pi-task-dispatcher 3.1.0 → 3.2.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-task-dispatcher",
-  "version": "3.1.0",
+  "version": "3.2.0",
   "description": "Run configurable Pi worker models as supervised, observable, resumable jobs.",
   "type": "module",
   "keywords": [
@@ -26,7 +26,7 @@
   "scripts": {
     "test:gate": "node tests/gate-probe.mjs",
     "test:fake-rpc": "node tests/fake-rpc-check.mjs",
-    "check": "node tests/static-check.mjs && node tests/fake-rpc-check.mjs"
+    "check": "node tests/static-check.mjs && node tests/strict-jsonl-check.mjs && node tests/fake-rpc-check.mjs"
   },
   "peerDependencies": {
     "@earendil-works/pi-coding-agent": "*",

package/tests/fake-rpc-check.mjs

@@ -1,4 +1,4 @@
-/** Exercise WorkerJob lifecycle transitions against a deterministic fake JSONL/IPC child. */
+/** Exercise worker lifecycle, manager-proxied capabilities, and cleanup against a deterministic fake child. */
 
 import assert from "node:assert/strict";
 import { access, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
@@ -17,14 +17,71 @@ const projectRoot = path.resolve(import.meta.dirname, "..");
 const temporaryDirectory = await mkdtemp(path.join(tmpdir(), "pi-task-dispatcher-fake-rpc-"));
 const configPath = path.join(temporaryDirectory, "task-dispatcher.json");
 const promptMarker = path.join(temporaryDirectory, "prompts.txt");
+const observationPath = path.join(temporaryDirectory, "observations.jsonl");
 const inheritedWorkerRole = process.env.PI_TASK_DISPATCHER_WORKER;
 
+function usage(input, output) {
+	return {
+		input,
+		output,
+		cacheRead: 0,
+		cacheWrite: 0,
+		totalTokens: input + output,
+		cost: { input: input / 1000, output: output / 1000, cacheRead: 0, cacheWrite: 0, total: (input + output) / 1000 },
+	};
+}
+
+function tool(name, label, description, annotations, executionMode = "parallel") {
+	return {
+		name,
+		label,
+		description,
+		parameters: { type: "object", additionalProperties: false, properties: { path: { type: "string" }, action: { type: "string" }, value: { type: "string" } } },
+		outputSchema: { type: "object", additionalProperties: true },
+		executionMode,
+		annotations,
+	};
+}
+
+function terminal(status) {
+	return new Set(["completed", "failed", "cancelled", "checkpointed", "taken_over", "hard_stopped"]).has(status);
+}
+
+async function observations() {
+	const text = await readFile(observationPath, "utf8").catch(() => "");
+	return text.split("\n").filter(Boolean).map((line) => JSON.parse(line));
+}
+
+async function clearObservations() {
+	await writeFile(observationPath, "", "utf8");
+}
+
+async function waitForObservation(predicate, label) {
+	const deadline = Date.now() + 3_000;
+	while (Date.now() < deadline) {
+		const match = (await observations()).find(predicate);
+		if (match) return match;
+		await new Promise((resolve) => setTimeout(resolve, 20));
+	}
+	throw new Error(`Timed out waiting for ${label}.`);
+}
+
+async function waitFor(predicate, label) {
+	const deadline = Date.now() + 3_000;
+	while (Date.now() < deadline) {
+		if (predicate()) return;
+		await new Promise((resolve) => setTimeout(resolve, 20));
+	}
+	throw new Error(`Timed out waiting for ${label}.`);
+}
+
 // The test loads the parent extension even when invoked from a delegated review worker.
