omegon 0.6.0

package/extensions/cleave/index.ts

@@ -0,0 +1,3121 @@
+/**
+ * cleave — Recursive task decomposition extension for pi.
+ *
+ * Provides:
+ *   - `cleave_assess` tool: Assess directive complexity (LLM-callable)
+ *   - `/assess` command: Code assessment toolkit (cleave, diff, spec, complexity)
+ *   - `/cleave` command: Full decomposition workflow
+ *   - Session-start handler: Surfaces active OpenSpec changes with task progress
+ *
+ * State machine: ASSESS → PLAN → CONFIRM → DISPATCH → HARVEST → REPORT
+ *
+ * Ported from styrene-lab/cleave (Python) — the pattern library, complexity
+ * formula, conflict detection, and worktree management are preserved.
+ * The Claude Code SDK calls are replaced with pi's extension API.
+ */
+
+import type { ExtensionAPI, ExtensionCommandContext, AgentToolUpdateCallback } from "@cwilson613/pi-coding-agent";
+import { truncateTail, DEFAULT_MAX_BYTES, DEFAULT_MAX_LINES, formatSize } from "@cwilson613/pi-coding-agent";
+
+import { Text } from "@cwilson613/pi-tui";
+import { Type } from "@sinclair/typebox";
+import { spawn, execFile } from "node:child_process";
+import { promisify } from "node:util";
+import { createHash } from "node:crypto";
+
+import { sharedState, DASHBOARD_UPDATE_EVENT } from "../shared-state.ts";
+import { sciCall, sciOk, sciErr, sciExpanded } from "../sci-ui.ts";
+import { debug } from "../debug.ts";
+import { emitOpenSpecState } from "../openspec/dashboard-state.ts";
+import { getSharedBridge, buildSlashCommandResult } from "../lib/slash-command-bridge.ts";
+import { buildAssessBridgeResult } from "./bridge.ts";
+import {
+	assessDirective,
+	PATTERNS,
+	runDesignStructuralCheck,
+	buildDesignAssessmentPrompt,
+	parseDesignAssessmentFindings,
+	type AssessCompletion,
+	type AssessEffect,
+	type AssessLifecycleHint,
+	type AssessLifecycleOutcome,
+	type AssessLifecycleRecord,
+	type AssessSpecScenarioResult,
+	type AssessSpecSummary,
+	type AssessStructuredResult,
+	type DesignAssessmentResult,
+	type DesignAssessmentFinding,
+} from "./assessment.ts";
+import { detectConflicts, parseTaskResult } from "./conflicts.ts";
+import { emitResolvedBugCandidate } from "./lifecycle-emitter.ts";
+import { dispatchChildren, resolveExecuteModel } from "./dispatcher.ts";
+import { DEFAULT_REVIEW_CONFIG, type ReviewConfig } from "./review.ts";
+import {
+	detectOpenSpec,
+	findExecutableChanges,
+	openspecChangeToSplitPlanWithContext,
+	buildOpenSpecContext,
+	writeBackTaskCompletion,
+	getActiveChangesStatus,
+	type OpenSpecContext,
+} from "./openspec.ts";
+import { buildPlannerPrompt, getRepoTree, parsePlanResponse } from "./planner.ts";
+import {
+	matchSkillsToAllChildren,
+	resolveSkillPaths,
+	getPreferredTier,
+} from "./skills.ts";
+import { discoverGuardrails, runGuardrails, formatGuardrailResults } from "./guardrails.ts";
+import type { CleaveState, ChildState, SplitPlan } from "./types.ts";
+import { DEFAULT_CONFIG } from "./types.ts";
+import {
+	buildCheckpointPlan,
+	classifyDirtyPaths as classifyPreflightDirtyPaths,
+	findIncompleteRuns,
+	initWorkspace,
+	loadState,
+	readTaskFiles,
+	saveState,
+	type ClassifiedDirtyPath,
+	type DirtyTreeClassification as WorkspaceDirtyTreeClassification,
+} from "./workspace.ts";
+import type { SkillDirective } from "./workspace.ts";
+import {
+	cleanupWorktrees,
+	createWorktree,
+	ensureCleanWorktree,
+	getCurrentBranch,
+	mergeBranch,
+	pruneWorktreeDirs,
+} from "./worktree.ts";
+import { inspectGitState } from "../lib/git-state.ts";
+
+// ─── Dashboard state emitter ────────────────────────────────────────────────
+
+/** Map internal ChildStatus to the dashboard's simplified status. */
+function mapChildStatus(status: string): "pending" | "running" | "done" | "failed" {
+	if (status === "completed") return "done";
+	if (status === "running" || status === "failed") return status;
+	return "pending"; // pending, needs_decomposition → pending
+}
+
+/**
+ * Emit cleave dashboard state to sharedState.cleave and fire the
+ * dashboard update event so the footer re-renders immediately.
+ *
+ * Called at lifecycle transitions so the unified dashboard can
+ * render live progress without polling.
+ */
+function emitCleaveState(
+	pi: ExtensionAPI,
+	status: string,
+	runId?: string,
+	children?: Array<{ label: string; status: string; durationSec?: number }>,
+): void {
+	(sharedState as any).cleave = {
+		status,
+		runId,
+		updatedAt: Date.now(),
+		children: children?.map((c) => ({
+			label: c.label,
+			status: mapChildStatus(c.status),
+			elapsed: c.durationSec,
+		})),
+	};
+	debug("cleave", "emitState", { status, runId, childCount: children?.length });
+	pi.events.emit(DASHBOARD_UPDATE_EVENT, { source: "cleave" });
+}
+
+// ─── Helpers ────────────────────────────────────────────────────────────────
+
+function generateRunId(): string {
+	return `clv-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 6)}`;
+}
+
+function formatAssessment(a: ReturnType<typeof assessDirective>): string {
+	const lines = [
+		"**Assessment**",
+		"",
+		`  Decision: **${a.decision}**`,
+		`  Complexity: ${a.complexity}`,
+		`  Systems: ${a.systems}`,
+		`  Modifiers: ${a.modifiers.length > 0 ? a.modifiers.join(", ") : "none"}`,
+		`  Method: ${a.method}`,
+	];
+	if (a.pattern) {
+		lines.push(`  Pattern: ${a.pattern} (${(a.confidence * 100).toFixed(0)}%)`);
+	}
+	lines.push("", `  ${a.reasoning}`);
+	return lines.join("\n");
+}
+
+function formatConflicts(conflicts: ReturnType<typeof detectConflicts>): string {
+	if (conflicts.length === 0) return "No conflicts detected. ✓";
+	return conflicts
+		.map(
+			(c, i) =>
+				`**Conflict ${i + 1}:** ${c.type}\n` +
+				`  ${c.description}\n` +
+				`  Involved: tasks ${c.involved.join(", ")}\n` +
+				`  Resolution: ${c.resolution}`,
+		)
+		.join("\n\n");
+}
+
+function formatSpecVerification(ctx: OpenSpecContext): string {
+	const lines = [
+		"### Spec Verification",
+		"",
+		"The following spec scenarios should now be satisfied. **Verify each one:**",
+		"",
+	];
+
+	for (const ss of ctx.specScenarios) {
+		lines.push(`**${ss.domain} → ${ss.requirement}**`);
+		for (const scenario of ss.scenarios) {
+			// Extract just the scenario name (first line) and the Given/When/Then
+			const scenarioLines = scenario.split("\n");
+			const name = scenarioLines[0];
+			lines.push(`- [ ] ${name}`);
+			// Include Given/When/Then as indented detail
+			const gwt = scenarioLines.slice(1).filter((l) => l.trim());
+			if (gwt.length > 0) {
+				for (const l of gwt) {
+					lines.push(`      ${l.trim()}`);
+				}
+			}
+		}
+		lines.push("");
+	}
+
+	if (ctx.apiContract) {
+		lines.push(
+			"**API Contract Conformance (`api.yaml`)**",
+			"- [ ] All contract paths/methods are implemented",
+			"- [ ] Request/response schemas match the contract",
+			"- [ ] Status codes and error responses match the contract",
+			"- [ ] No undocumented endpoints exist outside the contract",
+			"",
+		);
+	}
+
+	lines.push(
+		"---",
+		"Run tests, inspect the code, or manually verify each scenario above.",
+		"If all pass, the change is ready for `/opsx:archive`.",
+	);
+
+	return lines.join("\n");
+}
+
+interface DirtyTreePreflightOptions {
+	repoPath: string;
+	openspecChangePath?: string;
+	onUpdate?: AgentToolUpdateCallback<Record<string, unknown>>;
+	ui?: {
+		/** Text input — used for commit-message approval in the checkpoint flow. */
+		input?: (prompt: string, initial?: string) => Promise<string | undefined>;
+		/** Modal select — used to choose the preflight action. Falls back to input when absent. */
+		select?: (title: string, options: string[]) => Promise<string | undefined>;
+	};
+}
+
+/** Labelled preflight actions shown in the modal select UI. */
+const PREFLIGHT_ACTION_OPTIONS = [
+	"checkpoint        — commit related changes and continue  [use when work is ready to save before cleaving]",
+	"stash-unrelated   — stash unrelated/unknown files and continue  [use when dirty files are not part of this change]",
+	"stash-volatile    — stash volatile artifacts only  [use when only build/cache artifacts are dirty]",
+	"proceed-without-cleave — skip cleave and continue working  [use when you want to defer parallel dispatch]",
+	"cancel            — abort cleave  [use when you want to exit without making any changes]",
+] as const;
+
+/** Extract the action keyword from a labelled option string (text before first whitespace run ending in ' — '). */
+function parsePreflightAction(selected: string | undefined): string | undefined {
+	if (!selected) return undefined;
+	const normalized = normalizePreflightInput(selected);
+	if (!normalized) return undefined;
+	// Trim leading spaces then grab the first non-space token.
+	return normalized.trim().split(/\s+/)[0]?.toLowerCase();
+}
+
+const TRANSIENT_CLIPBOARD_ATTACHMENT_PATH =
+	/^\/var\/folders\/[A-Za-z0-9_-]+\/[A-Za-z0-9_-]+\/T\/pi-clipboard-[A-Fa-f0-9-]+\.(?:png|jpe?g|gif|webp)$/;
+
+function normalizePreflightInput(response: string | undefined): string | undefined {
+	const trimmed = response?.trim();
+	if (!trimmed) return undefined;
+	if (TRANSIENT_CLIPBOARD_ATTACHMENT_PATH.test(trimmed)) return undefined;
+	return trimmed;
+}
+
+function formatDirtyTreeSummary(classification: WorkspaceDirtyTreeClassification, suggestedMessage: string | null): string {
+	const renderGroup = (title: string, entries: ClassifiedDirtyPath[], empty: string): string[] => [
+		title,
+		...(entries.length > 0
+			? entries.map((entry) => `- [${entry.confidence}] \`${entry.path}\` — ${entry.reason}`)
+			: [`- ${empty}`]),
+		"",
+	];
+	const unrelatedOrUnknown = [...classification.unrelated, ...classification.unknown];
+	const lines = [
+		"### Dirty Tree Preflight",
+		"",
+		"Cleave requires an explicit preflight decision before worktree creation.",
+		"",
+		...renderGroup("**Related changes**", classification.related, "none detected"),
+		...renderGroup("**Unrelated / unknown changes**", unrelatedOrUnknown, "none detected"),
+		...renderGroup("**Volatile artifacts**", classification.volatile, "none detected"),
+		"**Actions:** `checkpoint`, `stash-unrelated`, `stash-volatile`, `proceed-without-cleave`, `cancel`",
+		...(suggestedMessage ? ["", `Suggested checkpoint commit: \`${suggestedMessage}\``] : []),
+	];
+	return lines.join("\n");
+}
+
+async function stashPaths(pi: ExtensionAPI, repoPath: string, label: string, entries: ClassifiedDirtyPath[]): Promise<void> {
+	if (entries.length === 0) return;
+	const args = ["stash", "push", "-u", "-m", label, "--", ...entries.map((entry) => entry.path)];
+	const result = await pi.exec("git", args, { cwd: repoPath, timeout: 15_000 });
+	if (result.code !== 0) throw new Error(result.stderr.trim() || `Failed to stash ${label}`);
+}
+
+async function checkpointRelatedChanges(
+	pi: ExtensionAPI,
+	repoPath: string,
+	classification: WorkspaceDirtyTreeClassification,
+	checkpointMessage: string | null,
+	ui?: { input?: (prompt: string, initial?: string) => Promise<string | undefined> },
+): Promise<void> {
+	// When the user explicitly chooses "checkpoint", commit all non-volatile dirty
+	// files — not just those confidently classified as related. The conservative
+	// classification is for automatic decisions; an explicit user choice overrides it.
+	const allNonVolatile = [
+		...classification.related,
+		...classification.unrelated,
+		...classification.unknown,
+	].map((f) => f.path);
+
+	const filesToCommit = classification.checkpointFiles.length > 0
+		? classification.checkpointFiles
+		: allNonVolatile;
+
+	if (filesToCommit.length === 0) {
+		throw new Error(
+			"Checkpoint scope is empty — no dirty files found to commit (only volatile artifacts are dirty). " +
+			"Choose a different preflight action.",
+		);
+	}
+
+	// Patch classification so the rest of the function uses the resolved file list.
+	classification = { ...classification, checkpointFiles: filesToCommit };
+	if (typeof ui?.input !== "function") {
+		throw new Error("Checkpoint requires interactive approval, but input is unavailable.");
+	}
+	const suggested = checkpointMessage ?? "chore(cleave): checkpoint before cleave";
+	const response = normalizePreflightInput(await ui.input(
+		[
+			`Checkpoint ${classification.checkpointFiles.length} related file(s).`,
+			`Press Enter to approve the suggested message, type a custom commit message to approve with edits, or type 'cancel' to decline.`,
+			`Suggested message: ${suggested}`,
+		].join("\n"),
+		suggested,
+	));
+	if (!response) {
+		// Accept the suggested message when the operator confirms with Enter.
+	} else if (response.toLowerCase() === "cancel") {
+		throw new Error("Checkpoint cancelled before commit approval.");
+	}
+	const commitMessage = response && response.length > 0 ? response : suggested;
+	const addResult = await pi.exec("git", ["add", "--", ...classification.checkpointFiles], { cwd: repoPath, timeout: 15_000 });
+	if (addResult.code !== 0) {
+		throw new Error(
+			`git add failed during checkpoint — ${addResult.stderr.trim() || "unknown error staging checkpoint files"}. ` +
+			"The checkpoint was not created. Choose a different preflight action or resolve the staging error first.",
+		);
+	}
+	const commitResult = await pi.exec("git", ["commit", "-m", commitMessage, "--", ...classification.checkpointFiles], {
+		cwd: repoPath,
+		timeout: 20_000,
+	});
+	if (commitResult.code !== 0) {
+		throw new Error(
+			`git commit failed during checkpoint — ${commitResult.stderr.trim() || "unknown error creating checkpoint commit"}. ` +
+			"The checkpoint was not created. Resolve the git error and try again.",
+		);
+	}
+}
+
+export async function runDirtyTreePreflight(pi: ExtensionAPI, options: DirtyTreePreflightOptions): Promise<"continue" | "skip_cleave" | "cancelled"> {
+	const status = await pi.exec("git", ["status", "--porcelain"], {
+		cwd: options.repoPath,
+		timeout: 5_000,
+	});
+	const gitState = inspectGitState(status.stdout);
+	if (gitState.entries.length === 0) return "continue";
+
+	const openspecContext = options.openspecChangePath
+		? (() => {
+			try {
+				return buildOpenSpecContext(options.openspecChangePath!);
+			} catch {
+				return null;
+			}
+		})()
+		: null;
+	const changeName = options.openspecChangePath?.replace(/\\/g, "/").split("/").pop() ?? null;
+	const classification = classifyPreflightDirtyPaths(gitState.entries.map((entry) => entry.path), {
+		changeName,
+		openspecContext,
+	});
+	// Compute initial checkpoint plan for the summary display only.
