opencode-swarm-plugin 0.43.0 → 0.44.1

package/evals/fixtures/compaction-cases.ts

@@ -1,350 +0,0 @@
-/**
- * Test cases for compaction hook coordinator resumption
- *
- * Each case simulates a different swarm state and verifies that
- * the compaction hook injects the correct context for resumption.
- */
-
-import type { Cell } from "swarm-mail";
-
-/**
- * Compaction test case structure
- */
-export interface CompactionTestCase {
-  name: string;
-  description: string;
-  /**
-   * Simulated hive state (cells to create)
-   */
-  hiveCells: Array<Omit<Cell, "created_at" | "updated_at" | "closed_at">>;
-  /**
-   * Simulated swarm-mail state
-   */
-  swarmMailState: {
-    agents: number;
-    reservations: number;
-    messages: number;
-  };
-  /**
-   * Expected detection confidence
-   */
-  expected: {
-    confidence: "high" | "medium" | "low" | "none";
-    contextInjected: boolean;
-    contextType: "full" | "fallback" | "none";
-    /**
-     * Patterns that MUST appear in injected context (if injected)
-     */
-    mustContain?: string[];
-    /**
-     * Patterns that MUST NOT appear
-     */
-    mustNotContain?: string[];
-  };
-}
-
-export const compactionCases: CompactionTestCase[] = [
-  // ============================================================================
-  // HIGH CONFIDENCE: Active swarm with in_progress epic
-  // ============================================================================
-  {
-    name: "Active swarm with in_progress epic",
-    description:
-      "Compaction happens mid-swarm with an active epic and subtasks. Should inject full context with specific epic ID.",
-    hiveCells: [
-      {
-        id: "test-project-lf2p4u-epic123",
-        project_key: "/test/project",
-        type: "epic",
-        status: "in_progress",
-        title: "Add user authentication",
-        description: "Implement OAuth with NextAuth.js",
-        priority: 2,
-        parent_id: null,
-        assignee: "coordinator",
-        closed_reason: null,
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-      {
-        id: "test-project-lf2p4u-epic123.1",
-        project_key: "/test/project",
-        type: "task",
-        status: "closed",
-        title: "OAuth provider config",
-        description: "Configure GitHub OAuth provider",
-        priority: 2,
-        parent_id: "test-project-lf2p4u-epic123",
-        assignee: "BlueLake",
-        closed_reason: "Done: configured GitHub provider",
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-      {
-        id: "test-project-lf2p4u-epic123.2",
-        project_key: "/test/project",
-        type: "task",
-        status: "in_progress",
-        title: "Auth middleware",
-        description: "Create middleware for protecting routes",
-        priority: 2,
-        parent_id: "test-project-lf2p4u-epic123",
-        assignee: "RedMountain",
-        closed_reason: null,
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-      {
-        id: "test-project-lf2p4u-epic123.3",
-        project_key: "/test/project",
-        type: "task",
-        status: "open",
-        title: "Session management",
-        description: "Implement session persistence with Redis",
-        priority: 2,
-        parent_id: "test-project-lf2p4u-epic123",
-        assignee: null,
-        closed_reason: null,
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-    ],
-    swarmMailState: {
-      agents: 2,
-      reservations: 1,
-      messages: 5,
-    },
-    expected: {
-      confidence: "high",
-      contextInjected: true,
-      contextType: "full",
-      mustContain: [
-        "SWARM ACTIVE",
-        "COORDINATOR",
-        "swarm_status",
-        "swarmmail_inbox",
-        "Keep Cooking",
-      ],
-      mustNotContain: [
-        "bd-xxx", // Should NOT contain placeholder IDs
-        "Check Your Context", // Should NOT be fallback detection
-      ],
-    },
-  },
-
-  // ============================================================================
-  // MEDIUM CONFIDENCE: Multiple epics, need to identify active one
-  // ============================================================================
-  {
-    name: "Multiple epics with one in_progress",
-    description:
-      "Multiple epics exist, but only one is in_progress. Should detect and inject context for the active one.",
-    hiveCells: [
-      {
-        id: "test-project-lf2p4u-epic100",
-        project_key: "/test/project",
-        type: "epic",
-        status: "closed",
-        title: "Refactor auth system",
-        description: "Old completed epic",
-        priority: 2,
-        parent_id: null,
-        assignee: null,
-        closed_reason: "Done",
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-      {
-        id: "test-project-lf2p4u-epic200",
-        project_key: "/test/project",
-        type: "epic",
-        status: "in_progress",
-        title: "Add rate limiting",
-        description: "Implement Redis-based rate limiting",
-        priority: 2,
-        parent_id: null,
-        assignee: "coordinator",
-        closed_reason: null,