 delete process.env.PI_TASK_DISPATCHER_WORKER;
 process.env.PI_TASK_DISPATCHER_TEST_MODE = "1";
 process.env.PI_TASK_DISPATCHER_CONFIG_PATH = configPath;
 process.env.PI_TASK_DISPATCHER_CLI_OVERRIDE = path.join(projectRoot, "tests", "fake-rpc-worker.mjs");
 process.env.PI_TASK_DISPATCHER_FAKE_PROMPT_MARKER = promptMarker;
+process.env.PI_TASK_DISPATCHER_FAKE_OBSERVATION_MARKER = observationPath;
 
 try {
 	await writeFile(configPath, "{}\n", "utf8");
@@ -38,22 +95,47 @@ try {
 	});
 	const loaded = await jiti.import(path.join(projectRoot, "extensions", "task-dispatcher.ts"));
 	const register = loaded.default ?? loaded;
-	const tools = new Map();
+	const registeredTools = new Map();
 	const handlers = new Map();
+	const managerTools = [
+		tool("read", "Read", "Manager-controlled read override", { readOnlyHint: true }, "parallel"),
+		tool("bash", "Bash", "Manager-controlled command", { destructiveHint: true }, "sequential"),
+		tool("edit", "Edit", "Manager-controlled edit", { destructiveHint: true }, "sequential"),
+		tool("write", "Write", "Manager-controlled write", { destructiveHint: true }, "sequential"),
+		tool("codemode", "Codemode", "Child-local codemode capability", { readOnlyHint: false }, "sequential"),
+		tool("extension_tool", "Extension tool", "An extension-defined callable tool", { destructiveHint: true }, "parallel"),
+		tool("rich_tool", "Rich tool", "A proxy serialization exercise", { destructiveHint: true }, "parallel"),
+	];
+	const managerMetadata = managerTools.map((entry) => ({
+		name: entry.name,
+		description: entry.description,
+		parameters: entry.parameters,
+		namespace: entry.name === "extension_tool" ? { name: "extension", description: "Manager extension namespace", instructions: "Use exact manager arguments." } : undefined,
+		annotations: entry.annotations,
+		promptGuidelines: [`${entry.name} is controlled by the manager.`],
+		sourceInfo: { path: `<fake:${entry.name}>`, source: "fake", scope: "temporary", origin: "top-level", baseDir: undefined },
+	}));
+	managerMetadata.push({ name: "registered_only", description: "Registered but excluded from this invocation", annotations: { readOnlyHint: true }, promptGuidelines: [], sourceInfo: { path: "<fake:registered-only>", source: "fake", scope: "temporary", origin: "top-level" } });
 	const pi = {
-		registerTool(tool) { tools.set(tool.name, tool); },
-		getActiveTools() { return ["read", "bash", "edit", "write", "codemode"]; },
+		registerTool(entry) { registeredTools.set(entry.name, entry); },
+		getActiveTools() { return managerTools.map((entry) => entry.name); },
+		getAllTools() { return managerMetadata; },
 		on(event, handler) { handlers.set(event, handler); },
 	};
 	register(pi);
 
 	const activeModel = { provider: "fake-provider", id: "active-model", contextWindow: 272_000, maxTokens: 128_000 };
 	const statusUpdates = [];
+	const parentExecutions = [];
+	const transitiveDecisions = [];
+	let activeParentExecutions = 0;
+	let maximumConcurrentParentExecutions = 0;
 	const context = {
 		cwd: projectRoot,
 		hasUI: true,
 		model: activeModel,
 		thinkingLevel: "medium",
+		get tools() { return managerTools.filter((tool) => tool.name !== "codemode"); },
 		ui: {
 			theme: { fg(_color, text) { return text; } },
 			setWidget() {},
@@ -63,20 +145,79 @@ try {
 			find(provider, id) { return { provider, id, contextWindow: 272_000, maxTokens: 128_000 }; },
 			hasConfiguredAuth() { return true; },
 		},
+		async executeTool(name, args, options) {
+			const call = { name, args, options, aborted: false, usage: undefined };
+			parentExecutions.push(call);
+			activeParentExecutions++;
+			maximumConcurrentParentExecutions = Math.max(maximumConcurrentParentExecutions, activeParentExecutions);
+			try {
+				if (process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO === "transitive" && name === "extension_tool") {
+					const authorize = handlers.get("tool_call");
+					transitiveDecisions.push(
+						authorize?.({ type: "tool_call", toolCallId: "start-transitive/1", parentToolCallId: "start-transitive", toolName: "extension_tool", input: args }, context),
+						authorize?.({ type: "tool_call", toolCallId: "start-transitive/1/1", parentToolCallId: "start-transitive/1", toolName: "read", input: { path: "allowed.txt" } }, context),
+						authorize?.({ type: "tool_call", toolCallId: "start-transitive/1/2", parentToolCallId: "start-transitive/1", toolName: "edit", input: { path: "blocked.txt" } }, context),
+					);
+				}
+				if (args?.path === "wait-for-abort.txt") {
+					await new Promise((resolve) => {
+						if (options.signal.aborted) return resolve();
+						options.signal.addEventListener("abort", resolve, { once: true });
+					});
+					call.aborted = true;
+					return { toolCall: { id: "aborted" }, isError: true, result: { content: [{ type: "text", text: "manager operation cancelled" }] } };
+				}
+				if (args?.path === "presentation.txt") {
+					options.onUpdate?.({ content: [{ type: "text", text: "partial read result that must be replaced" }], details: { internal: "hidden" }, usage: usage(99, 99) });
+					await new Promise((resolve) => setTimeout(resolve, 150));