+	// The plan is rebuilt from currentClassification inside the loop on each attempt (C4).
+	const initialCheckpointPlan = buildCheckpointPlan(classification, { changeName, openspecContext });
+	const summary = formatDirtyTreeSummary(classification, initialCheckpointPlan.message);
+	options.onUpdate?.({ content: [{ type: "text", text: summary }], details: { phase: "preflight" } });
+
+	if (gitState.nonVolatile.length === 0) {
+		if (classification.volatile.length > 0) {
+			await stashPaths(pi, options.repoPath, "cleave-preflight-volatile", classification.volatile);
+			options.onUpdate?.({
+				content: [{ type: "text", text: "Volatile-only dirty tree detected — stashed volatile artifacts automatically before cleave." }],
+				details: { phase: "preflight", autoResolved: "volatile_only_stash" },
+			});
+		}
+		return "continue";
+	}
+
+	const hasSelect = typeof options.ui?.select === "function";
+	const hasInput = typeof options.ui?.input === "function";
+	if (!hasSelect && !hasInput) {
+		throw new Error(summary + "\n\nInteractive input is unavailable, so cleave cannot resolve the dirty tree automatically.");
+	}
+
+	// Mutable classification — refreshed after each checkpoint attempt (C1/W1).
+	let currentClassification = classification;
+
+	while (true) {
+		let answer: string | undefined;
+		if (hasSelect) {
+			// Preferred path: show a modal select with labelled options and descriptions.
+			const selected = await options.ui!.select!(
+				"Dirty tree detected — choose a preflight action to proceed:",
+				[...PREFLIGHT_ACTION_OPTIONS],
+			);
+			answer = parsePreflightAction(selected);
+		} else {
+			// Fallback: raw text input (headless / test environments).
+			answer = normalizePreflightInput(
+				await options.ui!.input!("Dirty tree action [checkpoint|stash-unrelated|stash-volatile|proceed-without-cleave|cancel]:"),
+			)?.toLowerCase();
+		}
+		try {
+			switch (answer) {
+				case "checkpoint": {
+					// Rebuild the checkpoint plan from the current (possibly refreshed) classification (C4).
+					const currentCheckpointPlan = buildCheckpointPlan(currentClassification, { changeName, openspecContext });
+					const committedFiles = new Set(currentClassification.checkpointFiles);
+					await checkpointRelatedChanges(pi, options.repoPath, currentClassification, currentCheckpointPlan.message, options.ui);
+					// Re-verify cleanliness after the checkpoint commit.
+					const postCheckpointStatus = await pi.exec("git", ["status", "--porcelain"], {
+						cwd: options.repoPath,
+						timeout: 5_000,
+					});
+					const postState = inspectGitState(postCheckpointStatus.stdout);
+					if (postState.entries.length === 0) {
+						// Tree is clean — checkpoint fully resolved the dirty tree.
+						return "continue";
+					}
+
+					// Re-derive classification from the post-checkpoint state (C1).
+					currentClassification = classifyPreflightDirtyPaths(
+						postState.entries.map((e) => e.path),
+						{ changeName, openspecContext },
+					);
+
+					// C2: If only volatile files remain, auto-stash and continue.
+					if (postState.nonVolatile.length === 0 && currentClassification.volatile.length > 0) {
+						await stashPaths(pi, options.repoPath, "cleave-preflight-volatile", currentClassification.volatile);
+						options.onUpdate?.({
+							content: [{ type: "text", text: "Checkpoint succeeded. Remaining volatile artifacts stashed automatically — cleave continuing." }],
+							details: { phase: "preflight", autoResolved: "volatile_only_stash" },
+						});
+						return "continue";
+					}
+
+					// Remaining dirty files — emit precise diagnosis (W1: distinguish committed-but-still-dirty vs excluded-from-scope).
+					const remainingPaths = postState.entries.map((e) => e.path);
+					const diagnosisLines = [
+						"Checkpoint committed successfully, but dirty files remain — cleave cannot continue yet:",
+						...currentClassification.related.map((f) =>
+							committedFiles.has(f.path)
+								? `  • ${f.path}  [was committed but remains dirty — file may have been modified after staging or only partially staged]`
+								: `  • ${f.path}  [related but excluded from checkpoint scope — confidence too low to commit automatically]`
+						),
+						...currentClassification.unrelated.map((f) => `  • ${f.path}  [unrelated: ${f.reason}]`),
+						...currentClassification.unknown.map((f) => `  • ${f.path}  [unknown — not in change scope, was not checkpointed]`),
+						...currentClassification.volatile.map((f) => `  • ${f.path}  [volatile artifact — will be auto-stashed]`),
+						"",
+						"Choose another preflight action to resolve the remaining files.",
+					];
+					options.onUpdate?.({
+						content: [{ type: "text", text: diagnosisLines.join("\n") }],
+						details: { phase: "preflight", postCheckpointDirty: remainingPaths },
+					});
+					break;
+				}
+				case "stash-unrelated":
+					// C1: Use currentClassification (refreshed after checkpoint) not the stale original.
+					await stashPaths(pi, options.repoPath, "cleave-preflight-unrelated", [...currentClassification.unrelated, ...currentClassification.unknown]);
+					return "continue";
+				case "stash-volatile":
+					await stashPaths(pi, options.repoPath, "cleave-preflight-volatile", currentClassification.volatile);
+					return "continue";
+				case "proceed-without-cleave":
+					return "skip_cleave";
+				case "cancel":
+				case "":
+					return "cancelled";
+				default:
+					options.onUpdate?.({
+						content: [{ type: "text", text: "Invalid preflight action. Choose checkpoint, stash-unrelated, stash-volatile, proceed-without-cleave, or cancel." }],
+						details: { phase: "preflight" },
+					});
+			}
+		} catch (error) {
+			const message = error instanceof Error ? error.message : String(error);
+			options.onUpdate?.({
+				content: [{ type: "text", text: `Preflight action failed: ${message}` }],
+				details: { phase: "preflight" },
+			});
+		}
+	}
+}
+
+interface AssessExecutionContext {
+	cwd: string;
+	bridgeInvocation?: boolean;
+	hasUI?: boolean;
+	model?: { id?: string };
+	waitForIdle?: (() => Promise<void>) | undefined;
+}
+
+interface AssessDiffContext {
+	ref: string;
+	diffStat: string;
+	diffContent: string;
+	recentLog: string;
+}
+
+function makeAssessResult<TData>(input: {
+	subcommand: AssessStructuredResult<TData>["subcommand"];
+	args: string;
+	ok: boolean;
+	summary: string;
+	humanText: string;
+	data: TData;
+	effects?: AssessEffect[];
+	nextSteps?: string[];
+	completion?: AssessCompletion;
+	lifecycle?: AssessLifecycleHint;
+	lifecycleRecord?: AssessLifecycleRecord;
+}): AssessStructuredResult<TData> {
+	return {
+		command: "assess",
+		subcommand: input.subcommand,
+		args: input.args,
+		ok: input.ok,
+		summary: input.summary,
+		humanText: input.humanText,
+		data: input.data,
+		effects: input.effects ?? [],
+		nextSteps: input.nextSteps ?? [],
+		completion: input.completion,
+		lifecycle: input.lifecycle,
+		lifecycleRecord: input.lifecycleRecord,
+	};
+}
+
+async function collectAssessmentSnapshot(pi: ExtensionAPI, cwd: string): Promise<{ gitHead: string | null; fingerprint: string }> {
+	let gitHead: string | null = null;
+	let status = "";
+
+	try {
+		const head = await pi.exec("git", ["rev-parse", "--short", "HEAD"], { cwd, timeout: 5_000 });
+		if (head.code === 0) gitHead = head.stdout.trim() || null;
+	} catch {
+		/* proceed with null gitHead */
+	}
+
+	try {
+		const diff = await pi.exec("git", ["status", "--short", "--untracked-files=all"], { cwd, timeout: 5_000 });
+		if (diff.code === 0) status = diff.stdout.trim();
+	} catch {
+		/* proceed with empty status */
+	}
+
+	const fingerprint = createHash("sha256")
+		.update(gitHead ?? "nogit")
+		.update("\n")
+		.update(status)
+		.digest("hex");
+
+	return { gitHead, fingerprint };
+}
+
+async function buildLifecycleRecord(
+	pi: ExtensionAPI,
+	cwd: string,
+	options: {
+		changeName: string;
+		assessmentKind: "spec" | "cleave";
+		outcome: AssessLifecycleOutcome;
+		recommendedAction: string | null;
+		changedFiles?: string[];
+		constraints?: string[];
+		snapshot?: { gitHead: string | null; fingerprint: string };
+	},
+): Promise<AssessLifecycleRecord> {
+	const snapshot = options.snapshot ?? await collectAssessmentSnapshot(pi, cwd);
+	return {
+		changeName: options.changeName,
+		assessmentKind: options.assessmentKind,
+		outcome: options.outcome,
+		timestamp: new Date().toISOString(),
+		snapshot,
+		reconciliation: {
+			reopen: options.outcome === "reopen",
+			changedFiles: [...new Set((options.changedFiles ?? []).map((file) => file.trim()).filter(Boolean))],
+			constraints: [...new Set((options.constraints ?? []).map((constraint) => constraint.trim()).filter(Boolean))],
+			recommendedAction: options.recommendedAction,
+		},
+	};
+}
+
+interface AssessSpecAgentResult {
+	summary: AssessSpecSummary;
+	scenarios: AssessSpecScenarioResult[];
+	changedFiles?: string[];
+	constraints?: string[];
+	overallNotes?: string;
+}
+
+interface SpecAssessmentRunnerInput {
+	repoPath: string;
+	changeName: string;
+	scenarioText: string;
+	designContext: string[];
+	apiContractContext: string[];
+	diffContent: string;
+	expectedScenarioCount: number;
+	modelId?: string;
+}
+
+interface SpecAssessmentRunnerOutput {
+	assessed: AssessSpecAgentResult;
+	snapshot?: { gitHead: string | null; fingerprint: string };
+}
+
+interface AssessExecutorOverrides {
+	runSpecAssessment?: (input: SpecAssessmentRunnerInput) => Promise<SpecAssessmentRunnerOutput>;
+}
+
+function isInteractiveAssessContext(ctx: AssessExecutionContext): ctx is AssessExecutionContext & ExtensionCommandContext {
+	return ctx.bridgeInvocation !== true && ctx.hasUI === true && typeof ctx.waitForIdle === "function";
+}
+
+function countSpecScenarios(specCtx: OpenSpecContext): number {
+	return specCtx.specScenarios.reduce((total, scenarioSet) => total + scenarioSet.scenarios.length, 0);
+}
+
+function determineSpecOutcome(summary: AssessSpecSummary): AssessLifecycleOutcome {
+	if (summary.fail > 0) return "reopen";
+	if (summary.unclear > 0) return "ambiguous";
+	return "pass";
+}
+
+function normalizeSpecAssessment(payload: AssessSpecAgentResult, expectedTotal: number): AssessSpecAgentResult {
+	const scenarios = payload.scenarios.map((scenario) => ({
+		...scenario,
+		evidence: [...new Set((scenario.evidence ?? []).map((entry) => entry.trim()).filter(Boolean))],
+		notes: scenario.notes?.trim() || undefined,
+	}));
+	const summary: AssessSpecSummary = {
+		total: payload.summary.total,
+		pass: payload.summary.pass,
+		fail: payload.summary.fail,
+		unclear: payload.summary.unclear,
+	};
+	if (summary.total !== expectedTotal || scenarios.length !== expectedTotal) {
+		throw new Error(`Assessment returned ${scenarios.length}/${expectedTotal} scenarios.`);
+	}
+	return {
+		summary,
+		scenarios,
+		changedFiles: [...new Set((payload.changedFiles ?? []).map((entry) => entry.trim()).filter(Boolean))],
+		constraints: [...new Set((payload.constraints ?? []).map((entry) => entry.trim()).filter(Boolean))],
+		overallNotes: payload.overallNotes?.trim() || undefined,
+	};
+}
+
+function extractJsonObject(text: string): string | null {
+	const fenced = text.match(/```json\s*([\s\S]*?)```/i);
+	if (fenced?.[1]) return fenced[1].trim();
+	const firstBrace = text.indexOf("{");
+	const lastBrace = text.lastIndexOf("}");
+	if (firstBrace === -1 || lastBrace === -1 || lastBrace <= firstBrace) return null;
+	return text.slice(firstBrace, lastBrace + 1).trim();
+}
+
+function extractAssistantText(content: unknown): string {
+	if (typeof content === "string") return content.trim();
+	if (!Array.isArray(content)) return "";
+	return content
+		.map((item) => {
+			if (typeof item === "string") return item;
+			if (!item || typeof item !== "object") return "";
+			return typeof (item as { text?: unknown }).text === "string"
+				? (item as { text: string }).text
+				: "";
+		})
+		.join("\n")
+		.trim();
+}
+
+function formatSpecOutcomeLabel(outcome: AssessLifecycleOutcome): string {
+	switch (outcome) {
+		case "pass":
+			return "PASS";
+		case "reopen":
+			return "REOPEN";
+		case "ambiguous":
+			return "AMBIGUOUS";
+	}
+}
+
+function buildSpecAssessmentHumanText(changeName: string, assessed: AssessSpecAgentResult, outcome: AssessLifecycleOutcome): string {
+	const lines = [
+		`**Spec Assessment Complete: \`${changeName}\`**`,
+		"",
+		`Outcome: **${formatSpecOutcomeLabel(outcome)}**`,
+		`Scenarios: ${assessed.summary.pass}/${assessed.summary.total} pass` +
+			(assessed.summary.fail > 0 ? `, ${assessed.summary.fail} fail` : "") +
+			(assessed.summary.unclear > 0 ? `, ${assessed.summary.unclear} unclear` : ""),
+	];
+
+	for (const scenario of assessed.scenarios) {
+		lines.push(
+			"",
+			`- [${scenario.status}] ${scenario.domain} → ${scenario.requirement}`,
+			`  ${scenario.scenario.replace(/\n/g, " ")}`,
+			...scenario.evidence.map((entry) => `  Evidence: ${entry}`),
+			...(scenario.notes ? [`  Notes: ${scenario.notes}`] : []),
+		);
+	}
+
+	if (assessed.overallNotes) {
+		lines.push("", `Overall notes: ${assessed.overallNotes}`);
+	}
+
+	return lines.join("\n");
+}
+
+async function runSpecAssessmentSubprocess(
+	input: SpecAssessmentRunnerInput,
+): Promise<SpecAssessmentRunnerOutput> {
+	const prompt = [
+		"You are performing a read-only OpenSpec compliance assessment.",
+		"Operate in read-only plan mode. Never call edit, write, or any workspace-mutating command.",
+		"Inspect the repository and determine whether the implementation satisfies every OpenSpec scenario below.",
+		"Return ONLY a JSON object with this exact shape:",
+		"{",
+		'  "summary": { "total": number, "pass": number, "fail": number, "unclear": number },',
+		'  "scenarios": [',
+		'    { "domain": string, "requirement": string, "scenario": string, "status": "PASS"|"FAIL"|"UNCLEAR", "evidence": string[], "notes"?: string }',
+		"  ],",
+		'  "changedFiles": string[],',
+		'  "constraints": string[],',
+		'  "overallNotes"?: string',
+		"}",
+		"Rules:",
+		`- Emit exactly ${input.expectedScenarioCount} scenario entries.`,
+		"- Use FAIL when the code clearly contradicts or omits the scenario.",
+		"- Use UNCLEAR only when code inspection cannot safely prove PASS or FAIL.",
+		"- Evidence must cite concrete files, symbols, or line references when possible.",
+		"- changedFiles should list files that would need modification if the result reopens work.",
+		"- constraints should list newly discovered implementation constraints.",
+		"- Do not wrap the JSON in explanatory prose.",
+		"",
+		`Change: ${input.changeName}`,
+		"",
+		"## Acceptance Criteria",
+		"",
+		input.scenarioText,
+		"",
+		...input.designContext,
+		...input.apiContractContext,
+		...(input.diffContent ? ["### Recent Changes", "", "```diff", input.diffContent, "```", ""] : []),
+	].join("\n");
+
+	const args = ["--mode", "json", "--plan", "-p", "--no-session"];
+	if (input.modelId) args.push("--model", input.modelId);
+
+	return await new Promise<SpecAssessmentRunnerOutput>((resolve, reject) => {
+		const proc = spawn("pi", args, {
+			cwd: input.repoPath,
+			shell: false,
+			stdio: ["pipe", "pipe", "pipe"],
+			env: {
+				...process.env,
+				PI_CHILD: "1",
+				TERM: process.env.TERM ?? "dumb",
+			},
+		});
+		let stdout = "";
+		let stderr = "";
+		let buffer = "";
+		let assistantText = "";
+		let settled = false;
+		const settleReject = (error: Error) => {
+			if (settled) return;
+			settled = true;
+			clearTimeout(timer);
+			reject(error);
+		};
+		const settleResolve = (value: SpecAssessmentRunnerOutput) => {
+			if (settled) return;
+			settled = true;
+			clearTimeout(timer);
+			resolve(value);
+		};
+		const timer = setTimeout(() => {
+			proc.kill("SIGTERM");
+			setTimeout(() => {
+				if (!proc.killed) proc.kill("SIGKILL");
+			}, 5_000);
+			settleReject(new Error(`Timed out after 120s while assessing ${input.changeName}.`));
+		}, 120_000);
+
+		const processLine = (line: string) => {
+			if (!line.trim()) return;
+			stdout += line + "\n";
