pi-harness-runtime 1.1.39 → 1.1.42

package/harness/langchain/daemon.ts

@@ -884,9 +884,13 @@ export class LoopDaemon {
 			const deps: LoopDeps =
 				this.config.deps ??
 				(this.config.dryRun
-					? buildDryRunDeps({ maxIterations: this.config.maxIterations })
-					: await buildRealLoopDeps({
+					? buildDryRunDeps({
 							maxIterations: this.config.maxIterations,
+							blackboardDir: this.bus.getWorkspace(),
+						})
+						: await buildRealLoopDeps({
+							maxIterations: this.config.maxIterations,
+							blackboardDir: this.bus.getWorkspace(),
 						}));
 
 			// Inject transition publishing into onStep (also drives the widget)
@@ -971,7 +975,7 @@ export class LoopDaemon {
 						: verdict === "blocked"
 							? "rejected"
 							: "max_iterations",
-					);
+				);
 			}
 
 			// ── Gate check after blocked verdict ────────────────────────────────

package/harness/langchain/graph.ts

@@ -24,6 +24,8 @@ import {
 	type ReviewVerdict,
 	ReviewVerdictSchema,
 } from "./agents.js";
+import { WriteReviewBlackboard } from "../../packages/write-review/src/blackboard.js";
+import type { Verdict } from "../../packages/write-review/src/types.js";
 import type { LoopWidget } from "./widget.js";
 
 // ─── State ──────────────────────────────────────────────────────────────────
@@ -176,6 +178,8 @@ export type WriteReviewLoop = ReturnType<typeof buildWriteReviewLoop>;
 export interface RealLoopOptions {
 	maxIterations?: number;
 	onStep?: (step: string, state: LoopState) => void;
+	/** Directory for .write-review/blackboard. Defaults to process.cwd(). */
+	blackboardDir?: string;
 	importAgents?: () => Promise<{
 		createPlannerAgent: () => {
 			invoke: (input: { messages: unknown[] }) => Promise<{
@@ -199,6 +203,10 @@ export interface RealLoopOptions {
 /**
  * Build LoopDeps backed by the real GPT/MiniMax agents from agents.ts.
  * Kept lazy (dynamic import) so dry-run mode never touches API keys.
+ *
+ * The WriteReviewBlackboard is updated at each step and its markdown is
+ * injected into every agent prompt — so each agent sees the shared scoreboard
+ * naturally, without explicit prompt injection.
  */
 export async function buildRealLoopDeps(
 	options: RealLoopOptions = {},
@@ -208,13 +216,67 @@ export async function buildRealLoopDeps(
 	const coder = mod.createCoderAgent();
 	const reviewer = mod.createReviewerAgent();
 
+	// Create and init the shared blackboard
+	const blackboard = new WriteReviewBlackboard(
+		options.blackboardDir ?? process.cwd(),
+	);
+	if (blackboard.exists()) {
+		blackboard.load();
+	} else {
+		blackboard.init();
+	}
+
+	/** Inject scoreboard markdown into a user message. */
+	const withScoreboard = (msg: string) => {
+		const scoreboard = blackboard.toMarkdown();
+		return msg + "\n\n## Current Progress\n" + scoreboard;
+	};
+
+	// Wrap onStep to also update the blackboard
+	const originalOnStep = options.onStep;
+	const onStep = (step: string, state: LoopState) => {
+		originalOnStep?.(step, state);
+
+		switch (step) {
+			case "plan": {
+				blackboard.init();
+				blackboard.startWriting(); // phase = "writing", iteration++
+				break;
+			}
+			case "write": {
+				// Extract file paths from the code output
+				const files = blackboard.extractFilePaths(state.code ?? "");
+				if (files.length > 0) {
+					blackboard.setCodeFiles(files);
+				}
+				blackboard.save();
+				break;
+			}
+			case "review": {
+				const verdict = state.review?.verdict as Verdict | undefined;
+				if (verdict) {
+					blackboard.setVerdict(verdict, state.review?.summary);
+				}
+				if (state.review?.comments && state.review.comments.length > 0) {
+					blackboard.setChangesRequested(
+						state.review.comments.map((c) =>
+							`[${c.severity ?? "minor"}] ${c.file ? `${c.file}: ` : ""}${c.comment}`,
+						),
+					);
+				}
+				blackboard.save();
+				break;
+			}
+		}
+	};
+
 	return {
 		maxIterations: options.maxIterations ?? 3,
-		onStep: options.onStep,
+		onStep,
 		plan: async (request) =>
 			lastMessage(
 				await planner.invoke({
-					messages: [{ role: "user", content: request }],
+					messages: [{ role: "user", content: withScoreboard(request) }],
 				}),
 			),
 		write: async (plan, review) => {
@@ -224,7 +286,9 @@ export async function buildRealLoopDeps(
 						.join("\n")}`
 				: `## Plan\n${plan}`;
 			return lastMessage(
-				await coder.invoke({ messages: [{ role: "user", content: userMsg }] }),
+				await coder.invoke({
+					messages: [{ role: "user", content: withScoreboard(userMsg) }],
+				}),
 			);
 		},
 		review: async (plan, code) => {
@@ -232,7 +296,9 @@ export async function buildRealLoopDeps(
 				messages: [
 					{
 						role: "user",
-						content: `## Plan\n${plan}\n\n## Code to review\n${code}`,
+						content: withScoreboard(
+							`## Plan\n${plan}\n\n## Code to review\n${code}`,
+						),
 					},
 				],
 			});
@@ -255,20 +321,78 @@ export async function buildRealLoopDeps(
 
