@iowarp/clio-coder 0.3.0 → 0.3.1

package/src/interactive/slash-commands.ts

@@ -0,0 +1,1319 @@
+import { BusChannels } from "../core/bus-events.js";
+import { THINKING_LEVELS } from "../core/defaults.js";
+import type { SafeEventBus } from "../core/event-bus.js";
+import type { AgentSpec } from "../domains/agents/spec.js";
+import type { ContextInitOptions } from "../domains/context/init-options.js";
+import type { DispatchContract, DispatchRequest } from "../domains/dispatch/contract.js";
+import { type AgentRoleFactsResolver, requestExecutionRole } from "../domains/dispatch/execution-role.js";
+import type { ReceiptIntegrityResult } from "../domains/dispatch/receipt-integrity.js";
+import type { RunReceipt } from "../domains/dispatch/types.js";
+import type { JobThinkingLevel } from "../domains/dispatch/validation.js";
+import type { InstalledExtension } from "../domains/extensions/index.js";
+import type { InteropAgentId, InteropProposal, InteropReport } from "../domains/interop/index.js";
+import type { ProvidersContract, ResolvedModelRef } from "../domains/providers/index.js";
+import { resolveModelReference } from "../domains/providers/index.js";
+import type { PromptTemplate, PromptTemplateExpansion, ResourceList } from "../domains/resources/index.js";
+import { parseSkillCommand } from "../domains/resources/index.js";
+import type { ShareImportPlan } from "../domains/share/index.js";
+import { isToolProfileName, TOOL_PROFILE_NAMES, type ToolProfileName } from "../tools/profiles.js";
+import type { NoticeLevel } from "./command-output.js";
+import { SETTINGS_SECTIONS, type SettingsCenterRowId, type SettingsSectionId } from "./overlays/settings.js";
+import type { CommandArgsSpec, CommandPositionalSpec, ParsedArgs } from "./slash-spec.js";
+import { matchFromSpec, usageLine } from "./slash-spec.js";
+import { formatWorkerShareNote, selectWorkerRunToShare, workerShareFactsFromEntry } from "./worker-share.js";
+import type { WorkerEntryState } from "./worker-stream.js";
+
+/**
+ * Ported from pi-coding-agent's BUILTIN_SLASH_COMMANDS registry. Each entry owns
+ * one user-facing slash command: how it parses, which SlashCommand kinds it
+ * produces, and how those kinds execute. Adding a new command is one entry
+ * rather than extending two parallel switches.
+ */
+
+type ShareCommandVariant =
+	| { kind: "share"; action: "export"; path: string }
+	| { kind: "share"; action: "import"; path: string; dryRun: boolean; force: boolean }
+	/** Put a finished worker run's answer into the main agent's context. Bare `/share` takes the newest one. */
+	| { kind: "share"; action: "worker-run"; runId?: string }
+	| { kind: "share"; action: "usage"; subcommand?: "export" | "import"; error?: string };
+
+type SlashCommandVariant =
+	| { kind: "quit" }
+	| { kind: "help"; query?: string }
+	| { kind: "init"; options: InitCommandOptions }
+	| { kind: "context-clear"; options: ContextClearCommandOptions }
+	| { kind: "context-refresh" }
+	| { kind: "skill-selector" }
+	| { kind: "skill-invocation"; text: string }
+	| { kind: "prompts" }
+	| { kind: "extensions" }
+	| { kind: "interop" }
+	| ShareCommandVariant
+	/** `source` is the line the operator typed, echoed above the run's transcript block. */
+	| { kind: "run"; agentId: string; task: string; options: RunCommandOptions; source: string }
+	| { kind: "run-usage" }
+	| { kind: "delegate"; agentId: string; task: string; source: string; share?: boolean }
+	| { kind: "delegate-usage" }
+	| { kind: "agents" }
+	| { kind: "cost" }
+	| { kind: "context-view" }
+	| { kind: "tasks" }
+	| { kind: "memory" }
+	| { kind: "memory-seed" }
+	| { kind: "view"; filter?: string }
+	| { kind: "view-verify"; runId: string }
+	| { kind: "view-usage" }
+	| { kind: "thinking-set"; level: string }
+	| { kind: "output-set"; verbosity: string }
+	| { kind: "model" }
+	| { kind: "model-set"; pattern: string }
+	| { kind: "settings"; section?: SettingsSectionId; rowId?: SettingsCenterRowId }
+	| { kind: "resume" }
+	| { kind: "new" }
+	| { kind: "tree" }
+	| { kind: "fork" }
+	| { kind: "compact"; instructions: string | undefined }
+	| { kind: "export"; path: string | undefined }
+	| { kind: "unknown"; text: string }
+	/** `text` is the whole line, so a token that turns out to name a prompt template keeps its arguments. */
+	| { kind: "unknown-command"; token: string; text: string }
+	| { kind: "usage-error"; command: string; reason: string }
+	| { kind: "empty" };
+
+export type SlashCommand = SlashCommandVariant;
+
+export type SlashCommandKind = SlashCommand["kind"];
+
+export interface RunIo {
+	stdout: (s: string) => void;
+	stderr: (s: string) => void;
+}
+
+export type SetThinkingLevelResult =
+	| { status: "applied"; level: string; display: string }
+	| { status: "unsupported"; level: string; supported: ReadonlyArray<string> }
+	| { status: "unavailable" };
+
+export type SetOutputVerbosityResult =
+	| { status: "applied"; verbosity: string }
+	| { status: "unsupported"; verbosity: string; supported: ReadonlyArray<string> }
+	| { status: "unavailable" };
+
+export type TaskMemorySeedCommandResult =
+	| { status: "seeded"; seeded: number; skipped: number; source: string }
+	| { status: "disabled" }
+	| { status: "not-found" };
+
+export type InitCommandOptions = ContextInitOptions;
+
+export interface ContextClearCommandOptions {
+	all?: boolean;
+	confirmed?: boolean;
+	confirmedAll?: boolean;
+}
+
+export interface RunCommandOptions {
+	workerProfile?: string;
+	workerRuntime?: string;
+	target?: string;
+	model?: string;
+	thinkingLevel?: JobThinkingLevel;
+	toolProfile?: ToolProfileName;
+	requiredCapabilities?: string[];
+	/** Hand the receipt's answer to the main agent when the run finishes. Off by default. */
+	share?: boolean;
+}
+
+export interface HandleRunDeps {
+	dispatch: DispatchContract;
+	/** Strict recipe facts the run's execution role is derived from. */
+	getAgentRoleFacts?: AgentRoleFactsResolver;
+	io: RunIo;
+	notice: (level: NoticeLevel, text: string) => void;
+	/**
+	 * Optional bus for forwarding per-event worker output. When supplied,
+	 * every non-heartbeat event is re-emitted on `BusChannels.DispatchProgress`
+	 * so UI surfaces (dispatch-board overlay) can update their row as the
+	 * stream arrives instead of waiting for the terminal receipt.