+			let event: unknown;
+			try {
+				event = JSON.parse(line);
+			} catch {
+				return;
+			}
+			if (!event || typeof event !== "object") return;
+			const typed = event as { type?: string; message?: { role?: string; content?: unknown } };
+			if (typed.type === "message_end" && typed.message?.role === "assistant") {
+				assistantText = extractAssistantText(typed.message.content);
+			}
+		};
+
+		proc.stdout.on("data", (data) => {
+			buffer += data.toString();
+			const lines = buffer.split("\n");
+			buffer = lines.pop() || "";
+			for (const line of lines) processLine(line);
+		});
+		proc.stderr.on("data", (data) => {
+			stderr += data.toString();
+		});
+		proc.on("error", (error) => {
+			settleReject(error);
+		});
+		proc.on("close", (code) => {
+			if (buffer.trim()) processLine(buffer.trim());
+			if ((code ?? 1) !== 0) {
+				settleReject(new Error(stderr.trim() || `Assessment subprocess exited with code ${code ?? 1}.`));
+				return;
+			}
+			const sourceText = assistantText || stdout;
+			const jsonText = extractJsonObject(sourceText);
+			if (!jsonText) {
+				settleReject(new Error(`Assessment subprocess did not return parseable JSON.\n${stderr || stdout}`));
+				return;
+			}
+			try {
+				const parsed = JSON.parse(jsonText) as AssessSpecAgentResult;
+				settleResolve({
+					assessed: normalizeSpecAssessment(parsed, input.expectedScenarioCount),
+				});
+			} catch (error) {
+				const message = error instanceof Error ? error.message : String(error);
+				settleReject(new Error(`Assessment JSON was invalid: ${message}`));
+			}
+		});
+
+		proc.stdin.write(prompt);
+		proc.stdin.end();
+	});
+}
+
+function applyAssessEffects(pi: ExtensionAPI, result: AssessStructuredResult): void {
+	for (const effect of result.effects) {
+		if (effect.type === "view") {
+			pi.sendMessage({
+				customType: "view",
+				content: effect.content,
+				display: effect.display ?? true,
+			});
+			continue;
+		}
+		if (effect.type === "follow_up") {
+			pi.sendUserMessage(effect.content, { deliverAs: "followUp" });
+		}
+	}
+}
+
+async function collectAssessDiffContext(
+	pi: ExtensionAPI,
+	cwd: string,
+	ref: string,
+	fallbackToUnstaged: boolean,
+): Promise<AssessDiffContext | null> {
+	let effectiveRef = ref;
+	let diffStat = "";
+	let diffContent = "";
+	let recentLog = "";
+
+	if (!effectiveRef && fallbackToUnstaged) {
+		for (const candidate of ["HEAD~3", "HEAD~2", "HEAD~1"]) {
+			try {
+				const test = await pi.exec("git", ["rev-parse", "--verify", candidate], { cwd, timeout: 3_000 });
+				if (test.code === 0) {
+					effectiveRef = candidate;
+					break;
+				}
+			} catch {
+				/* try next */
+			}
+		}
+	}
+
+	if (effectiveRef) {
+		try {
+			const stat = await pi.exec("git", ["diff", "--stat", effectiveRef], { cwd, timeout: 5_000 });
+			diffStat = stat.stdout.trim();
+			const diff = await pi.exec("git", ["diff", effectiveRef], { cwd, timeout: 10_000 });
+			diffContent = diff.stdout.slice(0, 40_000);
+			const log = await pi.exec("git", ["log", "--oneline", "-10"], { cwd, timeout: 5_000 });
+			recentLog = log.stdout.trim();
+		} catch {
+			/* fall through */
+		}
+	}
+
+	if (!diffStat && !diffContent && fallbackToUnstaged) {
+		try {
+			const stat = await pi.exec("git", ["diff", "--stat"], { cwd, timeout: 5_000 });
+			diffStat = stat.stdout.trim();
+			const diff = await pi.exec("git", ["diff"], { cwd, timeout: 10_000 });
+			diffContent = diff.stdout.slice(0, 40_000);
+			effectiveRef = "unstaged";
+		} catch {
+			/* proceed without git diff */
+		}
+	}
+
+	if (!diffStat && !diffContent) return null;
+	return { ref: effectiveRef, diffStat, diffContent, recentLog };
+}
+
+function buildGuardrailPreamble(cwd: string): string {
+	try {
+		const checks = discoverGuardrails(cwd);
+		if (checks.length === 0) return "";
+		const suite = runGuardrails(cwd, checks);
+		return formatGuardrailResults(suite);
+	} catch {
+		return "";
+	}
+}
+
+async function executeAssessCleave(
+	pi: ExtensionAPI,
+	ctx: AssessExecutionContext,
+	args: string,
+): Promise<AssessStructuredResult> {
+	const diffContext = await collectAssessDiffContext(pi, ctx.cwd, args.trim(), true);
+	if (!diffContext) {
+		return makeAssessResult({
+			subcommand: "cleave",
+			args,
+			ok: false,
+			summary: "No recent changes found",
+			humanText: "No recent changes found. Nothing to assess.",
+			data: { reason: "no_changes" },
+			effects: [{ type: "view", content: "No recent changes found. Nothing to assess." }],
+		});
+	}
+
+	const activeOpenSpec = getActiveChangesStatus(ctx.cwd)
+		.filter((status) => status.totalTasks > 0)
+		.sort((a, b) => b.lastModifiedMs - a.lastModifiedMs);
+	const targetChange = activeOpenSpec[0]?.name;
+	const lifecycle = targetChange
+		? { changeName: targetChange, assessmentKind: "cleave" as const, outcomes: ["pass", "reopen", "ambiguous"] as const }
+		: undefined;
+	const postAssessInstruction = targetChange
+		? [
+			"",
+			`After review/fixes/tests, call \`openspec_manage\` with action \`reconcile_after_assess\`, change_name \`${targetChange}\`, assessment_kind \`cleave\`, and outcome:`,
+			"- `pass` if all Critical/Warning work is resolved cleanly",
+			"- `reopen` if remaining work or follow-up fixes reopen implementation",
+			"- `ambiguous` if you cannot safely map reviewer findings back to task state",
+			"Include `changed_files` for any follow-up fix files and `constraints` for new implementation constraints discovered during review.",
+		]
+		: [];
+	const guardrailPreamble = buildGuardrailPreamble(ctx.cwd);
+	const prompt = [
+		"## Adversarial Review → Auto-Fix Pipeline",
+		"",
+		"You are doing an adversarial code review of recent changes.",
+		"Your job is to find real issues, then fix them automatically.",
+		"",
+		...(guardrailPreamble ? [
+			"### Deterministic Analysis",
+			"",
+			guardrailPreamble,
+			"",
+			"The above are compiler/linter findings — treat failures as Critical issues.",
+			"",
+		] : []),
+		"### Step 1: Review",
+		"",
+		"Analyze these recent changes for:",
+		"- **Critical bugs**: logic errors, race conditions, missing error handling",
+		"- **Warnings**: misleading names, missing edge cases, fragile patterns",
+		"- **Nits**: dead code, style inconsistencies (low priority)",
+		"",
+		"Recent commits:",
+		"```",
+		diffContext.recentLog,
+		"```",
+		"",
+		"Diff stat:",
+		"```",
+		diffContext.diffStat,
+		"```",
+		"",
+		"Full diff (truncated to 40KB):",
+		"```diff",
+		diffContext.diffContent,
+		"```",
+		"",
+		"### Step 2: Categorize",
+		"",
+		"Present findings as a numbered list grouped by severity:",
+		"- **C1, C2...** for critical issues",
+		"- **W1, W2...** for warnings",
+		"- **N1, N2...** for nits",
+		"",
+		"### Step 3: Fix",
+		"",
+		"After presenting the list, **immediately fix all Critical and Warning issues**.",
+		"Do NOT wait for confirmation — the user invoked `/assess cleave` which means",
+		'"assess and fix in one shot". Work through C and W items systematically.',
+		"Nits are optional — fix them if trivial, skip if not.",
+		"",
+		"After all fixes, run the test suite to verify nothing broke.",
+		"Then commit with a conventional commit message summarizing all fixes.",
+		...postAssessInstruction,
+	].join("\n");
+	const lifecycleRecord = targetChange
+		? await buildLifecycleRecord(pi, ctx.cwd, {
+			changeName: targetChange,
+			assessmentKind: "cleave",
+			outcome: "ambiguous",
+			recommendedAction: `Run openspec_manage reconcile_after_assess ${targetChange} with outcome pass, reopen, or ambiguous after review completes.`,
+		})
+		: undefined;
+	const humanText = [
+		"**Assess → Cleave pipeline starting...**",
+		"",
+		`Reviewing changes since \`${diffContext.ref}\`:`,
+		"```",
+		diffContext.diffStat,
+		"```",
+	].join("\n");
+	const nextSteps = ["Review findings", "Apply all Critical and Warning fixes", "Run verification and reconcile lifecycle state if needed"];
+	return makeAssessResult({
+		subcommand: "cleave",
+		args,
+		ok: true,
+		summary: `Prepared adversarial review for ${diffContext.ref}`,
+		humanText,
+		data: {
+			ref: diffContext.ref,
+			diffStat: diffContext.diffStat,
+			recentLog: diffContext.recentLog,
+			hasGuardrails: Boolean(guardrailPreamble),
+			reconcileChange: targetChange ?? null,
+			snapshot: lifecycleRecord?.snapshot ?? null,
+		},
+		effects: [
+			{ type: "view", content: humanText },
+			{ type: "follow_up", content: prompt },
+			...(lifecycle ? [{ type: "reconcile_hint" as const, ...lifecycle }] : []),
+		],
+		nextSteps,
+		lifecycle,
+		lifecycleRecord,
+	});
+}
+
+async function executeAssessDiff(
+	pi: ExtensionAPI,
+	ctx: AssessExecutionContext,
+	args: string,
+): Promise<AssessStructuredResult> {
+	const requestedRef = args.trim() || "HEAD~1";
+	const diffContext = await collectAssessDiffContext(pi, ctx.cwd, requestedRef, false);
+	if (!diffContext) {
+		const humanText = `No changes found relative to \`${requestedRef}\`.`;
+		return makeAssessResult({
+			subcommand: "diff",
+			args,
+			ok: false,
+			summary: `No diff found for ${requestedRef}`,
+			humanText,
+			data: { ref: requestedRef, reason: "no_changes" },
+			effects: [{ type: "view", content: humanText }],
+		});
+	}
+
+	const guardrailPreamble = buildGuardrailPreamble(ctx.cwd);
+	const humanText = [
+		`**Assessing diff since \`${diffContext.ref}\`...**`,
+		"```",
+		diffContext.diffStat,
+		"```",
+	].join("\n");
+	const prompt = [
+		`## Code Review: diff since \`${diffContext.ref}\``,
+		"",
+		"Do an adversarial code review of these changes.",
+		"Find bugs, fragile patterns, missing edge cases, and style issues.",
+		"",
+		...(guardrailPreamble ? [
+			"### Deterministic Analysis",
+			"",
+			guardrailPreamble,
+			"",
+			"The above are compiler/linter findings — treat failures as Critical issues.",
+			"",
+		] : []),
+		"Categorize findings as:",
+		"- **C1, C2...** Critical (logic errors, security, data loss)",
+		"- **W1, W2...** Warning (fragile, misleading, missing cases)",
+		"- **N1, N2...** Nit (style, dead code, minor)",
+		"",
+		"Diff stat:",
+		"```",
+		diffContext.diffStat,
+		"```",
+		"",
+		"```diff",
+		diffContext.diffContent,
+		"```",
+		"",
+		"Present findings only — do NOT fix anything unless I ask.",
+	].join("\n");
+	return makeAssessResult({
+		subcommand: "diff",
+		args,
+		ok: true,
+		summary: `Prepared review for ${diffContext.ref}`,
+		humanText,
+		data: {
+			ref: diffContext.ref,
+			diffStat: diffContext.diffStat,
+			hasGuardrails: Boolean(guardrailPreamble),
+		},
+		effects: [
+			{ type: "view", content: humanText },
+			{ type: "follow_up", content: prompt },
+		],
+		nextSteps: ["Read the review findings", "Decide whether to fix issues or continue implementation"],
+	});
+}
+
+async function executeAssessSpec(
+	pi: ExtensionAPI,
+	ctx: AssessExecutionContext,
+	args: string,
+	overrides?: AssessExecutorOverrides,
+): Promise<AssessStructuredResult> {
+	const repoPath = ctx.cwd;
+	const openspecDir = detectOpenSpec(repoPath);
+	if (!openspecDir) {
+		const humanText = "No `openspec/` directory found. Nothing to assess against.";
+		return makeAssessResult({
+			subcommand: "spec",
+			args,
+			ok: false,
+			summary: "OpenSpec directory not found",
+			humanText,
+			data: { reason: "openspec_missing" },
+			effects: [{ type: "view", content: humanText }],
+		});
+	}
+
+	const changes = findExecutableChanges(openspecDir);
+	if (changes.length === 0) {
+		const humanText = "No OpenSpec changes with tasks.md found.";
+		return makeAssessResult({
+			subcommand: "spec",
+			args,
+			ok: false,
+			summary: "No executable OpenSpec changes found",
+			humanText,
+			data: { reason: "no_executable_changes" },
+			effects: [{ type: "view", content: humanText }],
+		});
+	}
+
+	const requestedChange = args.trim();
+	let target = requestedChange
+		? changes.find((change) => change.name === requestedChange || change.name.includes(requestedChange))
+		: null;
+	if (!target) {
+		const status = getActiveChangesStatus(repoPath);
+		const withTasks = status.filter((entry) => entry.totalTasks > 0);
+		if (withTasks.length > 0) {
+			const byRecency = [...withTasks].sort((a, b) => b.lastModifiedMs - a.lastModifiedMs);
+			target = changes.find((change) => change.name === byRecency[0].name) ?? changes[0];
+		} else {
+			target = changes[0];
+		}
+	}
+
+	const specCtx = buildOpenSpecContext(target.path);
+	if (specCtx.specScenarios.length === 0) {
+		const humanText = `Change \`${target.name}\` has no delta spec scenarios to assess against.\n\nUse \`/assess diff\` for general code review instead.`;
+		return makeAssessResult({
+			subcommand: "spec",
+			args,
+			ok: false,
+			summary: `No delta scenarios found for ${target.name}`,
+			humanText,
+			data: { changeName: target.name, reason: "no_scenarios" },
+			effects: [{ type: "view", content: humanText }],
+		});
+	}
+
+	const scenarioText = specCtx.specScenarios.map((scenarioSet) => {
+		const renderedScenarios = scenarioSet.scenarios.map((scenario) => {
+			const lines = scenario.split("\n").map((line) => `    ${line}`).join("\n");
+			return lines;
+		}).join("\n");
+		return `**${scenarioSet.domain} → ${scenarioSet.requirement}**\n${renderedScenarios}`;
+	}).join("\n\n");
+
+	let diffContent = "";
+	try {
+		const diff = await pi.exec("git", ["diff", "HEAD~5", "--", "."], { cwd: repoPath, timeout: 10_000 });
+		diffContent = diff.stdout.slice(0, 30_000);
+	} catch {
+		try {
+			const diff = await pi.exec("git", ["diff", "--", "."], { cwd: repoPath, timeout: 10_000 });
+			diffContent = diff.stdout.slice(0, 30_000);
+		} catch {
+			/* proceed without diff */
+		}
+	}
+
+	const designContext = specCtx.decisions.length > 0
+		? [
+			"### Design Decisions",
+			"",
+			"The implementation should also reflect these decisions from design.md:",
+			"",
+			...specCtx.decisions.map((decision) => `- ${decision}`),
+			"",
+		]
+		: [];
+	const apiContractContext = specCtx.apiContract
+		? [
+			"### API Contract",
+			"",
+			"The implementation must conform to this OpenAPI/AsyncAPI contract (`api.yaml`).",
+			"Verify that:",
+			"- All paths/methods defined in the contract are implemented",
+			"- Request/response schemas match the contract exactly",
+			"- Status codes and error responses match the contract",
+			"- Security schemes are applied as specified",
+			"- Any endpoint in the code but NOT in the contract is flagged as undocumented",
+			"",
+			"```yaml",
+			specCtx.apiContract.length > 15_000
+				? specCtx.apiContract.slice(0, 15_000) + "\n# ... (truncated)"
+				: specCtx.apiContract,
+			"```",
+			"",
+		]
+		: [];
+	const lifecycle: AssessLifecycleHint = {
+		changeName: target.name,
+		assessmentKind: "spec",
+		outcomes: ["pass", "reopen", "ambiguous"],
+	};
+	const totalScenarioCount = countSpecScenarios(specCtx);
+	const introText = [
+		`**Spec Assessment: \`${target.name}\`**`,
+		"",
+		`Evaluating implementation against ${totalScenarioCount} spec scenarios`
+			+ (specCtx.decisions.length > 0 ? ` and ${specCtx.decisions.length} design decisions` : "")
+			+ (specCtx.apiContract ? " and API contract (`api.yaml`)" : "")
+			+ "...",
+	].join("\n");
+	const prompt = [
+		`## Spec-Driven Assessment: \`${target.name}\``,
+		"",
+		"Assess whether the current implementation satisfies these OpenSpec scenarios.",
+		"For each scenario, determine: **PASS**, **FAIL**, or **UNCLEAR**.",
+		...(specCtx.apiContract ? ["", "Also verify implementation conformance to the API contract."] : []),
+		"",
+		"### Acceptance Criteria",
+		"",
+		scenarioText,
+		"",
+		...designContext,
+		...apiContractContext,
+		"### Instructions",
+		"",
+		"1. Read the relevant source files to check each scenario",
+		...(specCtx.apiContract ? ["   - Also check route definitions, schemas, and status codes against the API contract"] : []),
+		"2. For each scenario, report:",
+		"   - **PASS** — implementation clearly satisfies the Given/When/Then",
+		"   - **FAIL** — implementation contradicts or is missing",