 /** Deterministic stubs: iteration 1 requests changes, iteration 2 approves. */
 export function buildDryRunDeps(
-	options: { maxIterations?: number; onStep?: LoopDeps["onStep"] } = {},
+	options: {
+		maxIterations?: number;
+		onStep?: LoopDeps["onStep"];
+		/** Directory for .write-review/blackboard. Defaults to process.cwd(). */
+		blackboardDir?: string;
+	} = {},
 ): LoopDeps {
 	let writeCount = 0;
-	return {
+
+	// Create and init the shared blackboard (even in dry-run, so widget can read it)
+	const blackboard = new WriteReviewBlackboard(
+		options.blackboardDir ?? process.cwd(),
+	);
+	if (blackboard.exists()) {
+		blackboard.load();
+	} else {
+		blackboard.init();
+	}
+
+	const withScoreboard = (msg: string) => {
+		const scoreboard = blackboard.toMarkdown();
+		return msg + "\n\n## Current Progress\n" + scoreboard;
+	};
+
+	// Wrap onStep to also update the blackboard
+	const originalOnStep = options.onStep;
+	const onStep = (step: string, state: LoopState) => {
+		originalOnStep?.(step, state);
+
+		switch (step) {
+			case "plan": {
+				blackboard.init();
+				blackboard.startWriting();
+				break;
+			}
+			case "write": {
+				const files = blackboard.extractFilePaths(state.code ?? "");
+				if (files.length > 0) blackboard.setCodeFiles(files);
+				blackboard.save();
+				break;
+			}
+			case "review": {
+				const verdict = state.review?.verdict as Verdict | undefined;
+				if (verdict) blackboard.setVerdict(verdict, state.review?.summary);
+				if (state.review?.comments && state.review.comments.length > 0) {
+					blackboard.setChangesRequested(
+						state.review.comments.map((c) =>
+							`[${c.severity ?? "minor"}] ${c.file ? `${c.file}: ` : ""}${c.comment}`,
+						),
+					);
+				}
+				blackboard.save();
+				break;
+			}
+		}
+	};
+
+		return {
 		maxIterations: options.maxIterations ?? 3,
-		onStep: options.onStep,
+		onStep,
 		plan: async (request) =>
-			`# Plan (dry-run)\n\nRequest: ${request}\n\n1. Stub step one\n2. Stub step two`,
+			withScoreboard(
+				`# Plan (dry-run)\n\nRequest: ${request}\n\n1. Stub step one\n2. Stub step two`,
+			),
 		write: async (plan, review) => {
 			writeCount += 1;
 			const fixNote = review
 				? `\n// addressing: ${review.comments.map((c) => c.comment).join("; ")}`
 				: "";
-			return `\`\`\`ts\n// stub code, iteration ${writeCount}\nexport const plan = ${JSON.stringify(plan.slice(0, 60))};${fixNote}\n\`\`\``;
+			return withScoreboard(
+				`\`\`\`ts\n// stub code, iteration ${writeCount}\nexport const plan = ${JSON.stringify(plan.slice(0, 60))};${fixNote}\n\`\`\``,
+			);
 		},
 		review: async (_plan, _code) => {
 			if (writeCount < 2) {