+	 */
+	bus?: SafeEventBus;
+	/**
+	 * Operator-turn sink for `--share`. Absent means the flag has nowhere to put
+	 * the answer, which is reported rather than silently dropped: an operator who
+	 * asked for the model to see a result must not be told it did.
+	 */
+	submitOperatorNote?: (text: string) => void;
+}
+
+/**
+ * Hand a finished run's sealed answer to the main agent. The receipt is the
+ * terminal truth here rather than the transcript block: the block is a view of
+ * this same text, and the caller already holds the receipt.
+ */
+function shareReceipt(agentId: string, receipt: RunReceipt, deps: HandleRunDeps): void {
+	if (!deps.submitOperatorNote) {
+		deps.notice("error", "--share is not wired in this session; the result stayed local");
+		return;
+	}
+	const note = formatWorkerShareNote({
+		agentId,
+		runId: receipt.runId,
+		outcome: receipt.outcome,
+		text: receipt.output?.text ?? "",
+	});
+	if (note === null) {
+		deps.notice("warn", `--share: run ${receipt.runId} produced no text to share`);
+		return;
+	}
+	deps.submitOperatorNote(note);
+}
+
+/**
+ * The attributed half every operator-started run shares, whether `/run` or
+ * `/delegate` built the request. Dispatch it, forward its events for surfaces
+ * that do not share the domain's bus, wait for the receipt, keep the notice bar
+ * quiet on success because the transcript block is the success signal, and
+ * share the sealed answer only when the operator asked. A failed run is still
+ * worth sharing when they did: the failure text is what they would want the
+ * main agent to work from.
+ */
+async function runAttributed(
+	command: "run" | "delegate",
+	request: DispatchRequest,
+	deps: HandleRunDeps,
+	share: boolean,
+): Promise<void> {
+	const { dispatch, notice, bus } = deps;
+	const progressBus = dispatch.ownsProgressBus?.(bus) === true ? undefined : bus;
+	try {
+		const handle = await dispatch.dispatch(request);
+		for await (const event of handle.events) {
+			const e = event as { type?: string };
+			if (!e.type || e.type === "heartbeat") continue;
+			progressBus?.emit(BusChannels.DispatchProgress, { runId: handle.runId, agentId: request.agentId, event });
+		}
+		const receipt = await handle.finalPromise;
+		if (receipt.exitCode !== 0 || receipt.failureMessage) {
+			const failure = receipt.failureMessage ? ` ${receipt.failureMessage}` : "";
+			notice("error", `${command} failed: exit=${receipt.exitCode}${failure}`);
+		}
+		if (share) shareReceipt(request.agentId, receipt, deps);
+	} catch (err) {
+		const msg = err instanceof Error ? err.message : String(err);
+		notice("error", `${command} failed: ${msg}`);
+	}
+}
+
+/**
+ * Dispatches /run through the dispatch contract and streams events to stdout.
+ * Target + model are resolved by the dispatch domain from request overrides,
+ * worker profiles, recipe hints, and `settings.workers.default`.
+ */
+export async function handleRun(
+	agentId: string,
+	task: string,
+	deps: HandleRunDeps,
+	options: RunCommandOptions = {},
+): Promise<void> {
+	if (options.target && options.workerProfile) {
+		deps.notice("warn", `--target ${options.target} takes precedence; --worker ${options.workerProfile} will be ignored`);
+	}
+	if (options.target && options.workerRuntime) {
+		deps.notice(
+			"warn",
+			`--target ${options.target} takes precedence; --runtime ${options.workerRuntime} will be ignored`,
+		);
+	}
+	await runAttributed(
+		"run",
+		{
+			agentId,
+			task,
+			executionRole: requestExecutionRole({
+				agentId,
+				...(deps.getAgentRoleFacts ? { resolveFacts: deps.getAgentRoleFacts } : {}),
+			}),
+			requestOrigin: "user",
+			...(options.workerProfile ? { workerProfile: options.workerProfile } : {}),
+			...(options.workerRuntime ? { workerRuntime: options.workerRuntime } : {}),
+			...(options.target ? { target: options.target } : {}),
+			...(options.model ? { model: options.model } : {}),
+			...(options.thinkingLevel ? { thinkingLevel: options.thinkingLevel } : {}),
+			...(options.toolProfile ? { toolProfile: options.toolProfile } : {}),
+			...(options.requiredCapabilities && options.requiredCapabilities.length > 0
+				? { requiredCapabilities: options.requiredCapabilities }
+				: {}),
+		},
+		deps,
+		options.share === true,
+	);
+}
+
+async function handleDelegate(
+	agentId: string,
+	task: string,
+	deps: HandleRunDeps,
+	options: { share?: boolean } = {},
+): Promise<void> {
+	await runAttributed(
+		"delegate",
+		{ agentId, delegationAgentId: agentId, executionRole: "builder", requestOrigin: "user", task },
+		deps,
+		options.share === true,
+	);
+}
+
+/**
+ * `/share [runId]`: hand a finished worker run's answer to the main agent.
+ *
+ * Nothing here is implicit. The operator either typed the command or passed
+ * `--share`, and the note travels the ordinary user-turn path, so the model
+ * receives it exactly as it receives anything else the operator writes.
+ */
+function shareWorkerRun(runId: string | undefined, ctx: SlashCommandContext): void {
+	if (!ctx.submitOperatorNote) {
+		ctx.notice("error", "share is not wired; the worker transcript is unavailable in this session");
+		return;
+	}
+	const entry = selectWorkerRunToShare(ctx.listWorkerRuns?.() ?? [], runId);
+	if (!entry) {
+		ctx.notice(
+			"error",
+			runId === undefined
+				? "no finished /run or /delegate result to share yet"
+				: `no finished run ${runId} in this session`,
+		);
+		return;
+	}
+	const facts = workerShareFactsFromEntry(entry);
+	const note = facts === null ? null : formatWorkerShareNote(facts);
+	if (note === null) {
+		ctx.notice("error", `run ${entry.runId} produced no text to share`);
+		return;
+	}
+	ctx.submitOperatorNote(note);
+}
+
+/**
+ * Runtime dependencies every slash-command handler may need. Every field is
+ * injected at startInteractive construction time; handlers never reach into
+ * the TUI, chat loop, or overlay module graph directly.