+					const result = { content: [{ type: "text", text: "final read result six red" }], details: { internal: "hidden" }, usage: usage(3, 1) };
+					call.usage = result.usage;
+					return { toolCall: { id: "presentation" }, isError: false, result };
+				}
+				if (name === "rich_tool") {
+					options.onUpdate?.({ content: [{ type: "text", text: "partial rich result" }] });
+					const result = {
+						content: [{ type: "text", text: "rich manager result" }, { type: "image", data: "ZmFrZQ==", mimeType: "image/png" }],
+						structuredContent: { source: "manager", count: 1 },
+						terminate: true,
+						usage: usage(3, 1),
+					};
+					call.usage = result.usage;
+					return { toolCall: { id: "rich" }, isError: true, result };
+				}
+				if (process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO === "concurrent") await new Promise((resolve) => setTimeout(resolve, 40));
+				const result = {
+					content: [{ type: "text", text: name === "extension_tool" ? "manager extension override" : `manager read override: ${args?.path ?? args?.action ?? "unknown"}` }],
+					details: { managerHook: true },
+					usage: usage(3, 1),
+				};
+				call.usage = result.usage;
+				return { toolCall: { id: `${name}-${parentExecutions.length}` }, isError: false, result };
+			} finally {
+				activeParentExecutions--;
+			}
+		},
 	};
 	const workload = { investigationUnits: 1, changeUnits: 0, verificationUnits: 1, expectedLongRunningSeconds: 0 };
-	const start = tools.get("delegate_engineering");
-	const designerStart = tools.get("delegate_designer");
-	const control = tools.get("delegate_control");
+	const start = registeredTools.get("delegate_engineering");
+	const designerStart = registeredTools.get("delegate_designer");
+	const control = registeredTools.get("delegate_control");
 	assert(start && designerStart && control, "Expected supervised tools to register.");
-	assert(tools.has("delegate_mechanical") && tools.has("delegate_assistant"), "Expected every worker role to register.");
+	assert(registeredTools.has("delegate_mechanical") && registeredTools.has("delegate_assistant"), "Expected every worker role to register.");
+	const schemaGuidance = registeredTools.get("delegate_mechanical").parameters.properties.tools.description;
+	assert.match(schemaGuidance, /codemode.*allowlist.*direct tool.*otherwise omit/i, "Schema guidance no longer describes the codemode delegation rule.");
+	for (const name of ["delegate_mechanical", "delegate_assistant", "delegate_engineering", "delegate_designer"]) {
+		const delegate = registeredTools.get(name);
+		assert.equal(delegate.parameters.properties.tools.description, schemaGuidance, `${name} does not expose shared codemode guidance.`);
+		assert(delegate.promptGuidelines.includes(schemaGuidance), `${name} does not expose manager codemode guidance.`);
+	}
 	const theme = { fg(_color, text) { return text; }, bold(text) { return text; } };
 	let rendererInvalidations = 0;
 	const renderContext = { args: {}, state: {}, invalidate() { rendererInvalidations++; }, cwd: projectRoot };
 	const callComponent = start.renderCall({}, theme, renderContext);
 	assert.doesNotThrow(() => callComponent.render(80), "Partial streamed arguments must never crash the renderer.");
 
-	// A normal fake job must pause, account assistant and tool-result usage once, then resume to completion.
+	// A normal job proves built-in names resolve through the manager context, without merging nested usage into child output.
 	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "normal";
 	let result = await start.execute("start-normal", { task: "fake normal job", tools: ["read"], workload }, undefined, undefined, context);
 	assert.equal(result.isError, false);
@@ -86,34 +227,26 @@ try {
 	assert.equal(result.details.sequence, 1);
 	assert.equal(result.details.turns, 1);
 	assert.equal(result.details.actionLedger.length, 1);
-	assert.equal(result.usage.input, 13);
-	assert.equal(result.usage.output, 3);
-	assert.equal(result.usage.totalTokens, 16);
+	assert.equal(result.details.actionLedger[0].mutating, false, "Read-only annotations must classify read as non-mutating.");
+	assert.equal(result.usage.input, 10, "Nested manager usage must not be sent back into child accounting.");
+	assert.equal(result.usage.output, 2);
+	assert.equal(parentExecutions.at(-1).name, "read");
+	assert.equal(parentExecutions.at(-1).usage.totalTokens, 4, "The parent manager hook must observe nested usage exactly once.");
+	assert.equal(parentExecutions.at(-1).args.path, "package.json", "Built-in name must route through manager override arguments.");
 	const partialRender = start.renderResult(result, { expanded: false, isPartial: true }, theme, renderContext).render(100).join("\n");
 	assert.match(partialRender, /Last: read package\.json/);
 	assert.match(partialRender, /Elapsed:/);
 	assert.equal(statusUpdates.at(-1)?.text, "engineer-1 · 11.0%/1k • active-model • medium");
 	const resultComponent = start.renderResult(result, { expanded: false, isPartial: false }, theme, renderContext);
 	renderContext.lastComponent = resultComponent;
-	const reusedResultComponent = start.renderResult(result, { expanded: false, isPartial: false }, theme, renderContext);
