opencode-swarm-plugin 0.44.0 → 0.44.2

package/src/compaction-prompt-scoring.ts

@@ -1,300 +0,0 @@
-/**
- * Compaction Prompt Quality Scoring - Pure Functions
- *
- * Evaluates the quality of continuation prompts generated after context compaction.
- * **Problem**: Post-compaction coordinators often "wake up" confused, forget their role,
- * and start editing files instead of checking worker status.
- *
- * **Solution**: Score prompts on 5 dimensions that predict coordinator success:
- *
- * 1. **Epic ID Specificity (0.20)**: Real IDs (`mjkw...`) not placeholders (`<epic-id>`, `bd-xxx`)
- *    - Placeholders = coordinator can't check actual swarm status
- *
- * 2. **Actionability (0.20)**: Tool calls with real values (e.g., `swarm_status(epic_id='mjkw81rkq4c')`)
- *    - Generic instructions like "check status" don't work
- *
- * 3. **Coordinator Identity (0.25)**: ASCII header + strong mandates (NEVER/ALWAYS)
- *    - Visual + semantic cues reinforce role post-compaction
- *
- * 4. **Forbidden Tools Listed (0.15)**: Explicitly lists Edit, Write, swarmmail_reserve, git commit
- *    - Naming forbidden tools reduces violations
- *
- * 5. **Post-Compaction Discipline (0.20)**: First suggested tool is swarm_status or inbox (not Edit)
- *    - First tool sets the pattern - "check status" vs "dive into code"
- *
- * **Pure functions**: These can be tested without evalite. The evalite wrappers are in
- * `evals/scorers/compaction-prompt-scorers.ts`.
- *
- * **Data source**: Captured from `captureCompactionEvent()` with `compaction_type: "prompt_generated"`.
- * The payload includes the FULL prompt content (not truncated) for scoring.
- *
- * **Integration**: `compaction-prompt.eval.ts` uses these scorers to track prompt quality over time.
- * Progressive gates enforce quality: bootstrap → stabilization → production.
- *
- * @module compaction-prompt-scoring
- */
-
-/**
- * Compaction prompt structure (from LLM generation)
- */
-export interface CompactionPrompt {
-	content: string;
-}
-
-/**
- * Scorer result type
- */
-export interface ScorerResult {
-	score: number;
-	message: string;
-}
-
-// ====== Shared Regex Patterns ======
-
-/** Matches real epic/cell IDs (mjkw prefix + 7+ base36 chars) */
-export const REAL_EPIC_ID = /mjkw[a-z0-9]{7,}/;
-
-/** Matches common placeholder patterns */
-export const PLACEHOLDERS = [
-	/<epic-id>/i,
-	/bd-xxx/,
-	/<path>/i,
-	/<project>/i,
-];
-
-/** Matches ASCII box-drawing characters (for headers) */
-export const ASCII_BOX = /[┌┐└┘─│]{3,}/;
-
-/** Matches strong mandate language */
-export const STRONG_LANGUAGE = [/\bNEVER\b/, /\bALWAYS\b/, /\bNON-NEGOTIABLE\b/];
-
-// ====== Pure Scoring Functions ======
-
-/**
- * Score epic ID specificity
- *
- * Validates that epic IDs are REAL, not placeholders.
- * Placeholders like <epic-id>, bd-xxx, <path> indicate
- * the prompt generator failed to inject actual values.
- *
- * @returns 1.0 if real IDs, 0.0 if placeholders found
- */
-export function scoreEpicIdSpecificity(prompt: CompactionPrompt): ScorerResult {
-	// Check for placeholder patterns
-	for (const pattern of PLACEHOLDERS) {
-		if (pattern.test(prompt.content)) {
-			return {
-				score: 0.0,
-				message: `Found placeholder: ${pattern.source}`,
-			};
-		}
-	}
-
-	// Check for real epic ID pattern
-	if (REAL_EPIC_ID.test(prompt.content)) {
-		return {
-			score: 1.0,
-			message: "Contains real epic ID",
-		};
-	}
-
-	return {
-		score: 0.0,
-		message: "No epic ID found",
-	};
-}
-
-/**
- * Score actionability of tool calls
- *
- * Validates that the prompt includes SPECIFIC actionable tool calls.