+ */
+export interface SlashCommandContext {
+	io: RunIo;
+	notice: (level: NoticeLevel, text: string) => void;
+	dispatch: DispatchContract;
+	bus: SafeEventBus;
+	/** Strict recipe facts used to keep interactive /run role-equivalent with the CLI. */
+	getAgentRoleFacts?: AgentRoleFactsResolver;
+	/**
+	 * Echo the typed command line into the transcript, dimmed, before the block
+	 * it starts. `/run` and `/delegate` use it so a worker's attributed block has
+	 * the request above it. Transcript-only; the line never enters model context.
+	 */
+	echoOperatorCommand?: (text: string) => void;
+	/**
+	 * Put operator-authored text into the session the way typed text enters it:
+	 * a user turn, persisted in the ledger, visible in the transcript. Used only
+	 * by `--share` and `/share`, and never with expansion, because a worker's
+	 * answer is literal text and must not be re-read as file or skill syntax.
+	 */
+	submitOperatorNote?: (text: string) => void;
+	/** Worker blocks this session folded, oldest first. `/share` picks from these. */
+	listWorkerRuns?: () => ReadonlyArray<WorkerEntryState>;
+	/** Fire-and-forget shutdown. Handler must not await. */
+	shutdown: () => void;
+	runInit: (options: InitCommandOptions) => void;
+	/** Write the current session transcript (all tool segments expanded, ANSI-stripped) to a Markdown file. */
+	exportTranscript: (path?: string) => void;
+	runContextClear: (options: ContextClearCommandOptions) => void;
+	/**
+	 * Re-index the codewiki and refresh `.clio-coder` state without touching CLIO-CODER.md.
+	 * Optional until the host wires onContextRefresh.
+	 */
+	runContextRefresh?: () => void;
+	openSkillsHub?: () => void;
+	listPrompts: () => ResourceList<PromptTemplate>;
+	/**
+	 * Resolve a `/name` against the loaded prompt templates. Absent when the host
+	 * wired no resources, in which case a command-shaped token that names no
+	 * builtin is simply not a command.
+	 */
+	expandPromptTemplate?: (text: string) => PromptTemplateExpansion;
+	listExtensions?: () => ReadonlyArray<InstalledExtension>;
+	/** Detection report, pending proposals, and the two consent actions `/interop` drives. */
+	interop?: {
+		report: () => InteropReport | null;
+		proposals: () => ReadonlyArray<InteropProposal>;
+		configured: () => ReadonlyArray<{ id: string; command: string; args: ReadonlyArray<string> }>;
+		accept: (kind: InteropAgentId) => void;
+		decline: (kind: InteropAgentId) => void;
+	};
+	listAgents: () => ReadonlyArray<AgentSpec>;
+	listDelegationAgents: () => ReadonlyArray<{
+		id: string;
+		command: string;
+		args: ReadonlyArray<string>;
+		toolGovernance?: string;
+		labels?: Record<string, string>;
+	}>;
+	exportShareArchive?: (outPath: string) => { fileCount: number; path: string };
+	importShareArchive?: (path: string, options: { dryRun?: boolean; force?: boolean }) => ShareImportPlan;
+	openCost: () => void;
+	/** Open the read-only `/context` overlay: categorized context-window ledger. */
+	openContextView: () => void;
+	/** Open the read-only `/tasks` overlay: the session task board with receipts. */
+	openTasks: () => void;
+	/** Open the read-only `/memory` overlay: approved lessons and the live task bank. */
+	openMemory: () => void;
+	/** Opt-in import from the newest structured handoff; absent when the host has no task bank. */
+	seedTaskMemory?: () => TaskMemorySeedCommandResult;
+	/** Open `/view`, the full observability artifact viewer. */
+	openView: (filter?: string) => void;
+	/**
+	 * Apply a thinking level named on the command line. Returns why it was
+	 * refused so the caller can say so, because the level the operator typed may
+	 * not be one the active target supports.
+	 */
+	setThinkingLevel?: (level: string) => SetThinkingLevelResult;
+	/** Apply a transcript verbosity named on the command line; refused values are reported by the caller. */
+	setOutputVerbosity?: (verbosity: string) => SetOutputVerbosityResult;
+	openModel: () => void;
+	/** Live providers contract used by `/model <pattern>` to resolve directly. */
+	providers: ProvidersContract;
+	/** Apply a resolved model reference to settings (and optionally thinking level). */
+	applyModelRef: (ref: ResolvedModelRef) => void;
+	openSettings: (section?: SettingsSectionId, rowId?: SettingsCenterRowId) => void;
+	openResume: () => void;
+	startNewSession: () => void;
+	openTree: () => void;
+	openMessagePicker: () => void;
+	openHelp: (query?: string) => void;
+	openAgents: () => void;
+	openPrompts: () => void;
+	openExtensions: () => void;
+	openInterop?: () => void;
+	setEditorText?: (text: string) => void;
+	/**
+	 * Run compaction for the current session. Handler resolves the target
+	 * model, reads session entries, calls session/compaction/compact, and
+	 * appends a compactionSummary entry. No-op when no session is open, so the
+	 * handler prints an actionable stderr line instead.
+	 */
+	runCompact: (instructions: string | undefined) => void;
+	/**
+	 * Escape hatch for the `view verify` entry: verify a receipt file on disk
+	 * and emit a single status line. Kept on the context so the registry does
+	 * not import the overlay module.
+	 */
+	verifyReceipt: (runId: string) => ReceiptIntegrityResult;
+	/**
+	 * Handles the "unknown" case: append the text to the chat panel as a user
+	 * turn, submit to the chat loop, and schedule a re-render. Handlers for
+	 * {kind:"unknown"} defer to this so the chat-panel reference stays scoped
+	 * to startInteractive.
+	 */
+	submitChat: (text: string) => void;
+	/** Re-render request; wraps tui.requestRender so handlers do not import TUI. */
+	render: () => void;
+}
+
+/** The verb a command performs, in the order /help lists the groups. */
+export const SLASH_COMMAND_GROUPS = ["Run", "Inspect", "Configure", "Sessions"] as const;
+export type SlashCommandGroup = (typeof SLASH_COMMAND_GROUPS)[number];
+
+export interface BuiltinSlashCommand {
+	name: string;
+	description: string;
+	group: SlashCommandGroup;
+	aliases?: ReadonlyArray<string>;
+	/**
+	 * Argument text an alias stands in for, prepended to whatever the operator
+	 * typed after it. `/compact` names a subcommand of `/context`, not the
+	 * command itself, so it cannot be spelled as a bare alias.