-	assert.equal(reusedResultComponent, resultComponent, "The result renderer did not reuse Pi's last component.");
-	const firstRender = reusedResultComponent.render(100).join("\n");
-	assert.match(firstRender, /Took:/);
-	assert.doesNotMatch(firstRender, /Elapsed:/);
-	for (let index = 0; index < 1_000; index++) {
-		const repeatedComponent = start.renderResult(result, { expanded: false, isPartial: false }, theme, renderContext);
-		assert.equal(repeatedComponent.render(100).join("\n"), firstRender, "Repeated render output changed.");
-		assert.match(callComponent.render(100).join("\n"), /engineer-1: paused/);
-	}
-	assert.equal(rendererInvalidations, 0, "A renderer invalidated its own render pass, risking an infinite loop.");
-	assert.match(callComponent.render(100).join("\n"), /engineer-1: paused/, "The shared call header did not update to paused.");
+	assert.equal(start.renderResult(result, { expanded: false, isPartial: false }, theme, renderContext), resultComponent, "The result renderer did not reuse Pi's last component.");
+	const firstRender = resultComponent.render(100).join("\n");
+	for (let index = 0; index < 100; index++) assert.equal(start.renderResult(result, { expanded: false, isPartial: false }, theme, renderContext).render(100).join("\n"), firstRender);
+	assert.equal(rendererInvalidations, 0, "A renderer invalidated its own render pass.");
 	const normalJobId = result.details.jobId;
 	const managerArgs = { jobId: normalJobId, action: "continue", actionBatches: 1 };
 	const managerRenderContext = { args: managerArgs, state: {}, invalidate() { rendererInvalidations++; }, cwd: projectRoot };
-	const managerCall = control.renderCall(managerArgs, theme, managerRenderContext);
-	assert.match(managerCall.render(100).join("\n"), /manager: engineer-1 → continue/);
-	assert.doesNotMatch(managerCall.render(100).join("\n"), /continue\//);
-
+	assert.match(control.renderCall(managerArgs, theme, managerRenderContext).render(100).join("\n"), /manager: engineer-1 → continue/);
 	result = await control.execute("continue-normal", managerArgs, undefined, undefined, context);
 	assert.equal(result.isError, false);
 	assert.equal(result.details.status, "completed");
@@ -121,13 +254,8 @@ try {
 	assert.equal(result.details.workerText, "deterministic worker complete");
 	assert.equal(result.usage.input, 5, "Continuation must return only newly observed usage.");
 	assert.equal(result.usage.totalTokens, 8);
-	assert.equal(statusUpdates.at(-1)?.text, "engineer-1 · 12.0%/1k • active-model • medium");
-	assert(!statusUpdates.some((update) => update.text === undefined), "Worker status must remain visible after it first appears.");
-	control.renderResult(result, { expanded: false, isPartial: false }, theme, managerRenderContext).render(100);
-	assert.match(managerCall.render(100).join("\n"), /manager: engineer-1 → completed/);
-	assert.doesNotMatch(managerCall.render(100).join("\n"), /continue\/completed/);
 
-	// The designer must load only the selected native codemode-only extension and account nested usage once.
+	// Codemode remains local while only the manager-selected non-codemode capability is proxied.
 	await writeFile(configPath, JSON.stringify({ designer: { model: "media-provider/media-model", thinking: "high" } }), "utf8");
 	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "codemode";
 	result = await designerStart.execute("start-designer", { task: "fake image job", tools: ["read", "codemode"], workload }, undefined, undefined, context);
@@ -140,150 +268,216 @@ try {
 	assert.equal(result.details.actionCounts.mutations, 1);
 	assert.equal(result.details.recentActions.at(-1).toolName, "codemode");
 	assert.deepEqual(result.details.recentActions.at(-1).mediaOutputs, [{ mimeType: "image/png", bytes: 4 }]);
-	assert.match(result.details.recentActions.at(-1).label, /1 image/);
-	assert.doesNotMatch(JSON.stringify(result.details), /ZmFrZQ==/, "Snapshots must not retain base64 media payloads.");
-	assert.equal(result.usage.input, 20, "Each nested and parent execution must contribute its own pre-merge usage exactly once.");
-	assert.equal(result.usage.output, 6);
-	assert.equal(result.usage.totalTokens, 26);
-	assert.equal(
-		statusUpdates.at(-1)?.text,
-		"designer-2 · 11.0%/1k • media-model • high | engineer-1 · 12.0%/1k • active-model • medium",
-		"Starting another role must retain the previous role's footer status.",
-	);
-	result = await control.execute(
-		"continue-designer",
-		{ jobId: result.details.jobId, action: "continue", actionBatches: 1 },
-		undefined,
-		undefined,
-		context,
-	);
+	assert.equal(result.usage.input, 17, "Child-local codemode usage remains child usage; proxied manager usage does not duplicate.");
+	assert.equal(result.usage.output, 5);
+	result = await control.execute("continue-designer", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
 	assert.equal(result.details.status, "completed");
-	assert.match(statusUpdates.at(-1)?.text ?? "", /designer-2.*\|.*engineer-1/);
 	await writeFile(configPath, "{}\n", "utf8");
 
-	// Streaming RPC updates must replace their partial text, preserve full UI details, and keep model output bounded.