- * Generic instructions like "check status" are useless.
- * Good: swarm_status(epic_id='mjkw81rkq4c', project_key='/path')
- * Bad: "Check the status of workers"
- *
- * @returns 1.0 if actionable tool calls with real values, 0.0 otherwise
- */
-export function scoreActionability(prompt: CompactionPrompt): ScorerResult {
-	// Check for actionable tool patterns
-	const actionableTools = [
-		/swarm_status\([^)]*epic_id\s*=\s*['"]mjkw[a-z0-9]{7,}['"]/,
-		/swarmmail_inbox\(\)/,
-	];
-
-	for (const pattern of actionableTools) {
-		if (pattern.test(prompt.content)) {
-			return {
-				score: 1.0,
-				message: "Contains actionable tool call with real values",
-			};
-		}
-	}
-
-	// Check if tool is mentioned but with placeholders
-	if (
-		/swarm_status\([^)]*<epic-id>/.test(prompt.content) ||
-		/swarm_status\([^)]*<path>/.test(prompt.content)
-	) {
-		return {
-			score: 0.0,
-			message: "Tool call has placeholders",
-		};
-	}
-
-	return {
-		score: 0.0,
-		message: "No actionable tool calls found",
-	};
-}
-
-/**
- * Score coordinator identity reinforcement
- *
- * Validates that the prompt has STRONG coordinator identity reinforcement.
- * Post-compaction coordinators lose their identity without visual+semantic cues.
- *
- * Checks:
- * 1. ASCII box header (visual anchor)
- * 2. Strong language (NEVER/ALWAYS, not "should"/"consider")
- *
- * @returns 1.0 for ASCII header + strong mandates, 0.5 for header only, 0.0 otherwise
- */
-export function scoreCoordinatorIdentity(
-	prompt: CompactionPrompt,
-): ScorerResult {
-	// Check for ASCII box header (uses box-drawing characters)
-	const hasAsciiHeader =
-		ASCII_BOX.test(prompt.content) &&
-		/(YOU ARE THE COORDINATOR|COORDINATOR MODE)/i.test(prompt.content);
-
-	if (!hasAsciiHeader) {
-		return {
-			score: 0.0,
-			message: "No ASCII header found",
-		};
-	}
-
-	// Check for strong mandate language
-	const hasStrongLanguage = STRONG_LANGUAGE.some((pattern) =>
-		pattern.test(prompt.content),
-	);
-
-	if (!hasStrongLanguage) {
-		return {
-			score: 0.5,
-			message: "ASCII header present but weak language",
-		};
-	}
-
-	return {
-		score: 1.0,
-		message: "ASCII header + strong mandates present",
-	};
-}
-
-/**
- * Score forbidden tools listing
- *
- * Validates that the prompt LISTS forbidden tools by name.
- * Coordinators must know exactly which tools to avoid.
- *
- * Required forbidden tools:
- * 1. Edit
- * 2. Write
- * 3. swarmmail_reserve (only workers reserve)
- * 4. git commit (workers commit)
- * 5. bash (for file modifications)
- *
- * @returns ratio of forbidden tools mentioned (0.0 to 1.0)
- */
-export function scoreForbiddenToolsPresent(
-	prompt: CompactionPrompt,
-): ScorerResult {
-	// Check for forbidden tool mentions
-	const forbiddenTools = [
-		/\bEdit\b/i,
-		/\bWrite\b/i,
-		/swarmmail_reserve/,
-		/git commit/,
-		/\bbash\b/i,
-	];
-
-	const foundTools = forbiddenTools.filter((pattern) =>
-		pattern.test(prompt.content),
-	);
-
-	const score = foundTools.length / forbiddenTools.length;
-
-	if (score === 1.0) {
-		return {
-			score: 1.0,
-			message: "All 5 forbidden tools listed",
-		};
-	}
-
-	if (score === 0) {
-		return {
-			score: 0.0,
-			message: "No forbidden tools listed (0/5)",
-		};
-	}
-
-	return {
-		score,
-		message: `${foundTools.length}/5 forbidden tools listed`,
-	};
-}
-
-/**
- * Score post-compaction discipline (first tool correctness)
- *
- * Validates that the FIRST suggested tool is correct.
