pi-harness-runtime 1.1.440 → 1.1.441

package/harness/langchain/model-router.test.ts

@@ -1,191 +0,0 @@
-import { describe, it, beforeEach, afterEach } from "node:test";
-import assert from "node:assert";
-import {
-	routeRequest,
-	routeRequestSmart,
-	resolveModel,
-} from "./model-router.js";
-
-const TEST_ENV = {
-	PLANNER_API_KEY: "pk-test-planner",
-	PLANNER_MODEL: "gpt-5.6",
-	PLANNER_BASE_URL: "https://api.openai.com/v1",
-	GLM_API_KEY: "pk-test-glm",
-	GLM_MODEL: "GLM-5.2",
-	GLM_BASE_URL: "https://api.z.ai/api/v1",
-	MINIMAX_API_KEY: "pk-test-minimax",
-	MINIMAX_MODEL: "MiniMax-M2",
-	MINIMAX_BASE_URL: "https://api.minimaxi.com/v1",
-};
-
-describe("resolveModel", () => {
-	const original = process.env;
-
-	beforeEach(() => {
-		process.env = { ...original, ...TEST_ENV };
-	});
-
-	afterEach(() => {
-		process.env = original;
-	});
-
-	it("resolves GLM family from env", () => {
-		const r = resolveModel("glm");
-		assert.equal(r.model, "GLM-5.2");
-		assert.equal(r.baseURL, "https://api.z.ai/api/v1");
-		assert.equal(r.apiKey, "pk-test-glm");
-	});
-
-	it("resolves PLANNER family from env", () => {
-		const r = resolveModel("planner");
-		assert.equal(r.model, "gpt-5.6");
-		assert.equal(r.apiKey, "pk-test-planner");
-	});
-
-	it("resolves MINIMAX family from env", () => {
-		const r = resolveModel("minimax");
-		assert.equal(r.model, "MiniMax-M2");
-		assert.equal(r.apiKey, "pk-test-minimax");
-	});
-
-	it("is case-insensitive", () => {
-		const r = resolveModel("GLM");
-		assert.equal(r.model, "GLM-5.2");
-	});
-
-	it("throws if API key is missing", () => {
-		delete process.env.GLM_API_KEY;
-		assert.throws(() => resolveModel("glm"), /GLM_API_KEY is not set/);
-	});
-
-	it("throws if MODEL is missing", () => {
-		delete process.env.GLM_MODEL;
-		assert.throws(() => resolveModel("glm"), /GLM_MODEL is not set/);
-	});
-
-	it("throws if BASE_URL is missing", () => {
-		delete process.env.GLM_BASE_URL;
-		assert.throws(() => resolveModel("glm"), /GLM_BASE_URL is not set/);
-	});
-});
-
-describe("routeRequest", () => {
-	const original = process.env;
-
-	beforeEach(() => {
-		process.env = { ...original, ...TEST_ENV };
-	});
-
-	afterEach(() => {
-		process.env = original;
-	});
-
-	it("default: planner→PLANNER, reviewer→GLM, coder→MINIMAX", () => {
-		const r = routeRequest("Implement a rate limiter");
-		assert.equal(r.planner.model, "gpt-5.6");
-		assert.equal(r.reviewer.model, "GLM-5.2");
-		assert.equal(r.coder.model, "MiniMax-M2");
-		assert.equal(r.cleanRequest, "Implement a rate limiter");
-	});
-
-	it("strips [role: family] directives from cleanRequest", () => {
-		const r = routeRequest(
-			"Fix the bug  [planner: gpt]  [reviewer: GLM]  [coder: MiniMax]",
-		);
-		assert.equal(r.cleanRequest, "Fix the bug");
-	});
-
-	it("override reviewer with GPT directive → PLANNER env", () => {
-		const r = routeRequest("Implement feature X  [reviewer: gpt]");
-		assert.equal(r.reviewer.model, "gpt-5.6");
-		assert.equal(r.cleanRequest, "Implement feature X");
-	});
-
-	it("override planner with GLM directive", () => {
-		const r = routeRequest("Plan this task  [planner: glm]");
-		assert.equal(r.planner.model, "GLM-5.2");
-	});
-
-	it("override coder with GPT directive → PLANNER env", () => {
-		const r = routeRequest("Write the code  [coder: gpt]");
-		assert.equal(r.coder.model, "gpt-5.6");
-	});
-
-	it("multiple directives together", () => {
-		const r = routeRequest(
-			"Implement X  [planner: gpt]  [reviewer: GLM]  [coder: MiniMax]",
-		);
-		assert.equal(r.planner.model, "gpt-5.6");
-		assert.equal(r.reviewer.model, "GLM-5.2");
-		assert.equal(r.coder.model, "MiniMax-M2");
-	});
-
-	it("partial directive — only reviewer overridden", () => {
