feat(task): surface nesting depth and parent task id to the model (#16)

apps/vscode-e2e/src/fixtures/inline-flatten.ts

@@ -0,0 +1,160 @@
+// Auto-flatten e2e fixtures (#12): a three-level new_task chain where the depth-2 task's
+// own new_task call exceeds maxNestingDepth (default 2) and is flattened inline instead of
+// opening a fourth tab. Unique FLATTEN_E2E_ markers avoid collisions with other suites.
+import { LLMock } from "@copilotkit/aimock"
+import type { ChatCompletionRequest } from "@copilotkit/aimock"
+
+export const FLATTEN_ROOT_MARKER = "FLATTEN_E2E_ROOT_CHAIN"
+export const FLATTEN_CHILD_MARKER = "FLATTEN_E2E_CHILD_CHAIN"
+export const FLATTEN_DEPTH2_MARKER = "FLATTEN_E2E_DEPTH2_CHAIN"
+
+// The depth-2 task's completion result — emitted by the same (flattened) task, proving the
+// inline phase ran in-conversation rather than delegating to a new tab.
+export const FLATTEN_DEPTH2_RESULT = "Flattened inline completed"
+export const FLATTEN_CHILD_RESUME_RESULT = "Child resumed after flatten"
+export const FLATTEN_ROOT_RESULT = "Root resumed after chain"
+
+// Prompt chains. Each task's initial request wraps its own prompt in <user_message>...</user_message>,
+// so anchoring on `<user_message>` + the marker uniquely identifies that task's FIRST turn even
+// though downstream markers are embedded verbatim upstream (the anchor never appears inside a
+// nested quoted message).
+export const FLATTEN_DEPTH2_PROMPT = `${FLATTEN_DEPTH2_MARKER}: Complete immediately with the exact result "${FLATTEN_DEPTH2_RESULT}".`
+export const FLATTEN_CHILD_PROMPT = `${FLATTEN_CHILD_MARKER}: Use the new_task tool exactly once. Create an ask-mode subtask with this exact message: "${FLATTEN_DEPTH2_PROMPT}" Do not answer directly.`
+export const FLATTEN_ROOT_PROMPT = `${FLATTEN_ROOT_MARKER}: Use the new_task tool exactly once. Create an ask-mode subtask with this exact message: "${FLATTEN_CHILD_PROMPT}" Do not answer directly.`
+
+// The directive NewTaskTool pushes when it flattens (inlineSubtask.ts buildInlineDirective).
+// It appears only in the depth-2 task's own requests — no other conversation ever sees it.
+const FLATTEN_DIRECTIVE = "auto-flattened"
+
+// reopenParentFromDelegation injects `Subtask <childId> completed.\n\nResult:\n<summary>` into the
+// resumed parent's history. In raw JSON bodies newlines are escaped, so match the serialized form.
+const INJECTION_PREFIX_RAW = "completed.\\n\\nResult:"
+
+const requestContains = (req: ChatCompletionRequest, expected: string[]) => {
+	const rawRequest = JSON.stringify(req)
+	return expected.every((text) => rawRequest.includes(text))
+}
+
+// aimock's `userMessage` matcher only inspects the LAST user message and joins only the
+// `type: "text"` content parts. Replicate that scoping inside predicates so resume turns — whose
+// last user message is a fresh environment block with no markers at all — never match initial-turn
+// fixtures.
+const lastUserMessageContains = (req: ChatCompletionRequest, text: string) => {
+	const userMessages = req.messages?.filter((message) => message.role === "user") ?? []
+	const last = userMessages.at(-1)
+	if (!last) return false
+	const content =
+		typeof last.content === "string"
+			? last.content
+			: (last.content ?? [])