+	// Parent streaming updates replace partial presentation text and preserve the bounded/expanded display contract.
 	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "presentation";
 	const presentationUpdates = [];
-	result = await start.execute(
-		"start-presentation",
-		{ task: "fake presentation job", tools: ["read"], workload },
-		undefined,
-		(update) => presentationUpdates.push(update.details?.presentationText ?? ""),
-		context,
-	);
+	result = await start.execute("start-presentation", { task: "fake presentation job", tools: ["read"], workload }, undefined, (update) => presentationUpdates.push(update.details?.presentationText ?? ""), context);
 	assert(presentationUpdates.some((text) => text.includes("stream delta that must be replaced")), "Live assistant stream text was not presented.");
-	assert(presentationUpdates.some((text) => text.includes("partial read result that must be replaced")), "Live tool partial text was not presented.");
-	assert.equal(result.isError, false);
+	assert(presentationUpdates.some((text) => text.includes("partial read result that must be replaced")), "Manager-proxied streaming update was not presented.");
 	assert.equal(result.details.status, "paused");
 	assert.match(result.details.presentationText, /authoritative first assistant message/);
-	assert.match(result.details.presentationText, /tool start: read presentation\.txt/);
-	assert.match(result.details.presentationText, /tool completed: read presentation\.txt/);
 	assert.match(result.details.presentationText, /final read result six red/);
 	assert.doesNotMatch(result.details.presentationText, /stream delta that must be replaced|stream block that must be replaced|partial read result that must be replaced/);
-	assert.doesNotMatch(result.details.presentationText, /\u001b/, "Presentation text must not retain terminal control sequences.");
 	const presentationRenderContext = { args: {}, state: {}, invalidate() { rendererInvalidations++; }, cwd: projectRoot };
 	const collapsedComponent = start.renderResult(result, { expanded: false, isPartial: false }, theme, presentationRenderContext);
 	presentationRenderContext.lastComponent = collapsedComponent;
 	const collapsed = collapsedComponent.render(100).join("\n");
 	assert.match(collapsed, /earlier visual rows omitted/);
-	assert.match(collapsed, /to expand/);
 	assert.match(collapsed, /final read result six red/);
-	assert.doesNotMatch(collapsed, /authoritative first assistant message/, "Collapsed output must retain only the latest visual rows.");
-
-	result = await control.execute(
-		"continue-presentation",
-		{ jobId: result.details.jobId, action: "continue", actionBatches: 1 },
-		undefined,
-		undefined,
-		context,
-	);
-	assert.equal(result.isError, false);
+	result = await control.execute("continue-presentation", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
 	assert.equal(result.details.status, "completed");
 	assert(result.details.workerTextDisplay.length <= 5_000, "Model-facing worker text was not bounded.");
 	assert(result.details.presentationText.length > 5_000, "Structured presentation text was unexpectedly bounded.");
 	assert.match(result.details.presentationText, /MIDDLE-ONLY-IN-EXPANDED-PRESENTATION/);
-	assert.doesNotMatch(result.content[0].text, /MIDDLE-ONLY-IN-EXPANDED-PRESENTATION/, "Model-facing state exposed the omitted middle of a long report.");
-	const expandedComponent = control.renderResult(result, { expanded: true, isPartial: false }, theme, presentationRenderContext);
-	presentationRenderContext.lastComponent = expandedComponent;
-	const expanded = expandedComponent.render(100).join("\n");
-	assert.match(expanded, /MIDDLE-ONLY-IN-EXPANDED-PRESENTATION/);
-	assert.match(expanded, /latest-presentation-marker/);
-	const reusedExpandedComponent = control.renderResult(result, { expanded: true, isPartial: false }, theme, presentationRenderContext);
-	assert.equal(reusedExpandedComponent, expandedComponent, "Expanded renderer state was not reused.");
-	assert.equal(reusedExpandedComponent.render(100).join("\n"), expanded, "Repeated expanded rendering changed output.");
+	assert.doesNotMatch(result.content[0].text, /MIDDLE-ONLY-IN-EXPANDED-PRESENTATION/);
 
-	// A tool-free output-limit stop must preserve the session so the manager can recover the report.
+	// Extension-defined tools are accepted when callable in ctx.tools and expose only the exact selected proxy surface.