-		const r = routeRequest("Do it  [reviewer: gpt]");
-		assert.equal(r.reviewer.model, "gpt-5.6");
-		assert.equal(r.planner.model, "gpt-5.6");
-		assert.equal(r.coder.model, "MiniMax-M2");
-	});
-
-	it("unknown family throws — must be a recognised family name", () => {
-		// FAMILY_MAP only knows: gpt, glm, minimax
-		// unknown families are passed through as-is, causing readEnv to throw
-		assert.throws(
-			() => routeRequest("Do it  [reviewer: unknown-provider]"),
-			/UNKNOWN-PROVIDER_API_KEY is not set/,
-		);
-	});
-});
-
-describe("routeRequestSmart — natural language hints", () => {
-	const original = process.env;
-
-	beforeEach(() => {
-		process.env = { ...original, ...TEST_ENV };
-	});
-
-	afterEach(() => {
-		process.env = original;
-	});
-
-	it("'use GLM as master reviewer' → reviewer→GLM", () => {
-		const r = routeRequestSmart(
-			"Implement a rate limiter. Use GLM as master reviewer.",
-		);
-		assert.equal(r.reviewer.model, "GLM-5.2");
-	});
-
-	it("'use GPT as reviewer' → reviewer→PLANNER", () => {
-		const r = routeRequestSmart("Implement a rate limiter. Use GPT as reviewer.");
-		assert.equal(r.reviewer.model, "gpt-5.6");
-	});
-
-	it("'[reviewer: GLM]' directive beats natural hint", () => {
-		const r = routeRequestSmart(
-			"Implement X. Use GPT as master reviewer.  [reviewer: GLM]",
-		);
-		// directive wins
-		assert.equal(r.reviewer.model, "GLM-5.2");
-		assert.equal(r.cleanRequest, "Implement X. Use GPT as master reviewer.");
-	});
-
-	it("'GPT as planner' → planner→PLANNER", () => {
-		const r = routeRequestSmart("Plan the implementation. GPT as planner.");
-		assert.equal(r.planner.model, "gpt-5.6");
-	});
-
-	it("strips directive comments but keeps natural hints", () => {
-		const r = routeRequestSmart(
-			"Fix the bug. Use GLM as master reviewer.  [reviewer: gpt]",
-		);
-		assert.equal(r.cleanRequest, "Fix the bug. Use GLM as master reviewer.");
-		assert.equal(r.reviewer.model, "gpt-5.6"); // directive wins
-	});
-
-	it("defaults all three when no hints or directives", () => {
-		const r = routeRequestSmart("Just do it.");
-		assert.equal(r.planner.model, "gpt-5.6");
-		assert.equal(r.reviewer.model, "GLM-5.2");
-		assert.equal(r.coder.model, "MiniMax-M2");
-	});
-});

package/harness/langchain/ping-pong-middleware.test.ts

@@ -1,69 +0,0 @@
-import { describe, it } from "node:test";
-import { equal, ok } from "node:assert";
-import { mkdtempSync, readFileSync, rmSync } from "node:fs";
-import { join } from "node:path";
-import { tmpdir } from "node:os";
-import { PingPongMiddleware } from "./ping-pong-middleware.js";
-
-describe("PingPongMiddleware", () => {
-	it("records explicit PING-PONG decisions in blackboard and inbox", () => {
-		const projectRoot = mkdtempSync(join(tmpdir(), "ping-pong-project-"));
-		const workspace = mkdtempSync(join(tmpdir(), "ping-pong-workspace-"));
-		try {
-			const result = new PingPongMiddleware().submit(
-				"Good plan, then write down plan to wiki/ then I will ask Minimax to write code then you review",
-				{
-					projectRoot,
-					workspace,
-					actorId: "test-agent",
-					now: () => new Date("2026-09-10T00:00:00.000Z"),
-				},
-			);
-
-			equal(result.enqueued, true);
-			ok(result.taskFile);
-			ok(result.taskId?.endsWith(".md"));
-			const taskBody = readFileSync(result.taskFile!, "utf8");
-			ok(taskBody.includes("kind: ping-pong"));
-			ok(taskBody.includes("ask Minimax to write code"));
-
-			const blackboard = JSON.parse(
-				readFileSync(join(projectRoot, ".write-review", "status.json"), "utf8"),
-			);
-			equal(blackboard.mode, "ping_pong");
-			equal(blackboard.phase, "writing");
-			equal(blackboard.pingPong.decision, "run_ping_pong");
-			equal(blackboard.pingPong.taskId, result.taskId);
-