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-      {
-        id: "test-project-lf2p4u-epic200.1",
-        project_key: "/test/project",
-        type: "task",
-        status: "open",
-        title: "Rate limit middleware",
-        description: "Create Express middleware",
-        priority: 2,
-        parent_id: "test-project-lf2p4u-epic200",
-        assignee: null,
-        closed_reason: null,
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-      {
-        id: "test-project-lf2p4u-epic300",
-        project_key: "/test/project",
-        type: "epic",
-        status: "open",
-        title: "Future epic",
-        description: "Not started yet",
-        priority: 1,
-        parent_id: null,
-        assignee: null,
-        closed_reason: null,
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-    ],
-    swarmMailState: {
-      agents: 1,
-      reservations: 0,
-      messages: 2,
-    },
-    expected: {
-      confidence: "medium",
-      contextInjected: true,
-      contextType: "full",
-      mustContain: ["SWARM ACTIVE", "COORDINATOR"],
-      mustNotContain: ["bd-xxx"],
-    },
-  },
-
-  // ============================================================================
-  // LOW CONFIDENCE: Cells exist but no active work
-  // ============================================================================
-  {
-    name: "Cells exist but no active swarm",
-    description:
-      "Hive has some cells but no in_progress work. Should inject fallback detection prompt.",
-    hiveCells: [
-      {
-        id: "test-project-lf2p4u-task001",
-        project_key: "/test/project",
-        type: "task",
-        status: "open",
-        title: "Fix typo in README",
-        description: null,
-        priority: 0,
-        parent_id: null,
-        assignee: null,
-        closed_reason: null,
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: null,
-      },
-    ],
-    swarmMailState: {
-      agents: 0,
-      reservations: 0,
-      messages: 0,
-    },
-    expected: {
-      confidence: "low",
-      contextInjected: true,
-      contextType: "fallback",
-      mustContain: [
-        "Swarm Detection",
-        "Check Your Context",
-        "swarm_decompose",
-        "swarmmail_init",
-      ],
-      mustNotContain: ["SWARM ACTIVE", "COORDINATOR"],
-    },
-  },
-
-  // ============================================================================
-  // NONE: Empty hive, no swarm activity
-  // ============================================================================
-  {
-    name: "Empty hive - no swarm activity",
-    description:
-      "No cells, no swarm-mail activity. Should NOT inject any context.",
-    hiveCells: [],
-    swarmMailState: {
-      agents: 0,
-      reservations: 0,
-      messages: 0,
-    },
-    expected: {
-      confidence: "none",
-      contextInjected: false,
-      contextType: "none",
-      mustContain: [],
-      mustNotContain: ["SWARM", "COORDINATOR", "swarm_status"],
-    },
-  },
-
-  // ============================================================================
-  // EDGE CASE: Blocked epic (should still detect as active swarm)
-  // ============================================================================
-  {
-    name: "Blocked epic with subtasks",
-    description:
-      "Epic is blocked but has in_progress subtasks. Should detect as active swarm.",
-    hiveCells: [
-      {
-        id: "test-project-lf2p4u-epic400",
-        project_key: "/test/project",
-        type: "epic",
-        status: "blocked",
-        title: "Migration to TypeScript",
-        description: "Full codebase migration",
-        priority: 3,
-        parent_id: null,
-        assignee: "coordinator",
-        closed_reason: null,
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-      {
-        id: "test-project-lf2p4u-epic400.1",
-        project_key: "/test/project",
-        type: "task",
-        status: "in_progress",
-        title: "Migrate utils",
-        description: "Convert utils to TypeScript",
-        priority: 2,
-        parent_id: "test-project-lf2p4u-epic400",
-        assignee: "GreenValley",
-        closed_reason: null,
-        deleted_at: null,
-        deleted_by: null,
-        delete_reason: null,
-        created_by: "coordinator",
-      },
-    ],
-    swarmMailState: {
-      agents: 1,
-      reservations: 1,
-      messages: 3,
-    },
-    expected: {
-      confidence: "high",
-      contextInjected: true,
-      contextType: "full",
-      mustContain: ["SWARM ACTIVE", "COORDINATOR"],
-      mustNotContain: ["bd-xxx"],
-    },
-  },
-];

package/evals/fixtures/compaction-prompt-cases.ts

@@ -1,311 +0,0 @@
-/**
- * Test cases for compaction prompt quality evaluation
- *
- * Each case represents a continuation prompt that should be generated
- * after context compaction. Tests validate that prompts have:
- * - Real epic IDs (not placeholders)
- * - Actionable tool calls with specific values
- * - Strong coordinator identity
- * - Explicit forbidden tools list
- * - Correct first tool suggestion
- */
