pi-harness-runtime 1.1.41 → 1.1.43

package/harness/langchain/agents.ts

@@ -47,27 +47,60 @@ export const ReviewVerdictSchema = z.object({
 export type ReviewVerdict = z.infer<typeof ReviewVerdictSchema>;
 
 // ─── Prompts ────────────────────────────────────────────────────────────────
+//
+// PROMPTS ARE WRITTEN FOR DAEMON AUDIENCE, NOT HUMAN READERS.
+//
+// Key rules:
+// - NEVER ask for confirmation or permission
+// - NEVER output "Next Steps", "Ready for review", or any waiting-for-human signal
+// - NEVER "stop and summarize" — the loop routes automatically
+// - If stuck, fail with a clear error, do not wait silently
+// - Output is consumed by the loop, not a human reading for guidance
 
-const PLANNER_PROMPT = `You are the Planning Agent (GPT). Given a feature request, produce a
-concise implementation plan: goals, non-goals, ordered steps, files to touch,
-risks, and a definition-of-done the reviewer can check against.
-Output markdown only.`;
-
-const CODER_PROMPT = `You are the Coding Agent (MiniMax). You receive a plan (and possibly
-review comments from a previous iteration) and produce complete, runnable code.
-Always output fenced code blocks with file-path headers. Address every review
-comment explicitly.`;
-
-const REVIEWER_PROMPT = `You are the Code Review Agent (GPT). You review the coder's output
-against the plan and the review checklist. Be strict but fair: request changes
-only for real problems. Respond via the structured schema.`;
-
-const SUPERVISOR_PROMPT = `You are the Supervisor (GPT). Coordinate the workflow:
-1. Call delegate_to_coder with the instruction (include review comments when fixing).
-2. Call delegate_to_reviewer with the plan + code to review.
-3. If the reviewer requests changes, call delegate_to_coder again with the comments.
-4. Repeat until approved or you judge further iterations unproductive (then stop and summarize).
-Keep the loop tight: no more than 4 coder-review rounds unless the task clearly needs more.`;
+
+const PLANNER_PROMPT = `You are the Planning Agent (GPT) in an autonomous daemon loop.
+
+Given a feature request, produce a concise implementation plan: goals, non-goals,
+ordered steps, files to touch, risks, and a definition-of-done the reviewer can
+check against.
+
+Output markdown only. Do NOT ask for confirmation. Do NOT output "Next Steps".
+The loop will route automatically.`;
+
+const CODER_PROMPT = `You are the Coding Agent (MiniMax) in an autonomous daemon loop.
+
+You receive a plan (and possibly review comments from a previous iteration) and
+produce complete, runnable code. Output fenced code blocks with file-path headers.
+Address every review comment explicitly.
+
+IMPORTANT:
+- Do NOT output "Next Steps", "Ready for review", or any waiting-for-human signal
+- Do NOT ask for confirmation or permission — write and finish
+- If you cannot proceed, fail with a clear error message
+- Your output is consumed by the loop, not read by a human for guidance`;
+
+
+const REVIEWER_PROMPT = `You are the Code Review Agent (GPT) in an autonomous daemon loop.
+
+You review the coder's output against the plan. Be strict: request changes only
+for real problems. Output a structured verdict — the loop reads it automatically.
+
+IMPORTANT:
+- Do NOT output "Next Steps" or "Looks good, waiting for approval"
+- Do NOT ask for confirmation — output the verdict and finish
+- If you need clarification, fail with a clear error, do not wait silently
+- Respond via the structured schema`;
+
+const SUPERVISOR_PROMPT = `You are the Supervisor (GPT) in an autonomous daemon loop.
+Coordinate the workflow automatically. The loop routes to the reviewer and back
+without human input.
+
+IMPORTANT:
+- Do NOT output "Next Steps" or any waiting-for-human signal
+- Do NOT ask for permission to continue — loop continuously
+- If truly blocked, call delegate_to_reviewer with verdict="blocked" and stop
+- Do NOT output a summary for a human — the loop reads the verdict
+- Keep iterating until approved or truly blocked, then output a verdict`;
 
 // ─── Agents ─────────────────────────────────────────────────────────────────
 

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)

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/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-harness-runtime",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "description": "[BETA] Codex-style /usage status + autonomous coding harness for pi. Not production ready — expect breaking changes.",
   "type": "module",
   "scripts": {

package/packages/a2a-adapter/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/a2a-adapter",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "type": "module",
   "description": "Google Agent-to-Agent (A2A) protocol adapter for pi-harness",
   "main": "dist/index.js",

package/packages/architecture-generator/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-harness/architecture-generator",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "description": "Generate ADR and Mermaid diagrams from project analysis",
   "type": "module",
   "main": "./dist/index.js",

package/packages/auth/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-harness/auth",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "description": "pi-harness-runtime — auth",
   "type": "module",
   "main": "./dist/index.js",

package/packages/autonomous-refactor/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/autonomous-refactor",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "type": "module",
   "description": "Autonomous refactoring engine",
   "main": "dist/index.js",

package/packages/autonomous-runtime/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/autonomous-runtime",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "description": "Autonomous Operations Runtime — task inbox, lease management, and worker lifecycle (RFC-0101)",
   "type": "module",
   "main": "./dist/index.js",

package/packages/cache-strategy/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/cache-strategy",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "type": "module",
   "description": "Smart caching with TTL and invalidation strategies (RFC-0022)",
   "main": "dist/index.js",

package/packages/capability-registry/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-harness/capability-registry",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "description": "Capability Registry for pi-harness-runtime — tracks model capabilities and enables capability-based routing",
   "type": "module",
   "main": "./dist/index.js",

package/packages/checkpoint/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/checkpoint",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "description": "Enhanced checkpoint system with incremental snapshots, diff-based storage, and recovery strategies",
   "type": "module",
   "main": "./dist/index.js",

package/packages/cli-plugin-sdk/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/cli-plugin-sdk",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "type": "module",
   "description": "CLI Plugin SDK for pi-harness — install, list, remove, update, invoke plugins",
   "main": "dist/index.js",

package/packages/clipboard/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/clipboard",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "type": "module",
   "main": "src/index.ts",
   "exports": {

package/packages/clipboard-plugin/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/clipboard-plugin",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "type": "module",
   "main": "src/index.ts",
   "exports": {

package/packages/code-generation/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/code-generation",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "description": "Code generation pipeline for pi-harness-runtime",
   "type": "module",
   "main": "./dist/index.js",

package/packages/code-review/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/code-review",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "description": "Automated code review engine for pi-harness-runtime",
   "type": "module",
   "main": "./dist/index.js",

package/packages/codex-adapter/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi/codex-adapter",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "type": "module",
   "description": "OpenAI Codex provider adapter for pi-harness",
   "main": "dist/index.js",

package/packages/config-capture/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@moocoding/config-capture",
-  "version": "1.1.41",
+  "version": "1.1.43",
   "description": "Auto-capture API-fetched configuration values to AGENTS.md",
   "type": "module",
   "main": "src/index.ts",