+	 */
+	aliasArgs?: Readonly<Record<string, string>>;
+	/** Excluded from /help, autocomplete, and the docs command table. */
+	hidden?: boolean;
+	args?: CommandArgsSpec;
+	/** Stable copy shown beside first-token subcommand completions. */
+	subcommandDescriptions?: Readonly<Record<string, string>>;
+	/** The set of SlashCommand kinds this entry is responsible for dispatching. */
+	kinds: ReadonlyArray<SlashCommandKind>;
+	/** Return the parsed SlashCommand for `trimmed` or null if this entry does not match. */
+	match?(trimmed: string): SlashCommand | null;
+	fromArgs?(parsed: ParsedArgs, trimmed: string): SlashCommand;
+	/** Execute `command` against `ctx`. Called only for kinds declared in `kinds`. */
+	handle(command: SlashCommand, ctx: SlashCommandContext): void;
+}
+
+const RUN_THINKING_LEVELS: ReadonlyArray<JobThinkingLevel> = THINKING_LEVELS;
+
+function isRunThinkingLevel(value: string): value is JobThinkingLevel {
+	return RUN_THINKING_LEVELS.some((level) => level === value);
+}
+
+function isSettingsSectionId(value: string): value is SettingsSectionId {
+	return SETTINGS_SECTIONS.some((section) => section.id === value);
+}
+
+/** `/context compact` alone keeps a free-form optional instruction tail. */
+const COMPACT_POSITIONALS: ReadonlyArray<CommandPositionalSpec> = [
+	{ name: "instructions", required: false, rest: true },
+];
+
+/**
+ * A registered command whose arguments do not parse is a mistake to report,
+ * not a chat message to send. Falling through to `unknown` submitted the raw
+ * text to the model, which then answered it conversationally while the
+ * operator believed a setting had changed and nothing had. `unknown` stays
+ * reserved for input that matched no command at all, so pasted text beginning
+ * with a slash still reaches the model.
+ */
+function fromArgsOrUsage(name: string, command: SlashCommand): (parsed: ParsedArgs) => SlashCommand {
+	return (parsed) => (parsed.error ? { kind: "usage-error", command: name, reason: parsed.error } : command);
+}
+
+/**
+ * A configuration-shaped command that is a shortcut into one settings section.
+ * It parses to the same `settings` command `/settings <section>` produces, so
+ * the settings entry dispatches it and this entry owns no kind of its own.
+ */
+function settingsDeepLink(
+	name: string,
+	section: SettingsSectionId,
+	description: string,
+	rowId?: SettingsCenterRowId,
+): BuiltinSlashCommand {
+	return {
+		name,
+		description,
+		group: "Configure",
+		kinds: [],
+		args: {},
+		fromArgs: fromArgsOrUsage(name, { kind: "settings", section, ...(rowId ? { rowId } : {}) }),
+		handle: () => undefined,
+	};
+}
+
+function usageNotice(entry: BuiltinSlashCommand, subcommand?: string): string {
+	return usageLine(entry, subcommand).trim();
+}
+
+export const BUILTIN_SLASH_COMMANDS: ReadonlyArray<BuiltinSlashCommand> = [
+	{
+		name: "quit",
+		description: "Exit Clio Coder",
+		group: "Sessions",
+		aliases: ["exit"],
+		kinds: ["quit"],
+		args: {},
+		fromArgs: fromArgsOrUsage("quit", { kind: "quit" }),
+		handle(_command, ctx) {
+			ctx.shutdown();
+		},
+	},
+	{
+		name: "help",
+		description: "Open the interactive help center showing commands and keys",
+		group: "Inspect",
+		kinds: ["help"],
+		args: {
+			positionals: [{ name: "query", required: false, rest: true }],
+		},
+		fromArgs(parsed) {
+			return { kind: "help", ...(parsed.rest ? { query: parsed.rest } : {}) };
+		},
+		handle(command, ctx) {
+			ctx.openHelp(command.kind === "help" ? command.query : undefined);
+		},
+	},
+	{
+		name: "skill",
+		description: "Open the Skills Hub or invoke a skill",
+		group: "Run",
+		aliases: ["skill:", "skills:"],
+		kinds: ["skill-selector", "skill-invocation"],
+		args: {
+			positionals: [
+				{ name: "name", required: false },
+				{ name: "task", required: false, rest: true },
+			],
+		},
+		match(trimmed) {
+			if (trimmed === "/skill" || trimmed === "/skill:" || trimmed === "/skills:") {
+				return { kind: "skill-selector" };
+			}
+			const command = parseSkillCommand(trimmed);
+			if (command) {
+				return { kind: "skill-invocation", text: trimmed };
+			}
+			return null;
+		},
+		handle(command, ctx) {
+			if (command.kind === "skill-selector") {
+				ctx.openSkillsHub?.();
+			} else if (command.kind === "skill-invocation") {
+				ctx.submitChat(command.text);
+			}
+		},
+	},
+	{
+		name: "prompts",
+		description: "List prompt templates",
+		group: "Inspect",
+		kinds: ["prompts"],
+		args: {},
+		fromArgs: fromArgsOrUsage("prompts", { kind: "prompts" }),
+		handle(_command, ctx) {
+			ctx.openPrompts();
+		},
+	},
+	{
+		name: "extensions",
+		description: "List installed extensions",
+		group: "Inspect",
+		kinds: ["extensions"],
+		args: {},
+		fromArgs: fromArgsOrUsage("extensions", { kind: "extensions" }),
+		handle(_command, ctx) {
+			ctx.openExtensions?.();
+		},
+	},
+	{
+		name: "interop",
+		description: "Review other coding agents detected on this machine",
+		group: "Inspect",
+		kinds: ["interop"],
+		args: {},
+		fromArgs: fromArgsOrUsage("interop", { kind: "interop" }),
+		handle(_command, ctx) {
+			ctx.openInterop?.();
+		},
+	},
+	{
+		name: "share",
+		description: "Share a worker result with the main agent, or export and import Clio archives",
+		group: "Sessions",
+		kinds: ["share"],
+		args: {
+			positionals: [{ name: "runId", required: false }],
+			subcommands: {
+				export: {
+					positionals: [{ name: "path", required: true }],
+				},
+				import: {
+					flags: [{ name: "--dry-run" }, { name: "--force" }],
+					positionals: [{ name: "path", required: true }],
+				},
+			},
+		},
+		fromArgs(parsed) {
+			const subcommand = parsed.subcommand === "export" || parsed.subcommand === "import" ? parsed.subcommand : undefined;
+			if (parsed.error) {
+				return {
+					kind: "share",
+					action: "usage",
+					...(subcommand !== undefined ? { subcommand } : {}),
+					error: parsed.error,
+				};
+			}
+			const path = parsed.positionals[0];
+			if (subcommand === "export" && path !== undefined) return { kind: "share", action: "export", path };
+			if (subcommand === "import" && path !== undefined) {
+				return {
+					kind: "share",
+					action: "import",
+					path,
+					dryRun: parsed.flags.has("--dry-run"),
+					force: parsed.flags.has("--force"),
+				};
+			}
+			// No subcommand means the worker-result sense of the word, with or
+			// without a run id. `export` and `import` are the archive senses and are
+			// matched above, so neither reading can swallow the other.