- * Coordinators should check status FIRST, not edit files.
- *
- * Good first tools:
- * - swarm_status
- * - swarmmail_inbox
- *
- * Bad first tools:
- * - Edit
- * - Write
- * - Read (should check status first)
- *
- * @returns 1.0 if first tool is swarm_status or inbox, 0.0 otherwise
- */
-export function scorePostCompactionDiscipline(
-	prompt: CompactionPrompt,
-): ScorerResult {
-	// Extract first tool call (look for function-like patterns)
-	const toolCallPattern =
-		/\b(swarm_status|swarmmail_inbox|Edit|Write|Read)\b/i;
-	const match = prompt.content.match(toolCallPattern);
-
-	if (!match) {
-		return {
-			score: 0.0,
-			message: "No tool calls found",
-		};
-	}
-
-	const firstTool = match[1].toLowerCase();
-
-	if (firstTool === "swarm_status") {
-		return {
-			score: 1.0,
-			message: "First tool is swarm_status (correct)",
-		};
-	}
-
-	if (firstTool === "swarmmail_inbox") {
-		return {
-			score: 1.0,
-			message: "First tool is inbox (correct)",
-		};
-	}
-
-	return {
-		score: 0.0,
-		message: `First tool is ${match[1]} (should be swarm_status or inbox)`,
-	};
-}

package/src/contributor-tools.test.ts

@@ -1,133 +0,0 @@
-/**
- * Contributor Tools Integration Tests
- *
- * Tests for contributor_lookup tool that fetches GitHub profiles
- * and generates changeset credits.
- */
-
-import { afterAll, describe, expect, test } from "bun:test";
-import { closeAllSwarmMail } from "swarm-mail";
-import { contributorTools, resetContributorCache } from "./contributor-tools";
-import { resetMemoryCache } from "./memory-tools";
-
-interface ToolContext {
-	sessionID: string;
-}
-
-describe("contributor tools integration", () => {
-	afterAll(async () => {
-		resetContributorCache();
-		resetMemoryCache();
-		await closeAllSwarmMail();
-	});
-
-	test("contributor_lookup tool is registered", () => {
-		expect(contributorTools).toHaveProperty("contributor_lookup");
-		expect(typeof contributorTools.contributor_lookup.execute).toBe("function");
-	});
-
-	describe("contributor_lookup", () => {
-		test("returns formatted credit with name + twitter", async () => {
-			const tool = contributorTools.contributor_lookup;
-			const result = await tool.execute(
-				{
-					login: "kentcdodds",
-					issue: 42,
-				},
-				{ sessionID: "test-session" } as ToolContext,
-			);
-
-			expect(typeof result).toBe("string");
-			const parsed = JSON.parse(result);
-			
-			expect(parsed.login).toBe("kentcdodds");
-			expect(parsed.name).toBeDefined();
-			expect(parsed.twitter).toBeDefined();
-			expect(parsed.credit_line).toContain("Thanks to");
-			expect(parsed.credit_line).toContain("reporting #42");
-			expect(parsed.credit_line).toContain("@kentcdodds");
-			expect(parsed.credit_line).toContain("https://x.com/");
-			expect(parsed.memory_stored).toBe(true);
-		});
-
-		test("handles missing twitter gracefully", async () => {
-			const tool = contributorTools.contributor_lookup;
-			
-			// Use a user that likely has name but no twitter
-			// (we'll test the format logic mainly)
-			const result = await tool.execute(
-				{
-					login: "torvalds", // Linus Torvalds - has name, might not have twitter
-					issue: 123,
-				},
-				{ sessionID: "test-session" } as ToolContext,
-			);
-
-			const parsed = JSON.parse(result);
-			
-			expect(parsed.login).toBe("torvalds");
-			expect(parsed.name).toBeDefined();
-			expect(parsed.credit_line).toContain("Thanks to");
-			expect(parsed.credit_line).toContain("reporting #123");
-			// Should have GitHub mention if no Twitter
-			if (!parsed.twitter) {
-				expect(parsed.credit_line).toContain("on GitHub");
