@desplega.ai/agent-swarm 1.20.0 → 1.49.0

package/src/tests/scheduler-backoff.test.ts

@@ -0,0 +1,166 @@
+import { afterAll, beforeAll, describe, expect, test } from "bun:test";
+import { unlink } from "node:fs/promises";
+import {
+  closeDb,
+  createScheduledTask,
+  getScheduledTaskById,
+  initDb,
+  updateScheduledTask,
+} from "../be/db";
+
+const TEST_DB_PATH = "./test-scheduler-backoff.sqlite";
+
+describe("scheduler exponential backoff", () => {
+  beforeAll(() => {
+    initDb(TEST_DB_PATH);
+  });
+
+  afterAll(async () => {
+    closeDb();
+    try {
+      await unlink(TEST_DB_PATH);
+      await unlink(`${TEST_DB_PATH}-wal`);
+      await unlink(`${TEST_DB_PATH}-shm`);
+    } catch {
+      // ignore
+    }
+  });
+
+  test("new scheduled tasks have consecutiveErrors = 0", () => {
+    const schedule = createScheduledTask({
+      name: "backoff-test-1",
+      taskTemplate: "Test task",
+      intervalMs: 60000,
+    });
+
+    expect(schedule.consecutiveErrors).toBe(0);
+    expect(schedule.lastErrorAt).toBeUndefined();
+    expect(schedule.lastErrorMessage).toBeUndefined();
+  });
+
+  test("updateScheduledTask can set error tracking fields", () => {
+    const schedule = createScheduledTask({
+      name: "backoff-test-2",
+      taskTemplate: "Test task",
+      intervalMs: 60000,
+    });
+
+    const now = new Date().toISOString();
+    const updated = updateScheduledTask(schedule.id, {
+      consecutiveErrors: 3,
+      lastErrorAt: now,
+      lastErrorMessage: "Connection refused",
+    });
+
+    expect(updated).not.toBeNull();
+    expect(updated!.consecutiveErrors).toBe(3);
+    expect(updated!.lastErrorAt).toBe(now);
+    expect(updated!.lastErrorMessage).toBe("Connection refused");
+  });
+
+  test("error tracking can be reset to 0 on success", () => {
+    const schedule = createScheduledTask({
+      name: "backoff-test-3",
+      taskTemplate: "Test task",
+      intervalMs: 60000,
+    });
+
+    // Simulate errors
+    updateScheduledTask(schedule.id, {
+      consecutiveErrors: 4,
+      lastErrorAt: new Date().toISOString(),
+      lastErrorMessage: "Some error",
+    });
+
+    // Simulate successful execution — reset errors
+    const updated = updateScheduledTask(schedule.id, {
+      consecutiveErrors: 0,
+      lastErrorAt: null,
+      lastErrorMessage: null,
+    });
+
+    expect(updated!.consecutiveErrors).toBe(0);
+    expect(updated!.lastErrorAt).toBeUndefined();
+    expect(updated!.lastErrorMessage).toBeUndefined();
+  });
+
+  test("schedule can be auto-disabled via enabled = false", () => {
+    const schedule = createScheduledTask({
+      name: "backoff-test-4",
+      taskTemplate: "Test task",
+      intervalMs: 60000,
+    });
+
+    expect(schedule.enabled).toBe(true);
+
+    // Simulate auto-disable after MAX_CONSECUTIVE_ERRORS
+    const updated = updateScheduledTask(schedule.id, {
+      consecutiveErrors: 5,
+      lastErrorAt: new Date().toISOString(),
+      lastErrorMessage: "Repeated failure",
+      enabled: false,
+    });
+
+    expect(updated!.enabled).toBe(false);
+    expect(updated!.consecutiveErrors).toBe(5);
+  });
+
+  test("nextRunAt can be pushed forward for backoff", () => {
+    const schedule = createScheduledTask({
+      name: "backoff-test-5",
+      taskTemplate: "Test task",
+      intervalMs: 60000,
+    });
+
+    const now = new Date();
+    const backoffMs = 300_000; // 5 minutes
+    const backoffTime = new Date(now.getTime() + backoffMs).toISOString();
+
+    const updated = updateScheduledTask(schedule.id, {
+      consecutiveErrors: 2,
+      lastErrorAt: now.toISOString(),
+      lastErrorMessage: "Timeout",
+      nextRunAt: backoffTime,
+    });
+
+    expect(updated!.nextRunAt).toBe(backoffTime);
+    expect(updated!.consecutiveErrors).toBe(2);
+  });
+
+  test("error message is truncated to 500 chars", () => {
+    const schedule = createScheduledTask({
+      name: "backoff-test-6",
+      taskTemplate: "Test task",