+		"   - **UNCLEAR** — can't determine without running tests",
+		"3. Summarize with a count: N/M scenarios passing",
+		"4. For any FAIL items, explain what's wrong and suggest fixes",
+		"5. Do NOT auto-fix — this is assessment only",
+		`6. After the assessment, if the result reopens work or reveals new constraints/file-scope drift, call \`openspec_manage\` with action \`reconcile_after_assess\`, change_name \`${target.name}\`, assessment_kind \`spec\`, and outcome \`reopen\` or \`ambiguous\` as appropriate. If all scenarios pass cleanly, call it with outcome \`pass\` to refresh lifecycle state.`,
+		...(diffContent ? ["", "### Recent Changes (for context)", "", "```diff", diffContent, "```"] : []),
+	].join("\n");
+
+	if (isInteractiveAssessContext(ctx)) {
+		const lifecycleRecord = await buildLifecycleRecord(pi, ctx.cwd, {
+			changeName: target.name,
+			assessmentKind: "spec",
+			outcome: "ambiguous",
+			recommendedAction: `Run openspec_manage reconcile_after_assess ${target.name} with outcome pass, reopen, or ambiguous after scenario evaluation completes.`,
+		});
+		return makeAssessResult({
+			subcommand: "spec",
+			args,
+			ok: true,
+			summary: `Prepared spec assessment for ${target.name}`,
+			humanText: introText,
+			data: {
+				changeName: target.name,
+				scenarioCount: totalScenarioCount,
+				decisionCount: specCtx.decisions.length,
+				hasApiContract: Boolean(specCtx.apiContract),
+				snapshot: lifecycleRecord.snapshot,
+			},
+			effects: [
+				{ type: "view", content: introText },
+				{ type: "follow_up", content: prompt },
+				{ type: "reconcile_hint" as const, ...lifecycle },
+			],
+			nextSteps: ["Assess each scenario", "Reconcile lifecycle state based on the assessment outcome"],
+			completion: { completed: false, completedInBand: false, requiresFollowUp: true },
+			lifecycle,
+			lifecycleRecord,
+		});
+	}
+
+	const runSpecAssessment = overrides?.runSpecAssessment ?? runSpecAssessmentSubprocess;
+	const completed = await runSpecAssessment({
+		repoPath,
+		changeName: target.name,
+		scenarioText,
+		designContext,
+		apiContractContext,
+		diffContent,
+		expectedScenarioCount: totalScenarioCount,
+		modelId: ctx.model?.id,
+	});
+	const assessed = normalizeSpecAssessment(completed.assessed, totalScenarioCount);
+	const outcome = determineSpecOutcome(assessed.summary);
+	const snapshot = completed.snapshot ?? await collectAssessmentSnapshot(pi, ctx.cwd);
+	const recommendedAction = `Call openspec_manage reconcile_after_assess for ${target.name} with assessment_kind spec and outcome ${outcome}.`;
+	const lifecycleRecord = await buildLifecycleRecord(pi, ctx.cwd, {
+		changeName: target.name,
+		assessmentKind: "spec",
+		outcome,
+		recommendedAction,
+		changedFiles: assessed.changedFiles,
+		constraints: assessed.constraints,
+		snapshot,
+	});
+	const humanText = buildSpecAssessmentHumanText(target.name, assessed, outcome);
+	const summary = `Completed spec assessment for ${target.name}: ${assessed.summary.pass}/${assessed.summary.total} pass, ${assessed.summary.fail} fail, ${assessed.summary.unclear} unclear`;
+	const nextSteps = [
+		`Call openspec_manage reconcile_after_assess for ${target.name} with outcome ${outcome}`,
+		...(outcome === "pass" ? [`If archive gates clear, run /opsx:archive ${target.name}`] : ["Address findings before archive"]),
+	];
+	return makeAssessResult({
+		subcommand: "spec",
+		args,
+		ok: true,
+		summary,
+		humanText,
+		data: {
+			changeName: target.name,
+			outcome,
+			scenarioSummary: assessed.summary,
+			scenarios: assessed.scenarios,
+			changedFiles: assessed.changedFiles ?? [],
+			constraints: assessed.constraints ?? [],
+			overallNotes: assessed.overallNotes ?? null,
+			snapshot,
+			recommendedReconcileOutcome: outcome,
+		},
+		effects: [{ type: "reconcile_hint" as const, ...lifecycle }],
+		nextSteps,
+		completion: { completed: true, completedInBand: true, requiresFollowUp: false, outcome },
+		lifecycle,
+		lifecycleRecord,
+	});
+}
+
+async function executeAssessComplexity(args: string): Promise<AssessStructuredResult> {
+	const directive = args.trim();
+	if (!directive) {
+		const humanText = "Usage: `/assess complexity <directive>`\n\nAssess whether a task should be decomposed or executed directly.";
+		return makeAssessResult({
+			subcommand: "complexity",
+			args,
+			ok: false,
+			summary: "Missing directive for complexity assessment",
+			humanText,
+			data: { reason: "missing_directive" },
+			effects: [{ type: "view", content: humanText }],
+		});
+	}
+
+	const assessment = assessDirective(directive);
+	const humanText = [
+		formatAssessment(assessment),
+		"",
+		assessment.decision === "cleave"
+			? `**→ Decomposition recommended.** Use \`/cleave ${directive}\` to proceed.`
+			: assessment.decision === "execute"
+				? "**→ Execute directly.** Task is below complexity threshold."
+				: "**→ Manual assessment needed.** No pattern matched.",
+	].join("\n");
+	return makeAssessResult({
+		subcommand: "complexity",
+		args,
+		ok: true,
+		summary: `Complexity decision: ${assessment.decision}`,
+		humanText,
+		data: assessment,
+		effects: [{ type: "view", content: humanText }],
+		nextSteps: assessment.decision === "cleave" ? [`Run /cleave ${directive}`] : ["Execute directly"],
+	});
+}
+
+async function runDesignAssessmentSubprocess(
+	repoPath: string,
+	nodeId: string,
+	modelId?: string,
+): Promise<{ findings: DesignAssessmentFinding[]; nodeTitle: string; structuralPass: boolean }> {
+	const prompt = [
+		"You are performing a read-only design-tree node assessment.",
+		"Operate in read-only plan mode. Never call edit, write, or any workspace-mutating command.",
+		"",
+		`## Task`,
+		"",
+		`1. Call design_tree with action='node', node_id='${nodeId}' to load the node.`,
+		"2. Run the structural pre-check:",
+		"   - open_questions must be empty — if not, emit a structural finding for each",
+		"   - decisions must have at least one entry — if not, emit a structural finding",
+		"   - acceptanceCriteria must have at least one scenario, falsifiability, or constraint — if not, emit a structural finding",
+		"3. If structural pre-check fails, output ONLY the JSON result below and stop.",
+		"4. Otherwise, evaluate each acceptance criterion against the document body:",
+		"   - For each Scenario (Given/When/Then): does the document body address the Then clause?",
+		"   - For each Falsifiability condition: is it addressed, ruled out, or acknowledged as a known risk?",
+		"   - For each Constraint: is it satisfied by the document content?",
+		"",
+		"## Output Format",
+		"",
+		"Output ONLY a single JSON object (no prose, no markdown, no code blocks):",
+		"{",
+		'  "nodeTitle": "<title from node>",',
+		'  "structuralPass": true|false,',
+		'  "findings": [',
+		'    {"type":"scenario"|"falsifiability"|"constraint"|"structural","index":N,"pass":true|false,"finding":"<reason>"}',
+		"  ]",
+		"}",
+	].join("\n");
+
+	const args = ["--mode", "json", "--plan", "-p", "--no-session"];
+	if (modelId) args.push("--model", modelId);
+
+	return await new Promise<{ findings: DesignAssessmentFinding[]; nodeTitle: string; structuralPass: boolean }>(
+		(resolve, reject) => {
+			const proc = spawn("pi", args, {
+				cwd: repoPath,
+				shell: false,
+				stdio: ["pipe", "pipe", "pipe"],
+				env: { ...process.env, PI_CHILD: "1", TERM: process.env.TERM ?? "dumb" },
+			});
+			let buffer = "";
+			let assistantText = "";
+			let settled = false;
+			const settleReject = (error: Error) => {
+				if (settled) return;
+				settled = true;
+				clearTimeout(timer);
+				reject(error);
+			};
+			const settleResolve = (value: { findings: DesignAssessmentFinding[]; nodeTitle: string; structuralPass: boolean }) => {
+				if (settled) return;
+				settled = true;
+				clearTimeout(timer);
+				resolve(value);
+			};
+			const timer = setTimeout(() => {
+				proc.kill("SIGTERM");
+				setTimeout(() => { if (!proc.killed) proc.kill("SIGKILL"); }, 5_000);
+				settleReject(new Error(`Timed out after 120s while assessing design node ${nodeId}.`));
+			}, 120_000);
+			const processLine = (line: string) => {
+				if (!line.trim()) return;
+				let event: unknown;
+				try { event = JSON.parse(line); } catch { return; }
+				if (!event || typeof event !== "object") return;
+				const typed = event as { type?: string; message?: { role?: string; content?: unknown } };
+				if (typed.type === "message_end" && typed.message?.role === "assistant") {
+					assistantText = extractAssistantText(typed.message.content);
+				}
+			};
+			proc.stdout.on("data", (data) => {
+				buffer += (data as Buffer).toString();
+				const lines = buffer.split("\n");
+				buffer = lines.pop() || "";
+				for (const line of lines) processLine(line);
+			});
+			let stderr = "";
+			proc.stderr.on("data", (data) => { stderr += (data as Buffer).toString(); });
+			proc.on("error", (error) => settleReject(error));
+			proc.on("close", (code) => {
+				if (buffer.trim()) processLine(buffer.trim());
+				if ((code ?? 1) !== 0) {
+					settleReject(new Error(stderr.trim() || `Design assessment subprocess exited with code ${code ?? 1}.`));
+					return;
+				}
+				const jsonText = extractJsonObject(assistantText || buffer);
+				if (!jsonText) {
+					settleReject(new Error(`Design assessment subprocess did not return parseable JSON.\n${stderr}`));
+					return;
+				}
+				try {
+					const parsed = JSON.parse(jsonText) as { nodeTitle?: string; structuralPass?: boolean; findings?: DesignAssessmentFinding[] };
+					settleResolve({
+						nodeTitle: parsed.nodeTitle ?? nodeId,
+						structuralPass: parsed.structuralPass ?? true,
+						findings: Array.isArray(parsed.findings) ? parsed.findings : [],
+					});
+				} catch (err) {
+					settleReject(new Error(`Design assessment JSON was invalid: ${String(err)}`));
+				}
+			});
+			proc.stdin.write(prompt + "\n");
+			proc.stdin.end();
+		},
+	);
+}
+
+async function executeAssessDesign(
+	pi: ExtensionAPI,
+	ctx: AssessExecutionContext,
+	args: string,
+): Promise<AssessStructuredResult> {
+	const cwd = ctx.cwd;
+	// Resolve: explicit arg → focused node → error
+	const nodeId = args.trim() || sharedState.designTree?.focusedNode?.id || null;
+
+	if (!nodeId) {
+		const humanText = "Usage: `/assess design <node-id>`\n\nProvide a design-tree node ID, or set a focused node via `design_tree_update` with action 'focus' and run `/assess design` without arguments.";
+		return makeAssessResult({
+			subcommand: "design",
+			args,
+			ok: false,
+			summary: "Missing node-id for design assessment",
+			humanText,
+			data: { reason: "missing_node_id" },
+			effects: [{ type: "view", content: humanText }],
+		});
+	}
+
+	// Build the interactive follow-up prompt
+	const interactivePrompt = [
+		`## Design Assessment: \`${nodeId}\``,
+		"",
+		"Assess this design-tree node for readiness to be marked as 'decided'.",
+		"",
+		"### Steps",
+		"",
+		`1. Call \`design_tree\` with \`action='node'\`, \`node_id='${nodeId}'\` to load the node and its document body.`,
+		"2. **Structural pre-check** (fail fast with specific finding per gap):",
+		"   - If `open_questions.length > 0`: FAIL — list each unresolved question",
+		"   - If `decisions.length === 0`: FAIL — no decisions recorded",
+		"   - If `acceptanceCriteria` has no scenarios, falsifiability, or constraints: FAIL — empty acceptance criteria",
+		"   - If any structural check fails, stop here and report findings.",
+		"3. **Acceptance criteria evaluation** (against the document body):",
+		"   - For each **Scenario** (Given/When/Then): does the document body address the Then clause?",
+		"   - For each **Falsifiability** condition: is it addressed, ruled out, or noted as a known risk?",
+		"   - For each **Constraint**: is it satisfied by the document content?",
+		"4. **Write `assessment.json`** to `openspec/design/${nodeId}/assessment.json` with structure:",
+		"   ```json",
+		`   {"nodeId":"${nodeId}","pass":true|false,"structuralPass":true|false,"findings":[...]}`,
+		"   ```",
+		"   Each finding: `{\"type\":\"scenario\"|\"falsifiability\"|\"constraint\"|\"structural\",\"index\":N,\"pass\":true|false,\"finding\":\"<reason>\"}`",
+		"5. **Report** overall PASS/FAIL with per-finding details.",
+		"   - If PASS: suggest `design_tree_update` with `set_status(decided)` for this node.",
+		"   - If FAIL: list each failing finding with an actionable fix.",
+	].join("\n");
+
+	if (isInteractiveAssessContext(ctx)) {
+		const introText = `Running design assessment for node \`${nodeId}\`…`;
+		return makeAssessResult({
+			subcommand: "design",
+			args,
+			ok: true,
+			summary: `Prepared design assessment for ${nodeId}`,
+			humanText: introText,
+			data: { nodeId },
+			effects: [
+				{ type: "view", content: introText },
+				{ type: "follow_up", content: interactivePrompt },
+			],
+			nextSteps: ["Evaluate acceptance criteria", "Write assessment.json", "Set status to decided if pass"],
+			completion: { completed: false, completedInBand: false, requiresFollowUp: true },
+		});
+	}
+
+	// Bridged / subprocess mode
+	let subResult: { findings: DesignAssessmentFinding[]; nodeTitle: string; structuralPass: boolean };
+	try {
+		subResult = await runDesignAssessmentSubprocess(cwd, nodeId, ctx.model?.id);
+	} catch (err) {
+		const msg = `Design assessment subprocess failed: ${String(err)}`;
+		return makeAssessResult({
+			subcommand: "design",
+			args,
+			ok: false,
+			summary: msg,
+			humanText: msg,
+			data: { reason: "subprocess_failed", nodeId },
+			effects: [{ type: "view", content: msg }],
+		});
+	}
+
+	const { findings, nodeTitle, structuralPass } = subResult;
+	const overallPass = structuralPass && findings.length > 0 && findings.every((f) => f.pass);
+
+	const result: DesignAssessmentResult = { nodeId, pass: overallPass, structuralPass, findings };
+
+	// Write assessment.json
+	await writeDesignAssessment(cwd, nodeId, result);
+
+	// Build human text
+	const failFindings = findings.filter((f) => !f.pass);
+	const passFindings = findings.filter((f) => f.pass);
+	const humanLines: string[] = [
+		`## Design Assessment: ${nodeTitle} (${nodeId})`,
+		"",
+		overallPass
+			? `**✅ PASS** — ${passFindings.length}/${findings.length} criteria satisfied. Ready to set status → decided.`
+			: structuralPass
+				? `**❌ FAIL** — ${failFindings.length}/${findings.length} criteria not satisfied.`
+				: "**❌ Structural pre-check failed** — resolve these issues before assessing.",
+		"",
+	];
+	if (failFindings.length > 0) {
+		humanLines.push("### Issues to Resolve");
+		for (const f of failFindings) humanLines.push(`- [${f.type}#${f.index}] ${f.finding}`);
+		humanLines.push("");
+	}
+	if (passFindings.length > 0) {
+		humanLines.push("### Satisfied");
+		for (const f of passFindings) humanLines.push(`- ✓ [${f.type}#${f.index}] ${f.finding}`);
+	}
+
+	const humanText = humanLines.join("\n");
+	const nextSteps = overallPass
+		? [`Run design_tree_update with action 'set_status', node_id '${nodeId}', status 'decided'`]
+		: failFindings.map((f) => f.finding.split(".")[0] ?? f.finding);
+
+	return makeAssessResult({
+		subcommand: "design",
+		args,
+		ok: overallPass,
+		summary: overallPass
+			? `Design node '${nodeId}' passed — ready to decide`
+			: `Design node '${nodeId}' failed — ${failFindings.length} issue(s) to resolve`,
+		humanText,
+		data: result,
+		effects: [{ type: "view", content: humanText }],
+		nextSteps,
+	});
+}
+
+async function writeDesignAssessment(cwd: string, nodeId: string, result: DesignAssessmentResult): Promise<void> {
+	try {
+		const { writeFile } = await import("node:fs/promises");
+		const { join } = await import("node:path");
+		const { existsSync } = await import("node:fs");
+		const dir = join(cwd, "openspec", "design", nodeId);
+		// Do NOT create the directory — if it doesn't exist the node has no design change
+		// scaffolded yet, and creating assessment.json here would trigger the "active not
+		// archived" gate on set_status(decided) and implement. Write only if already scaffolded.