-
-import type { CompactionPrompt } from "../../src/compaction-prompt-scoring.js";
-
-/**
- * Compaction prompt test case structure
- */
-export interface CompactionPromptTestCase {
-	name: string;
-	description: string;
-	/**
-	 * The generated continuation prompt
-	 */
-	prompt: CompactionPrompt;
-	/**
-	 * Expected scoring outcomes
-	 */
-	expected: {
-		/**
-		 * Should have real epic IDs (not placeholders)
-		 */
-		hasRealEpicId: boolean;
-		/**
-		 * Should have actionable tool calls
-		 */
-		isActionable: boolean;
-		/**
-		 * Should have strong coordinator identity
-		 */
-		hasCoordinatorIdentity: boolean;
-		/**
-		 * Should list forbidden tools by name
-		 */
-		listsForbiddenTools: boolean;
-		/**
-		 * First suggested tool should be correct
-		 */
-		hasCorrectFirstTool: boolean;
-	};
-}
-
-export const compactionPromptCases: CompactionPromptTestCase[] = [
-	// ============================================================================
-	// PERFECT PROMPT: All criteria met
-	// ============================================================================
-	{
-		name: "Perfect coordinator resumption prompt",
-		description:
-			"Ideal continuation prompt with all quality criteria met: real IDs, actionable tools, strong identity, forbidden list, correct first tool",
-		prompt: {
-			content: `
-┌─────────────────────────────────────────────────────────────┐
-│                 🐝 COORDINATOR RESUMPTION                   │
-│                   Context Compacted                         │
-└─────────────────────────────────────────────────────────────┘
-
-You are the COORDINATOR of swarm epic mjkweh2p4u5.
-
-## IMMEDIATE ACTIONS (Do These FIRST)
-
-1. swarm_status(epic_id="mjkweh2p4u5", project_key="/Users/joel/Code/myapp")
-2. swarmmail_inbox(limit=5)
-3. Review any completed work
-
-## FORBIDDEN TOOLS (NEVER Use These)
-
-Coordinators do NOT edit code directly. These tools are FORBIDDEN:
-- edit
-- write
-- bash (for file modifications)
-- swarmmail_reserve (only workers reserve)
-- git commit (workers commit)
-
-Use swarm_spawn_subtask to delegate work to workers.
-
-## Your Role
-
-You orchestrate. You do NOT implement. Spawn workers, monitor progress, unblock, ship.
-
-ALWAYS spawn workers for file modifications.
-NEVER edit files yourself.
-NON-NEGOTIABLE: Check status and inbox before making decisions.
-`,
-		},
-		expected: {
-			hasRealEpicId: true,
-			isActionable: true,
-			hasCoordinatorIdentity: true,
-			listsForbiddenTools: true,
-			hasCorrectFirstTool: true,
-		},
-	},
-
-	// ============================================================================
-	// BAD PROMPT: Placeholder epic ID
-	// ============================================================================
-	{
-		name: "Prompt with placeholder epic ID",
-		description:
-			"Contains placeholder <epic-id> instead of real ID - fails specificity check",
-		prompt: {
-			content: `
-## Coordinator Resumption
-
-You are coordinating epic <epic-id>.
-
-Check the status with:
-1. swarm_status(epic_id="<epic-id>", project_key="<path>")
-2. swarmmail_inbox()
-
-Continue orchestrating the swarm.
-`,
-		},
-		expected: {
-			hasRealEpicId: false, // <epic-id> is a placeholder
-			isActionable: false, // Has placeholders in tool calls
-			hasCoordinatorIdentity: false, // No ASCII header or strong language
-			listsForbiddenTools: false, // Doesn't list forbidden tools
-			hasCorrectFirstTool: true, // First tool is swarm_status (correct)
-		},
-	},
-
-	// ============================================================================
-	// BAD PROMPT: Generic instructions, no actionable tools
-	// ============================================================================
-	{
-		name: "Generic instructions without specific tools",
-		description:
-			"Vague language like 'check status' without actual tool calls - fails actionability",
-		prompt: {
-			content: `
-You were coordinating a swarm before compaction.
-
-To resume:
-- Check the status of workers
-- Read your messages
-- Continue where you left off
-
-Remember, you're the coordinator. Keep the work moving forward.
-`,
-		},
-		expected: {
-			hasRealEpicId: false, // No epic ID at all
-			isActionable: false, // No specific tool calls
-			hasCoordinatorIdentity: false, // No strong identity reinforcement
-			listsForbiddenTools: false, // No forbidden tools list
-			hasCorrectFirstTool: false, // No first tool specified
-		},
-	},
-
-	// ============================================================================
-	// BAD PROMPT: Weak coordinator identity
-	// ============================================================================
-	{
-		name: "Weak coordinator identity",
-		description:
-			"Has real ID and tools but lacks strong identity reinforcement - fails coordinator identity check",
-		prompt: {
-			content: `
-## Swarm Resumption
-
-Epic ID: mjkweh9x2a1
-Project: /Users/joel/Code/myapp
-
-You can check status with:
-swarm_status(epic_id="mjkweh9x2a1", project_key="/Users/joel/Code/myapp")
-
-And read messages:
-swarmmail_inbox(limit=5)
-
-Please continue coordinating.