+      intervalMs: 60000,
+    });
+
+    const longMessage = "x".repeat(1000);
+    const updated = updateScheduledTask(schedule.id, {
+      consecutiveErrors: 1,
+      lastErrorAt: new Date().toISOString(),
+      lastErrorMessage: longMessage.slice(0, 500),
+    });
+
+    expect(updated!.lastErrorMessage!.length).toBe(500);
+  });
+
+  test("consecutiveErrors persists across reads", () => {
+    const schedule = createScheduledTask({
+      name: "backoff-test-7",
+      taskTemplate: "Test task",
+      intervalMs: 60000,
+    });
+
+    updateScheduledTask(schedule.id, {
+      consecutiveErrors: 3,
+      lastErrorAt: new Date().toISOString(),
+      lastErrorMessage: "Error 3",
+    });
+
+    // Read back from DB
+    const reloaded = getScheduledTaskById(schedule.id);
+    expect(reloaded).not.toBeNull();
+    expect(reloaded!.consecutiveErrors).toBe(3);
+    expect(reloaded!.lastErrorMessage).toBe("Error 3");
+  });
+});

package/src/tests/self-improvement.test.ts

@@ -0,0 +1,540 @@
+import { afterAll, beforeAll, describe, expect, test } from "bun:test";
+import { unlink } from "node:fs/promises";
+import {
+  closeDb,
+  completeTask,
+  createAgent,
+  createEpic,
+  createMemory,
+  createTaskExtended,
+  failTask,
+  getAgentById,
+  initDb,
+  searchMemoriesByVector,
+  updateMemoryEmbedding,
+} from "../be/db";
+import { serializeEmbedding } from "../be/embedding";
+import { getBasePrompt } from "../prompts/base-prompt";
+
+const TEST_DB_PATH = "./test-self-improvement.sqlite";
+
+describe("Self-Improvement Mechanisms", () => {
+  const leadId = "aaaa0000-0000-4000-8000-000000000001";
+  const workerId = "bbbb0000-0000-4000-8000-000000000002";
+  const otherWorkerId = "cccc0000-0000-4000-8000-000000000003";
+
+  beforeAll(async () => {
+    for (const suffix of ["", "-wal", "-shm"]) {
+      try {
+        await unlink(TEST_DB_PATH + suffix);
+      } catch {
+        // File doesn't exist
+      }
+    }
+
+    initDb(TEST_DB_PATH);
+
+    createAgent({ id: leadId, name: "Test Lead", isLead: true, status: "idle" });
+    createAgent({ id: workerId, name: "Test Worker", isLead: false, status: "idle" });
+    createAgent({ id: otherWorkerId, name: "Other Worker", isLead: false, status: "idle" });
+  });
+
+  afterAll(async () => {
+    closeDb();
+    for (const suffix of ["", "-wal", "-shm"]) {
+      try {
+        await unlink(TEST_DB_PATH + suffix);
+      } catch {
+        // File doesn't exist
+      }
+    }
+  });
+
+  // ==========================================================================
+  // P2: store-progress memory indexing for completed and failed tasks
+  // ==========================================================================
+
+  describe("store-progress memory indexing", () => {
+    test("completed task creates agent-scoped memory with output", () => {
+      const task = createTaskExtended("Test task for completion", {
+        agentId: workerId,
+        source: "mcp",
+        priority: 50,
+      });
+
+      const output = "Successfully completed the task with great results";
+      completeTask(task.id, output);
+
+      // Simulate what store-progress does: create memory for completed task
+      const taskContent = `Task: ${task.task}\n\nOutput:\n${output}`;
+      const memory = createMemory({
+        agentId: workerId,
+        content: taskContent,
+        name: `Task: ${task.task.slice(0, 80)}`,
+        scope: "agent",
+        source: "task_completion",
+        sourceTaskId: task.id,
+      });
+
+      expect(memory.scope).toBe("agent");
+      expect(memory.source).toBe("task_completion");
+      expect(memory.sourceTaskId).toBe(task.id);
+      expect(memory.content).toContain("Output:");
+      expect(memory.content).toContain(output);
+      expect(memory.content).not.toContain("undefined");
+    });
+
+    test("completed task with undefined output uses fallback", () => {
+      const task = createTaskExtended("Task without output", {
+        agentId: workerId,
