opencode-swarm-plugin 0.43.0 → 0.44.1

package/evals/fixtures/coordinator-sessions.ts

@@ -1,328 +0,0 @@
-/**
- * Coordinator Session Test Fixtures
- *
- * Synthetic coordinator sessions for testing coordinator-discipline scorers.
- * Each fixture demonstrates good or bad coordinator behavior.
- */
-
-import type { CoordinatorSession } from "../../src/eval-capture.js";
-
-/**
- * PERFECT COORDINATOR
- *
- * - No violations (no direct edits, tests, or reservations)
- * - 100% spawn efficiency (3/3 workers spawned)
- * - 100% review thoroughness (all workers reviewed)
- * - Fast time to first spawn (30s)
- */
-export const perfectCoordinator: CoordinatorSession = {
-  session_id: "test-session-perfect",
-  epic_id: "test-epic-perfect",
-  start_time: "2025-01-01T10:00:00.000Z",
-  end_time: "2025-01-01T10:30:00.000Z",
-  events: [
-    // 1. Decomposition complete
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:00:00.000Z",
-      event_type: "DECISION",
-      decision_type: "decomposition_complete",
-      payload: { subtask_count: 3 },
-    },
-    // 2. First spawn (30s after decomp)
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:00:30.000Z",
-      event_type: "DECISION",
-      decision_type: "worker_spawned",
-      payload: { worker: "BlueLake", bead_id: "test-epic-perfect.1" },
-    },
-    // 3. Second spawn
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:01:00.000Z",
-      event_type: "DECISION",
-      decision_type: "worker_spawned",
-      payload: { worker: "GreenMountain", bead_id: "test-epic-perfect.2" },
-    },
-    // 4. Third spawn
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:01:30.000Z",
-      event_type: "DECISION",
-      decision_type: "worker_spawned",
-      payload: { worker: "RedForest", bead_id: "test-epic-perfect.3" },
-    },
-    // 5. First worker completes
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:10:00.000Z",
-      event_type: "OUTCOME",
-      outcome_type: "subtask_success",
-      payload: { bead_id: "test-epic-perfect.1", worker: "BlueLake" },
-    },
-    // 6. First review
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:11:00.000Z",
-      event_type: "DECISION",
-      decision_type: "review_completed",
-      payload: {
-        bead_id: "test-epic-perfect.1",
-        approved: true,
-        issues: [],
-      },
-    },
-    // 7. Second worker completes
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:15:00.000Z",
-      event_type: "OUTCOME",
-      outcome_type: "subtask_success",
-      payload: { bead_id: "test-epic-perfect.2", worker: "GreenMountain" },
-    },
-    // 8. Second review
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:16:00.000Z",
-      event_type: "DECISION",
-      decision_type: "review_completed",
-      payload: {
-        bead_id: "test-epic-perfect.2",
-        approved: true,
-        issues: [],
-      },
-    },
-    // 9. Third worker completes
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:20:00.000Z",
-      event_type: "OUTCOME",
-      outcome_type: "subtask_success",
-      payload: { bead_id: "test-epic-perfect.3", worker: "RedForest" },
-    },
-    // 10. Third review
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:21:00.000Z",
-      event_type: "DECISION",
-      decision_type: "review_completed",
-      payload: {
-        bead_id: "test-epic-perfect.3",
-        approved: true,
-        issues: [],
-      },
-    },
-    // 11. Epic complete
-    {
-      session_id: "test-session-perfect",
-      epic_id: "test-epic-perfect",
-      timestamp: "2025-01-01T10:30:00.000Z",
-      event_type: "OUTCOME",
-      outcome_type: "epic_complete",
-      payload: { epic_id: "test-epic-perfect", total_subtasks: 3 },
-    },
-  ],
-};
-
-/**
- * BAD COORDINATOR - Multiple Violations
- *
- * - 3 violations (edited file, ran tests, reserved files)
- * - 33% spawn efficiency (only 1/3 workers spawned)
- * - 0% review thoroughness (no reviews)
- * - Slow time to first spawn (10 minutes)
- */
-export const badCoordinator: CoordinatorSession = {
-  session_id: "test-session-bad",
-  epic_id: "test-epic-bad",
-  start_time: "2025-01-01T10:00:00.000Z",
-  end_time: "2025-01-01T11:00:00.000Z",
-  events: [
-    // 1. Decomposition complete
-    {
-      session_id: "test-session-bad",