package/harness/langchain/scoreboard.test.ts

@@ -0,0 +1,222 @@
+/**
+ * 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: [],
+		});
+
+		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: [],
+		});
+
+		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: [],
+		});
+
+		deps.onStep?.("write", {
+			request: "test",
+			plan: "# plan",
+			iteration: 1,
+			code: "``` harness/langchain/surge.ts\nconst x = 1;\n```",
+			review: null as any,
+			log: [],
+		});
+
+		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: [],
+		});
+
+		deps.onStep?.("write", {
+			request: "test",
+			plan: "# plan",
+			iteration: 1,
+			code: '```harness/langchain/surge.ts\nconst x = 1;\n```',
+			review: null as any,
+			log: [],
+		});
+
+		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: [],
+		});
+
+		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: [],
+		});
+
+		// 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: [],
+		});
+		deps.onStep?.("write", {
+			request: "test",
+			plan: "",
+			iteration: 1,
+			code: "```a.ts\n1\n```",
+			review: null as any,
+			log: [],
+		});
+		deps.onStep?.("review", {
+			request: "test",
+			plan: "",
+			iteration: 1,
+			code: "```a.ts\n1\n```",
+			review: { verdict: "approved", summary: "ok", comments: [] },
+			log: [],
+		});
+
+		const content = JSON.parse(
+			readFileSync(join(blackboardDir, "status.json"), "utf8"),
+		);
+		assert.equal(content.phase, "approved");
+		assert.equal(content.iteration, 1);
+	});
+
+});

package/harness/langchain/widget.test.ts

@@ -25,7 +25,6 @@ const stubTheme = {
 };
 
 describe("T14 — LoopWidget", () => {
-
 	let widget: LoopWidget;
 
 	beforeEach(() => {
@@ -58,14 +57,14 @@ describe("T14 — LoopWidget", () => {
 		widget.setSurgePause(new Date(Date.now() + 30 * 60_000)); // 30 min
 		const lines = widget.renderWidget(80, stubTheme);
 		// Label is "surge (in 30m)"
-		assert.ok(lines.some(l => l.includes("surge") && l.includes("in 30m")));
+		assert.ok(lines.some((l) => l.includes("surge") && l.includes("in 30m")));
 	});
 
 	it("setError shows error detail", () => {
 		widget.setError("model timeout");
 		const lines = widget.renderWidget(80, stubTheme);
-		assert.ok(lines.some(l => l.includes("error")));
-		assert.ok(lines.some(l => l.includes("model timeout")));
+		assert.ok(lines.some((l) => l.includes("error")));
+		assert.ok(lines.some((l) => l.includes("model timeout")));
 	});
 
 	it("setComplete approved shows ✓ loop complete", () => {
@@ -101,10 +100,13 @@ describe("T14 — LoopWidget", () => {
 		// Files accumulate across iterations (same codebase); only counters reset.
 		// After reset, a.ts shows with 0 write steps and 0 blockers.
 		const lines = widget.renderWidget(80, stubTheme);
-		const fileLines = lines.filter(l => l.includes("a.ts"));
+		const fileLines = lines.filter((l) => l.includes("a.ts"));
 		assert.ok(fileLines.length > 0, "file should still appear");
 		// No blocker count in the file row (reset to 0)
-		assert.ok(!fileLines.some(l => l.includes("!2")), "blocker count should be reset");
+		assert.ok(
+			!fileLines.some((l) => l.includes("!2")),
+			"blocker count should be reset",
+		);
 	});
 
 	// ─── File record tracking ──────────────────────────────────────────────
@@ -112,29 +114,33 @@ describe("T14 — LoopWidget", () => {
 	it("recordWrite shows file in widget", () => {
 		widget.recordWrite("src/foo.ts");
 		const lines = widget.renderWidget(80, stubTheme);
-		assert.ok(lines.some(l => l.includes("foo.ts")));
+		assert.ok(lines.some((l) => l.includes("foo.ts")));
 	});
 
 	it("recordWrite accumulates across calls", () => {
 		widget.recordWrite("src/bar.ts");
 		widget.recordWrite("src/bar.ts");
 		const lines = widget.renderWidget(80, stubTheme);
-		assert.ok(lines.some(l => l.includes("bar.ts") && l.includes("✍2")));
+		assert.ok(lines.some((l) => l.includes("bar.ts") && l.includes("✍2")));
 	});
 