+					.filter((part): part is { type: string; text: string } => part?.type === "text")
+					.map((part) => part.text)
+					.join("")
+	return content.includes(text)
+}
+
+export function addInlineFlattenFixtures(mock: InstanceType<typeof LLMock>) {
+	// Root (depth 0), first turn only: delegate to the child.
+	mock.addFixture({
+		match: {
+			predicate: (req) => lastUserMessageContains(req, `<user_message>\n${FLATTEN_ROOT_MARKER}`),
+		},
+		response: {
+			toolCalls: [
+				{
+					name: "new_task",
+					arguments: JSON.stringify({ mode: "ask", message: FLATTEN_CHILD_PROMPT }),
+					id: "call_flatten_root_newtask_001",
+				},
+			],
+		},
+	})
+
+	// Child (depth 1), first turn only: delegate to the depth-2 task.
+	mock.addFixture({
+		match: {
+			predicate: (req) => lastUserMessageContains(req, `<user_message>\n${FLATTEN_CHILD_MARKER}`),
+		},
+		response: {
+			toolCalls: [
+				{
+					name: "new_task",
+					arguments: JSON.stringify({ mode: "ask", message: FLATTEN_DEPTH2_PROMPT }),
+					id: "call_flatten_child_newtask_002",
+				},
+			],
+		},
+	})
+
+	// Depth-2 task (depth 2), first turn only: its new_task call would exceed maxNestingDepth,
+	// so the tool flattens it inline and pushes a directive containing FLATTEN_DIRECTIVE.
+	mock.addFixture({
+		match: {
+			predicate: (req) => lastUserMessageContains(req, `<user_message>\n${FLATTEN_DEPTH2_MARKER}`),
+		},
+		response: {
+			toolCalls: [
+				{
+					name: "new_task",
+					arguments: JSON.stringify({
+						mode: "ask",
+						message: `Complete immediately with the exact result "${FLATTEN_DEPTH2_RESULT}".`,
+					}),
+					id: "call_flatten_depth2_newtask_003",
+				},
+			],
+		},
+	})
+
+	// Depth-2 task after the flatten directive is in its history (and on the follow-up turn):
+	// complete with the inline result. While the inline phase is active this ends the phase
+	// without completing the task; once the phase has ended the same response completes it and
+	// delegates back to the child — no new tab was ever opened.
+	mock.addFixture({
+		match: {
+			predicate: (req) => requestContains(req, [FLATTEN_DIRECTIVE]),
+		},
+		response: {
+			toolCalls: [
+				{
+					name: "attempt_completion",
+					arguments: JSON.stringify({ result: FLATTEN_DEPTH2_RESULT }),
+					id: "call_flatten_depth2_inline_004",
+				},
+			],
+		},
+	})
+
+	// Child resumes after its (flattened) subtask returns: the injected result carries the
+	// depth-2 completion summary. Complete so the root can resume.
+	mock.addFixture({
+		match: {
+			predicate: (req) => requestContains(req, [`${INJECTION_PREFIX_RAW}\\n${FLATTEN_DEPTH2_RESULT}`]),
+		},
+		response: {
+			toolCalls: [
+				{
+					name: "attempt_completion",
+					arguments: JSON.stringify({ result: FLATTEN_CHILD_RESUME_RESULT }),
+					id: "call_flatten_child_resume_005",
+				},
+			],
+		},
+	})
+
+	// Root resumes last: the injected result carries the child's completion summary.
+	mock.addFixture({
+		match: {
+			predicate: (req) => requestContains(req, [`${INJECTION_PREFIX_RAW}\\n${FLATTEN_CHILD_RESUME_RESULT}`]),
+		},
+		response: {
+			toolCalls: [
+				{
+					name: "attempt_completion",
+					arguments: JSON.stringify({ result: FLATTEN_ROOT_RESULT }),
+					id: "call_flatten_root_resume_006",
+				},
+			],
+		},
+	})
+}