-      epic_id: "test-epic-bad",
-      timestamp: "2025-01-01T10:00:00.000Z",
-      event_type: "DECISION",
-      decision_type: "decomposition_complete",
-      payload: { subtask_count: 3 },
-    },
-    // 2. VIOLATION: Coordinator edited file directly
-    {
-      session_id: "test-session-bad",
-      epic_id: "test-epic-bad",
-      timestamp: "2025-01-01T10:01:00.000Z",
-      event_type: "VIOLATION",
-      violation_type: "coordinator_edited_file",
-      payload: { file: "src/auth.ts", reason: "should spawn worker instead" },
-    },
-    // 3. VIOLATION: Coordinator ran tests
-    {
-      session_id: "test-session-bad",
-      epic_id: "test-epic-bad",
-      timestamp: "2025-01-01T10:02:00.000Z",
-      event_type: "VIOLATION",
-      violation_type: "coordinator_ran_tests",
-      payload: { command: "bun test", reason: "workers do verification" },
-    },
-    // 4. VIOLATION: Coordinator reserved files
-    {
-      session_id: "test-session-bad",
-      epic_id: "test-epic-bad",
-      timestamp: "2025-01-01T10:03:00.000Z",
-      event_type: "VIOLATION",
-      violation_type: "coordinator_reserved_files",
-      payload: { paths: ["src/**"], reason: "only workers reserve" },
-    },
-    // 5. First spawn (10 minutes after decomp - way too slow)
-    {
-      session_id: "test-session-bad",
-      epic_id: "test-epic-bad",
-      timestamp: "2025-01-01T10:10:00.000Z",
-      event_type: "DECISION",
-      decision_type: "worker_spawned",
-      payload: { worker: "BlueLake", bead_id: "test-epic-bad.1" },
-    },
-    // 6. Worker completes (but no review!)
-    {
-      session_id: "test-session-bad",
-      epic_id: "test-epic-bad",
-      timestamp: "2025-01-01T10:20:00.000Z",
-      event_type: "OUTCOME",
-      outcome_type: "subtask_success",
-      payload: { bead_id: "test-epic-bad.1", worker: "BlueLake" },
-    },
-    // 7. VIOLATION: No worker spawned for subtask 2
-    {
-      session_id: "test-session-bad",
-      epic_id: "test-epic-bad",
-      timestamp: "2025-01-01T10:30:00.000Z",
-      event_type: "VIOLATION",
-      violation_type: "no_worker_spawned",
-      payload: { bead_id: "test-epic-bad.2", reason: "coordinator did work directly" },
-    },
-    // 8. VIOLATION: No worker spawned for subtask 3
-    {
-      session_id: "test-session-bad",
-      epic_id: "test-epic-bad",
-      timestamp: "2025-01-01T10:40:00.000Z",
-      event_type: "VIOLATION",
-      violation_type: "no_worker_spawned",
-      payload: { bead_id: "test-epic-bad.3", reason: "coordinator did work directly" },
-    },
-  ],
-};
-
-/**
- * DECENT COORDINATOR - Some Issues
- *
- * - 1 violation (ran tests once)
- * - 100% spawn efficiency (2/2 workers spawned)
- * - 50% review thoroughness (reviewed only 1/2)
- * - Good time to first spawn (45s)
- */
-export const decentCoordinator: CoordinatorSession = {
-  session_id: "test-session-decent",
-  epic_id: "test-epic-decent",
-  start_time: "2025-01-01T10:00:00.000Z",
-  end_time: "2025-01-01T10:25:00.000Z",
-  events: [
-    // 1. Decomposition complete
-    {
-      session_id: "test-session-decent",
-      epic_id: "test-epic-decent",
-      timestamp: "2025-01-01T10:00:00.000Z",
-      event_type: "DECISION",
-      decision_type: "decomposition_complete",
-      payload: { subtask_count: 2 },
-    },
-    // 2. First spawn (45s - acceptable)
-    {
-      session_id: "test-session-decent",
-      epic_id: "test-epic-decent",
-      timestamp: "2025-01-01T10:00:45.000Z",
-      event_type: "DECISION",
-      decision_type: "worker_spawned",
-      payload: { worker: "BlueLake", bead_id: "test-epic-decent.1" },
-    },
-    // 3. Second spawn
-    {
-      session_id: "test-session-decent",
-      epic_id: "test-epic-decent",
-      timestamp: "2025-01-01T10:01:00.000Z",
-      event_type: "DECISION",
-      decision_type: "worker_spawned",
-      payload: { worker: "GreenMountain", bead_id: "test-epic-decent.2" },
-    },
-    // 4. First worker completes
-    {
-      session_id: "test-session-decent",
-      epic_id: "test-epic-decent",
-      timestamp: "2025-01-01T10:10:00.000Z",
-      event_type: "OUTCOME",
-      outcome_type: "subtask_success",
-      payload: { bead_id: "test-epic-decent.1", worker: "BlueLake" },
-    },
-    // 5. First review
-    {
-      session_id: "test-session-decent",