-			const audit = readFileSync(result.auditFile, "utf8");
-			ok(audit.includes("run_ping_pong"));
-			ok(audit.includes("test-agent"));
-		} finally {
-			rmSync(projectRoot, { recursive: true, force: true });
-			rmSync(workspace, { recursive: true, force: true });
-		}
-	});
-
-	it("records inline decisions without enqueueing", () => {
-		const projectRoot = mkdtempSync(join(tmpdir(), "ping-pong-project-"));
-		const workspace = mkdtempSync(join(tmpdir(), "ping-pong-workspace-"));
-		try {
-			const result = new PingPongMiddleware().submit(
-				"analyze why the loop is quiet, do not edit code",
-				{ projectRoot, workspace, actorId: "test-agent" },
-			);
-
-			equal(result.enqueued, false);
-			equal(result.taskFile, undefined);
-			const blackboard = JSON.parse(
-				readFileSync(join(projectRoot, ".write-review", "status.json"), "utf8"),
-			);
-			equal(blackboard.mode, "ping_pong");
-			equal(blackboard.phase, "idle");
-			equal(blackboard.pingPong.decision, "handle_inline");
-		} finally {
-			rmSync(projectRoot, { recursive: true, force: true });
-			rmSync(workspace, { recursive: true, force: true });
-		}
-	});
-});

package/harness/langchain/scoreboard.test.ts

@@ -1,267 +0,0 @@
-/**
- * Scoreboard integration tests — T15
- *
- * Tests that the WriteReviewBlackboard is correctly wired into the LangGraph loop:
- * 1. `buildRealLoopDeps` creates and updates the blackboard at each step
- * 2. Agent prompts receive the scoreboard markdown
- * 3. `status.json` reflects the current phase/verdict/files
- *
- * Run:
- *   bun test harness/langchain/scoreboard.test.ts
- */
-
-import { describe, it, beforeEach, afterEach } from "node:test";
-import assert from "node:assert";
-import { existsSync, readFileSync, rmSync } from "fs";
-import { join } from "path";
-
-import { buildDryRunDeps } from "./graph.js";
-import { createLoopCheckpointer } from "./checkpointer.js";
-
-describe("T15 — Scoreboard blackboard wiring", () => {
-	const tmpDir = `/tmp/scoreboard-test-${Date.now()}`;
-	const blackboardDir = join(tmpDir, ".write-review");
-
-	beforeEach(() => {
-		// Ensure clean slate
-		rmSync(tmpDir, { recursive: true, force: true });
-	});
-
-	afterEach(() => {
-		rmSync(tmpDir, { recursive: true, force: true });
-	});
-
-	// ─── Basic wiring ───────────────────────────────────────────────────────────
-
-	it("buildDryRunDeps creates a status.json file", async () => {
-		const deps = buildDryRunDeps({
-			maxIterations: 2,
-			blackboardDir: tmpDir,
-		});
-
-		// Trigger onStep to initialize the blackboard
-		deps.onStep?.("plan", {
-			request: "test request",
-			plan: "",
-			iteration: 0,
-			code: "",
-			review: null as any,
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-
-		const statusPath = join(blackboardDir, "status.json");
-		assert.ok(existsSync(statusPath), "status.json should be created");
-	});
-
-	it("status.json shows phase=writing after plan step", async () => {
-		const deps = buildDryRunDeps({
-			maxIterations: 2,
-			blackboardDir: tmpDir,
-		});
-
-		deps.onStep?.("plan", {
-			request: "test",
-			plan: "",
-			iteration: 0,
-			code: "",
-			review: null as any,
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-
-		const content = JSON.parse(
-			readFileSync(join(blackboardDir, "status.json"), "utf8"),
-		);
-		assert.equal(content.phase, "writing");
-		assert.equal(content.iteration, 1);
-	});
-
-	it("status.json shows code files after write step", async () => {
-		const deps = buildDryRunDeps({
-			maxIterations: 2,
-			blackboardDir: tmpDir,
-		});
-
-		deps.onStep?.("plan", {
-			request: "test",
-			plan: "",
-			iteration: 0,
-			code: "",
-			review: null as any,
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-