-			}
-		});
-
-		test("works without issue number", async () => {
-			const tool = contributorTools.contributor_lookup;
-			const result = await tool.execute(
-				{
-					login: "kentcdodds",
-				},
-				{ sessionID: "test-session" } as ToolContext,
-			);
-
-			const parsed = JSON.parse(result);
-			
-			expect(parsed.login).toBe("kentcdodds");
-			expect(parsed.credit_line).toContain("Thanks to");
-			// Should NOT contain "reporting #"
-			expect(parsed.credit_line).not.toContain("reporting #");
-		});
-
-		test("stores contributor info in semantic-memory", async () => {
-			const tool = contributorTools.contributor_lookup;
-			const result = await tool.execute(
-				{
-					login: "gaearon", // Dan Abramov
-					issue: 99,
-				},
-				{ sessionID: "test-session" } as ToolContext,
-			);
-
-			const parsed = JSON.parse(result);
-			
-			// Just verify memory_stored flag - embedding search may be async
-			expect(parsed.memory_stored).toBe(true);
-		});
-
-		test("returns all expected fields", async () => {
-			const tool = contributorTools.contributor_lookup;
-			const result = await tool.execute(
-				{
-					login: "kentcdodds",
-				},
-				{ sessionID: "test-session" } as ToolContext,
-			);
-
-			const parsed = JSON.parse(result);
-			
-			// Required fields
-			expect(parsed).toHaveProperty("login");
-			expect(parsed).toHaveProperty("credit_line");
-			expect(parsed).toHaveProperty("memory_stored");
-			
-			// Optional fields (may be null but should be present)
-			expect(parsed).toHaveProperty("name");
-			expect(parsed).toHaveProperty("twitter");
-			expect(parsed).toHaveProperty("bio");
-		});
-	});
-});

package/src/contributor-tools.ts

@@ -1,201 +0,0 @@
-/**
- * Contributor Tools - GitHub profile extraction for changeset credits
- *
- * Provides contributor_lookup tool for fetching GitHub profiles and
- * generating formatted changeset credit lines. Automatically stores
- * contributor info in semantic-memory for future reference.
- *
- * Based on patterns from gh-issue-triage skill.
- */
-
-import { tool } from "@opencode-ai/plugin";
-import { z } from "zod";
-import { getMemoryAdapter } from "./memory-tools";
-
-// ============================================================================
-// Types
-// ============================================================================
-
-const GitHubUserSchema = z.object({
-	login: z.string(),
-	name: z.string().nullable(),
-	twitter_username: z.string().nullable(),
-	blog: z.string().nullable(),
-	bio: z.string().nullable(),
-	avatar_url: z.string(),
-	html_url: z.string(),
-	public_repos: z.number().optional(),
-	followers: z.number().optional(),
-});
-
-type GitHubUser = z.infer<typeof GitHubUserSchema>;
-
-interface ContributorResult {
-	login: string;
-	name: string | null;
-	twitter: string | null;
-	bio: string | null;
-	credit_line: string;
-	memory_stored: boolean;
-}
-
-interface ToolContext {
-	sessionID: string;
-}
-
-// ============================================================================
-// Core Functions
-// ============================================================================
-
-/**
- * Fetch GitHub user profile via gh CLI
- */
-async function fetchGitHubUser(login: string): Promise<GitHubUser> {
-	const result = await Bun.$`gh api users/${login}`.json();
-	return GitHubUserSchema.parse(result);
-}
-
-/**
- * Format changeset credit line based on available data
- *
- * Hierarchy:
- * 1. Name + Twitter: "Thanks to {Name} ([@twitter](...)) for reporting #{issue}!"
- * 2. Name only: "Thanks to {Name} (@{login} on GitHub) for reporting #{issue}!"