+        source: "mcp",
+        priority: 50,
+      });
+
+      const output: string | undefined = undefined;
+      completeTask(task.id, output);
+
+      // Simulate store-progress logic with undefined guard
+      const taskContent = `Task: ${task.task}\n\nOutput:\n${output || "(no output)"}`;
+
+      expect(taskContent).toContain("(no output)");
+      expect(taskContent).not.toContain("undefined");
+    });
+
+    test("failed task creates memory with failure reason", () => {
+      const task = createTaskExtended("Task that will fail", {
+        agentId: workerId,
+        source: "mcp",
+        priority: 50,
+      });
+
+      const failureReason = "Could not connect to external API";
+      failTask(task.id, failureReason);
+
+      // Simulate store-progress failed task memory creation
+      const taskContent = `Task: ${task.task}\n\nFailure reason:\n${failureReason}\n\nThis task failed. Learn from this to avoid repeating the mistake.`;
+      const memory = createMemory({
+        agentId: workerId,
+        content: taskContent,
+        name: `Task: ${task.task.slice(0, 80)}`,
+        scope: "agent",
+        source: "task_completion",
+        sourceTaskId: task.id,
+      });
+
+      expect(memory.source).toBe("task_completion");
+      expect(memory.content).toContain("Failure reason:");
+      expect(memory.content).toContain(failureReason);
+      expect(memory.content).toContain("Learn from this");
+    });
+
+    test("failed task with undefined failureReason uses fallback", () => {
+      const failureReason: string | undefined = undefined;
+
+      // Simulate store-progress logic with undefined guard
+      const taskContent = `Task: Some task\n\nFailure reason:\n${failureReason || "No reason provided"}\n\nThis task failed.`;
+
+      expect(taskContent).toContain("No reason provided");
+      expect(taskContent).not.toContain("undefined");
+    });
+
+    test("short task content is skipped (< 30 chars)", () => {
+      // Simulate the length check in store-progress
+      const shortContent = "Task: X\n\nOutput:\n";
+      expect(shortContent.length).toBeLessThan(30);
+      // In store-progress, this would return early without creating memory
+    });
+  });
+
+  // ==========================================================================
+  // P3: Swarm memory auto-promotion
+  // ==========================================================================
+
+  describe("swarm memory auto-promotion", () => {
+    test("research task type promotes to swarm scope", () => {
+      const task = createTaskExtended("Research best practices for testing", {
+        agentId: workerId,
+        source: "mcp",
+        priority: 50,
+        taskType: "research",
+      });
+
+      completeTask(task.id, "Found several useful patterns");
+
+      // Simulate the shouldShareWithSwarm logic
+      const shouldShareWithSwarm =
+        task.taskType === "research" ||
+        task.tags?.includes("knowledge") ||
+        task.tags?.includes("shared");
+
+      expect(shouldShareWithSwarm).toBe(true);
+
+      // Verify swarm memory can be created
+      const swarmMemory = createMemory({
+        agentId: workerId,
+        scope: "swarm",
+        name: `Shared: ${task.task.slice(0, 80)}`,
+        content: `Task completed by agent ${workerId}:\n\nTask: ${task.task}\n\nOutput:\nFound several useful patterns`,
+        source: "task_completion",
+        sourceTaskId: task.id,
+      });
+
+      expect(swarmMemory.scope).toBe("swarm");
+      expect(swarmMemory.source).toBe("task_completion");
+    });
+
+    test("knowledge-tagged task promotes to swarm scope", () => {
+      const task = createTaskExtended("Document API conventions", {
+        agentId: workerId,
+        source: "mcp",
+        priority: 50,
+        tags: ["knowledge"],
+      });
+
+      const shouldShareWithSwarm =
+        task.taskType === "research" ||
+        task.tags?.includes("knowledge") ||
+        task.tags?.includes("shared");