-		deps.onStep?.("write", {
-			request: "test",
-			plan: "# plan",
-			iteration: 1,
-			code: "``` harness/langchain/surge.ts\nconst x = 1;\n```",
-			review: null as any,
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-
-		const content = JSON.parse(
-			readFileSync(join(blackboardDir, "status.json"), "utf8"),
-		);
-		assert.ok(
-			content.codeFiles?.some((f: string) => f.includes("surge.ts")),
-			"codeFiles should include surge.ts",
-		);
-	});
-
-	it("status.json shows verdict after review step", async () => {
-		const deps = buildDryRunDeps({
-			maxIterations: 2,
-			blackboardDir: tmpDir,
-		});
-
-		deps.onStep?.("plan", {
-			request: "test",
-			plan: "",
-			iteration: 0,
-			code: "",
-			review: null as any,
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-
-		deps.onStep?.("write", {
-			request: "test",
-			plan: "# plan",
-			iteration: 1,
-			code: "```harness/langchain/surge.ts\nconst x = 1;\n```",
-			review: null as any,
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-
-		deps.onStep?.("review", {
-			request: "test",
-			plan: "# plan",
-			iteration: 1,
-			code: "```test.ts\nconst x = 1;\n```",
-			review: {
-				verdict: "changes_requested",
-				summary: "needs error handling",
-				comments: [
-					{ file: "test.ts", comment: "add try/catch", severity: "critical" },
-				],
-			},
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-
-		const content = JSON.parse(
-			readFileSync(join(blackboardDir, "status.json"), "utf8"),
-		);
-		assert.equal(content.phase, "changes_requested");
-		assert.equal(content.verdict, "changes_requested");
-		assert.ok(
-			content.changesRequested?.length > 0,
-			"changesRequested should be set",
-		);
-	});
-
-	it("dry-run plan prompt includes scoreboard markdown", async () => {
-		const deps = buildDryRunDeps({
-			maxIterations: 2,
-			blackboardDir: tmpDir,
-		});
-
-		// Initialize blackboard
-		deps.onStep?.("plan", {
-			request: "test",
-			plan: "",
-			iteration: 0,
-			code: "",
-			review: null as any,
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-
-		// Call plan() — the stub returns a string that includes scoreboard context
-		const _result = await deps.plan("implement X");
-
-		// The stub output should include the plan text; the scoreboard is in the
-		// agent context, not the stub return value. Verify status.json was written.
-		const statusPath = join(blackboardDir, "status.json");
-		assert.ok(existsSync(statusPath), "status.json should exist after plan()");
-	});
-
-	it("blackboard persists across multiple onStep calls", async () => {
-		const deps = buildDryRunDeps({
-			maxIterations: 3,
-			blackboardDir: tmpDir,
-		});
-
-		deps.onStep?.("plan", {
-			request: "test",
-			plan: "",
-			iteration: 0,
-			code: "",
-			review: null as any,
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-		deps.onStep?.("write", {
-			request: "test",
-			plan: "",
-			iteration: 1,
-			code: "```a.ts\n1\n```",
-			review: null as any,
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-		deps.onStep?.("review", {
-			request: "test",
-			plan: "",
-			iteration: 1,
-			code: "```a.ts\n1\n```",
-			review: { verdict: "approved", summary: "ok", comments: [] },
-			log: [],
-			prevCode: undefined,
-			writtenFiles: {},
-			userInput: null,
-			userResponse: null,
-		});
-
-		const content = JSON.parse(
-			readFileSync(join(blackboardDir, "status.json"), "utf8"),
-		);
-		assert.equal(content.phase, "approved");
-		assert.equal(content.iteration, 1);
-	});
-});

package/harness/langchain/status-line.test.ts

@@ -1,106 +0,0 @@
-/**
- * StatusLineManager tests — T16
- *
- * Tests the pi-lens detection, quota formatting, and status file writing
- * when pi-lens is not available.