+			if (subcommand === undefined) {
+				return { kind: "share", action: "worker-run", ...(path !== undefined ? { runId: path } : {}) };
+			}
+			return { kind: "share", action: "usage", ...(subcommand !== undefined ? { subcommand } : {}) };
+		},
+		handle(command, ctx) {
+			if (command.kind !== "share") return;
+			if (command.action === "worker-run") {
+				shareWorkerRun(command.runId, ctx);
+				return;
+			}
+			if (command.action === "usage") {
+				const entry = BUILTIN_SLASH_COMMANDS.find((candidate) => candidate.name === "share");
+				if (!entry) return;
+				const usage = usageNotice(entry, command.subcommand);
+				ctx.notice("info", command.error ? `${command.error}\n${usage}` : usage);
+				return;
+			}
+			if (command.action === "export") {
+				if (!ctx.exportShareArchive) {
+					ctx.notice("error", "share export is not wired");
+					return;
+				}
+				const result = ctx.exportShareArchive(command.path);
+				ctx.notice("success", `exported ${result.fileCount} item(s) to ${result.path}`);
+				return;
+			}
+			if (command.action === "import") {
+				if (!ctx.importShareArchive) {
+					ctx.notice("error", "share import is not wired");
+					return;
+				}
+				const plan = ctx.importShareArchive(command.path, { dryRun: command.dryRun, force: command.force });
+				for (const diag of plan.diagnostics) {
+					const detail = diag.path ? `${diag.message}: ${diag.path}` : diag.message;
+					ctx.notice("warn", `${diag.type}: ${detail}`);
+				}
+				const write = plan.actions.filter((action) => action.action === "write").length;
+				const overwrite = plan.actions.filter((action) => action.action === "overwrite").length;
+				const skip = plan.actions.filter((action) => action.action === "skip").length;
+				ctx.notice(
+					command.dryRun ? "info" : "success",
+					`${command.dryRun ? "dry-run" : "import"} write=${write} overwrite=${overwrite} skip=${skip} settings=${plan.actions.filter((action) => action.action === "settings").length}`,
+				);
+			}
+		},
+	},
+	{
+		name: "run",
+		description: "Run a fleet agent",
+		group: "Run",
+		kinds: ["run", "run-usage"],
+		args: {
+			parseFlagsBeforeRest: true,
+			flags: [
+				{ name: "--agent-profile", aliases: ["--worker-profile", "--worker"], takesValue: true, valueName: "profile" },
+				{ name: "--runtime", aliases: ["--agent-runtime", "--worker-runtime"], takesValue: true, valueName: "runtimeId" },
+				{ name: "--target", takesValue: true, valueName: "id" },
+				{ name: "--model", takesValue: true, valueName: "id" },
+				{ name: "--thinking", takesValue: true, values: RUN_THINKING_LEVELS, valueName: "level" },
+				{ name: "--tool-profile", takesValue: true, values: TOOL_PROFILE_NAMES },
+				{ name: "--require", takesValue: true, repeatable: true, valueName: "cap" },
+				{ name: "--share" },
+			],
+			positionals: [
+				{ name: "agent", required: true },
+				{ name: "task", required: true, rest: true },
+			],
+		},
+		fromArgs(parsed, trimmed) {
+			if (parsed.error) return { kind: "run-usage" };
+			const options: RunCommandOptions = {};
+
+			const workerProfile = parsed.flags.get("--agent-profile");
+			if (typeof workerProfile === "string") options.workerProfile = workerProfile;
+
+			const workerRuntime = parsed.flags.get("--runtime");
+			if (typeof workerRuntime === "string") options.workerRuntime = workerRuntime;
+
+			const target = parsed.flags.get("--target");
+			if (typeof target === "string") options.target = target;
+
+			const model = parsed.flags.get("--model");
+			if (typeof model === "string") options.model = model;
+
+			const thinking = parsed.flags.get("--thinking");
+			if (typeof thinking === "string") {
+				if (!isRunThinkingLevel(thinking)) return { kind: "run-usage" };
+				options.thinkingLevel = thinking;
+			}
+
+			const toolProfile = parsed.flags.get("--tool-profile");
+			if (typeof toolProfile === "string") {
+				if (!isToolProfileName(toolProfile)) return { kind: "run-usage" };
+				options.toolProfile = toolProfile;
+			}
+
+			const requiredCapabilities = parsed.flagValues.get("--require");
+			if (requiredCapabilities && requiredCapabilities.length > 0) {
+				options.requiredCapabilities = [...requiredCapabilities];
+			}
+
+			if (parsed.flags.has("--share")) options.share = true;
+
+			const agentId = parsed.positionals[0] ?? "";
+			const task = parsed.positionals[1] ?? "";
+
+			return { kind: "run", agentId, task, options, source: trimmed };
+		},
+		handle(command, ctx) {
+			if (command.kind === "run-usage") {
+				const entry = BUILTIN_SLASH_COMMANDS.find((e) => e.name === "run");
+				if (entry) {
+					ctx.notice("info", usageNotice(entry));
+				}
+				return;
+			}
+			if (command.kind !== "run") return;
+			const { agentId, task, options } = command;
+			ctx.echoOperatorCommand?.(command.source);
+			void (async () => {
+				await handleRun(
+					agentId,
+					task,
+					{
+						dispatch: ctx.dispatch,
+						...(ctx.getAgentRoleFacts ? { getAgentRoleFacts: ctx.getAgentRoleFacts } : {}),
+						io: ctx.io,
+						notice: ctx.notice,
+						bus: ctx.bus,
+						...(ctx.submitOperatorNote ? { submitOperatorNote: ctx.submitOperatorNote } : {}),
+					},
+					options,
+				);
+				ctx.render();
+			})();
+		},
+	},
+	{
+		name: "delegate",
+		description: "Run an ACP delegation agent",
+		group: "Run",
+		kinds: ["delegate", "delegate-usage"],
+		args: {
+			parseFlagsBeforeRest: true,
+			flags: [{ name: "--share" }],
+			positionals: [
+				{ name: "agent-id", required: true },
+				{ name: "task", required: true, rest: true },
+			],
+		},
+		fromArgs(parsed, trimmed) {
+			if (parsed.error) return { kind: "delegate-usage" };
+			const agentId = parsed.positionals[0] ?? "";
+			const task = parsed.positionals[1] ?? "";
+			return { kind: "delegate", agentId, task, source: trimmed, ...(parsed.flags.has("--share") ? { share: true } : {}) };
+		},
+		handle(command, ctx) {
+			if (command.kind === "delegate-usage") {
+				const entry = BUILTIN_SLASH_COMMANDS.find((e) => e.name === "delegate");
+				if (entry) {
+					ctx.notice("info", usageNotice(entry));
+				}
+				return;
+			}
+			if (command.kind !== "delegate") return;
+			ctx.echoOperatorCommand?.(command.source);