+		if (!existsSync(dir)) return;
+		await writeFile(join(dir, "assessment.json"), JSON.stringify(result, null, 2), "utf8");
+	} catch {
+		// non-fatal — assessment result still returned to caller
+	}
+}
+
+export function createAssessStructuredExecutors(pi: ExtensionAPI, overrides?: AssessExecutorOverrides) {
+	return {
+		cleave: (args: string, ctx: AssessExecutionContext) => executeAssessCleave(pi, ctx, args),
+		diff: (args: string, ctx: AssessExecutionContext) => executeAssessDiff(pi, ctx, args),
+		spec: (args: string, ctx: AssessExecutionContext) => executeAssessSpec(pi, ctx, args, overrides),
+		complexity: (args: string) => executeAssessComplexity(args),
+		design: (args: string, ctx: AssessExecutionContext) => executeAssessDesign(pi, ctx, args),
+	} as const;
+}
+
+// ─── Extension ──────────────────────────────────────────────────────────────
+
+export default function cleaveExtension(pi: ExtensionAPI) {
+	// ── Initialize dashboard state ──────────────────────────────────
+	emitCleaveState(pi, "idle");
+
+	// ── Agent start: inject OpenSpec status into context ─────────────
+	// Uses before_agent_start (not session_start) so the status message
+	// enters the agent's conversation context, not just the TUI display.
+	let openspecFirstTurn = true;
+
+	pi.on("before_agent_start", (_event, ctx) => {
+		if (!openspecFirstTurn) return;
+		openspecFirstTurn = false;
+
+		try {
+			const status = getActiveChangesStatus(ctx.cwd);
+			if (status.length === 0) return;
+
+			const lines = ["**OpenSpec Changes**", ""];
+			for (const s of status) {
+				const progress = s.totalTasks > 0
+					? `${s.doneTasks}/${s.totalTasks} tasks`
+					: "no tasks";
+				const artifacts: string[] = [];
+				if (s.hasProposal) artifacts.push("proposal");
+				if (s.hasDesign) artifacts.push("design");
+				if (s.hasSpecs) artifacts.push("specs");
+				const artStr = artifacts.length > 0 ? ` [${artifacts.join(", ")}]` : "";
+
+				const icon = s.totalTasks > 0 && s.doneTasks >= s.totalTasks ? "✓" : "◦";
+				lines.push(`  ${icon} **${s.name}** — ${progress}${artStr}`);
+			}
+
+			const incomplete = status.filter((s) => s.totalTasks > 0 && s.doneTasks < s.totalTasks);
+			if (incomplete.length > 0) {
+				lines.push("", `Use \`/opsx:apply\` to continue or \`/cleave\` to parallelize.`);
+			}
+
+			const withTasks = status.filter((s) => s.totalTasks > 0);
+			const allDone = withTasks.length > 0 && withTasks.every((s) => s.doneTasks >= s.totalTasks);
+			if (allDone) {
+				lines.push("", `All tasks complete. Run \`/opsx:verify\` → \`/opsx:archive\` to finalize.`);
+			}
+
+			const content = lines.join("\n");
+			return {
+				message: {
+					customType: "openspec-status",
+					content,
+					display: true,
+				},
+			};
+		} catch {
+			// Non-fatal — don't block agent start
+		}
+	});
+
+
+	// ── cleave_assess tool ───────────────────────────────────────────────
+	pi.registerTool({
+		name: "cleave_assess",
+		label: "Cleave Assess",
+		description:
+			"Assess the complexity of a task directive to determine if it should be " +
+			"decomposed (cleaved) into subtasks or executed directly. Returns complexity " +
+			"score, matched pattern, confidence, and decision (execute/cleave).\n\n" +
+			"Use before attempting complex multi-system tasks to decide whether decomposition is warranted.",
+		promptSnippet:
+			"Assess task complexity for decomposition — returns pattern match, complexity score, and execute/cleave decision",
+		promptGuidelines: [
+			"Call cleave_assess before starting any multi-system or cross-cutting task to determine if decomposition is needed",
+			"If decision is 'execute', proceed directly. If 'cleave', use /cleave to decompose. If 'needs_assessment', proceed directly — it means no pattern matched but the task is likely simple enough for in-session execution.",
+			"Complexity formula: (1 + systems) × (1 + 0.5 × modifiers). Threshold default: 2.0.",
+			"The /assess command provides code assessment: `/assess cleave` (adversarial review + auto-fix), `/assess diff [ref]` (review only), `/assess spec [change]` (validate against OpenSpec scenarios), `/assess design [node-id]` (evaluate design-tree node readiness before set_status(decided)).",
+			"When the repo has openspec/ with active changes, suggest `/assess spec` after implementation and before `/opsx:archive`.",
+			"Run `/assess design <node-id>` before calling design_tree_update with set_status(decided) to verify acceptance criteria are satisfied.",
+		],
+
+		parameters: Type.Object({
+			directive: Type.String({ description: "The task directive to assess" }),
+			threshold: Type.Optional(Type.Number({ description: "Complexity threshold (default: 2.0)" })),
+		}),
+
+		renderCall(args, theme) {
+			const dir = args.directive.length > 55
+				? args.directive.slice(0, 52) + "…"
+				: args.directive;
+			return sciCall("cleave_assess", dir, theme);
+		},
+
+		renderResult(result, { expanded }, theme) {
+			const d = result.details as {
+				score?: number; complexity?: number; decision?: string;
+				systems?: number; modifiers?: string[]; pattern?: string; method?: string;
+			} | undefined;
+			const score = d?.score ?? d?.complexity;
+			const decision = d?.decision;
+			const scoreStr = score != null ? `complexity ${score.toFixed(1)}` : "";
+			const decisionStr = decision ? `→ ${decision}` : "";
+			const summary = [scoreStr, decisionStr].filter(Boolean).join("  ");
+
+			if (expanded) {
+				const lines: string[] = [];
+				// Structured breakdown
+				if (d?.pattern) lines.push(`${theme.fg("accent", "Pattern")} ${theme.fg("muted", d.pattern)}`);
+				if (d?.method) lines.push(`${theme.fg("accent", "Method")} ${theme.fg("muted", d.method)}`);
+				if (d?.systems != null) lines.push(`${theme.fg("accent", "Systems")} ${theme.fg("muted", String(d.systems))}`);
+				if (d?.modifiers?.length) lines.push(`${theme.fg("accent", "Modifiers")} ${theme.fg("muted", d.modifiers.join(", "))}`);
+				if (score != null) {
+					const color = score >= 2.0 ? "warning" : "success";
+					lines.push(`${theme.fg("accent", "Score")} ${theme.fg(color as any, score.toFixed(1))}`);
+				}
+				if (decision) {
+					const color = decision === "cleave" ? "warning" : "success";
+					lines.push(`${theme.fg("accent", "Decision")} ${theme.fg(color as any, decision)}`);
+				}
+				if (lines.length === 0) {
+					// Fallback to raw text
+					const text = (result.content?.[0] && "text" in result.content[0] ? result.content[0].text : "") ?? "";
+					lines.push(...text.split("\n"));
+				}
+				return sciExpanded(lines, summary, theme);
+			}
+
+			if (!summary) {
+				const first = result.content?.[0];
+				return sciOk(((first && "text" in first ? first.text : null) ?? "").split("\n")[0].slice(0, 80), theme);
+			}
+
+			return sciOk(summary, theme);
+		},
+
+		async execute(_toolCallId, params, _signal, _onUpdate, _ctx) {
+			const assessment = assessDirective(params.directive, params.threshold ?? DEFAULT_CONFIG.threshold);
+			const text = formatAssessment(assessment);
+
+			return {
+				content: [{ type: "text", text }],
+				details: {
+					...assessment,
+					availablePatterns: Object.values(PATTERNS).map((p) => p.name),
+				},
+			};
+		},
+	});
+
+	// ── /assess command ──────────────────────────────────────────────────
+	const ASSESS_SUBS = [
+		{ value: "cleave", label: "cleave", description: "Adversarial review → auto-fix (optional: ref)" },
+		{ value: "diff", label: "diff", description: "Assess uncommitted or recent changes for issues" },
+		{ value: "spec", label: "spec", description: "Assess implementation against OpenSpec scenarios" },
+		{ value: "complexity", label: "complexity", description: "Assess directive complexity (cleave_assess)" },
+		{ value: "design", label: "design", description: "Assess design-tree node readiness before set_status(decided)" },
+	];
+	const assessExecutors = createAssessStructuredExecutors(pi);
+	const slashCommandBridge = getSharedBridge();
+	const toBridgeAssessResult = (
+		bridgedArgs: string[],
+		result: AssessStructuredResult,
+	): ReturnType<typeof buildSlashCommandResult> => buildAssessBridgeResult(bridgedArgs, result);
+
+	slashCommandBridge.register(pi, {
+		name: "assess",
+		description: "Adversarial review + auto-fix (default), or: /assess <diff|spec|complexity> [args]",
+		bridge: {
+			agentCallable: true,
+			sideEffectClass: "workspace-write",
+			resultContract: "cleave.assess.v1",
+			summary: "Lifecycle-safe assessment commands for spec, diff, cleave, and complexity",
+		},
+		getArgumentCompletions: (prefix: string) => {
+			const parts = prefix.split(" ");
+			if (parts.length <= 1) {
+				const partial = parts[0] || "";
+				const filtered = ASSESS_SUBS.filter((s) => s.value.startsWith(partial));
+				return filtered.length > 0 ? filtered : null;
+			}
+			return null;
+		},
+		structuredExecutor: async (args, ctx) => {
+			const trimmed = (args || "").trim();
+			if (!trimmed) {
+				return buildSlashCommandResult("assess", [], {
+					ok: false,
+					summary: "/assess requires an explicit bridged subcommand",
+					humanText: "Bare /assess remains interactive-only in v1. Use one of: /assess spec, /assess diff, /assess cleave, or /assess complexity.",
+					data: { supportedSubcommands: ASSESS_SUBS.map((sub) => sub.value) },
+					effects: { sideEffectClass: "workspace-write" },
+					nextSteps: ASSESS_SUBS.map((sub) => ({ label: `Run /assess ${sub.value}` })),
+				});
+			}
+
+			const parts = trimmed.split(/\s+/);
+			const sub = parts[0] || "";
+			const rest = parts.slice(1).join(" ");
+			const assessCtx: AssessExecutionContext = {
+				cwd: ctx.cwd,
+				bridgeInvocation: (ctx as { bridgeInvocation?: boolean }).bridgeInvocation,
+				hasUI: ctx.hasUI,
+				model: ctx.model ? { id: ctx.model.id } : undefined,
+				waitForIdle: "waitForIdle" in ctx && typeof ctx.waitForIdle === "function"
+					? ctx.waitForIdle.bind(ctx)
+					: undefined,
+			};
+			switch (sub) {
+				case "cleave":
+					return toBridgeAssessResult(parts, await assessExecutors.cleave(rest, assessCtx));
+				case "diff":
+					return toBridgeAssessResult(parts, await assessExecutors.diff(rest, assessCtx));
+				case "spec":
+					return toBridgeAssessResult(parts, await assessExecutors.spec(rest, assessCtx));
+				case "complexity":
+					return toBridgeAssessResult(parts, await assessExecutors.complexity(rest));
+				case "design":
+					return toBridgeAssessResult(parts, await assessExecutors.design(rest, assessCtx));
+				default:
+					return buildSlashCommandResult("assess", parts, {
+						ok: false,
+						summary: `Unsupported bridged /assess target: ${sub}`,
+						humanText: `Bridged /assess currently supports only: ${ASSESS_SUBS.map((item) => item.value).join(", ")}. Freeform adversarial review remains interactive-only in v1.`,
+						data: { supportedSubcommands: ASSESS_SUBS.map((item) => item.value) },
+						effects: { sideEffectClass: "workspace-write" },
+						nextSteps: ASSESS_SUBS.map((item) => ({ label: `Run /assess ${item.value}` })),
+					});
+			}
+		},
+		interactiveHandler: async (result, args) => {
+			const trimmed = (args || "").trim();
+			const sub = trimmed.split(/\s+/)[0] || "";
+			if (!trimmed || !ASSESS_SUBS.some((item) => item.value === sub)) {
+				pi.sendUserMessage([
+					"# Adversarial Assessment",
+					"",
+					trimmed
+						? "You are now operating as a hostile reviewer. Your job is to find everything wrong with the work completed in this session."
