@dxos/assistant-toolkit 0.10.0 → 0.11.1

package/src/skills/memory/skill.test.ts

@@ -1,138 +0,0 @@
-//
-// Copyright 2026 DXOS.org
-//
-
-import { describe, expect, it } from '@effect/vitest';
-import * as Effect from 'effect/Effect';
-import * as Layer from 'effect/Layer';
-
-import { OpaqueToolkit } from '@dxos/ai';
-import { Skill } from '@dxos/compute';
-import { Database, Feed, Filter, Obj, Query } from '@dxos/echo';
-import { TestHelpers } from '@dxos/effect/testing';
-import { AgentService } from '@dxos/functions-runtime';
-import { AssistantTestLayer } from '@dxos/functions-runtime/testing';
-import { EntityId } from '@dxos/keys';
-
-import { Memory } from '../../types/Memory';
-import { WebSearchToolkit } from '../websearch';
-import { MemoryHandlers } from './operations';
-import MemorySkill from './skill';
-
-EntityId.dangerouslyDisableRandomness();
-
-const TestLayer = AssistantTestLayer({
-  operationHandlers: MemoryHandlers,
-  types: [Memory, Skill.Skill, Feed.Feed],
-  skills: [MemorySkill.make()],
-  tracing: 'pretty',
-});
-
-const TestLayerWithWebSearch = AssistantTestLayer({
-  operationHandlers: MemoryHandlers,
-  toolkits: [OpaqueToolkit.make(WebSearchToolkit, Layer.empty)],
-  types: [Memory, Skill.Skill, Feed.Feed],
-  skills: [MemorySkill.make()],
-  tracing: 'pretty',
-});
-
-describe('Memory Skill', () => {
-  it.effect(
-    'save: saves a memory',
-    Effect.fnUntraced(
-      function* (_) {
-        const agent = yield* AgentService.createSession({
-          skills: [MemorySkill.make()],
-        });
-        yield* agent.submitPrompt('Remember that my favorite programming language is TypeScript.');
-        yield* agent.waitForCompletion();
-        const memories = yield* Database.query(Query.select(Filter.type(Memory))).run;
-        expect(memories.length).toBeGreaterThanOrEqual(1);
-      },
-      Effect.provide(TestLayer),
-      TestHelpers.provideTestContext,
-    ),
-    { timeout: 60_000 },
-  );
-
-  it.effect(
-    'query: searches memories by text',
-    Effect.fnUntraced(
-      function* (_) {
-        yield* Database.add(
-          Obj.make(Memory, {
-            title: 'Favorite color',
-            content: 'The user prefers blue.',
-          }),
-        );
-        yield* Database.add(
-          Obj.make(Memory, {
-            title: 'Meeting notes',
-            content: 'Discussed project timeline with Alice.',
-          }),
-        );
-        const agent = yield* AgentService.createSession({
-          skills: [MemorySkill.make()],
-        });
-        yield* agent.submitPrompt('Search your memories for anything about colors.');
-        yield* agent.waitForCompletion();
-      },
-      Effect.provide(TestLayer),
-      TestHelpers.provideTestContext,
-    ),
-    { timeout: 60_000 },
-  );
-
-  it.effect(
-    'delete: removes a memory',
-    Effect.fnUntraced(
-      function* (_) {
-        yield* Database.add(
-          Obj.make(Memory, {
-            title: 'Outdated fact',
-            content: 'The sky is green.',
-          }),
-        );
-        const agent = yield* AgentService.createSession({
-          skills: [MemorySkill.make()],
-        });
-        yield* agent.submitPrompt('Delete the memory about "Outdated fact".');
-        yield* agent.waitForCompletion();
-        const memories = yield* Database.query(Query.select(Filter.type(Memory))).run;
-        const found = memories.find((memory) => memory.title === 'Outdated fact');
-        expect(found).toBeUndefined();
-      },
-      Effect.provide(TestLayer),
-      TestHelpers.provideTestContext,
-    ),
-    { timeout: 60_000 },
-  );
-
-  // TODO(dmaretskyi): Flaky. The model does not reliably call save-memory after a provider-executed web search.
-  it.effect.skip(
-    'natural: saves memories from a conversation with web search',
-    Effect.fnUntraced(
-      function* (_) {
-        const agent = yield* AgentService.createSession({
-          skills: [MemorySkill.make()],
-        });
-        yield* agent.submitPrompt(
-          "I'm going to LA next week. Find me some good hotels and remember the recommendations.",
-        );
-        yield* agent.waitForCompletion();
-        const memories = yield* Database.query(Query.select(Filter.type(Memory))).run;
-        expect(memories.length).toBeGreaterThanOrEqual(1);
-        const hasLAMemory = memories.some(
-          (memory) =>
-            memory.title.toLowerCase().includes('la') ||
-            memory.title.toLowerCase().includes('los angeles') ||
-            memory.content.toLowerCase().includes('hotel'),
-        );
-        expect(hasLAMemory).toBe(true);
-      },
-      Effect.provide(TestLayerWithWebSearch),
-      TestHelpers.provideTestContext,
-    ),
-    { timeout: 120_000 },
-  );
-});