apps/vscode-e2e/src/runTest.ts

@@ -9,6 +9,7 @@ import { LLMock } from "@copilotkit/aimock"
 import { addApplyDiffResultFixtures } from "./fixtures/apply-diff"
 import { addDeepSeekV4Fixtures } from "./fixtures/deepseek-v4"
 import { addExecuteCommandResultFixtures } from "./fixtures/execute-command"
+import { addInlineFlattenFixtures } from "./fixtures/inline-flatten"
 import { addFastExitShellRaceResultFixtures } from "./fixtures/fast-exit-shell-race"
 import { addZeroChunkShellRaceResultFixtures } from "./fixtures/zero-chunk-shell-race"
 import { addTerminalReuseShellRaceFixtures } from "./fixtures/terminal-reuse-shell-race"
@@ -139,6 +140,7 @@ async function main() {
 				addListFilesResultFixtures(mock)
 				addReadFileResultFixtures(mock)
 				addSearchFilesResultFixtures(mock)
+				addInlineFlattenFixtures(mock)
 				addSubtaskFixtures(mock)
 				addUseMcpToolResultFixtures(mock)
 				addWriteToFileResultFixtures(mock)

apps/vscode-e2e/src/suite/inline-flatten.test.ts

@@ -0,0 +1,107 @@
+import * as assert from "assert"
+
+import { RooCodeEventName, type ClineMessage } from "@roo-code/types"
+
+import {
+	FLATTEN_CHILD_RESUME_RESULT,
+	FLATTEN_DEPTH2_RESULT,
+	FLATTEN_ROOT_PROMPT,
+	FLATTEN_ROOT_RESULT,
+} from "../fixtures/inline-flatten"
+import { setDefaultSuiteTimeout } from "./test-utils"
+import { sleep, waitUntilCompleted } from "./utils"
+
+suite("Roo Code Inline Flatten", function () {
+	setDefaultSuiteTimeout(this)
+
+	// A three-level new_task chain: root (depth 0) -> child (depth 1) -> depth-2 task.
+	// The depth-2 task's own new_task call would exceed maxNestingDepth (default 2), so the
+	// tool flattens it inline instead of opening a fourth tab. This crosses real extension-
+	// host boundaries (task instances, delegation links, webview events) that unit tests
+	// cannot represent.
+	test("depth-limit new_task flattens inline: no fourth task opens", async () => {
+		const api = globalThis.api
+		const says: Record<string, ClineMessage[]> = {}
+		let maxStackLength = 0
+
+		const messageHandler = ({ taskId, message }: { taskId: string; message: ClineMessage }) => {
+			if (message.type === "say" && message.partial === false) {
+				says[taskId] = says[taskId] || []
+				says[taskId].push(message)
+			}
+
+			const stackLength = api.getCurrentTaskStack().length
+			if (stackLength > maxStackLength) {
+				maxStackLength = stackLength
+			}
+		}
+
+		api.on(RooCodeEventName.Message, messageHandler)
+
+		try {
+			const rootTaskId = await waitUntilCompleted({
+				api,
+				start: () =>
+					api.startNewTask({
+						configuration: {
+							mode: "ask",
+							alwaysAllowModeSwitch: true,
+							alwaysAllowSubtasks: true,
+							autoApprovalEnabled: true,
+							enableCheckpoints: false,
+						},
+						text: FLATTEN_ROOT_PROMPT,
+					}),
+			})
+
+			const taskIds = Object.keys(says)
+			assert.strictEqual(
+				taskIds.length,
+				3,
+				`Expected exactly 3 tasks (root, child, depth-2); observed ${taskIds.join(", ")}`,
+			)
+
+			// The flattened work ran in the depth-2 task's own conversation — its completion
+			// carries the inline result. No fourth task exists to carry it.
+			const depth2TaskId = taskIds.find(
+				(taskId) =>
+					taskId !== rootTaskId &&
+					says[taskId]?.some(
+						({ say, text }) => say === "completion_result" && text?.trim() === FLATTEN_DEPTH2_RESULT,
+					),
+			)
+			assert.ok(depth2TaskId, `Depth-2 task should complete with the inline result "${FLATTEN_DEPTH2_RESULT}"`)
+
+			// The child resumed after its (flattened) subtask returned.
+			const childTaskId = taskIds.find((taskId) => taskId !== rootTaskId && taskId !== depth2TaskId)
+			assert.ok(
+				says[childTaskId!]?.some(
+					({ say, text }) => say === "completion_result" && text?.trim() === FLATTEN_CHILD_RESUME_RESULT,
+				),
+				"Child should resume and complete after the flattened subtask returns",
+			)
+
+			// The root resumed last with the chain result.
+			assert.strictEqual(
+				says[rootTaskId]
+					?.filter(({ say }) => say === "completion_result")
+					.map(({ text }) => text?.trim())
+					.find((text): text is string => !!text),
+				FLATTEN_ROOT_RESULT,
+				"Root should resume with the chain result after the child returns",
+			)
+
+			// The task stack never held more than the three real tasks — flattening opened no tab.
+			assert.ok(
+				maxStackLength <= 3,
+				`Task stack should never exceed 3 (flattened work opens no new task); observed max ${maxStackLength}`,
+			)
+		} finally {
+			api.off(RooCodeEventName.Message, messageHandler)
+			while (api.getCurrentTaskStack().length > 0) {
+				await api.clearCurrentTask()
+			}
+			await sleep(1_500)
+		}
+	})
+})