+						: "You are now operating as a hostile reviewer. Your job is to find everything wrong with the work completed in this session. Do not be polite. Do not hedge. If something is broken, say it's broken.",
+					...(trimmed ? ["", "**User instructions:** " + trimmed] : []),
+					"",
+					"Follow the user's instructions above for tone and scope, but still perform a thorough review.",
+					"Read every file that was changed. Be specific. Cite line numbers.",
+					"",
+					"## Output Format",
+					"",
+					"### Verdict",
+					"One of: `PASS` | `PASS WITH CONCERNS` | `NEEDS REWORK` | `REJECT`",
+					"",
+					"### Critical Issues",
+					"### Warnings",
+					"### Nitpicks",
+					"### Omissions",
+					"### What Actually Worked",
+				].join("\n"));
+				return;
+			}
+			const assessResult: AssessStructuredResult = {
+				command: "assess",
+				subcommand: (result.data as any)?.subcommand ?? "diff",
+				args: trimmed,
+				ok: result.ok,
+				summary: result.summary,
+				humanText: result.humanText,
+				data: (result.data as any)?.data,
+				effects: (result.data as any)?.assessEffects ?? [],
+				nextSteps: (result.nextSteps ?? []).map((step) => step.label),
+				completion: (result.data as any)?.completion,
+				lifecycle: (result.data as any)?.lifecycleHint,
+				lifecycleRecord: result.lifecycle as AssessLifecycleRecord | undefined,
+			};
+			applyAssessEffects(pi, assessResult);
+		},
+		agentHandler: async (result, args) => {
+			const trimmed = (args || "").trim();
+			const assessResult: AssessStructuredResult = {
+				command: "assess",
+				subcommand: (result.data as any)?.subcommand ?? "diff",
+				args: trimmed,
+				ok: result.ok,
+				summary: result.summary,
+				humanText: result.humanText,
+				data: (result.data as any)?.data,
+				effects: (result.data as any)?.assessEffects ?? [],
+				nextSteps: (result.nextSteps ?? []).map((step) => step.label),
+				completion: (result.data as any)?.completion,
+				lifecycle: (result.data as any)?.lifecycleHint,
+				lifecycleRecord: result.lifecycle as AssessLifecycleRecord | undefined,
+			};
+			applyAssessEffects(pi, assessResult);
+		},
+	});
+	pi.registerTool(slashCommandBridge.createToolDefinition());
+
+	// ── /cleave command ──────────────────────────────────────────────────
+	pi.registerCommand("cleave", {
+		description: "Recursive task decomposition (usage: /cleave <directive>)",
+		handler: async (args, ctx) => {
+			const directive = (args || "").trim();
+
+			if (!directive) {
+				pi.sendMessage({
+					customType: "view",
+					content: [
+						"**Cleave — Recursive Task Decomposition**",
+						"",
+						"Usage: `/cleave <directive>`",
+						"",
+						"Example: `/cleave Implement JWT authentication with refresh tokens`",
+						"",
+						"The directive will be assessed for complexity. If it exceeds the",
+						"threshold, it will be decomposed into 2-4 child tasks executed",
+						"in parallel via git worktrees.",
+						"",
+						"Available patterns: " + Object.values(PATTERNS).map((p) => p.name).join(", "),
+					].join("\n"),
+					display: true,
+				});
+				return;
+			}
+
+			// Delegate the full workflow to the LLM via a structured prompt.
+			// This allows the LLM to handle the interactive confirm gates
+			// and adapt to user feedback, while we provide all the mechanical
+			// infrastructure via tools.
+
+			const assessment = assessDirective(directive);
+			const assessmentText = formatAssessment(assessment);
+
+			if (assessment.decision === "execute" || assessment.decision === "needs_assessment") {
+				pi.sendMessage({
+					customType: "view",
+					content: [
+						assessmentText,
+						"",
+						assessment.decision === "needs_assessment"
+							? "**→ Execute directly** — no pattern matched; heuristic suggests in-session execution."
+							: "**→ Execute directly** — complexity is below threshold.",
+						"Proceeding with the task in-session.",
+					].join("\n"),
+					display: true,
+				});
+
+				// Hand off to the LLM to execute directly
+				pi.sendUserMessage(
+					`Execute this task directly (cleave assessment says it's simple enough):\n\n${directive}`,
+					{ deliverAs: "followUp" },
+				);
+				return;
+			}
+
+			// Task needs cleaving — check for OpenSpec first, then fall back to LLM
+			const repoPath = ctx.cwd;
+
+			// ── OpenSpec fast path ─────────────────────────────────────
+			const openspecDir = detectOpenSpec(repoPath);
+			if (openspecDir) {
+				const executableChanges = findExecutableChanges(openspecDir);
+				if (executableChanges.length > 0) {
+					// Try to find a change whose name matches the directive.
+					// Three strategies: exact slug containment, word overlap, partial prefix.
+					const directiveSlug = directive.toLowerCase().replace(/[^\w]+/g, "-");
+					const directiveWords = new Set(
+						directive.toLowerCase().replace(/[^\w\s]/g, "").split(/\s+/).filter((w) => w.length > 2),
+					);
+
+					const matched = executableChanges.find((c) => {
+						// Strategy 1: slug containment (either direction)
+						if (directiveSlug.includes(c.name) || c.name.includes(directiveSlug.slice(0, 20))) return true;
+						// Strategy 2: word overlap — change name words appear in directive
+						const changeWords = c.name.split("-").filter((w) => w.length > 2);
+						const overlap = changeWords.filter((w) => directiveWords.has(w)).length;
+						if (changeWords.length > 0 && overlap >= Math.ceil(changeWords.length * 0.5)) return true;
+						return false;
+					});
+
+					// Only use OpenSpec if we found a matching change — never silently
+					// pick an unrelated change
+					if (!matched) {
+						// No match — mention available changes but fall through to LLM planner
+						pi.sendMessage({
+							customType: "view",
+							content: [
+								`OpenSpec changes found but none matched the directive.`,
+								`Available: ${executableChanges.map((c) => c.name).join(", ")}`,
+								`Falling back to LLM planner.`,
+							].join("\n"),
+							display: true,
+						});
+					}
+					const change = matched!;
+					const result = change ? openspecChangeToSplitPlanWithContext(change.path) : null;
+
+					if (result) {
+						const { plan, context } = result;
+						const planJson = JSON.stringify(plan, null, 2);
+
+						// Report what OpenSpec artifacts we found
+						const artifactNotes: string[] = [];
+						if (context.designContent) artifactNotes.push(`design.md (${context.decisions.length} decisions, ${context.fileChanges.length} file changes)`);
+						if (context.specScenarios.length > 0) artifactNotes.push(`specs (${context.specScenarios.length} scenarios for post-merge verification)`);
+
+						pi.sendMessage({
+							customType: "view",
+							content: [
+								assessmentText,
+								"",
+								`**→ OpenSpec plan detected** from \`${change.name}/tasks.md\``,
+								...(artifactNotes.length > 0 ? [`**Artifacts:** ${artifactNotes.join("; ")}`] : []),
+								"",
+								`**Rationale:** ${plan.rationale}`,
+								`**Children:** ${plan.children.map((c) => c.label).join(", ")}`,
+								"",
+								"Review the plan and confirm to execute via `cleave_run`.",
+							].join("\n"),
+							display: true,
+						});
+
+						pi.sendUserMessage(
+							[
+								"## Cleave Decomposition (OpenSpec)",
+								"",
+								`OpenSpec change \`${change.name}\` provides a pre-built split plan.`,
+								"",
+								"### Split Plan",
+								"",
+								"```json",
+								planJson,
+								"```",
+								"",
+								"Present this plan to the user for review. After confirmation,",
+								`use the \`cleave_run\` tool with this plan_json, the original directive,`,
+								`and \`openspec_change_path\` set to \`${change.path}\`.`,
+								"",
+								"### Original Directive",
+								"",
+								directive,
+							].join("\n"),
+							{ deliverAs: "followUp" },
+						);
+						return;
+					}
+				}
+			}
+
+			// ── LLM planning fallback ──────────────────────────────────
+			let repoTree: string;
+			try {
+				repoTree = await getRepoTree(pi, repoPath);
+			} catch {
+				repoTree = "(unable to read repo structure)";
+			}
+
+			const plannerPrompt = buildPlannerPrompt(directive, repoTree, []);
+
+			pi.sendMessage({
+				customType: "view",
+				content: [
+					assessmentText,
+					"",
+					"**→ Decomposition needed.** Generating split plan...",
+				].join("\n"),
+				display: true,
+			});
+
+			// Delegate to the LLM to:
+			// 1. Generate a split plan (can use ask_local_model or think about it)
+			// 2. Present the plan for confirmation
+			// 3. Execute via cleave_run tool
+			pi.sendUserMessage(
+				[
+					`## Cleave Decomposition`,
+					"",
+					`The directive needs decomposition (complexity ${assessment.complexity}, pattern: ${assessment.pattern || "none"}).`,
+					"",
+					"### Step 1: Generate a split plan",
+					"",
+					"Use `ask_local_model` with this planning prompt to generate a JSON split plan:",
+					"",
+					"```",
+					plannerPrompt,
+					"```",
+					"",
+					"Parse the JSON response and present the plan to me for review.",
+					"",
+					"### Step 2: After I confirm",
+					"",
+					"Use the `cleave_run` tool with the plan to execute the decomposition.",
+					"",
+					"### Original Directive",
+					"",
+					directive,
+				].join("\n"),
+				{ deliverAs: "followUp" },
+			);
+		},
+	});
+
+	// ── /cleave resume command ────────────────────────────────────────
+	/**
+	 * Resume an interrupted cleave run.
+	 *
+	 * When a cleave session is killed before the harvest/merge phase the
+	 * workspace state file is left with `phase: "dispatch"` and some children
+	 * still `pending`.  This command:
+	 *
+	 *   1. Finds the most recent interrupted run for the current repo.
+	 *   2. Re-dispatches pending children (dispatchChildren skips completed ones).
+	 *   3. Runs the harvest/merge phase and cleans up worktrees.
+	 *
+	 * The full harvest is intentionally the same code path as cleave_run
+	 * (minus OpenSpec write-back which requires the original change path).
+	 */
+	pi.registerCommand("cleave resume", {
+		description: "Resume an interrupted cleave run — re-dispatch pending children and complete the harvest/merge phase",
+		handler: async (_args, ctx) => {
+			const repoPath = ctx.cwd;
+			const signal: AbortSignal | undefined = (ctx as any).signal;
+
+			const incomplete = findIncompleteRuns(repoPath);
+			if (incomplete.length === 0) {
+				pi.sendMessage({
+					customType: "view",
+					content: "**Cleave Resume** — no interrupted runs found for this repository.",
+					display: true,
+				});
+				return;
+			}
+
+			const state = incomplete[0]!;
+
+			const emit = (text: string) => pi.sendMessage({ customType: "view", content: text, display: true });
+
+			const completedBefore = state.children.filter((c) => c.status === "completed").length;
+			const pendingChildren = state.children.filter(
+				(c) => c.status !== "completed" && c.status !== "failed",
+			);
+
+			const header = [
+				`**Cleave Resume** — \`${state.runId}\``,
+				`Directive: ${state.directive}`,
+				`Base branch: \`${state.baseBranch}\``,
+				"",
+				...state.children.map((c) => {
+					const icon = c.status === "completed" ? "✅" : c.status === "failed" ? "❌" : "⏳";
+					return `  ${icon} [${c.childId}] \`${c.label}\` — ${c.status}`;
+				}),
+				"",
+				`${completedBefore} already completed, ${pendingChildren.length} to dispatch`,
+			].join("\n");
+			emit(header);
+
+			// ── Re-dispatch any pending children ──────────────────────
+			if (pendingChildren.length > 0) {
+				emit(`Resuming dispatch for ${pendingChildren.length} pending child(ren)…`);
+				emitCleaveState(pi, "dispatching", state.runId, state.children);
+				await dispatchChildren(
+					pi,
+					state,
+					4, // maxParallel
+					120 * 60 * 1000,
+					undefined,
+					signal,
+					(msg) => emit(msg),
+					DEFAULT_REVIEW_CONFIG,
+				);
+				saveState(state);
+			}
+
+			// ── Harvest + merge ────────────────────────────────────────
+			emitCleaveState(pi, "merging", state.runId, state.children);
+			state.phase = "harvest";
+			saveState(state);
+
+			const taskContents = readTaskFiles(state.workspacePath);
+			const taskResults = [...taskContents.entries()].map(([id, content]) =>
+				parseTaskResult(content, `${id}-task.md`),
+			);
+			const conflicts = detectConflicts(taskResults);
+
+			state.phase = "reunify";
+			saveState(state);
+
+			const completedChildren2 = state.children.filter((c) => c.status === "completed");
+			const mergeResults: Array<{ label: string; branch: string; success: boolean; conflicts: string[] }> = [];
+
+			for (const child of completedChildren2) {
+				const result = await mergeBranch(pi, repoPath, child.branch, state.baseBranch);
+				mergeResults.push({ label: child.label, branch: child.branch, success: result.success, conflicts: result.conflictFiles });
+				if (!result.success) break;
+			}
+
+			const mergeFailures = mergeResults.filter((m) => !m.success);
+
+			if (mergeResults.length > 0 && mergeFailures.length === 0) {
+				await cleanupWorktrees(pi, repoPath);
+			} else {
+				await pruneWorktreeDirs(pi, repoPath);
+			}
+
+			// ── Finalise state ─────────────────────────────────────────
+			const allOk =
+				state.children.every((c) => c.status === "completed") &&
+				mergeResults.every((m) => m.success) &&
+				conflicts.length === 0;
+
+			state.phase = allOk ? "complete" : "failed";
+			state.completedAt = new Date().toISOString();
+			state.totalDurationSec = Math.round(
+				(new Date(state.completedAt).getTime() - new Date(state.createdAt).getTime()) / 1000,
+			);
+			emitCleaveState(pi, allOk ? "done" : "failed", state.runId, state.children);
+			saveState(state);
+
+			const completedCount = state.children.filter((c) => c.status === "completed").length;
+			const failedCount = state.children.filter((c) => c.status === "failed").length;
+
+			const report = [
+				`## Cleave Resume Report: ${state.runId}`,
+				"",
+				`**Status:** ${allOk ? "✓ COMPLETE" : "✗ ISSUES"}`,
+				`**Children:** ${completedCount} completed, ${failedCount} failed of ${state.children.length}`,
+				`**Merges:** ${mergeResults.filter((m) => m.success).length} succeeded, ${mergeFailures.length} failed`,
+				conflicts.length > 0 ? `\n${formatConflicts(conflicts)}` : "",
+				mergeFailures.length > 0
+					? `\n**Merge failures:**\n${mergeFailures.map((m) => `  • \`${m.branch}\`: ${m.conflicts.join(", ") || "unknown error"}`).join("\n")}`
+					: "",
+			].filter(Boolean).join("\n");
+			emit(report);
+		},
+	});
+
+	// ── cleave_run tool ──────────────────────────────────────────────────
+	pi.registerTool({
+		name: "cleave_run",
+		label: "Cleave Run",
+		description:
+			"Execute a cleave decomposition plan. Creates git worktrees for each child, " +
+			"dispatches child pi processes, harvests results, detects conflicts, and " +
+			"merges branches back. Requires a split plan (from cleave_assess + planning).\n\n" +
+			"Each child runs in an isolated git worktree on its own branch.",
+		promptSnippet:
+			"Execute a cleave decomposition plan — parallel child dispatch in git worktrees, conflict detection, merge, and report",
+		promptGuidelines: [
+			"When an OpenSpec change was used to generate the plan, ALWAYS pass `openspec_change_path` so child tasks get design context and tasks.md is reconciled on completion.",