package/src/skills/planning/skill.conversations.json

@@ -1 +0,0 @@
-{"conversations":[{"parameters":{"model":"dxn:com.anthropic.model.claudeSonnet46","stream":false,"tools":[]},"prompt":{"content":[{"role":"system","content":"You decide whether an agent should stop or continue working on its plan, given the conversation\nso far and the agent's remaining plan tasks.\nReply with exactly one word: \"stop\" or \"continue\". Do not use tools. Do not add explanation.","options":{}},{"role":"user","content":"The agent is about to finish, but its plan still has incomplete tasks:\n\n<plan>\n- **TODO**: Buy eggs <!-- id=task-1 -->\n- **IN-PROGRESS**: Crack eggs <!-- id=task-2 -->\n</plan>\n\nShould the agent STOP now (no more work needed) or CONTINUE working on the plan?\nReply with exactly one word: \"stop\" or \"continue\".","options":{}}]},"response":[{"type":"response-metadata","id":"msg_015FvDvL8TEw6YqGBNLs8yRs","modelId":"claude-sonnet-4-5-20250929","timestamp":"1970-01-01T00:00:00.000Z","metadata":{}},{"type":"text","text":"continue","metadata":{}},{"type":"finish","reason":"stop","usage":{"inputTokens":151,"outputTokens":4,"totalTokens":155,"cachedInputTokens":0},"metadata":{"anthropic":{"usage":{"cache_creation":{"ephemeral_1h_input_tokens":0,"ephemeral_5m_input_tokens":0},"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"input_tokens":151,"output_tokens":4,"server_tool_use":null,"service_tier":"standard"}}}}]},{"parameters":{"model":"dxn:com.anthropic.model.claude-sonnet-4-6.default","stream":false,"tools":[]},"prompt":{"content":[{"role":"system","content":"You decide whether an agent should stop or continue working on its plan, given the conversation\nso far and the agent's remaining plan tasks.\nReply with exactly one word: \"stop\" or \"continue\". Do not use tools. Do not add explanation.","options":{}},{"role":"user","content":"The agent is about to finish, but its plan still has incomplete tasks:\n\n<plan>\n- **TODO**: Buy eggs <!-- id=task-1 -->\n- **IN-PROGRESS**: Crack eggs <!-- id=task-2 -->\n</plan>\n\nShould the agent STOP now (no more work needed) or CONTINUE working on the plan?\nReply with exactly one word: \"stop\" or \"continue\".","options":{}}]},"response":[{"type":"response-metadata","id":"msg_01NSkcUieedKCREEajLSJWYs","modelId":"claude-sonnet-4-6","timestamp":"1970-01-01T00:00:00.000Z","metadata":{}},{"type":"text","text":"continue","metadata":{}},{"type":"finish","reason":"stop","usage":{"inputTokens":152,"outputTokens":4,"totalTokens":156,"cachedInputTokens":0},"metadata":{"anthropic":{"usage":{"cache_creation":{"ephemeral_1h_input_tokens":0,"ephemeral_5m_input_tokens":0},"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"input_tokens":152,"output_tokens":4,"server_tool_use":null,"service_tier":"standard"}}}}]}]}