packages/types/src/global-settings.ts

@@ -48,6 +48,25 @@ export const DEFAULT_DIFF_FUZZY_THRESHOLD = 1.0
 
 export const DEFAULT_DESTRUCTIVE_COMMAND_GUARD_ENABLED = false
 
+/**
+ * Default maximum task nesting depth (root = 0).
+ *
+ * A value of `0` disables delegation entirely — every `new_task` call is executed
+ * inline in the current conversation instead of opening a child tab. The range is
+ * clamped to 0–5 by the settings UI and validated here.
+ */
+export const DEFAULT_MAX_NESTING_DEPTH = 2
+
+/**
+ * Default for auto-flattening when the nesting limit is reached.
+ *
+ * When `true`, a `new_task` call that would exceed `maxNestingDepth` is executed inline
+ * in the current conversation (same Task instance, phase marker) rather than opening a
+ * child tab. When `false`, such a call is rejected with an error result so the model
+ * continues working directly.
+ */
+export const DEFAULT_AUTO_FLATTEN_ON_LIMIT = true
+
 /**
  * Terminal output preview size options for persisted command output.
  *
@@ -151,6 +170,18 @@ export const globalSettingsSchema = z.object({
 	autoCondenseContext: z.boolean().optional(),
 	autoCondenseContextPercent: z.number().optional(),
 
+	/**
+	 * Maximum task nesting depth (root = 0). Range 0–5; `0` disables delegation entirely.
+	 * @default 2
+	 */
+	maxNestingDepth: z.number().int().min(0).max(5).optional(),
+	/**
+	 * When the nesting limit is reached, execute the subtask inline in the current
+	 * conversation instead of opening a child tab. When `false`, such calls are rejected.
+	 * @default true
+	 */
+	autoFlattenOnLimit: z.boolean().optional(),
+
 	/**
 	 * Whether to include current time in the environment details
 	 * @default true

packages/types/src/history.ts

@@ -24,6 +24,7 @@ export const historyItemSchema = z.object({
 	delegatedToId: z.string().optional(), // Last child this parent delegated to
 	childIds: z.array(z.string()).optional(), // All children spawned by this task
 	awaitingChildId: z.string().optional(), // Child currently awaited (set when delegated)
+	depth: z.number().int().min(0).optional(), // Nesting level; root = 0, child = parent.depth + 1
 	completedByChildId: z.string().optional(), // Child that completed and resumed this parent
 	completionResultSummary: z.string().optional(), // Summary from completed child
 })

packages/types/src/task.ts

@@ -94,6 +94,8 @@ export interface CreateTaskOptions {
 	/** Whether to start the task loop immediately (default: true).
 	 *  When false, the caller must invoke `task.start()` manually. */
 	startTask?: boolean
+	/** Nesting level for newly created tasks; root = 0, child = parent.depth + 1. */
+	depth?: number
 }
 
 export enum TaskStatus {
@@ -116,6 +118,8 @@ export interface TaskLike {
 	readonly rootTaskId?: string
 	readonly parentTaskId?: string
 	readonly childTaskId?: string
+	/** Nesting level; root = 0, child = parent.depth + 1. */
+	readonly depth: number
 	readonly metadata: TaskMetadata
 	readonly taskStatus: TaskStatus
 	readonly taskAsk: ClineMessage | undefined

packages/types/src/vscode-extension-host.ts

@@ -341,6 +341,8 @@ export type ExtensionState = Pick<
 
 	writeDelayMs: number
 	diffFuzzyThreshold: number
+	maxNestingDepth?: number // Maximum task nesting depth (root = 0); default 2, range 0–5
+	autoFlattenOnLimit?: boolean // Execute subtasks inline when the nesting limit is reached; default true
 
 	enableCheckpoints: boolean
 	checkpointTimeout: number // Timeout for checkpoint initialization in seconds (default: 15)