blocks.filter(ContentBlock.is('toolCall')).map((block) => ({ block, message: _ }))),
+      Array.filter((_) => !_.block.providerExecuted),
+      Array.reverse,
+    );
+
+  /**
+   * Initialize a session: set up history, perform summarization if needed, and submit the user prompt.
+   * Must be called before `runTurn()`.
+   */
+  begin = ({
+    prompt,
+    system,
+    history = [],
+    blueprints = [],
+    objects = [],
+  }: AiRequestBeginProps): Effect.Effect<void, AiRequestRunError, AiRequestRunRequirements> =>
+    Effect.gen(this, function* () {
+      this._started = Date.now();
+      this._history = [...history];
+      this._pending = [];
+
+      const systemPrompt = yield* formatSystemPrompt({ system, blueprints, objects }).pipe(Effect.orDie);
+
+      if (this._options.summarizationThreshold !== undefined) {
+        const tokenCount = yield* AiPreprocessor.estimateTokens(
+          yield* AiPreprocessor.preprocessPrompt([...this._history], {
+            system: systemPrompt,
+          }),
+        );
+        if (tokenCount > this._options.summarizationThreshold) {
+          const summary = yield* AiSummarizer.summarize([...this._history]);
+          yield* this._submitMessage(summary);
+        }
+      }
+
+      yield* this._submitMessage(yield* formatUserPrompt({ prompt, history }));
+    }).pipe(Effect.withSpan('AiRequest.begin'));
+
+  /**
+   * Execute a single turn: one LLM generation followed by tool execution.
+   * The toolkit and system prompt can be updated between turns to reflect context changes (e.g. dynamically enabled blueprints).
+   */
+  runAgentTurn = <const R = never>({
+    system,
+    toolkit: opaqueToolkit,
+  }: AiRequestTurnProps<R>): Effect.Effect<AiRequestTurnResult, AiRequestRunError, AiRequestRunRequirements | R> =>
+    Effect.gen(this, function* () {
+      log.info('request', {
+        system: { snippet: createSnippet(system), length: system.length },
+        pending: this._pending.length,
+        history: this._history.length,
+      });
+
+      const prompt = yield* AiPreprocessor.preprocessPrompt([...this._history, ...this._pending], {
+        system,
+        cacheControl: 'ephemeral',
+      });
+
+      const toolkit = opaqueToolkit ? yield* opaqueToolkit.handlers : undefined;
+
+      const observer = this._observer;
+      let currentMessageId: Obj.ID | null = null;
+
+      const messages = yield* LanguageModel.streamText({
+        prompt,
+        toolkit,
+        disableToolCallResolution: true,
+      }).pipe(
+        withoutToolCallParising,
+        AiParser.parseResponse({
+          emitPartial: true,
+          onBegin: () => observer.onBegin(),
+          onBlock: (block) => observer.onBlock(block),
+          onPart: (part) => observer.onPart(part as any),
+          onEnd: (summary) => observer.onEnd(summary),
+        }),
+        Stream.mapEffect(
+          (block) =>
+            Effect.gen(this, function* () {
+              if (block.pending) {
+                currentMessageId ??= Obj.ID.random();
+                log('emit ephemeral message', { id: currentMessageId, type: block._tag });
+                yield* Trace.write(PartialBlock, {
+                  messageId: currentMessageId,
+                  role: 'assistant',
+                  block,
+                });
+                return Option.none();
+              } else {
+                currentMessageId ??= Obj.ID.random();
+                const id = currentMessageId;
+                currentMessageId = null;
+                log('emit complete message', { id, type: block._tag });
+                const message = Obj.make(Message.Message, {
+                  id,
+                  created: new Date().toISOString(),
+                  sender: { role: 'assistant' },
+                  blocks: [block],
+                });
+                return Option.some(yield* this._submitMessage(message));