package/src/skills/planning/skill.test.ts

@@ -1,123 +0,0 @@
-//
-// Copyright 2026 DXOS.org
-//
-
-import { describe, it, test } from '@effect/vitest';
-import * as Effect from 'effect/Effect';
-
-import { Operation, Process, Skill } from '@dxos/compute';
-import { ProcessManager } from '@dxos/compute-runtime';
-import { getSession } from '@dxos/compute/AgentService';
-import { Database, Feed, Obj } from '@dxos/echo';
-import { TestHelpers } from '@dxos/effect/testing';
-import { AssistantTestLayer } from '@dxos/functions-runtime/testing';
-import { invariant } from '@dxos/invariant';
-import { EntityId } from '@dxos/keys';
-
-import { Agent, Chat, Plan } from '../../types';
-import { PlanningHandlers, PlanningOperations } from './operations';
-import PlanningSkill from './skill';
-
-EntityId.dangerouslyDisableRandomness();
-
-const TestLayer = AssistantTestLayer({
-  operationHandlers: PlanningHandlers,
-  types: [Agent.Agent, Plan.Plan, Chat.Chat, Chat.CompanionTo, Skill.Skill, Feed.Feed],
-  skills: [PlanningSkill.make()],
-});
-
-describe('Planning skill', () => {
-  // The hook fires the plan-reminder operation only while the plan has open tasks; cover that
-  // predicate directly so the branch is verified without an agent turn.
-  describe('hasIncompleteTasks', () => {
-    test('is true while any task is todo or in-progress', ({ expect }) => {
-      expect(Plan.hasIncompleteTasks(makePlan(['todo']))).toBe(true);
-      expect(Plan.hasIncompleteTasks(makePlan(['done', 'in-progress']))).toBe(true);
-    });
-
-    test('is false when every task is done', ({ expect }) => {
-      expect(Plan.hasIncompleteTasks(makePlan(['done', 'done']))).toBe(false);
-      expect(Plan.hasIncompleteTasks(makePlan([]))).toBe(false);
-    });
-  });
-
-  // The end-request hook reads the plan (Tier A), runs an ephemeral stop/continue check over the
-  // conversation history (LanguageModel, memoized), and on "continue" reaches the live host through
-  // HarnessService Tier B (like the alarm skill's set-alarm) to enqueue a reminder. Spawning a
-  // host on the agent's chat feed and invoking the operation against that conversation dispatches
-  // over the process RPC loopback; the host staying live proves the read, the check, and the enqueue
-  // resolved end-to-end without failing the process.
-  describe('plan-reminder hook (end-request, Tier B)', () => {
-    it.scoped(
-      'enqueues a continuation reminder when the plan has incomplete tasks',
-      Effect.fnUntraced(
-        function* ({ expect }) {
-          const { conversation, host } = yield* setupAgentWithLiveHost([
-            { id: Plan.TaskId.make('task-1'), title: 'Buy eggs', status: 'todo' },
-            { id: Plan.TaskId.make('task-2'), title: 'Crack eggs', status: 'in-progress' },
-          ]);
-
-          yield* Operation.invoke(PlanningOperations.PlanReminder, {}).pipe(
-            Effect.provide(Operation.withInvocationOptions({ conversation })),
-          );
-
-          expect(host.status.state).not.toBe(Process.State.FAILED);
-          expect(host.status.state).not.toBe(Process.State.TERMINATED);
-        },
-        Effect.provide(TestLayer),
-        TestHelpers.provideTestContext,
-      ),
-    );
-
-    it.scoped(
-      'is a no-op when the plan is complete',
-      Effect.fnUntraced(
-        function* ({ expect }) {
-          const { conversation, host } = yield* setupAgentWithLiveHost([
-            { id: Plan.TaskId.make('task-1'), title: 'Buy eggs', status: 'done' },
-          ]);
-
-          yield* Operation.invoke(PlanningOperations.PlanReminder, {}).pipe(
-            Effect.provide(Operation.withInvocationOptions({ conversation })),
-          );
-
-          expect(host.status.state).not.toBe(Process.State.FAILED);
-          expect(host.status.state).not.toBe(Process.State.TERMINATED);
-        },
-        Effect.provide(TestLayer),
-        TestHelpers.provideTestContext,
-      ),
-    );
-  });
-});
-
-const makePlan = (statuses: readonly Plan.TaskStatus[]): Plan.Plan =>
-  Plan.makePlan({ tasks: statuses.map((status, index) => ({ title: `Task ${index}`, status })) });
-
-/**
- * Creates an agent (with the planning skill and a seeded plan) bound to its chat feed, then
- * spawns a live host process on that feed so HarnessService Tier B is reachable for the conversation.
- */
-const setupAgentWithLiveHost = (tasks: readonly Plan.Task[]) =>
-  Effect.gen(function* () {
-    const agent = yield* Agent.makeInitialized({ name: 'Planner', instructions: 'Test.' }, PlanningSkill.make());
-    const chat = yield* Database.load(agent.chat!);
-    const plan = yield* Chat.ensurePlan(chat);
-    Obj.update(plan, (plan) => {
-      plan.tasks.push(...tasks);
-    });
-    yield* Database.flush();
-
-    const chatFeed = agent.chat?.target?.feed?.target;
-    invariant(chatFeed, 'Agent chat feed not found.');
-
-    // Spawns (and stamps) the host process that owns the conversation's harness.
-    yield* getSession(chatFeed);
-
-    const conversation = Obj.getURI(chatFeed);
-    const processManager = yield* ProcessManager.Service;
-    const [host] = yield* processManager.list({ target: conversation });
-    invariant(host, 'Host process not found.');
-
-    return { conversation, host };
-  });