- * 3. Twitter only: "Thanks to [@twitter](...) for reporting #{issue}!"
- * 4. Fallback: "Thanks to @{login} for reporting #{issue}!"
- */
-function formatCreditLine(user: GitHubUser, issueNumber?: number): string {
-	const issueText = issueNumber ? `reporting #${issueNumber}` : "the report";
-
-	// PREFERRED: Full name + Twitter (best for engagement)
-	if (user.name && user.twitter_username) {
-		return `Thanks to ${user.name} ([@${user.twitter_username}](https://x.com/${user.twitter_username})) for ${issueText}!`;
-	}
-
-	// Twitter only (no name available)
-	if (user.twitter_username) {
-		return `Thanks to [@${user.twitter_username}](https://x.com/${user.twitter_username}) for ${issueText}!`;
-	}
-
-	// Name only (no Twitter)
-	if (user.name) {
-		return `Thanks to ${user.name} (@${user.login} on GitHub) for ${issueText}!`;
-	}
-
-	// Fallback: GitHub username only
-	return `Thanks to @${user.login} for ${issueText}!`;
-}
-
-/**
- * Store contributor info in semantic-memory
- */
-async function storeContributorMemory(
-	user: GitHubUser,
-	issueNumber?: number,
-): Promise<boolean> {
-	try {
-		const adapter = await getMemoryAdapter();
-
-		const twitterPart = user.twitter_username
-			? ` (@${user.twitter_username} on Twitter)`
-			: "";
-		const issuePart = issueNumber ? `. Filed issue #${issueNumber}` : "";
-		const bioPart = user.bio ? `. Bio: '${user.bio}'` : "";
-
-		const tags = [
-			"contributor",
-			user.login,
-			issueNumber ? `issue-${issueNumber}` : null,
-		]
-			.filter(Boolean)
-			.join(",");
-
-		const information = `Contributor @${user.login}: ${user.name || user.login}${twitterPart}${issuePart}${bioPart}`;
-
-		await adapter.store({
-			information,
-			tags,
-		});
-
-		return true;
-	} catch (error) {
-		console.error("Failed to store contributor memory:", error);
-		return false;
-	}
-}
-
-// ============================================================================
-// Cache Management
-// ============================================================================
-
-/**
- * Reset cache for testing
- */
-export function resetContributorCache(): void {
-	// Currently no cache, but keeping this for consistency with other tools
-}
-
-// ============================================================================
-// Plugin Tools
-// ============================================================================
-
-/**
- * Look up GitHub contributor and generate changeset credit
- */
-export const contributor_lookup = tool({
-	description:
-		"Fetch GitHub contributor profile and generate formatted changeset credit. Automatically stores contributor info in semantic-memory. Returns login, name, twitter, bio, and ready-to-paste credit_line.",
-	args: {
-		login: tool.schema.string().describe("GitHub username (required)"),
-		issue: tool.schema
-			.number()
-			.optional()
-			.describe("Issue number for context (optional)"),
-	},
-	async execute(
-		args: { login: string; issue?: number },
-		_ctx: ToolContext,
-	): Promise<string> {
-		try {
-			// Fetch GitHub profile
-			const user = await fetchGitHubUser(args.login);
-
-			// Format credit line
-			const creditLine = formatCreditLine(user, args.issue);
-
-			// Store in semantic-memory
-			const memoryStored = await storeContributorMemory(user, args.issue);
-
-			// Build result
-			const result: ContributorResult = {
-				login: user.login,
-				name: user.name,
-				twitter: user.twitter_username,
-				bio: user.bio,
-				credit_line: creditLine,
-				memory_stored: memoryStored,
-			};
-
-			return JSON.stringify(result, null, 2);
-		} catch (error) {
-			if (error instanceof Error) {
-				return JSON.stringify({
-					error: error.message,
-					login: args.login,
-				});
-			}
-			return JSON.stringify({
-				error: "Unknown error fetching contributor",
-				login: args.login,
-			});
-		}
-	},
-});
-
-// ============================================================================
-// Exports
-// ============================================================================
-
-export const contributorTools = {
-	contributor_lookup,
-} as const;