+              }
+            }),
+          { concurrency: 1, unordered: false },
+        ),
+        Stream.filterMap((_) => _),
+        Stream.runCollect,
+        Effect.map(Chunk.toArray),
+      );
+      log.info('messages', { messages });
+
+      const toolCalls = this.getToolCalls();
+
+      if (toolCalls.length === 0) {
+        this._ended = Date.now();
+        return { messages, done: true };
+      } else if (!toolkit) {
+        throw new Error('No toolkit provided');
+      }
+
+      return { messages, done: false };
+    }).pipe(Effect.withSpan('AiRequest.runAgentTurn'));
+
+  runTools = <const R = never>({
+    toolkit: opaqueToolkit,
+  }: {
+    toolkit?: OpaqueToolkit.OpaqueToolkit<R>;
+  }): Effect.Effect<void, AiRequestRunError, AiRequestRunRequirements | R> =>
+    Effect.gen(this, function* () {
+      const toolkit = opaqueToolkit ? yield* opaqueToolkit.handlers : undefined;
+      const toolCalls = this.getToolCalls();
+      const toolResults = yield* Effect.forEach(toolCalls, ({ block, message }) => {
+        if (!toolkit) {
+          throw new Error('No toolkit provided');
+        }
+        return callTool(toolkit, block);
+      });
+
+      yield* this._submitMessage(
+        Obj.make(Message.Message, {
+          created: new Date().toISOString(),
+          sender: { role: 'tool' },
+          blocks: toolResults,
+        }),
+      );
+
+      this._toolCalls += toolResults.length;
+    }).pipe(Effect.withSpan('AiRequest.runTools'));
+
+  /**
+   * Run a full conversation turn loop. Equivalent to calling `begin()` then `runTurn()` in a loop.
+   */
+  run = <const R = never>({
+    prompt,
+    system: systemTemplate,
+    history = [],
+    objects = [],
+    blueprints = [],
+    toolkit,
+  }: AiRequestRunProps<R>): Effect.Effect<Message.Message[], AiRequestRunError, AiRequestRunRequirements | R> =>
+    Effect.gen(this, function* () {
+      yield* this.begin({ prompt, system: systemTemplate, history, objects, blueprints });
+
+      const system = yield* formatSystemPrompt({ system: systemTemplate, blueprints, objects }).pipe(Effect.orDie);
+
+      do {
+        const { done } = yield* this.runAgentTurn({ system, toolkit });
+        if (done) {
+          break;
+        }
+        yield* this.runTools({ toolkit });
+      } while (true);
+
+      log('done', { pending: this._pending.length, duration: this.duration, tools: this._toolCalls });
+      return this._pending;
+    }).pipe(this._semaphore.withPermits(1), Effect.withSpan('AiRequest.run'));
+}
+
+const createSnippet = (text: string, len = 32) =>
+  text.length <= len * 2 ? text : [text.slice(0, len), '...', text.slice(-len)].join('');
+
+// TODO(dmaretskyi): Extract as a general util.
+const logDump = (message: string, data: unknown) => {
+  // eslint-disable-next-line @typescript-eslint/no-require-imports
+  const { writeFileSync } = require('node:fs');
+  const path = `/tmp/log-data-${Date.now()}.json`;
+  writeFileSync(path, JSON.stringify(data, null, 2));
+  log.info(message, { path });
+};

package/src/session/session-ollama.test.ts

@@ -0,0 +1,81 @@
+//
+// Copyright 2026 DXOS.org
+//
+
+import { describe, it } from '@effect/vitest';
+import * as Effect from 'effect/Effect';
+import * as Layer from 'effect/Layer';
+
+import { OpaqueToolkit } from '@dxos/ai';
+import { CalculatorLayer, CalculatorToolkit } from '@dxos/ai/testing';
+import { TestHelpers } from '@dxos/effect/testing';
+import { dbg, log } from '@dxos/log';
+import { ContentBlock } from '@dxos/types';
+
+import { ToolExecutionServices } from '../functions';
+import { AssistantTestLayer } from '../testing';
+import { AiRequest } from './request';
+
+/**
+ * Exercises the full AiRequest tool-call loop against a local Ollama instance.
+ *
+ * Prerequisites:
+ * - `ollama serve` running on `localhost:11434` with the `gpt-oss:20b` model pulled.