package/src/types/Plan.ts

@@ -1,123 +0,0 @@
-//
-// Copyright 2026 DXOS.org
-//
-
-// @import-as-namespace
-
-import * as Schema from 'effect/Schema';
-
-import { Process } from '@dxos/compute';
-import { Annotation, DXN, Obj, Type } from '@dxos/echo';
-
-export const TaskId = Schema.String.pipe(Schema.brand('@dxos/assistant-toolkit/TaskId'));
-export type TaskId = Schema.Schema.Type<typeof TaskId>;
-
-export const Task = Schema.Struct({
-  /**
-   * Short task ID unique within the plan.
-   */
-  id: TaskId,
-
-  title: Schema.String.annotations({
-    description: 'Task title and description.',
-  }),
-
-  status: Schema.Literal('todo', 'in-progress', 'done', 'failed'),
-
-  /**
-   * Whether this task was handed to a sub-agent via the delegation tool. Only delegated tasks are
-   * picked up by the supervisor's reconcile loop, so a task created via ordinary planning
-   * (`update-tasks`) is not double-spawned as a sub-agent.
-   */
-  delegated: Schema.optional(Schema.Boolean),
-
-  /**
-   * Sub-agent process id; set when the supervisor spawns a delegated task.
-   */
-  agentPid: Schema.optional(Process.ID).annotations({
-    description: 'Sub-agent process id for delegated tasks.',
-  }),
-
-  /**
-   * Parent task ID.
-   */
-  parent: Schema.optional(TaskId).annotations({
-    description: 'Parent task ID.',
-  }),
-});
-
-export type Task = Schema.Schema.Type<typeof Task>;
-
-/** Lifecycle status of a {@link Task}. */
-export type TaskStatus = Task['status'];
-
-/**
- * Short tag label for a delegated sub-agent process id (e.g. `agent-a1b2`).
- */
-export const formatAgentPidTag = (pid: Process.ID): string => `agent-${String(pid).replaceAll('-', '').slice(0, 4)}`;
-
-/**
- * Hierarchical collection of tasks for humans and agents to track progress.
- */
-export class Plan extends Type.makeObject<Plan>(DXN.make('org.dxos.type.plan', '0.1.0'))(
-  Schema.Struct({
-    tasks: Schema.Array(Task),
-  }).pipe(Annotation.HiddenAnnotation.set(true)),
-) {}
-export const generateTaskId = (plan: Plan): TaskId => {
-  const existingIds = plan.tasks
-    .map((task) => {
-      const [num, letter] = task.id.split('-');
-      return parseInt(num);
-    })
-    .filter((_) => !isNaN(_));
-  let newId;
-  if (existingIds.length === 0) {
-    newId = 1;
-  } else {
-    newId = Math.max(...existingIds) + 1;
-  }
-  // Add random suffix to avoid collisions.
-  return TaskId.make(`${newId}-${crypto.randomUUID().slice(0, 2)}`);
-};
-
-/**
- * Add new tasks to a plan, generating IDs for new tasks.
- * Use inside an `Obj.update` callback.
- */
-export const addTasks = (
-  plan: Obj.Mutable<Plan>,
-  tasks: (Pick<Task, 'title'> & Partial<Pick<Task, 'status' | 'parent' | 'delegated'>>)[],
-) => {
-  for (const task of tasks) {
-    const taskId = generateTaskId(plan);
-    plan.tasks.push({ id: taskId, ...task, status: task.status ?? 'todo' });
-  }
-  return plan;
-};
-
-interface MakePlanProps {
-  tasks: (Pick<Task, 'title'> & Partial<Pick<Task, 'status'>>)[];
-}
-
-export const makePlan = (props: MakePlanProps): Plan => {
-  const plan = Obj.make(Plan, { tasks: [] });
-  Obj.update(plan, (plan) => {
-    addTasks(plan, props.tasks);
-  });
-  return plan;
-};
-
-/**
- * Formats plan to markdown format.
- */
-export const formatPlan = (plan: Plan): string => {
-  return plan.tasks
-    .map((task) => `- **${task.status?.toLocaleUpperCase()}**: ${task.title ?? 'No title'} <!-- id=${task.id} -->`)
-    .join('\n');
-};
-
-/**
- * True when the plan has any task that is not yet `done`.
- */
-export const hasIncompleteTasks = (plan: Plan): boolean => plan.tasks.some((task) => task.status !== 'done');