+
+      expect(shouldShareWithSwarm).toBe(true);
+    });
+
+    test("shared-tagged task promotes to swarm scope", () => {
+      const task = createTaskExtended("Build shared utility", {
+        agentId: workerId,
+        source: "mcp",
+        priority: 50,
+        tags: ["shared", "utility"],
+      });
+
+      const shouldShareWithSwarm =
+        task.taskType === "research" ||
+        task.tags?.includes("knowledge") ||
+        task.tags?.includes("shared");
+
+      expect(shouldShareWithSwarm).toBe(true);
+    });
+
+    test("epic-linked task promotes to swarm scope", () => {
+      const epic = createEpic({
+        name: "Test Epic for Promotion",
+        goal: "Test that epic-linked tasks promote to swarm scope",
+        createdByAgentId: leadId,
+      });
+
+      const task = createTaskExtended("Implement feature X for epic", {
+        agentId: workerId,
+        source: "mcp",
+        priority: 50,
+        taskType: "implementation",
+        epicId: epic.id,
+      });
+
+      const shouldShareWithSwarm =
+        task.taskType === "research" ||
+        task.tags?.includes("knowledge") ||
+        task.tags?.includes("shared") ||
+        task.epicId != null;
+
+      expect(shouldShareWithSwarm).toBe(true);
+    });
+
+    test("regular task without epicId does NOT promote to swarm scope", () => {
+      const task = createTaskExtended("Fix a typo", {
+        agentId: workerId,
+        source: "mcp",
+        priority: 50,
+        taskType: "quick-fix",
+        tags: ["bug-fix"],
+      });
+
+      const shouldShareWithSwarm =
+        task.taskType === "research" ||
+        task.tags?.includes("knowledge") ||
+        task.tags?.includes("shared") ||
+        task.epicId != null;
+
+      expect(shouldShareWithSwarm).toBe(false);
+    });
+
+    test("failed task does NOT promote to swarm scope", () => {
+      const task = createTaskExtended("Research something", {
+        agentId: workerId,
+        source: "mcp",
+        priority: 50,
+        taskType: "research",
+      });
+
+      const status = "failed";
+      // In store-progress, shouldShareWithSwarm only fires for status === "completed"
+      const shouldShareWithSwarm =
+        status === "completed" &&
+        (task.taskType === "research" ||
+          task.tags?.includes("knowledge") ||
+          task.tags?.includes("shared") ||
+          task.epicId != null);
+
+      expect(shouldShareWithSwarm).toBe(false);
+    });
+  });
+
+  // ==========================================================================
+  // P6: inject-learning tool
+  // ==========================================================================
+
+  describe("inject-learning tool logic", () => {
+    test("lead agent can inject learning into worker memory (swarm-scoped)", () => {
+      const callerAgent = getAgentById(leadId);
+      expect(callerAgent).not.toBeNull();
+      expect(callerAgent!.isLead).toBe(true);
+
+      const category = "best-practice";
+      const learning = "Always run lint before committing";
+      const content = `[Lead Feedback — ${category}]\n\n${learning}`;
+
+      const memory = createMemory({
+        agentId: workerId,
+        scope: "swarm",
+        name: `Lead feedback: ${category} — ${learning.slice(0, 60)}`,
+        content,
+        source: "manual",
+      });
+
+      expect(memory.agentId).toBe(workerId);
+      expect(memory.scope).toBe("swarm");
+      expect(memory.content).toContain("[Lead Feedback — best-practice]");
+      expect(memory.content).toContain(learning);
+    });
+
+    test("non-lead agent is rejected", () => {
+      const callerAgent = getAgentById(workerId);
+      expect(callerAgent).not.toBeNull();
+      expect(callerAgent!.isLead).toBe(false);
+
+      // In the tool handler, this check prevents non-leads from injecting
+      const canInject = callerAgent!.isLead;
+      expect(canInject).toBe(false);
+    });
+
+    test("injected learning is visible to target worker in memory search", () => {