- */
-
-import { describe, it, beforeEach, afterEach } from "node:test";
-import assert from "node:assert";
-import { existsSync, readFileSync, rmSync, writeFileSync, mkdirSync } from "fs";
-import { join } from "path";
-import {
-	StatusLineManager,
-	isPiLensAvailable,
-	formatQuotaLine,
-} from "./status-line.js";
-
-describe("isPiLensAvailable", () => {
-	it("returns false when no PI_LENS env vars are set", () => {
-		assert.ok(!isPiLensAvailable());
-	});
-});
-
-describe("formatQuotaLine", () => {
-	it("returns 'quota unknown' when no data", () => {
-		assert.ok(formatQuotaLine({}).startsWith("MiniMax: quota unknown"));
-	});
-
-	it("shows 5h usage when available", () => {
-		const line = formatQuotaLine({ fiveHourPercent: 89 });
-		assert.ok(line.includes("5h: 89% left"), line);
-		assert.ok(line.includes("MiniMax:"), line);
-	});
-
-	it("shows both 5h and weekly when both available", () => {
-		const line = formatQuotaLine({ fiveHourPercent: 89, weeklyPercent: 75 });
-		assert.ok(line.includes("5h: 89% left"), line);
-		assert.ok(line.includes("week: 75% left"), line);
-	});
-
-	it("shows 0% left when at 100%", () => {
-		const line = formatQuotaLine({ fiveHourPercent: 100 });
-		assert.ok(line.includes("0% left"), line);
-	});
-
-	it("shows ? for invalid negative percentage", () => {
-		// fmtPercent returns ? for negative values
-		const line = formatQuotaLine({ fiveHourPercent: -1 });
-		assert.ok(line.includes("5h: ?"), line);
-	});
-
-	it("shows reset countdown when resetAt is provided", () => {
-		const future = new Date(Date.now() + 2 * 60 * 60 * 1000).toISOString(); // 2h from now
-		const line = formatQuotaLine({
-			fiveHourPercent: 89,
-			fiveHourResetAt: future,
-		});
-		assert.ok(line.includes("resets in"), line);
-	});
-});
-
-describe("StatusLineManager", () => {
-	const tmpDir = join(
-		process.env.TMPDIR ?? "/tmp",
-		`harness-status-test-${Date.now()}`,
-	);
-	let manager: StatusLineManager;
-
-	beforeEach(() => {
-		mkdirSync(tmpDir, { recursive: true });
-		manager = new StatusLineManager(tmpDir);
-	});
-
-	afterEach(() => {
-		manager.stop();
-		try {
-			rmSync(tmpDir, { recursive: true, force: true });
-		} catch {
-			/* ignore */
-		}
-	});
-
-	it("getLine returns combined loop + quota status", () => {
-		manager.updateLoopStatus("[P] planning");
-		const line = manager.getLine();
-		assert.ok(line.includes("planning"), line);
-	});
-
-	it("getLine returns quota-only when no loop status", () => {
-		const line = manager.getLine();
-		assert.ok(line.includes("MiniMax"), line);
-	});
-
-	it("writes .harness-status file when pi-lens is not available", () => {
-		manager.updateLoopStatus("[W] coding");
-		const statusFile = join(tmpDir, ".harness-status");
-		assert.ok(existsSync(statusFile), ".harness-status should exist");
-		const content = readFileSync(statusFile, "utf8");
-		assert.ok(content.includes("coding"), content);
-	});
-
-	it("start() and stop() do not throw", () => {
-		assert.doesNotThrow(() => manager.start());
-		assert.doesNotThrow(() => manager.stop());
-	});
-});