+			"Treat lifecycle reconciliation as required: after cleave_run, ensure tasks.md, design-tree status, and dashboard-facing progress reflect the merged reality before archive.",
+			"After cleave_run completes with OpenSpec, follow the Next Steps in the report (typically `/assess spec` → `/opsx:verify` → `/opsx:archive`).",
+		],
+		parameters: Type.Object({
+			directive: Type.String({ description: "The original task directive" }),
+			plan_json: Type.String({
+				description:
+					'JSON string of the split plan: {"children": [{"label": "...", "description": "...", "scope": [...], "depends_on": [...]}], "rationale": "..."}',
+			}),
+			prefer_local: Type.Optional(
+				Type.Boolean({ description: "Use local model for leaf tasks when possible (default: true)" }),
+			),
+			max_parallel: Type.Optional(
+				Type.Number({ description: "Maximum parallel children (default: 4)" }),
+			),
+			openspec_change_path: Type.Optional(
+				Type.String({
+					description:
+						"Path to an OpenSpec change directory. When provided, child task files are " +
+						"enriched with design.md context (architecture decisions, file scope) and " +
+						"post-merge verification checks specs against implementation.",
+				}),
+			),
+			review: Type.Optional(
+				Type.Boolean({
+					description:
+						"Enable adversarial review loop after each child completes. " +
+						"Runs an gloriana-tier reviewer that checks for bugs, security issues, " +
+						"and spec compliance. Severity-gated fix iterations with churn detection. " +
+						"Default: false.",
+				}),
+			),
+			review_max_warning_fixes: Type.Optional(
+				Type.Number({
+					description: "Maximum fix iterations for warning-level issues (default: 1)",
+				}),
+			),
+			review_max_critical_fixes: Type.Optional(
+				Type.Number({
+					description: "Maximum fix iterations for critical issues before escalation (default: 2)",
+				}),
+			),
+			review_churn_threshold: Type.Optional(
+				Type.Number({
+					description: "Fraction of reappearing issues that triggers churn bail (default: 0.5)",
+				}),
+			),
+		}),
+
+		renderCall(args, theme) {
+			const plan = (() => {
+				try { return JSON.parse(args.plan_json); } catch { return null; }
+			})();
+			const n = Array.isArray(plan?.children) ? plan.children.length : "?";
+			const dir = args.directive.length > 50
+				? args.directive.slice(0, 47) + "…"
+				: args.directive;
+			return sciCall("cleave_run", `${n} children · ${dir}`, theme);
+		},
+
+		renderResult(result, { expanded, isPartial }, theme) {
+			if (isPartial) {
+				// Phase-aware child table from details
+				const d = result?.details as {
+					children?: Array<{ label: string; status: string; branch?: string }>;
+					phase?: string;
+				} | undefined;
+				const children = d?.children ?? [];
+				if (children.length === 0) {
+					const msg = result.content?.[0];
+					const txt = (msg && "text" in msg ? msg.text : null) ?? "running…";
+					const phase = d?.phase ? theme.fg("dim", ` [${d.phase}]`) : "";
+					return sciOk(txt.split("\n")[0].slice(0, 60) + phase, theme);
+				}
+				const done = children.filter((c) => c.status === "completed").length;
+				const failed = children.filter((c) => c.status === "failed").length;
+				const running = children.filter((c) => c.status === "running").length;
+				const total = children.length;
+				const phase = d?.phase ?? "running";
+
+				const parts = [`${done}/${total} done`];
+				if (running > 0) parts.push(`${running} running`);
+				if (failed > 0) parts.push(theme.fg("error", `${failed} failed`));
+				const footer = parts.join("  ") + theme.fg("dim", ` · ${phase}`);
+
+				const rows = children.slice(0, 10).map((c) => {
+					const icon =
+						c.status === "completed" ? theme.fg("success", "✓")
+						: c.status === "running" ? theme.fg("warning", "⟳")
+						: c.status === "failed" ? theme.fg("error", "✕")
+						: theme.fg("muted", "○");
+					const label = c.status === "running"
+						? theme.fg("accent", c.label)
+						: c.status === "failed"
+						? theme.fg("error", c.label)
+						: theme.fg("dim", c.label);
+					return `${icon} ${label}`;
+				});
+				if (children.length > 10) {
+					rows.push(theme.fg("muted", `… ${children.length - 10} more`));
+				}
+				return sciExpanded(rows, footer, theme);
+			}
+
+			// Final result
+			const first = result.content?.[0];
+			const text = (first && "text" in first ? first.text : null) ?? "";
+			const isError = (result as any).isError;
+			const hasConflicts = text.toLowerCase().includes("conflict");
+
+			// Extract structured info from details for expanded view
+			const d = result?.details as {
+				children?: Array<{ label: string; status: string }>;
+				merged?: number; failed?: number; conflicts?: number;
+				duration?: number; filesChanged?: number;
+			} | undefined;
+
+			if (expanded) {
+				const lines: string[] = [];
+				const children = d?.children ?? [];
+				if (children.length > 0) {
+					for (const c of children.slice(0, 12)) {
+						const icon =
+							c.status === "completed" ? theme.fg("success", "✓")
+							: c.status === "failed" ? theme.fg("error", "✕")
+							: theme.fg("muted", "○");
+						const label = c.status === "failed"
+							? theme.fg("error", c.label)
+							: theme.fg("muted", c.label);
+						lines.push(`${icon} ${label}`);
+					}
+					if (children.length > 12) {
+						lines.push(theme.fg("muted", `… ${children.length - 12} more`));
+					}
+				} else {
+					// Fall back to raw text
+					lines.push(...text.split("\n").slice(0, 15));
+				}
+
+				const merged = d?.merged ?? children.filter(c => c.status === "completed").length;
+				const failed = d?.failed ?? children.filter(c => c.status === "failed").length;
+				const total = children.length || "?";
+				const dur = d?.duration != null ? ` · ${(d.duration / 1000).toFixed(0)}s` : "";
+				const footer = failed > 0
+					? `${merged}/${total} merged  ${theme.fg("error", `${failed} failed`)}${dur}`
+					: `${merged}/${total} merged${dur}`;
+
+				return sciExpanded(lines, footer, theme);
+			}
+
+			// Collapsed
+			const firstLine = text.split("\n")[0];
+			if (isError) {
+				return sciErr("✕ " + firstLine.slice(0, 70), theme);
+			} else if (hasConflicts) {
+				return sciOk("⚠ " + firstLine.slice(0, 70), theme);
+			} else {
+				return sciOk("✓ " + firstLine.slice(0, 70), theme);
+			}
+		},
+
+		async execute(_toolCallId, params, signal, onUpdate, ctx) {
+			// Parse the plan
+			emitCleaveState(pi, "assessing");
+
+			let plan: SplitPlan;
+			try {
+				plan = parsePlanResponse(params.plan_json);
+			} catch (e: any) {
+				emitCleaveState(pi, "failed");
+				throw new Error(`Invalid split plan: ${e.message}`);
+			}
+
+			const repoPath = ctx.cwd;
+			const maxParallel = params.max_parallel ?? DEFAULT_CONFIG.maxParallel;
+			const preferLocal = params.prefer_local ?? DEFAULT_CONFIG.preferLocal;
+
+			emitCleaveState(pi, "planning");
+
+			// ── OPENSPEC CONTEXT ───────────────────────────────────────
+			let openspecCtx: OpenSpecContext | null = null;
+			if (params.openspec_change_path) {
+				try {
+					openspecCtx = buildOpenSpecContext(params.openspec_change_path);
+				} catch {
+					// Non-fatal — proceed without enrichment
+				}
+			}
+
+			// ── SKILL MATCHING ─────────────────────────────────────────
+			// Initialize skills on children (parsePlanResponse may not set them)
+			for (const child of plan.children) {
+				child.skills = child.skills ?? [];
+			}
+
+			// Auto-match skills from scope patterns for children without annotations
+			matchSkillsToAllChildren(plan.children);
+
+			// Resolve skill names to absolute SKILL.md paths
+			const allSkillNames = new Set(plan.children.flatMap((c) => c.skills));
+			const { resolved: resolvedPaths } = resolveSkillPaths([...allSkillNames]);
+
+			// Build per-child skill directive map
+			const resolvedSkillMap = new Map<number, SkillDirective[]>();
+			for (let i = 0; i < plan.children.length; i++) {
+				const child = plan.children[i];
+				const directives: SkillDirective[] = [];
+				for (const skillName of child.skills) {
+					const found = resolvedPaths.find((r) => r.skill === skillName);
+					if (found) {
+						directives.push({ skill: found.skill, path: found.path });
+					}
+				}
+				resolvedSkillMap.set(i, directives);
+			}
+
+			// ── PREFLIGHT ──────────────────────────────────────────────
+			const toolUi = (ctx as {
+				ui?: {
+					input?: (prompt: string, initial?: string) => Promise<string | undefined>;
+					select?: (title: string, options: string[]) => Promise<string | undefined>;
+				};
+			}).ui;
+			const preflightOutcome = await runDirtyTreePreflight(pi, {
+				repoPath,
+				openspecChangePath: params.openspec_change_path,
+				onUpdate,
+				ui: (toolUi?.input || toolUi?.select) ? {
+					...(typeof toolUi.input === "function" ? { input: toolUi.input.bind(toolUi) } : {}),
+					...(typeof toolUi.select === "function" ? { select: toolUi.select.bind(toolUi) } : {}),
+				} : undefined,
+			});
+			if (preflightOutcome === "skip_cleave") {
+				const message = "Dirty-tree preflight resolved to proceed without cleave. Worktree creation and dispatch were skipped.";
+				emitCleaveState(pi, "idle");
+				return {
+					content: [{ type: "text", text: message }],
+					details: { phase: "preflight", skipped: true, reason: "proceed_without_cleave" },
+				};
+			}
+			if (preflightOutcome === "cancelled") {
+				emitCleaveState(pi, "idle");
+				return {
+					content: [{ type: "text", text: "Cleave cancelled during dirty-tree preflight." }],
+					details: { phase: "preflight", cancelled: true },
+				};
+			}
+			await ensureCleanWorktree(pi, repoPath);
+			const baseBranch = await getCurrentBranch(pi, repoPath);
+
+			// ── MODEL RESOLUTION ───────────────────────────────────────
+			// Determine local model availability (needed for model resolution)
+			let localModelAvailable = false;
+			let localModel: string | undefined;
+			if (preferLocal) {
+				try {
+					const ollamaResult = await pi.exec("ollama", ["list", "--json"], { timeout: 5_000 });
+					if (ollamaResult.code === 0) {
+						const models = JSON.parse(ollamaResult.stdout);
+						if (Array.isArray(models?.models) && models.models.length > 0) {
+							// Prefer code-optimised models for leaf tasks; fall back in order
+							const available = models.models.map((m: { name: string }) => m.name);
+							// Code-biased preference from shared registry (extensions/lib/local-models.ts)
+							const { PREFERRED_ORDER_CODE: preferredCodeModels } = await import("../lib/local-models.ts");
+							localModel =
+								preferredCodeModels.find((id) => available.includes(id)) ?? available[0];
+							localModelAvailable = true;
+						}
+					}
+				} catch {
+					// No local model available
+				}
+			}
+
+			// Resolve execute model for each child
+			for (const child of plan.children) {
+				child.executeModel = resolveExecuteModel(
+					child,
+					preferLocal,
+					localModelAvailable,
+					getPreferredTier,
+				);
+			}
+
+			// ── INITIALIZE STATE ───────────────────────────────────────
+			const state: CleaveState = {
+				runId: generateRunId(),
+				phase: "dispatch",
+				directive: params.directive,
+				repoPath,
+				baseBranch,
+				assessment: assessDirective(params.directive),
+				plan,
+				children: plan.children.map((c, i) => ({
+					childId: i,
+					label: c.label,
+					dependsOn: c.dependsOn,
+					status: "pending" as const,
+					branch: `cleave/${i}-${c.label}`,
+					backend: c.executeModel === "local" ? "local" as const : "cloud" as const,
+					executeModel: c.executeModel,
+				})),
+				workspacePath: "",
+				totalDurationSec: 0,
+				createdAt: new Date().toISOString(),
+			};
+
+			// Create workspace — pass OpenSpec context and resolved skills to enrich child task files
+			const wsPath = initWorkspace(state, plan, repoPath, openspecCtx, resolvedSkillMap);
+			state.workspacePath = wsPath;
+
+			// ── CREATE WORKTREES ───────────────────────────────────────
+			onUpdate?.({
+				content: [{ type: "text", text: "Creating git worktrees..." }],
+				details: { phase: "dispatch", children: state.children },
+			});
+
+			for (const child of state.children) {
+				try {
+					const wt = await createWorktree(pi, repoPath, child.label, child.childId, baseBranch);
+					child.worktreePath = wt.path;
+					child.branch = wt.branch;
+				} catch (e: any) {
+					child.status = "failed";
+					child.error = `Worktree creation failed: ${e.message}`;
+				}
+			}
+
+			saveState(state);
+
+			// ── DISPATCH ───────────────────────────────────────────────
+			// localModel was already resolved in MODEL RESOLUTION section above
+
+			emitCleaveState(pi, "dispatching", state.runId, state.children);
+
+			onUpdate?.({
+				content: [{ type: "text", text: `Dispatching ${state.children.length} children...` }],
+				details: { phase: "dispatch", children: state.children },
+			});
+
+			// ── REVIEW CONFIG ──────────────────────────────────────
+			const reviewConfig: ReviewConfig = {
+				enabled: params.review ?? DEFAULT_REVIEW_CONFIG.enabled,
+				maxWarningFixes: params.review_max_warning_fixes ?? DEFAULT_REVIEW_CONFIG.maxWarningFixes,
+				maxCriticalFixes: params.review_max_critical_fixes ?? DEFAULT_REVIEW_CONFIG.maxCriticalFixes,
+				churnThreshold: params.review_churn_threshold ?? DEFAULT_REVIEW_CONFIG.churnThreshold,
+			};
+
+			await dispatchChildren(
+				pi,
+				state,
+				maxParallel,
+				120 * 60 * 1000, // 2 hour timeout per child
+				localModel,
+				signal ?? undefined,
+				(msg) => {
+					onUpdate?.({
+						content: [{ type: "text", text: msg }],
+						details: { phase: "dispatch", children: state.children },
+					});
+				},
+				reviewConfig,
+			);
+
+			// ── HARVEST + CONFLICTS ────────────────────────────────────
+			emitCleaveState(pi, "merging", state.runId, state.children);
+
+			state.phase = "harvest";
+			saveState(state);
+
+			const taskContents = readTaskFiles(wsPath);
+			const taskResults = [...taskContents.entries()].map(([id, content]) =>
+				parseTaskResult(content, `${id}-task.md`),
+			);
+			const conflicts = detectConflicts(taskResults);
+			const cleaveCandidates = taskResults.flatMap((result) =>
+				result.summary
+					? emitResolvedBugCandidate(result.summary, result.path)
+					: [],
+			);
+
+			// ── MERGE ──────────────────────────────────────────────────
+			state.phase = "reunify";
+			saveState(state);
+
+			const completedChildren = state.children.filter((c) => c.status === "completed");
+			const mergeResults: Array<{ label: string; branch: string; success: boolean; conflicts: string[] }> = [];
+
+			for (const child of completedChildren) {
+				const result = await mergeBranch(pi, repoPath, child.branch, baseBranch);
+				mergeResults.push({
+					label: child.label,
+					branch: child.branch,
+					success: result.success,
+					conflicts: result.conflictFiles,
+				});
+				// On merge failure, stop merging further children to avoid
+				// compounding a partially-merged state
+				if (!result.success) break;
+			}
+
+			// ── CLEANUP ────────────────────────────────────────────────
+			// Only clean up worktrees if all merges succeeded. On merge
+			// failure, preserve branches so the user can manually resolve.