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "extension";
+	result = await start.execute("start-extension", { task: "use extension capability", tools: ["extension_tool"], workload }, undefined, undefined, context);
+	assert.equal(result.isError, false);
+	assert.equal(result.details.status, "paused");
+	assert.equal(parentExecutions.at(-1).name, "extension_tool");
+	assert.match(result.details.presentationText, /manager extension override/);
+	await control.execute("continue-extension", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
+
+	// Transitive manager calls remain inside the exact selected grant, including through extension tools.
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "transitive";
+	result = await start.execute("start-transitive", { task: "enforce transitive capabilities", tools: ["extension_tool", "read"], workload }, undefined, undefined, context);
+	assert.equal(result.details.status, "paused");
+	transitiveDecisions.push(handlers.get("tool_call")?.({ type: "tool_call", toolCallId: "start-transitive/1/3", parentToolCallId: "start-transitive/1", toolName: "read", input: { path: "late.txt" } }, context));
+	assert.equal(transitiveDecisions.length, 4);
+	assert.equal(transitiveDecisions[0], undefined, "The selected first-level proxy call was blocked.");
+	assert.equal(transitiveDecisions[1], undefined, "A selected transitive manager call was blocked.");
+	assert.equal(transitiveDecisions[2]?.block, true, "An unselected transitive manager call was not blocked.");
+	assert.match(transitiveDecisions[2]?.reason ?? "", /outside this delegation's selected capability grant/i);
+	assert.equal(transitiveDecisions[3]?.block, true, "A delayed transitive call survived executor-lease revocation.");
+	assert.match(transitiveDecisions[3]?.reason ?? "", /executor lease is no longer active/i);
+	await control.execute("continue-transitive", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
+
+	// Proxy forwarding preserves streamed content, image and structured output, isError, terminate, and strips nested usage/details.
+	await clearObservations();
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "proxy-rich";
+	result = await start.execute("start-rich", { task: "exercise proxy result transport", tools: ["rich_tool"], workload }, undefined, undefined, context);
+	assert.equal(result.details.status, "paused");
+	assert.equal(result.details.recentActions.at(-1).status, "failed");
+	assert.deepEqual(result.details.recentActions.at(-1).mediaOutputs, [{ mimeType: "image/png", bytes: 4 }]);
+	const rich = await waitForObservation((entry) => entry.kind === "rich", "rich proxy observation");
+	assert.equal(rich.updates, 1);
+	assert.equal(rich.hasImage, true);
+	assert.deepEqual(rich.structuredContent, { source: "manager", count: 1 });
+	assert.equal(rich.isError, true);
+	assert.equal(rich.terminate, true);
+	assert.equal(rich.hasUsage, false, "Nested usage must be stripped before delivery to the child.");
+	await control.execute("continue-rich", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
+
+	// Multiple child proxy requests can execute concurrently through the manager hook.
+	maximumConcurrentParentExecutions = 0;
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "concurrent";
+	result = await start.execute("start-concurrent", { task: "run two manager tools", tools: ["read", "extension_tool"], workload }, undefined, undefined, context);
+	assert.equal(result.details.status, "paused");
+	assert.equal(maximumConcurrentParentExecutions, 2, "Independent proxy requests were serialized instead of using manager execution concurrency.");
+	await control.execute("continue-concurrent", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
+
+	// Reject spoofed job nonces, unknown names, replayed ids, and post-boundary requests without invoking unauthorized manager tools.
+	await clearObservations();
+	const securityCallsBefore = parentExecutions.length;
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "security";
+	result = await start.execute("start-security", { task: "probe private proxy validation", tools: ["read"], workload }, undefined, undefined, context);
+	assert.equal(result.details.status, "paused");
+	const activeSecurity = await waitForObservation((entry) => entry.kind === "security-active", "active security rejections");
+	assert.match(activeSecurity.cross, /nonce is invalid/i);
+	assert.match(activeSecurity.unknown, /not in this job's selected capability grant/i);
+	assert.match(activeSecurity.first, /manager read override: once\.txt/);
+	assert.match(activeSecurity.replay, /request id was replayed/i);
+	const pausedSecurity = await waitForObservation((entry) => entry.kind === "security-paused", "paused security rejection");
+	assert.match(pausedSecurity.text, /no bounded manager executor lease is active/i);
+	assert.equal(parentExecutions.length - securityCallsBefore, 1, "Only the first valid security probe may reach ctx.executeTool().");
+	await control.execute("continue-security", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
+
+	// Continuations fail closed if current manager callable definitions no longer match the original grant.