+			const share = command.share === true;
+			void (async () => {
+				await handleDelegate(
+					command.agentId,
+					command.task,
+					{
+						dispatch: ctx.dispatch,
+						io: ctx.io,
+						notice: ctx.notice,
+						bus: ctx.bus,
+						...(ctx.submitOperatorNote ? { submitOperatorNote: ctx.submitOperatorNote } : {}),
+					},
+					{ share },
+				);
+				ctx.render();
+			})();
+		},
+	},
+	{
+		name: "agents",
+		description: "List Clio agents and ACP delegation agents",
+		group: "Inspect",
+		kinds: ["agents"],
+		args: {},
+		fromArgs: fromArgsOrUsage("agents", { kind: "agents" }),
+		handle(_command, ctx) {
+			ctx.openAgents();
+		},
+	},
+	settingsDeepLink("targets", "targets", "Open Settings → Targets: health, use, connect, probe, remove"),
+	{
+		name: "cost",
+		description: "Show session token and cost totals",
+		group: "Inspect",
+		kinds: ["cost"],
+		args: {},
+		fromArgs: fromArgsOrUsage("cost", { kind: "cost" }),
+		handle(_command, ctx) {
+			ctx.openCost();
+		},
+	},
+	{
+		name: "context",
+		description: "Context hub: window overlay plus compact, init, refresh, and reset",
+		group: "Inspect",
+		aliases: ["ctx", "compact"],
+		aliasArgs: { compact: "compact" },
+		kinds: ["context-view", "compact", "init", "context-clear", "context-refresh"],
+		subcommandDescriptions: {
+			compact: "Compact session context",
+			init: "Initialize project context",
+			refresh: "Refresh project context",
+			reset: "Reset project context",
+		},
+		args: {
+			subcommands: {
+				compact: { positionals: [...COMPACT_POSITIONALS] },
+				init: {},
+				refresh: {},
+				reset: {},
+			},
+		},
+		fromArgs(parsed) {
+			if (parsed.error) return { kind: "usage-error", command: "context", reason: parsed.error };
+			switch (parsed.subcommand) {
+				case "compact":
+					return { kind: "compact", instructions: parsed.rest };
+				case "init":
+					return { kind: "init", options: {} };
+				case "refresh":
+					return { kind: "context-refresh" };
+				case "reset":
+					return { kind: "context-clear", options: {} };
+				default:
+					return { kind: "context-view" };
+			}
+		},
+		handle(command, ctx) {
+			switch (command.kind) {
+				case "context-view":
+					ctx.openContextView();
+					return;
+				case "compact":
+					ctx.runCompact(command.instructions);
+					return;
+				case "init":
+					ctx.runInit(command.options);
+					return;
+				case "context-clear":
+					ctx.runContextClear(command.options);
+					return;
+				case "context-refresh":
+					if (ctx.runContextRefresh) {
+						ctx.runContextRefresh();
+					} else {
+						ctx.notice("error", "context refresh is not wired; pass onContextRefresh to startInteractive");
+					}
+					return;
+				default:
+					return;
+			}
+		},
+	},
+	settingsDeepLink("fleet", "fleet", "Open Settings → Fleet: defaults, profiles, agent bindings, nodes"),
+	{
+		name: "tasks",
+		description: "Show the session task board the agent tracks with the tasks tool",
+		group: "Inspect",
+		kinds: ["tasks"],
+		args: {},
+		fromArgs: fromArgsOrUsage("tasks", { kind: "tasks" }),
+		handle(_command, ctx) {
+			ctx.openTasks();
+		},
+	},
+	{
+		name: "memory",
+		description: "Inspect task memory or seed it from the newest handoff",
+		group: "Inspect",
+		kinds: ["memory", "memory-seed"],
+		subcommandDescriptions: { seed: "Seed the task bank from the newest handoff" },
+		args: { subcommands: { seed: {} } },
+		fromArgs(parsed) {
+			if (parsed.error) return { kind: "usage-error", command: "memory", reason: parsed.error };
+			return parsed.subcommand === "seed" ? { kind: "memory-seed" } : { kind: "memory" };
+		},
+		handle(command, ctx) {
+			if (command.kind === "memory") {
+				ctx.openMemory();
+				return;
+			}
+			if (command.kind !== "memory-seed") return;
+			const result = ctx.seedTaskMemory?.() ?? { status: "not-found" as const };
+			switch (result.status) {
+				case "seeded":
+					ctx.notice(
+						"success",
+						`task memory seeded ${result.seeded} entr${result.seeded === 1 ? "y" : "ies"} from ${result.source}${result.skipped > 0 ? `; skipped ${result.skipped} duplicate${result.skipped === 1 ? "" : "s"}` : ""}`,
+					);
+					return;
+				case "disabled":
+					ctx.notice("warn", "task memory is disabled; enable memory.intervention before seeding");
+					return;
+				case "not-found":
+					ctx.notice("info", "no structured task-memory snapshot found in the newest handoff");
+			}
+		},
+	},
+	{
+		name: "view",
+		description: "Browse session artifacts and verify receipts",
+		group: "Inspect",
+		kinds: ["view", "view-verify", "view-usage"],
+		args: {
+			positionals: [{ name: "filter", required: false, rest: true }],
+			subcommands: {
+				verify: {
+					positionals: [{ name: "runId", required: true }],
+				},
+			},
+		},
+		fromArgs(parsed) {
+			if (parsed.error) return { kind: "view-usage" };
+			if (parsed.subcommand === "verify") {
+				const runId = parsed.positionals[0] ?? "";
+				return { kind: "view-verify", runId };
+			}
+			return parsed.rest ? { kind: "view", filter: parsed.rest } : { kind: "view" };
+		},
+		handle(command, ctx) {
+			const entry = BUILTIN_SLASH_COMMANDS.find((e) => e.name === "view");
+			if (!entry) return;
+			if (command.kind === "view") {
+				ctx.openView(command.filter);
+				return;
+			}
+			if (command.kind === "view-usage") {
+				ctx.notice("info", usageNotice(entry, "verify"));
+				return;
+			}
+			if (command.kind !== "view-verify") return;
+			const result = ctx.verifyReceipt(command.runId);
+			if (result.ok) {
+				ctx.notice("success", `verify ok ${command.runId}`);
+			} else {
+				ctx.notice("error", `verify fail ${command.runId} ${result.reason}`);
+			}
+		},
+	},
+	{
+		name: "thinking",
+		description: "Set the chat thinking level, or open Settings → Orchestrator",
+		group: "Configure",
+		kinds: ["thinking-set"],
+		args: { positionals: [{ name: "level", required: false }] },
+		fromArgs(parsed) {
+			if (parsed.error) return { kind: "usage-error", command: "thinking", reason: parsed.error };
+			const level = parsed.positionals[0];
+			return level
+				? { kind: "thinking-set", level }
+				: { kind: "settings", section: "orchestrator", rowId: "orchestrator.thinkingLevel" };