+      // Create memory with embedding for searchability
+      const content = "[Lead Feedback — mistake-pattern]\n\nNever force-push to main branch";
+      const memory = createMemory({
+        agentId: workerId,
+        scope: "agent",
+        name: "Lead feedback: mistake-pattern — Never force-push to main branch",
+        content,
+        source: "manual",
+      });
+
+      const embedding = new Float32Array([0.7, 0.3, 0.0]);
+      updateMemoryEmbedding(memory.id, serializeEmbedding(embedding));
+
+      // Worker can find it via search
+      const results = searchMemoriesByVector(new Float32Array([0.7, 0.3, 0.0]), workerId, {
+        isLead: false,
+        scope: "agent",
+      });
+
+      const found = results.find((r) => r.id === memory.id);
+      expect(found).toBeDefined();
+      expect(found!.content).toContain("Never force-push");
+    });
+
+    test("injected learning is NOT visible to other workers", () => {
+      const content = "[Lead Feedback — preference]\n\nUse bun instead of npm";
+      const memory = createMemory({
+        agentId: workerId,
+        scope: "agent",
+        name: "Lead feedback: preference — Use bun instead of npm",
+        content,
+        source: "manual",
+      });
+
+      const embedding = new Float32Array([0.2, 0.8, 0.1]);
+      updateMemoryEmbedding(memory.id, serializeEmbedding(embedding));
+
+      // Other worker should NOT see it
+      const results = searchMemoriesByVector(new Float32Array([0.2, 0.8, 0.1]), otherWorkerId, {
+        isLead: false,
+        scope: "agent",
+      });
+
+      const found = results.find((r) => r.id === memory.id);
+      expect(found).toBeUndefined();
+    });
+
+    test("learning categories are properly formatted", () => {
+      const categories = ["mistake-pattern", "best-practice", "codebase-knowledge", "preference"];
+
+      for (const category of categories) {
+        const content = `[Lead Feedback — ${category}]\n\nSome learning`;
+        expect(content).toContain(`[Lead Feedback — ${category}]`);
+      }
+    });
+  });
+
+  // ==========================================================================
+  // P7: Memory search agent ID security
+  // ==========================================================================
+
+  describe("memory search agent ID security", () => {
+    test("agent can only search their own memories (not others)", () => {
+      // Create private memories for worker and other worker
+      const workerMemory = createMemory({
+        agentId: workerId,
+        scope: "agent",
+        name: "Worker Private Secret",
+        content: "My secret API key pattern",
+        source: "manual",
+      });
+      updateMemoryEmbedding(workerMemory.id, serializeEmbedding(new Float32Array([0.5, 0.5, 0.0])));
+
+      const otherMemory = createMemory({
+        agentId: otherWorkerId,
+        scope: "agent",
+        name: "Other Worker Secret",
+        content: "Other agent's private data",
+        source: "manual",
+      });
+      updateMemoryEmbedding(otherMemory.id, serializeEmbedding(new Float32Array([0.5, 0.5, 0.0])));
+
+      // Worker searching with their own ID should see their memory but not other's
+      const workerResults = searchMemoriesByVector(new Float32Array([0.5, 0.5, 0.0]), workerId, {
+        isLead: false,
+        scope: "all",
+      });
+
+      const workerNames = workerResults.map((r) => r.name);
+      expect(workerNames).toContain("Worker Private Secret");
+      expect(workerNames).not.toContain("Other Worker Secret");
+    });
+
+    test("missing agent ID should be rejected (endpoint logic)", () => {
+      // Simulate the endpoint logic: searchAgentId = myAgentId (from header only)
+      const myAgentId: string | undefined = undefined;
+      const searchAgentId = myAgentId; // No fallback to body.agentId
+
+      // The endpoint requires both query and searchAgentId
+      const isValid = !!searchAgentId;