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "normal";
+	result = await start.execute("start-revalidate", { task: "revalidate manager definition", tools: ["read"], workload }, undefined, undefined, context);
+	const originalReadParameters = managerTools.find((entry) => entry.name === "read").parameters;
+	managerTools.find((entry) => entry.name === "read").parameters = { ...originalReadParameters, properties: { ...originalReadParameters.properties, changed: { type: "boolean" } } };
+	const revalidated = await control.execute("continue-revalidate", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
+	assert.equal(revalidated.isError, true);
+	assert.equal(revalidated.details.status, "hard_stopped");
+	assert.match(revalidated.content[0].text, /manifest revalidation failed/i);
+	managerTools.find((entry) => entry.name === "read").parameters = originalReadParameters;
+	result = await start.execute("start-revalidate-identity", { task: "revalidate manager identity", tools: ["read"], workload }, undefined, undefined, context);
+	const readIndex = managerTools.findIndex((entry) => entry.name === "read");
+	const originalReadTool = managerTools[readIndex];
+	managerTools[readIndex] = { ...originalReadTool };
+	const identityRevalidated = await control.execute("continue-revalidate-identity", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
+	assert.equal(identityRevalidated.isError, true);
+	assert.equal(identityRevalidated.details.status, "hard_stopped");
+	assert.match(identityRevalidated.content[0].text, /original callable identity/i);
+	managerTools[readIndex] = originalReadTool;
+
+	// Revalidate again at execution time so a replacement during an active lease cannot win a race.
+	await clearObservations();
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "identity-race";
+	const identityRace = start.execute("start-identity-race", { task: "reject a mid-lease tool replacement", tools: ["read"], workload }, undefined, undefined, context);
+	await waitForObservation((entry) => entry.kind === "identity-race-ready", "identity-race worker readiness");
+	managerTools[readIndex] = { ...originalReadTool };
+	result = await identityRace;
+	assert.equal(result.isError, true);
+	assert.equal(result.details.status, "hard_stopped");
+	assert.match(result.content[0].text, /grant changed before read execution|original callable identity/i);
+	managerTools[readIndex] = originalReadTool;
+
+	// Fail closed on non-JSON-safe callable schemas, recursive names, and registered-but-noncallable requests.
+	managerTools.find((entry) => entry.name === "read").parameters = { type: "object", invalid: () => undefined };
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "normal";
+	result = await start.execute("start-invalid-schema", { task: "reject unsafe schema", tools: ["read"], workload }, undefined, undefined, context);
+	assert.equal(result.isError, true);
+	assert.match(result.content[0].text, /not JSON-serializable/i);
+	managerTools.find((entry) => entry.name === "read").parameters = originalReadParameters;
+	result = await start.execute("start-recursive", { task: "reject recursive delegation", tools: ["delegate_engineering"], workload }, undefined, undefined, context);
+	assert.equal(result.isError, true);
+	assert.match(result.content[0].text, /cannot be delegated recursively/i);
+	result = await start.execute("start-noncallable", { task: "reject unavailable tool", tools: ["registered_only"], workload }, undefined, undefined, context);
+	assert.equal(result.isError, true);
+	assert.match(result.content[0].text, /registered but not callable/i);
+
+	// Existing lifecycle fail-closed paths remain enforced with the manager proxy present.
 	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "length";
 	result = await start.execute("start-length", { task: "fake truncated job", tools: ["read"], workload }, undefined, undefined, context);
-	assert.equal(result.isError, false);
 	assert.equal(result.details.status, "paused");
-	assert.equal(result.details.turns, 1);
-	assert.equal(result.details.workerText, "truncated worker report");
 	assert.match(result.details.phase, /output limit/i);
-	assert.match(result.details.recommendation, /continue to recover/i);
-	result = await control.execute(
-		"continue-length",
-		{ jobId: result.details.jobId, action: "continue", actionBatches: 1 },
-		undefined,
-		undefined,
-		context,
-	);
-	assert.equal(result.isError, false);
+	result = await control.execute("continue-length", { jobId: result.details.jobId, action: "continue", actionBatches: 1 }, undefined, undefined, context);
 	assert.equal(result.details.status, "completed");
-	assert.equal(result.details.workerText, "deterministic worker complete");
-
-	// A handled prompt starts no agent run and must fail promptly rather than waiting for settlement.
 	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "handled";
-	const handledStartedAt = Date.now();
 	result = await start.execute("start-handled", { task: "fake handled job", tools: ["read"], workload }, undefined, undefined, context);
 	assert.equal(result.isError, true);
-	assert.equal(result.details.status, "hard_stopped");
 	assert.match(result.content[0].text, /intercepted by an input handler/i);
-	assert(Date.now() - handledStartedAt < 5_000, "Handled prompt should not wait for agent settlement.");
-
-	// A tool-use turn without the private IPC signal must fail closed instead of being mistaken for completion.