-      epic_id: "test-epic-decent",
-      timestamp: "2025-01-01T10:11:00.000Z",
-      event_type: "DECISION",
-      decision_type: "review_completed",
-      payload: {
-        bead_id: "test-epic-decent.1",
-        approved: true,
-        issues: [],
-      },
-    },
-    // 6. VIOLATION: Ran tests (one slip-up)
-    {
-      session_id: "test-session-decent",
-      epic_id: "test-epic-decent",
-      timestamp: "2025-01-01T10:15:00.000Z",
-      event_type: "VIOLATION",
-      violation_type: "coordinator_ran_tests",
-      payload: { command: "bun test", reason: "should let worker verify" },
-    },
-    // 7. Second worker completes
-    {
-      session_id: "test-session-decent",
-      epic_id: "test-epic-decent",
-      timestamp: "2025-01-01T10:20:00.000Z",
-      event_type: "OUTCOME",
-      outcome_type: "subtask_success",
-      payload: { bead_id: "test-epic-decent.2", worker: "GreenMountain" },
-    },
-    // 8. No review for second worker (50% review rate)
-    // 9. Epic complete
-    {
-      session_id: "test-session-decent",
-      epic_id: "test-epic-decent",
-      timestamp: "2025-01-01T10:25:00.000Z",
-      event_type: "OUTCOME",
-      outcome_type: "epic_complete",
-      payload: { epic_id: "test-epic-decent", total_subtasks: 2 },
-    },
-  ],
-};
-
-/**
- * All test fixtures
- */
-export const coordinatorSessionFixtures = [
-  perfectCoordinator,
-  badCoordinator,
-  decentCoordinator,
-];

package/evals/fixtures/decomposition-cases.ts

@@ -1,105 +0,0 @@
-/**
- * Test cases for swarm task decomposition
- *
- * Each case includes:
- * - input: task description and optional context
- * - expected: validation criteria (min/max subtasks, required files)
- */
-
-export interface DecompositionTestCase {
-  input: {
-    task: string;
-    context?: string;
-  };
-  expected: {
-    minSubtasks: number;
-    maxSubtasks: number;
-    requiredFiles?: string[];
-  };
-}
-
-export const decompositionCases: DecompositionTestCase[] = [
-  {
-    input: {
-      task: "Add user authentication with OAuth",
-      context: "Next.js App Router application with existing user model",
-    },
-    expected: {
-      minSubtasks: 3,
-      maxSubtasks: 6,
-      requiredFiles: [
-        "src/auth/oauth.ts",
-        "src/auth/middleware.ts",
-        "app/api/auth/[...nextauth]/route.ts",
-      ],
-    },
-  },
-  {
-    input: {
-      task: "Implement rate limiting for API endpoints",
-      context: "Express.js API with Redis available",
-    },
-    expected: {
-      minSubtasks: 2,
-      maxSubtasks: 4,
-      requiredFiles: [
-        "src/middleware/rate-limit.ts",
-        "src/utils/redis-client.ts",
-      ],
-    },
-  },
-  {
-    input: {
-      task: "Add TypeScript strict mode to legacy JavaScript project",
-      context: "Large codebase with 50+ JS files, currently untyped",
-    },
-    expected: {
-      minSubtasks: 4,
-      maxSubtasks: 8,
-      requiredFiles: ["tsconfig.json"],
-    },
-  },
-  {
-    input: {
-      task: "Create admin dashboard for user management",
-      context: "React app with existing component library and API client",
-    },
-    expected: {
-      minSubtasks: 4,
-      maxSubtasks: 7,
-      requiredFiles: [
-        "src/pages/admin/Dashboard.tsx",
-        "src/components/admin/UserTable.tsx",
-        "src/api/admin.ts",
-      ],
-    },
-  },
-  {
-    input: {
-      task: "Fix memory leak in long-running background job",
-      context:
-        "Node.js worker that processes queue messages, memory grows over time",
-    },
-    expected: {
-      minSubtasks: 2,
-      maxSubtasks: 4,
-      requiredFiles: ["src/workers/queue-processor.ts"],
-    },
-  },
-  {
-    input: {
-      task: "Implement feature flag system with remote config",
-      context:
-        "Microservices architecture, need runtime toggles without deploys",
-    },
-    expected: {
-      minSubtasks: 3,
-      maxSubtasks: 6,
-      requiredFiles: [
-        "src/feature-flags/client.ts",
-        "src/feature-flags/middleware.ts",
-        "src/feature-flags/types.ts",
-      ],
-    },
-  },
-];

package/evals/lib/compaction-loader.test.ts

@@ -1,248 +0,0 @@
-/**
- * Compaction Data Loader Tests
- *
- * Tests loading COMPACTION events from session JSONL files.