+			const mergeFailures = mergeResults.filter((m) => !m.success);
+
+			// ── TASK WRITE-BACK ────────────────────────────────────────
+			// Mark completed child tasks as [x] done in OpenSpec tasks.md
+			let writeBackResult: { updated: number; totalTasks: number; allDone: boolean; unmatchedLabels: string[] } | null = null;
+			if (params.openspec_change_path && mergeFailures.length === 0) {
+				const completedLabels = state.children
+					.filter((c) => c.status === "completed")
+					.map((c) => c.label);
+				try {
+					writeBackResult = writeBackTaskCompletion(params.openspec_change_path, completedLabels);
+					emitOpenSpecState(repoPath, pi);
+				} catch {
+					// Non-fatal — report will note write-back wasn't possible
+				}
+			}
+
+			// ── SPEC VERIFICATION ──────────────────────────────────────
+			// If OpenSpec specs exist, check implementation against scenarios
+			let specVerification: string | null = null;
+			if (openspecCtx && openspecCtx.specScenarios.length > 0 && mergeFailures.length === 0) {
+				specVerification = formatSpecVerification(openspecCtx);
+			}
+			// ── POST-MERGE GUARDRAILS ──────────────────────────────────
+			let guardrailReport: string | null = null;
+			if (mergeFailures.length === 0) {
+				try {
+					const checks = discoverGuardrails(repoPath);
+					if (checks.length > 0) {
+						const suite = runGuardrails(repoPath, checks);
+						if (suite.allPassed) {
+							guardrailReport = "### Static Analysis\n\n✅ All deterministic checks passed after merge";
+						} else {
+							const failures = suite.results.filter((r) => !r.passed);
+							const lines = ["### Static Analysis", "", "⚠ **Post-merge regressions detected**", ""];
+							for (const f of failures) {
+								const capped = f.output.split("\n").slice(0, 20).join("\n");
+								lines.push(`**${f.check.name}** (exit ${f.exitCode}, ${f.durationMs}ms):`);
+								lines.push("```", capped, "```", "");
+							}
+							guardrailReport = lines.join("\n");
+						}
+					}
+				} catch { /* non-fatal */ }
+			}
+
+			if (cleaveCandidates.length > 0) {
+				(sharedState.lifecycleCandidateQueue ??= []).push({
+					source: "cleave",
+					context: `cleave run ${state.runId} final outcomes`,
+					candidates: cleaveCandidates,
+				});
+			}
+
+			if (mergeResults.length > 0 && mergeFailures.length === 0) {
+				// All merges succeeded — safe to clean up worktrees and branches
+				await cleanupWorktrees(pi, repoPath);
+			} else if (mergeResults.length === 0) {
+				// No merges attempted (e.g., all children misclassified or failed).
+				// Preserve branches — they may contain committed work.
+				// Only prune worktree directories to reclaim disk space.
+				await pruneWorktreeDirs(pi, repoPath);
+			} else {
+				// Partial merge failure — preserve branches for manual resolution
+				await pruneWorktreeDirs(pi, repoPath);
+			}
+
+			// ── REPORT ─────────────────────────────────────────────────
+			state.phase = "complete";
+			state.completedAt = new Date().toISOString();
+			state.totalDurationSec = Math.round(
+				(new Date(state.completedAt).getTime() - new Date(state.createdAt).getTime()) / 1000,
+			);
+
+			const allOk =
+				state.children.every((c) => c.status === "completed") &&
+				mergeResults.every((m) => m.success) &&
+				conflicts.length === 0;
+
+			if (!allOk) state.phase = "failed";
+			emitCleaveState(pi, allOk ? "done" : "failed", state.runId, state.children);
+			saveState(state);
+
+			// Build report
+			const completedCount = state.children.filter((c) => c.status === "completed").length;
+			const failedCount = state.children.filter((c) => c.status === "failed").length;
+
+			const reportLines = [
+				`## Cleave Report: ${state.runId}`,
+				"",
+				`**Directive:** ${params.directive}`,
+				`**Status:** ${allOk ? "✓ SUCCESS" : "✗ ISSUES DETECTED"}`,
+				`**Children:** ${completedCount} completed, ${failedCount} failed of ${state.children.length}`,
+				`**Duration:** ${state.totalDurationSec}s`,
+				`**Workspace:** \`${wsPath}\``,
+				"",
+			];
+
+			// Child details
+			for (const child of state.children) {
+				const icon = child.status === "completed" ? "✓" : child.status === "failed" ? "✗" : "⏳";
+				const dur = child.durationSec ? ` (${child.durationSec}s)` : "";
+				const reviewNote = child.reviewIterations && child.reviewIterations > 0
+					? ` [${child.reviewIterations} review${child.reviewIterations > 1 ? "s" : ""}: ${child.reviewDecision}]`
+					: "";
+				reportLines.push(`  ${icon} **${child.label}** [${child.backend ?? "cloud"}]${dur}: ${child.status}${reviewNote}`);
+				if (child.error) reportLines.push(`    Error: ${child.error}`);
+				if (child.reviewEscalationReason) reportLines.push(`    Review: ${child.reviewEscalationReason}`);
+			}
+
+			// Conflicts
+			if (conflicts.length > 0) {
+				reportLines.push("", "### Conflicts", "", formatConflicts(conflicts));
+			}
+
+			// Merge results (always show — makes partial merge state explicit)
+			if (mergeResults.length > 0) {
+				reportLines.push("", "### Merge Results");
+				const mergeSuccesses = mergeResults.filter((m) => m.success);
+				const notAttempted = completedChildren
+					.filter((c) => !mergeResults.some((m) => m.label === c.label))
+					.map((c) => c.label);
+				for (const m of mergeSuccesses) {
+					reportLines.push(`  ✓ ${m.label} merged`);
+				}
+				for (const m of mergeFailures) {
+					reportLines.push(`  ✗ ${m.label}: conflicts in ${m.conflicts.join(", ")}`);
+				}
+				for (const label of notAttempted) {
+					reportLines.push(`  ⏭ ${label}: skipped (earlier merge failed)`);
+				}
+			}
+
+			// Spec verification (post-merge)
+			if (specVerification) {
+				reportLines.push("", specVerification);
+			}
+
+			// Post-merge guardrail results
+			if (guardrailReport) {
+				reportLines.push("", guardrailReport);
+			}
+
+			// Task write-back status
+			if (writeBackResult && writeBackResult.updated > 0) {
+				reportLines.push(
+					"",
+					"### Task Write-Back",
+					`  ✓ Marked ${writeBackResult.updated} tasks as done in \`tasks.md\``,
+				);
+			}
+			if (writeBackResult && writeBackResult.unmatchedLabels.length > 0) {
+				reportLines.push(
+					"",
+					"### Lifecycle Reconciliation Warning",
+					"  ⚠ Completed cleave work could not be mapped back into `tasks.md` for:",
+					...writeBackResult.unmatchedLabels.map((label) => `  - ${label}`),
+					"",
+					"  tasks.md no longer matches the implementation plan. Reconcile the OpenSpec task groups before archive.",
+				);
+			}
+
+			// Next steps guidance
+			if (allOk && params.openspec_change_path) {
+				reportLines.push("", "### Next Steps");
+				if (writeBackResult?.allDone) {
+					reportLines.push(
+						"  All tasks complete. Ready to finalize:",
+						"  1. Run `/assess spec` to validate implementation against spec scenarios",
+						"  2. Run `/opsx:verify` for full verification",
+						"  3. Run `/opsx:archive` to merge delta specs and close the change",
+					);
+				} else {
+					reportLines.push(
+						"  Some tasks remain. Continue with:",
+						"  1. Run `/opsx:apply` to work on remaining tasks",
+						"  2. Or run `/cleave` again targeting the unfinished groups",
+					);
+				}
+			} else if (allOk && !params.openspec_change_path) {
+				reportLines.push(
+					"",
+					"### Next Steps",
+					"  Run tests and review the merged changes.",
+				);
+			}
+
+			const rawReport = reportLines.join("\n");
+			const truncation = truncateTail(rawReport, {
+				maxLines: DEFAULT_MAX_LINES,
+				maxBytes: DEFAULT_MAX_BYTES,
+			});
+			let report = truncation.content;
+			if (truncation.truncated) {
+				report += `\n\n[Report truncated: ${truncation.outputLines} of ${truncation.totalLines} lines` +
+					` (${formatSize(truncation.outputBytes)} of ${formatSize(truncation.totalBytes)})]`;
+			}
+
+			return {
+				content: [{ type: "text", text: report }],
+				details: {
+					runId: state.runId,
+					success: allOk,
+					childrenCompleted: completedCount,
+					childrenFailed: failedCount,
+					conflictsFound: conflicts.length,
+					mergeFailures: mergeFailures.length,
+					workspacePath: wsPath,
+				},
+			};
+		},
+	});
+
+	// ─── /cleave inspect ──────────────────────────────────────────────────────
+
+	/**
+	 * Run `git diff --stat` + `git status --short` in a worktree.
+	 * Returns a formatted string or an empty string if no changes / no worktree.
+	 * Result is cached for 2s to avoid subprocess spam on repeated renders.
+	 */
+	const execFileAsync = promisify(execFile);
+	const diffCache = new Map<string, { ts: number; result: string }>();
+
+	async function runGitDiff(worktreePath: string): Promise<string> {
+		const cached = diffCache.get(worktreePath);
+		if (cached && Date.now() - cached.ts < 2000) return cached.result;
+		try {
+			const [stat, status] = await Promise.all([
+				execFileAsync("git", ["diff", "--stat", "HEAD"], { cwd: worktreePath }).then(r => r.stdout.trim()).catch(() => ""),
+				execFileAsync("git", ["status", "--short"], { cwd: worktreePath }).then(r => r.stdout.trim()).catch(() => ""),
+			]);
+			const result = [stat, status].filter(Boolean).join("\n") || "(no changes yet)";
+			diffCache.set(worktreePath, { ts: Date.now(), result });
+			return result;
+		} catch {
+			return "(git unavailable)";
+		}
+	}
+
+	/** Open the inspect overlay for a specific child index. */
+	async function openInspectOverlay(childIdx: number, ctx: ExtensionCommandContext): Promise<void> {
+		const cl = (sharedState as any).cleave;
+		const children: any[] = cl?.children ?? [];
+
+		let selected = Math.max(0, Math.min(childIdx, children.length - 1));
+		let diffText = children.length === 0 ? "" : "(loading…)";
+		let diffLoaded = false;
+
+		const loadDiff = async () => {
+			const child = children[selected];
+			if (!child?.worktreePath) { diffText = "(no worktree assigned)"; diffLoaded = true; return; }
+			diffText = await runGitDiff(child.worktreePath);
+			diffLoaded = true;
+		};
+		// Load diff async — don't block overlay open
+		loadDiff();
+
+		await ctx.ui.custom((tui, theme, _kb, done) => {
+			const component = {
+				render(width: number): string[] {
+					const lines: string[] = [];
+					const child = children[selected];
+					const divFill = (label: string) => {
+						const prefix = `  ── ${label} `;
+						return theme.fg("dim", prefix + "─".repeat(Math.max(0, width - prefix.length)));
+					};
+
+					// ── Header ──────────────────────────────────────────────
+					const headerLabel = ` ⚡ cleave inspect `;
+					const headerFill = Math.max(0, width - headerLabel.length - 2);
+					lines.push(
+						theme.fg("dim", "──") +
+						theme.fg("accent", headerLabel) +
+						theme.fg("dim", "─".repeat(headerFill)),
+					);
+
+					if (children.length === 0) {
+						lines.push("");
+						lines.push(theme.fg("muted", "  No active cleave children."));
+						lines.push("");
+						lines.push(theme.fg("dim", "  q close"));
+						return lines;
+					}
+
+					// ── Child selector ───────────────────────────────────────
+					lines.push("");
+					for (let i = 0; i < children.length; i++) {
+						const c = children[i];
+						const icon =
+							c.status === "done" ? theme.fg("success", "✓") :
+							c.status === "failed" ? theme.fg("error", "✕") :
+							c.status === "running" ? theme.fg("warning", "⟳") :
+							theme.fg("dim", "○");
+						const elapsedSec = c.startedAt ? Math.floor((Date.now() - c.startedAt) / 1000) : null;
+						const elapsed = elapsedSec != null ? theme.fg("dim", `  ${elapsedSec}s`) : "";
+						const label = i === selected
+							? theme.bold(theme.fg("accent", `▶ ${c.label}`))
+							: theme.fg("muted", `  ${c.label}`);
+						lines.push(`  ${icon} ${label}${elapsed}`);
+					}
+
+					// ── Recent stdout ────────────────────────────────────────
+					lines.push("");
+					lines.push(divFill("activity"));
+					const recent: string[] = child?.recentLines ?? [];
+					if (recent.length === 0) {
+						lines.push(theme.fg("dim", "  (no output yet)"));
+					} else {
+						for (const l of recent.slice(-10)) {
+							lines.push(theme.fg("muted", `  ${l.slice(0, width - 4)}`));
+						}
+					}
+
+					// ── Git diff ─────────────────────────────────────────────
+					lines.push("");
+					lines.push(divFill("worktree diff"));
+					for (const l of diffText.split("\n").slice(0, 15)) {
+						lines.push(theme.fg("muted", `  ${l.slice(0, width - 4)}`));
+					}
+
+					// ── Footer ───────────────────────────────────────────────
+					lines.push("");
+					lines.push(theme.fg("dim", "  ↑↓ select child  ·  r refresh  ·  q close"));
+					return lines;
+				},
+				invalidate() {},
+				onKey(key: string) {
+					if (key === "escape" || key === "q") { done(undefined); return; }
+					if (key === "up" || key === "k") {
+						selected = Math.max(0, selected - 1);
+						diffLoaded = false; diffText = "(loading…)";
+						loadDiff().then(() => tui.requestRender());
+					}
+					if (key === "down" || key === "j") {
+						selected = Math.min(children.length - 1, selected + 1);
+						diffLoaded = false; diffText = "(loading…)";
+						loadDiff().then(() => tui.requestRender());
+					}
+					if (key === "r") {
+						const child = children[selected];
+						if (child?.worktreePath) diffCache.delete(child.worktreePath);
+						diffText = "(loading…)";
+						loadDiff().then(() => tui.requestRender());
+					}
+					tui.requestRender();
+				},
+			};
+			// Re-render when diff loads
+			const diffPoll = setInterval(() => {
+				if (diffLoaded) { clearInterval(diffPoll); tui.requestRender(); }
+			}, 100);
+			(component as any).dispose = () => clearInterval(diffPoll);
+			return component;
+		});
+	}
+
+	pi.registerCommand("cleave inspect", {
+		description: "Inspect a running cleave child — stdout ring buffer + live git diff",
+		handler: async (_args, ctx) => {
+			await openInspectOverlay(0, ctx);
+		},
+	});
+
+	// Register ctrl+i shortcut while any cleave run is active
+	let inspectShortcutActive = false;
+	pi.events.on(DASHBOARD_UPDATE_EVENT, (data: any) => {
+		const cl = (sharedState as any).cleave;
+		const isRunning = cl?.status === "running" || cl?.status === "dispatching";
+		if (isRunning && !inspectShortcutActive) {
+			inspectShortcutActive = true;
+			pi.registerShortcut("ctrl+i", {
+				description: "Inspect running cleave children",
+				handler: async (ctx) => {
+					await openInspectOverlay(0, ctx as any);
+				},
+			});
+		}
+	});
+}