 	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "missing-boundary";
 	result = await start.execute("start-missing-boundary", { task: "fake missing boundary job", tools: ["read"], workload }, undefined, undefined, context);
 	assert.equal(result.isError, true);
-	assert.equal(result.details.status, "hard_stopped");
 	assert.match(result.content[0].text, /without an enforced IPC boundary/i);
-
-	// A stats timeout after a valid boundary must remain terminal rather than resurrecting as paused.
 	await writeFile(configPath, JSON.stringify({ policy: { absoluteMaxSeconds: 5, noProgressSeconds: 1 } }), "utf8");
 	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "stats-timeout";
 	result = await start.execute("start-stats-timeout", { task: "fake stats timeout job", tools: ["read"], workload }, undefined, undefined, context);
 	assert.equal(result.isError, true);
-	assert.equal(result.details.status, "hard_stopped");
 	assert.notEqual(result.details.status, "paused");
-
-	// A silent prompt must terminate the worker at the configured active-time ceiling.
 	await writeFile(configPath, JSON.stringify({ policy: { absoluteMaxSeconds: 1, noProgressSeconds: 1 } }), "utf8");
 	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "timeout";
-	const timeoutStartedAt = Date.now();
 	result = await start.execute("start-timeout", { task: "fake timeout job", tools: ["read"], workload }, undefined, undefined, context);
 	assert.equal(result.isError, true);
 	assert.equal(result.details.status, "hard_stopped");
-	assert.match(result.content[0].text, /(deadline|timed out|closed)/i);
-	assert(Date.now() - timeoutStartedAt < 5_000, "Silent prompt should be bounded by test policy.");
-
-	// An already-aborted parent operation may initialize the child but must never submit task work.
 	await writeFile(configPath, "{}\n", "utf8");
+
+	// Parent cancellation aborts the nested manager execution and child exit cannot leave a live executor lease.
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "cancel-proxy";
+	const cancellation = new AbortController();
+	const cancellationResult = start.execute("start-cancel-proxy", { task: "cancel nested manager work", tools: ["read"], workload }, cancellation.signal, undefined, context);
+	await waitFor(() => parentExecutions.some((call) => call.args?.path === "wait-for-abort.txt"), "nested manager execution start");
+	cancellation.abort();
+	result = await cancellationResult;
+	assert.equal(result.isError, true);
+	assert.equal(result.details.status, "cancelled");
+	await waitFor(() => parentExecutions.some((call) => call.args?.path === "wait-for-abort.txt" && call.aborted), "nested manager execution abort");
+	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "child-exit";
+	result = await start.execute("start-child-exit", { task: "child exits", tools: ["read"], workload }, undefined, undefined, context);
+	assert.equal(result.isError, true);
+	assert(terminal(result.details.status), "Child exit left a nonterminal worker state.");
+
+	// An already-aborted operation never submits task work, and session shutdown terminates an idle child.
 	await rm(promptMarker, { force: true });
 	process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO = "normal";
-	const controller = new AbortController();
-	controller.abort();
-	result = await start.execute("start-aborted", { task: "must not be submitted", tools: ["read"], workload }, controller.signal, undefined, context);
+	const alreadyAborted = new AbortController();
+	alreadyAborted.abort();
+	result = await start.execute("start-aborted", { task: "must not be submitted", tools: ["read"], workload }, alreadyAborted.signal, undefined, context);
 	assert.equal(result.isError, true);
 	assert.equal(result.details.status, "cancelled");
 	await assert.rejects(access(promptMarker), /ENOENT/, "An already-aborted operation submitted a prompt.");
-
+	await clearObservations();
+	result = await start.execute("start-shutdown", { task: "session cleanup", tools: ["read"], workload }, undefined, undefined, context);
+	assert.equal(result.details.status, "paused");
+	const workerStart = await waitForObservation((entry) => entry.kind === "worker-start" && Number.isInteger(entry.pid), "session shutdown child identity");
 	await handlers.get("session_shutdown")?.({}, context);
-	console.log("Deterministic fake-RPC lifecycle checks passed.");
+	await waitFor(() => {
+		try {
+			process.kill(workerStart.pid, 0);
+			return false;
+		} catch {
+			return true;
+		}
+	}, "session shutdown child termination");
+
+	console.log("Deterministic fake-RPC lifecycle and manager-proxy checks passed.");
 } finally {
 	delete process.env.PI_TASK_DISPATCHER_TEST_MODE;
 	delete process.env.PI_TASK_DISPATCHER_CONFIG_PATH;
 	delete process.env.PI_TASK_DISPATCHER_CLI_OVERRIDE;
 	delete process.env.PI_TASK_DISPATCHER_FAKE_PROMPT_MARKER;
+	delete process.env.PI_TASK_DISPATCHER_FAKE_OBSERVATION_MARKER;
 	delete process.env.PI_TASK_DISPATCHER_FAKE_SCENARIO;
 	if (inheritedWorkerRole === undefined) delete process.env.PI_TASK_DISPATCHER_WORKER;
 	else process.env.PI_TASK_DISPATCHER_WORKER = inheritedWorkerRole;