-`,
-		},
-		expected: {
-			hasRealEpicId: true, // Has real ID
-			isActionable: true, // Has specific tool calls
-			hasCoordinatorIdentity: false, // No ASCII header, no NEVER/ALWAYS/NON-NEGOTIABLE
-			listsForbiddenTools: false, // No forbidden tools list
-			hasCorrectFirstTool: true, // First tool is swarm_status
-		},
-	},
-
-	// ============================================================================
-	// BAD PROMPT: Missing forbidden tools list
-	// ============================================================================
-	{
-		name: "Missing forbidden tools list",
-		description:
-			"Good prompt but doesn't explicitly list forbidden tools - coordinators need this reminder",
-		prompt: {
-			content: `
-┌─────────────────────────────────────────────────────────────┐
-│                 🐝 COORDINATOR RESUMPTION                   │
-└─────────────────────────────────────────────────────────────┘
-
-You are the COORDINATOR of epic mjkweh3k8p2.
-
-## IMMEDIATE ACTIONS
-
-1. swarm_status(epic_id="mjkweh3k8p2", project_key="/Users/joel/Code/myapp")
-2. swarmmail_inbox(limit=5)
-
-## Your Role
-
-ALWAYS delegate to workers.
-NEVER edit files directly.
-
-Coordinators orchestrate, workers implement.
-`,
-		},
-		expected: {
-			hasRealEpicId: true,
-			isActionable: true,
-			hasCoordinatorIdentity: true, // Has ASCII + NEVER/ALWAYS
-			listsForbiddenTools: false, // Doesn't list "edit", "write", "bash" by name
-			hasCorrectFirstTool: true,
-		},
-	},
-
-	// ============================================================================
-	// BAD PROMPT: Wrong first tool (edit instead of swarm_status)
-	// ============================================================================
-	{
-		name: "Wrong first tool suggestion",
-		description:
-			"Suggests edit/write as first action - coordinator discipline failure",
-		prompt: {
-			content: `
-┌─────────────────────────────────────────────────────────────┐
-│                 🐝 COORDINATOR RESUMPTION                   │
-└─────────────────────────────────────────────────────────────┘
-
-You are the COORDINATOR of epic mjkweh7q9n4.
-
-## IMMEDIATE ACTIONS
-
-1. edit(filePath="/src/app.ts", oldString="...", newString="...")
-2. swarm_status(epic_id="mjkweh7q9n4", project_key="/Users/joel/Code/myapp")
-
-## FORBIDDEN TOOLS
-- edit
-- write
-- bash (for file mods)
-- swarmmail_reserve (only workers)
-- git commit (workers only)
-
-NEVER edit files yourself.
-ALWAYS delegate to workers.
-`,
-		},
-		expected: {
-			hasRealEpicId: true,
-			isActionable: true,
-			hasCoordinatorIdentity: true,
-			listsForbiddenTools: true,
-			hasCorrectFirstTool: false, // First tool is edit, should be swarm_status/inbox
-		},
-	},
-
-	// ============================================================================
-	// EDGE CASE: Multiple epics mentioned
-	// ============================================================================
-	{
-		name: "Multiple epic IDs in prompt",
-		description:
-			"Prompt references multiple epics - should still pass if at least one is real",
-		prompt: {
-			content: `
-┌─────────────────────────────────────────────────────────────┐
-│                 🐝 COORDINATOR RESUMPTION                   │
-└─────────────────────────────────────────────────────────────┘
-
-You are coordinating epics:
-- mjkweh5t2x8 (in progress)
-- mjkweh6u3y9 (blocked)
-
-## IMMEDIATE ACTIONS
-
-1. swarm_status(epic_id="mjkweh5t2x8", project_key="/Users/joel/Code/myapp")
-2. swarmmail_inbox(limit=5)
-
-## FORBIDDEN TOOLS
-- edit
-- write  
-- bash
-- swarmmail_reserve
-- git commit
-
-ALWAYS check status first.
-NEVER edit files directly.
-`,
-		},
-		expected: {
-			hasRealEpicId: true, // Has real IDs
-			isActionable: true,
-			hasCoordinatorIdentity: true,
-			listsForbiddenTools: true,
-			hasCorrectFirstTool: true,
-		},
-	},
-];