 	it("recordBlockers shows blocker count", () => {
 		widget.recordBlockers("src/bad.ts", 2);
 		const lines = widget.renderWidget(80, stubTheme);
-		assert.ok(lines.some(l => l.includes("bad.ts") && l.includes("!2")));
+		assert.ok(lines.some((l) => l.includes("bad.ts") && l.includes("!2")));
 	});
 
 	it("recordReviewPass clears blockers and increments passes", () => {
 		widget.recordBlockers("src/ok.ts", 1);
 		widget.recordReviewPass("src/ok.ts");
 		const lines = widget.renderWidget(80, stubTheme);
-		assert.ok(lines.some(l => l.includes("ok.ts") && l.includes("🔍1")));
+		assert.ok(lines.some((l) => l.includes("ok.ts") && l.includes("🔍1")));
 		// No blocker count
-		assert.ok(!lines.some(l => l.includes("ok.ts") && l.includes("!") && !l.includes("🔍")));
+		assert.ok(
+			!lines.some(
+				(l) => l.includes("ok.ts") && l.includes("!") && !l.includes("🔍"),
+			),
+		);
 	});
 
 	// ─── Summary counts ────────────────────────────────────────────────────
@@ -147,7 +153,7 @@ describe("T14 — LoopWidget", () => {
 	it("errors shown as ●NE", () => {
 		widget.recordBlockers("src/err.ts", 3);
 		const lines = widget.renderWidget(80, stubTheme);
-		assert.ok(lines.some(l => l.includes("●3E")));
+		assert.ok(lines.some((l) => l.includes("●3E")));
 	});
 
 	// ─── Surge pause ──────────────────────────────────────────────────────
@@ -156,7 +162,7 @@ describe("T14 — LoopWidget", () => {
 		const resetAt = new Date(Date.now() + 90 * 60_000); // 90 minutes → "in 1h 30m"
 		widget.setSurgePause(resetAt);
 		const lines = widget.renderWidget(80, stubTheme);
-		const surgeLine = lines.find(l => l.includes("surge"));
+		const surgeLine = lines.find((l) => l.includes("surge"));
 		assert.ok(surgeLine !== undefined);
 		assert.match(surgeLine, /\(in \d+h/); // "in 1h 30m" matches
 	});
@@ -166,7 +172,7 @@ describe("T14 — LoopWidget", () => {
 		widget.setSurgePause(resetAt);
 		const lines = widget.renderWidget(80, stubTheme);
 		// Label format: "surge (now)"
-		assert.ok(lines.some(l => l.includes("surge") && l.includes("(now)")));
+		assert.ok(lines.some((l) => l.includes("surge") && l.includes("(now)")));
 	});
 
 	// ─── Console fallback ─────────────────────────────────────────────────
@@ -185,7 +191,9 @@ describe("T14 — LoopWidget", () => {
 	// ─── Width / truncation ───────────────────────────────────────────────
 
 	it("truncates long file names with …", () => {
-		widget.recordWrite("/very/long/path/to/a/very/long/file/name/that/exceeds/width.ts");
+		widget.recordWrite(
+			"/very/long/path/to/a/very/long/file/name/that/exceeds/width.ts",
+		);
 		const lines = widget.renderWidget(40, stubTheme);
 		for (const line of lines) {
 			const stripped = line.replace(/\x1b\[[0-9;]*m/g, "");
@@ -202,7 +210,7 @@ describe("T14 — LoopWidget", () => {
 		widget.reset();
 		const lines = widget.renderWidget(80, stubTheme);
 		assert.match(lines[0], /idle/);
-		assert.ok(!lines.some(l => l.includes("a.ts")));
+		assert.ok(!lines.some((l) => l.includes("a.ts")));
 	});
 
 	// ─── makeRenderer ─────────────────────────────────────────────────────
@@ -234,11 +242,10 @@ describe("T14 — LoopWidget", () => {
 
 	it("getLoopWidget returns the same instance", () => {
 		// Dynamic import to avoid module caching issues
-		import("./widget.js").then(m => {
+		import("./widget.js").then((m) => {
 			const a = m.getLoopWidget();
 			const b = m.getLoopWidget();
 			assert.strictEqual(a, b);
 		});
 	});
-
 });