+      expect(isValid).toBe(false);
+    });
+
+    test("agent ID from header is used, not from body", () => {
+      // Simulate the fixed logic
+      const headerAgentId = workerId;
+      const _bodyAgentId = otherWorkerId; // attacker trying to access other agent's memories
+
+      // Fixed code: searchAgentId = myAgentId (from header only)
+      const searchAgentId = headerAgentId; // NOT: headerAgentId || bodyAgentId
+
+      expect(searchAgentId).toBe(workerId);
+      expect(searchAgentId).not.toBe(otherWorkerId);
+    });
+  });
+
+  // ==========================================================================
+  // P2: Self-awareness in base prompt
+  // ==========================================================================
+
+  describe("base prompt self-awareness", () => {
+    test("base prompt includes 'How You Are Built' section", () => {
+      const prompt = getBasePrompt({
+        role: "worker",
+        agentId: workerId,
+        swarmUrl: "test.example.com",
+      });
+
+      expect(prompt).toContain("### How You Are Built");
+      expect(prompt).toContain("desplega-ai/agent-swarm");
+      expect(prompt).toContain("src/commands/runner.ts");
+      expect(prompt).toContain("src/hooks/hook.ts");
+    });
+
+    test("self-awareness section includes change proposal instructions", () => {
+      const prompt = getBasePrompt({
+        role: "worker",
+        agentId: workerId,
+        swarmUrl: "test.example.com",
+      });
+
+      expect(prompt).toContain("Proposing changes");
+      expect(prompt).toContain("@tarasyarema");
+    });
+
+    test("self-awareness is included for both worker and lead roles", () => {
+      const workerPrompt = getBasePrompt({
+        role: "worker",
+        agentId: workerId,
+        swarmUrl: "test.example.com",
+      });
+
+      const leadPrompt = getBasePrompt({
+        role: "lead",
+        agentId: leadId,
+        swarmUrl: "test.example.com",
+      });
+
+      expect(workerPrompt).toContain("### How You Are Built");
+      expect(leadPrompt).toContain("### How You Are Built");
+    });
+  });
+
+  // ==========================================================================
+  // P4: Session summary "no significant learnings" filter
+  // ==========================================================================
+
+  describe("session summary filtering", () => {
+    test("'no significant learnings' response is filtered out", () => {
+      const summary = "No significant learnings.";
+
+      const shouldIndex =
+        summary &&
+        summary.length > 20 &&
+        !summary.trim().toLowerCase().includes("no significant learnings");
+
+      expect(shouldIndex).toBe(false);
+    });
+
+    test("summary with actual learnings passes filter", () => {
+      const summary =
+        "- Discovered that the API requires Bearer prefix on auth headers\n- Found that bun test runs faster with --bail flag";
+
+      const shouldIndex =
+        summary &&
+        summary.length > 20 &&
+        !summary.trim().toLowerCase().includes("no significant learnings");
+
+      expect(shouldIndex).toBe(true);
+    });
+
+    test("very short summary is filtered out", () => {
+      const summary = "Done.";
+
+      const shouldIndex =
+        summary &&
+        summary.length > 20 &&
+        !summary.trim().toLowerCase().includes("no significant learnings");
+
+      expect(shouldIndex).toBe(false);
+    });
+
+    test("case-insensitive matching for 'no significant learnings'", () => {
+      const variants = [
+        "No Significant Learnings.",
+        "NO SIGNIFICANT LEARNINGS",
+        "no significant learnings",
+        "  No significant learnings.  ",
+      ];
+
+      for (const summary of variants) {
+        const shouldIndex =
+          summary &&
+          summary.length > 20 &&
+          !summary.trim().toLowerCase().includes("no significant learnings");
+
+        expect(shouldIndex).toBe(false);
+      }
+    });
+  });
+});