+		},
+		handle(command, ctx) {
+			if (command.kind !== "thinking-set") return;
+			const result = ctx.setThinkingLevel?.(command.level) ?? { status: "unavailable" as const };
+			if (result.status === "applied") {
+				ctx.notice("success", `thinking: ${result.display}`);
+			} else if (result.status === "unsupported") {
+				ctx.notice("error", `thinking level "${result.level}" is not available here; try ${result.supported.join(", ")}`);
+			} else {
+				ctx.notice("error", "thinking level cannot be set right now");
+			}
+			ctx.render();
+		},
+	},
+	{
+		name: "output",
+		description: "Set transcript detail (minimal, default, verbose), or open Settings → Terminal",
+		group: "Configure",
+		kinds: ["output-set"],
+		args: { positionals: [{ name: "verbosity", required: false }] },
+		fromArgs(parsed) {
+			if (parsed.error) return { kind: "usage-error", command: "output", reason: parsed.error };
+			const verbosity = parsed.positionals[0];
+			return verbosity
+				? { kind: "output-set", verbosity }
+				: { kind: "settings", section: "terminal", rowId: "terminal.outputVerbosity" };
+		},
+		handle(command, ctx) {
+			if (command.kind !== "output-set") return;
+			const result = ctx.setOutputVerbosity?.(command.verbosity) ?? { status: "unavailable" as const };
+			if (result.status === "applied") {
+				ctx.notice("success", `output detail: ${result.verbosity}`);
+			} else if (result.status === "unsupported") {
+				ctx.notice("error", `output detail "${result.verbosity}" is not one of ${result.supported.join(", ")}`);
+			} else {
+				ctx.notice("error", "output detail cannot be set right now");
+			}
+			ctx.render();
+		},
+	},
+	{
+		name: "model",
+		description: "Open model selector or set a model",
+		group: "Configure",
+		aliases: ["models"],
+		kinds: ["model", "model-set"],
+		args: {
+			positionals: [{ name: "pattern", required: false, rest: true }],
+		},
+		fromArgs(parsed) {
+			const pattern = parsed.positionals[0];
+			if (pattern) {
+				return { kind: "model-set", pattern };
+			}
+			return { kind: "model" };
+		},
+		handle(command, ctx) {
+			if (command.kind === "model") {
+				ctx.openModel();
+				return;
+			}
+			if (command.kind !== "model-set") return;
+			void (async () => {
+				try {
+					await ctx.providers.probeAllLive();
+				} catch (err) {
+					const msg = err instanceof Error ? err.message : String(err);
+					ctx.notice("warn", `model refresh failed; resolving cached catalog: ${msg}`);
+				}
+				const result = resolveModelReference(command.pattern, ctx.providers);
+				if (!result.ref) {
+					ctx.notice("error", result.error ?? `no match for "${command.pattern}"`);
+					ctx.render();
+					return;
+				}
+				if (result.warning) ctx.notice("warn", result.warning);
+				ctx.applyModelRef(result.ref);
+				const suffix = result.ref.thinkingLevel ? ` thinking=${result.ref.thinkingLevel}` : "";
+				ctx.notice("success", `active: ${result.ref.target}/${result.ref.model}${suffix}`);
+				ctx.render();
+			})();
+		},
+	},
+	settingsDeepLink(
+		"scoped-models",
+		"models",
+		"Open Settings → Models: the Alt+J / Alt+K cycle set and favorites",
+		"scope",
+	),
+	{
+		name: "settings",
+		description: "Open interactive settings",
+		group: "Configure",
+		aliases: ["config"],
+		kinds: ["settings"],
+		args: { positionals: [{ name: "section", required: false }] },
+		fromArgs(parsed) {
+			if (parsed.error) return { kind: "usage-error", command: "settings", reason: parsed.error };
+			const section = parsed.positionals[0];
+			if (section === undefined) return { kind: "settings" };
+			if (!isSettingsSectionId(section)) {
+				const known = SETTINGS_SECTIONS.map((entry) => entry.id).join(", ");
+				return { kind: "usage-error", command: "settings", reason: `Unknown section: ${section} (one of ${known})` };
+			}
+			return { kind: "settings", section };
+		},
+		handle(command, ctx) {
+			if (command.kind === "settings") ctx.openSettings(command.section, command.rowId);
+		},
+	},
+	{
+		name: "resume",
+		description: "Resume a past session",
+		group: "Sessions",
+		kinds: ["resume"],
+		args: {},
+		fromArgs: fromArgsOrUsage("resume", { kind: "resume" }),
+		handle(_command, ctx) {
+			ctx.openResume();
+		},
+	},
+	{
+		name: "new",
+		description: "Start a fresh session",
+		group: "Sessions",
+		kinds: ["new"],
+		args: {},
+		fromArgs: fromArgsOrUsage("new", { kind: "new" }),
+		handle(_command, ctx) {
+			ctx.startNewSession();
+		},
+	},
+	{
+		name: "tree",
+		description: "Open session tree navigator",
+		group: "Sessions",
+		kinds: ["tree"],
+		args: {},
+		fromArgs: fromArgsOrUsage("tree", { kind: "tree" }),
+		handle(_command, ctx) {
+			ctx.openTree();
+		},
+	},
+	{
+		name: "fork",
+		description: "Fork from an assistant turn",
+		group: "Sessions",
+		kinds: ["fork"],
+		args: {},
+		fromArgs: fromArgsOrUsage("fork", { kind: "fork" }),
+		handle(_command, ctx) {
+			ctx.openMessagePicker();
+		},
+	},
+	{
+		name: "export",
+		description: "Export the session transcript to Markdown",
+		group: "Sessions",
+		kinds: ["export"],
+		args: {
+			positionals: [{ name: "path", required: false }],
+		},
+		fromArgs(parsed) {
+			return { kind: "export", path: parsed.positionals[0] };
+		},
+		handle(command, ctx) {
+			if (command.kind !== "export") return;
+			ctx.exportTranscript(command.path);
+		},
+	},
+];
+
+const HANDLER_BY_KIND = new Map<SlashCommandKind, BuiltinSlashCommand>();
+const COMMAND_TERM_OWNER = new Map<string, string>();
+for (const entry of BUILTIN_SLASH_COMMANDS) {
+	for (const term of [entry.name, ...(entry.aliases ?? [])]) {
+		const owner = COMMAND_TERM_OWNER.get(term);
+		if (owner) {
+			throw new Error(`BUILTIN_SLASH_COMMANDS: command term "${term}" is owned by both "${owner}" and "${entry.name}"`);
+		}
+		COMMAND_TERM_OWNER.set(term, entry.name);
+	}
+	for (const kind of entry.kinds) {
+		if (HANDLER_BY_KIND.has(kind)) {
+			throw new Error(`BUILTIN_SLASH_COMMANDS: kind "${kind}" is owned by multiple entries`);
+		}
+		HANDLER_BY_KIND.set(kind, entry);
+	}
+}
+
+/**
+ * A token that could only have been meant as a command: one word of letters,
+ * digits, and hyphens after the leading slash.