- */
-import { afterAll, beforeAll, describe, expect, test } from "bun:test";
-import * as fs from "node:fs";
-import * as os from "node:os";
-import * as path from "node:path";
-import type { CoordinatorEvent } from "../../src/eval-capture.js";
-import {
-  loadCompactionEvents,
-  loadCompactionSessions,
-} from "./compaction-loader.js";
-
-// Test fixtures directory
-const TEST_SESSION_DIR = path.join(
-  os.tmpdir(),
-  `test-sessions-${Date.now()}`,
-);
-
-/**
- * Create a test session JSONL file
- */
-function createSessionFile(
-  sessionId: string,
-  events: CoordinatorEvent[],
-): void {
-  const sessionPath = path.join(TEST_SESSION_DIR, `${sessionId}.jsonl`);
-  const lines = events.map((e) => JSON.stringify(e)).join("\n");
-  fs.writeFileSync(sessionPath, `${lines}\n`, "utf-8");
-}
-
-beforeAll(() => {
-  // Create test session directory
-  fs.mkdirSync(TEST_SESSION_DIR, { recursive: true });
-
-  // Create test session files with COMPACTION events
-  createSessionFile("session-1", [
-    {
-      session_id: "session-1",
-      epic_id: "epic-1",
-      timestamp: "2025-01-01T10:00:00.000Z",
-      event_type: "DECISION",
-      decision_type: "decomposition_complete",
-      payload: { subtask_count: 3 },
-    },
-    {
-      session_id: "session-1",
-      epic_id: "epic-1",
-      timestamp: "2025-01-01T10:05:00.000Z",
-      event_type: "COMPACTION",
-      compaction_type: "detection_complete",
-      payload: {
-        confidence: "high",
-        context_type: "full",
-        epic_id: "epic-1",
-      },
-    },
-    {
-      session_id: "session-1",
-      epic_id: "epic-1",
-      timestamp: "2025-01-01T10:06:00.000Z",
-      event_type: "COMPACTION",
-      compaction_type: "prompt_generated",
-      payload: {
-        prompt_length: 5000,
-        full_prompt: "You are a coordinator...",
-        context_type: "full",
-      },
-    },
-  ]);
-
-  createSessionFile("session-2", [
-    {
-      session_id: "session-2",
-      epic_id: "epic-2",
-      timestamp: "2025-01-02T10:00:00.000Z",
-      event_type: "COMPACTION",
-      compaction_type: "context_injected",
-      payload: {
-        injection_point: "tool_call",
-        context_length: 3000,
-      },
-    },
-    {
-      session_id: "session-2",
-      epic_id: "epic-2",
-      timestamp: "2025-01-02T10:01:00.000Z",
-      event_type: "COMPACTION",
-      compaction_type: "resumption_started",
-      payload: {
-        epic_id: "epic-2",
-        resumption_type: "coordinator",
-      },
-    },
-  ]);
-
-  // Session with no COMPACTION events
-  createSessionFile("session-3", [
-    {
-      session_id: "session-3",
-      epic_id: "epic-3",
-      timestamp: "2025-01-03T10:00:00.000Z",
-      event_type: "DECISION",
-      decision_type: "worker_spawned",
-      payload: { worker: "BlueLake", bead_id: "epic-3.1" },
-    },
-  ]);
-});
-
-afterAll(() => {
-  // Clean up test session directory
-  if (fs.existsSync(TEST_SESSION_DIR)) {
-    fs.rmSync(TEST_SESSION_DIR, { recursive: true });
-  }
-});
-
-describe("loadCompactionEvents", () => {
-  test("loads all COMPACTION events from session directory", async () => {
-    const events = await loadCompactionEvents(TEST_SESSION_DIR);
-
-    expect(events.length).toBe(4);
-    expect(events.every((e) => e.event_type === "COMPACTION")).toBe(true);
-  });
-
-  test("filters by compaction_type", async () => {
-    const events = await loadCompactionEvents(TEST_SESSION_DIR, {
-      compaction_type: "detection_complete",
-    });
-
-    expect(events.length).toBe(1);
-    expect(events[0].compaction_type).toBe("detection_complete");
-  });
-
-  test("filters by session_ids", async () => {