+ *
+ * Run with:
+ * ```bash
+ * DX_TEST_TAGS=llm pnpm --filter @dxos/assistant exec vitest run session.ollama
+ * ```
+ */
+
+const TestLayer = Layer.empty.pipe(
+  Layer.provideMerge(ToolExecutionServices),
+  Layer.provideMerge(
+    AssistantTestLayer({
+      tracing: 'pretty',
+      aiServicePreset: 'ollama',
+      model: 'gpt-oss:20b',
+      disableLlmMemoization: true,
+    }),
+  ),
+  Layer.provideMerge(CalculatorLayer),
+);
+
+describe('AiRequest (ollama gpt-oss:20b)', () => {
+  it.effect(
+    'calculator tool-call loop',
+    Effect.fn(
+      function* ({ expect }) {
+        const request = new AiRequest();
+        const toolkit = yield* OpaqueToolkit.fromContext(CalculatorToolkit);
+
+        const messages = yield* request.run({
+          toolkit,
+          prompt: 'What is six times seven? Use the Calculator tool and reply with only the number.',
+          history: [],
+        });
+
+        dbg(messages);
+
+        const toolCalls = messages.flatMap((m) => m.blocks).filter(ContentBlock.is('toolCall'));
+        const toolResults = messages.flatMap((m) => m.blocks).filter(ContentBlock.is('toolResult'));
+
+        log.info('request result', {
+          messageCount: messages.length,
+          toolCalls: toolCalls.length,
+          toolResults: toolResults.length,
+          requestToolCalls: request.toolCalls,
+        });
+
+        expect(toolCalls.length).toBeGreaterThan(0);
+        expect(toolCalls.some((call) => call.name === 'Calculator')).toBe(true);
+        expect(toolResults.length).toBeGreaterThan(0);
+        expect(toolResults.some((result) => String(result.result).includes('42'))).toBe(true);
+      },
+      Effect.provide(TestLayer),
+      TestHelpers.provideTestContext,
+      TestHelpers.taggedTest('llm'),
+    ),
+    { timeout: 120_000 },
+  );
+});

package/src/session/toolkit.ts

@@ -0,0 +1,47 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import * as Toolkit from '@effect/ai/Toolkit';
+import * as Effect from 'effect/Effect';
+import * as Layer from 'effect/Layer';
+
+import { type AiToolNotFoundError, OpaqueToolkit, ToolExecutionService, ToolResolverService } from '@dxos/ai';
+import { type Blueprint } from '@dxos/blueprints';
+import { isTruthy } from '@dxos/util';
+
+export type CreateToolkitProps = {
+  toolkit?: OpaqueToolkit.Any;
+  blueprints?: readonly Blueprint.Blueprint[];
+  /**
+   * Self-contained with handlers toolkits.
+   */
+  opaqueToolkits?: readonly OpaqueToolkit.Any[];
+};
+
+/**
+ * Build a combined toolkit from the blueprint tools and the provided toolkit.
+ */
+export const createToolkit = ({
+  toolkit: toolkitProp,
+  blueprints = [],
+  opaqueToolkits = [],
+}: CreateToolkitProps): Effect.Effect<
+  OpaqueToolkit.OpaqueToolkit,
+  AiToolNotFoundError,
+  ToolResolverService | ToolExecutionService
+> =>
+  Effect.gen(function* () {
+    const blueprintToolkit = yield* ToolResolverService.resolveToolkit(blueprints.flatMap(({ tools }) => tools));
+    const blueprintToolHandler = yield* blueprintToolkit.toContext(ToolExecutionService.handlersFor(blueprintToolkit));
+    const opaqueToolkit = OpaqueToolkit.merge(...opaqueToolkits);
+
+    const toolkitDefs = [toolkitProp?.toolkit, blueprintToolkit, opaqueToolkit.toolkit].filter(isTruthy);
+    const mergedToolkit = Toolkit.merge(...toolkitDefs);
+    const combinedHandlerLayer = Layer.mergeAll(
+      Layer.succeedContext(blueprintToolHandler),
+      toolkitProp?.layer ?? OpaqueToolkit.empty.layer,
+      opaqueToolkit.layer,
+    );
+    return OpaqueToolkit.make(mergedToolkit, combinedHandlerLayer as any) as OpaqueToolkit.OpaqueToolkit;
+  }) as Effect.Effect<OpaqueToolkit.OpaqueToolkit, AiToolNotFoundError, ToolResolverService | ToolExecutionService>;