+ *
+ * A spelling that names no command used to match nothing and fall through to
+ * the model as prose. Measured: `/compact` was answered "/compact (completed)"
+ * in six output tokens while context stayed at 9%, no compaction summary was
+ * written, and no hook fired. The operator read a completion that never
+ * happened, which is the same failure `/thinking off` had. Only active commands
+ * run, so anything command-shaped that the registry does not claim fails here
+ * rather than being answered by a model that will invent a result.
+ *
+ * The shape test is what keeps that from swallowing ordinary text, because a
+ * leading slash is also how an absolute path starts. `/home/user/notes` and
+ * `/not/a/command` carry a separator, so they are not one word and still reach
+ * the model unchanged.
+ *
+ * One word followed by prose stays ambiguous: `/compact tidy up` is the defect
+ * and `/tmp is full` is a sentence, and nothing in the text separates them. It
+ * resolves as a command, so a sentence that has to open this way needs
+ * COMMAND_ESCAPE.
+ */
+const COMMAND_SHAPED_TOKEN = /^[A-Za-z][A-Za-z0-9-]*$/u;
+
+/**
+ * Prefix that sends a command-shaped line to the model as text.
+ *
+ * This used to be a leading space, guarded by testing the raw input rather than
+ * the trimmed one. No operator could ever reach it: the terminal engine trims
+ * the submitted line inside its own submit path, so the parser only ever sees
+ * text with the whitespace already gone. Typing ` /tmp is full` in the TUI
+ * failed with "/tmp is not a command" exactly like the unescaped spelling, and
+ * a contract test asserted the unreachable branch.
+ *
+ * A backslash survives the trim. It claims one character and only in front of a
+ * slash, so `\\server\share` and `\note to self` are still ordinary text.
+ */
+const COMMAND_ESCAPE = "\\/";
+
+/** Pure slash-command parser: no I/O, no side effects. Walks the registry in order. */
+export function parseSlashCommand(input: string): SlashCommand {
+	const trimmed = input.trim();
+	if (trimmed.length === 0) return { kind: "empty" };
+	if (trimmed.startsWith(COMMAND_ESCAPE)) return { kind: "unknown", text: trimmed.slice(1) };
+	for (const entry of BUILTIN_SLASH_COMMANDS) {
+		const match = entry.match ? entry.match(trimmed) : matchFromSpec(entry, trimmed);
+		if (match) return match;
+	}
+	if (trimmed.startsWith("/")) {
+		const token = trimmed.slice(1).split(/\s+/u)[0] ?? "";
+		if (COMMAND_SHAPED_TOKEN.test(token)) return { kind: "unknown-command", token, text: trimmed };
+	}
+	return { kind: "unknown", text: trimmed };
+}
+
+/**
+ * Dispatches a parsed SlashCommand to its owning registry entry. `empty` is a
+ * no-op and `unknown` falls through to chat submission; every other kind
+ * resolves to exactly one registry entry at module load.
+ */
+export function dispatchSlashCommand(command: SlashCommand, ctx: SlashCommandContext): void {
+	if (command.kind === "empty") return;
+	if (command.kind === "unknown") {
+		ctx.submitChat(command.text);
+		return;
+	}
+	if (command.kind === "unknown-command") {
+		// The registry does not own the token, but a prompt template might. The
+		// parser stays pure, so the lookup happens here, and only for a spelling
+		// that matched no command: `/name` is how every other agent invokes the
+		// commands sitting in the roots Clio reads, and answering "not a command"
+		// for one that is loaded made the whole foreign prompt surface unreachable.
+		const expansion = ctx.expandPromptTemplate?.(command.text);
+		if (expansion?.expanded === true) {
+			ctx.submitChat(command.text);
+			return;
+		}
+		// A template that exists and refused is not a typo. Its reason reaches the
+		// operator and nothing reaches the model.
+		const refusal = expansion?.expanded === false ? expansion.refusal : undefined;
+		ctx.notice("error", refusal ? refusal.message : `/${command.token} is not a command. Type /help for the list.`);
+		ctx.render();
+		return;
+	}
+	if (command.kind === "usage-error") {
+		const entry = BUILTIN_SLASH_COMMANDS.find((candidate) => candidate.name === command.command);
+		const usage = entry ? ` ${usageNotice(entry)}` : "";
+		ctx.notice("error", `${command.reason}.${usage}`);
+		ctx.render();
+		return;
+	}
+	const entry = HANDLER_BY_KIND.get(command.kind);
+	if (!entry) return;
+	entry.handle(command, ctx);
+}
+
+export interface CommandReferenceEntry {
+	name: string;
+	aliases: ReadonlyArray<string>;
+	/** Argument text each alias in `aliases` stands in for, when it stands for any. */
+	aliasArgs?: Readonly<Record<string, string>>;
+	usage: string;
+	description: string;
+	group: SlashCommandGroup;
+	/** The command's args grammar, carried through for argument autocomplete. */
+	args?: CommandArgsSpec;
+	/** Stable copy shown beside first-token subcommand completions. */
+	subcommandDescriptions?: Readonly<Record<string, string>>;
+}
+
+export function commandReference(): ReadonlyArray<CommandReferenceEntry> {
+	return BUILTIN_SLASH_COMMANDS.filter((entry) => entry.hidden !== true).map((entry) => {
+		const usage = usageLine(entry)
+			.replace(/^\nusage:\s*/, "")
+			.replace(/\n$/, "");
+		return {
+			name: entry.name,
+			aliases: entry.aliases ?? [],
+			...(entry.aliasArgs ? { aliasArgs: entry.aliasArgs } : {}),
+			usage,
+			description: entry.description,
+			group: entry.group,
+			...(entry.args ? { args: entry.args } : {}),
+			...(entry.subcommandDescriptions ? { subcommandDescriptions: entry.subcommandDescriptions } : {}),
+		};
+	});
+}