-    const events = await loadCompactionEvents(TEST_SESSION_DIR, {
-      sessionIds: ["session-1"],
-    });
-
-    expect(events.length).toBe(2);
-    expect(events.every((e) => e.session_id === "session-1")).toBe(true);
-  });
-
-  test("applies limit", async () => {
-    const events = await loadCompactionEvents(TEST_SESSION_DIR, {
-      limit: 2,
-    });
-
-    expect(events.length).toBe(2);
-  });
-
-  test("combines filters", async () => {
-    const events = await loadCompactionEvents(TEST_SESSION_DIR, {
-      compaction_type: "prompt_generated",
-      sessionIds: ["session-1"],
-      limit: 1,
-    });
-
-    expect(events.length).toBe(1);
-    expect(events[0].compaction_type).toBe("prompt_generated");
-    expect(events[0].session_id).toBe("session-1");
-  });
-
-  test("returns empty array for non-existent directory", async () => {
-    const events = await loadCompactionEvents("/non/existent/path");
-
-    expect(events).toEqual([]);
-  });
-
-  test("skips invalid JSONL lines", async () => {
-    // Create session with invalid JSON
-    const invalidPath = path.join(TEST_SESSION_DIR, "session-invalid.jsonl");
-    fs.writeFileSync(
-      invalidPath,
-      'invalid json\n{"session_id": "session-valid", "event_type": "COMPACTION", "compaction_type": "detection_complete", "epic_id": "epic-4", "timestamp": "2025-01-04T10:00:00.000Z", "payload": {}}\n',
-      "utf-8",
-    );
-
-    const events = await loadCompactionEvents(TEST_SESSION_DIR);
-
-    // Should skip invalid line but include valid one
-    expect(events.some((e) => e.session_id === "session-valid")).toBe(true);
-
-    // Clean up
-    fs.unlinkSync(invalidPath);
-  });
-});
-
-describe("loadCompactionSessions", () => {
-  test("groups events by session_id", async () => {
-    const sessions = await loadCompactionSessions(TEST_SESSION_DIR);
-
-    expect(sessions.length).toBe(2); // session-1 and session-2 (session-3 has no COMPACTION events)
-    expect(sessions[0].session_id).toBeDefined();
-    expect(sessions[0].events.length).toBeGreaterThan(0);
-  });
-
-  test("includes session metadata", async () => {
-    const sessions = await loadCompactionSessions(TEST_SESSION_DIR);
-
-    const session1 = sessions.find((s) => s.session_id === "session-1");
-    expect(session1).toBeDefined();
-    if (session1) {
-      expect(session1.epic_id).toBe("epic-1");
-      expect(session1.start_time).toBeDefined();
-      expect(session1.end_time).toBeDefined();
-    }
-  });
-
-  test("filters by compaction_type", async () => {
-    const sessions = await loadCompactionSessions(TEST_SESSION_DIR, {
-      compaction_type: "detection_complete",
-    });
-
-    expect(sessions.length).toBe(1);
-    expect(sessions[0].session_id).toBe("session-1");
-  });
-
-  test("filters by session_ids", async () => {
-    const sessions = await loadCompactionSessions(TEST_SESSION_DIR, {
-      sessionIds: ["session-2"],
-    });
-
-    expect(sessions.length).toBe(1);
-    expect(sessions[0].session_id).toBe("session-2");
-  });
-
-  test("applies limit", async () => {
-    const sessions = await loadCompactionSessions(TEST_SESSION_DIR, {
-      limit: 1,
-    });
-
-    expect(sessions.length).toBe(1);
-  });
-
-  test("returns empty array for non-existent directory", async () => {
-    const sessions = await loadCompactionSessions("/non/existent/path");
-
-    expect(sessions).toEqual([]);
-  });
-
-  test("excludes sessions with no COMPACTION events", async () => {
-    const sessions = await loadCompactionSessions(TEST_SESSION_DIR);
-
-    expect(sessions.every((s) => s.session_id !== "session-3")).toBe(true);
-  });
-});