@iowarp/clio-coder 0.3.0 → 0.3.2

package/src/entry/orchestrator.ts

@@ -0,0 +1,2011 @@
+import { join } from "node:path";
+import chalk from "chalk";
+import { modelBootstrapGenerate, resolveBootstrapRoute } from "../cli/bootstrap-generate.js";
+import { runHeadlessMainAgent } from "../cli/modes/print.js";
+import { formatBootTrace } from "../core/boot-trace.js";
+import { BusChannels } from "../core/bus-events.js";
+import { installBusTracer } from "../core/bus-trace.js";
+import { type ClioSettings, readSettings, type SettingsMutator, updateSettings } from "../core/config.js";
+import { DEFAULT_DELEGATION_PERMISSION_TIMEOUT_MS } from "../core/defaults.js";
+import { loadDomains } from "../core/domain-loader.js";
+import { expandInlineFileReferencesAsync } from "../core/file-references.js";
+import { setGitCommitAttributionEnabled } from "../core/git-commit-attribution.js";
+import { configureGuardrails } from "../core/guardrails.js";
+import { HEADLESS_PERMISSION_DENIED_REASON } from "../core/headless-permission.js";
+import { rememberRecentModel } from "../core/recent-models.js";
+import { protectedResidencyModels } from "../core/residency-protection.js";
+import {
+	applyOverrides,
+	applyRoutingPatch,
+	applySessionRouting,
+	diffRouting,
+	getAtPath,
+	isRoutingPath,
+	mergeRoutingPatchIntoSettings,
+	type RoutingPatch,
+	restoreRoutingFields,
+	routingChangeNotices,
+	routingPatchForId,
+	type SessionOverrides,
+	seedSessionRouting,
+	setAtPath,
+} from "../core/session-routing.js";
+import { getSharedBus } from "../core/shared-bus.js";
+import { isSkillActivation } from "../core/skill-activation.js";
+import { StartupTimer } from "../core/startup-timer.js";
+import { getTerminationCoordinator } from "../core/termination.js";
+import { clioDataDir, clioStateDir } from "../core/xdg.js";
+import { renderAgentCatalogSectionsFromSpecs } from "../domains/agents/catalog.js";
+import type { AgentsContract } from "../domains/agents/contract.js";
+import { AgentsDomainModule } from "../domains/agents/index.js";
+import type { ConfigContract } from "../domains/config/contract.js";
+import { ConfigDomainModule, createConfigDomainModule } from "../domains/config/index.js";
+import type { ContextContract } from "../domains/context/contract.js";
+import { bootstrapInputFromInitOptions } from "../domains/context/init-options.js";
+import { createContextDomainModule } from "../domains/context/runtime.js";
+import type { DispatchContract } from "../domains/dispatch/contract.js";
+import { createDispatchDedupRegistration } from "../domains/dispatch/dedup.js";
+import { agentRoleFactsResolver } from "../domains/dispatch/execution-role.js";
+import { readGateDecisionArtifacts, readPendingGateDecisions } from "../domains/dispatch/gate-decisions.js";
+import { createDispatchDomainModule } from "../domains/dispatch/index.js";
+import { type ExtensionsContract, ExtensionsDomainModule } from "../domains/extensions/index.js";
+import { type InteropContract, InteropDomainModule } from "../domains/interop/index.js";
+import {
+	describeUpgradeNotice,
+	ensureClioState,
+	LifecycleDomainModule,
+	takeUpgradeNotice,
+} from "../domains/lifecycle/index.js";
+import { getVersionInfo } from "../domains/lifecycle/version.js";
+import {
+	buildMemoryPromptSection,
+	canonicalMemoryRepositoryIdentity,
+	createTaskMemoryTelemetrySink,
+	createTaskMemoryTrace,
+	loadMemoryRecordsSync,
+	renderTaskMemoryHandoffSource,
+	seedTaskMemoryFromNewestHandoff,
+	type TaskMemoryModelClient,
+	taskMemoryBankSize,
+	taskMemoryHandoffSeedOffer,
+	taskMemoryTracePath,
+} from "../domains/memory/index.js";
+import { TaskMemoryBank } from "../domains/memory/task-bank.js";
+import {
+	createDetachedDispatchNudgeRegistration,
+	createReadOnlyExplorationNudgeRegistration,
+	createUnbackedWorkerClaimRegistration,
+	openDetachedBatchViews,
+} from "../domains/middleware/dispatch-nudge.js";
+import {
+	createHookReceiptLog,
+	createMiddlewareToolChoiceControl,
+	createSkillsReminderRegistration,
+	type ExtensionHookRoot,
+	installUserHooks,
+	type MiddlewareContract,
+	MiddlewareDomainModule,
+	writeMiddlewareDiagnosticToStderr,
+} from "../domains/middleware/index.js";
+import { createMemoryInterventionRegistration } from "../domains/middleware/memory-intervention.js";
+import { createTaskBoardReminderRegistration } from "../domains/middleware/task-board-reminder.js";
+import { createTaskNudgeRegistration } from "../domains/middleware/task-nudge.js";
+import type { ObservabilityContract } from "../domains/observability/index.js";
+import { ObservabilityDomainModule } from "../domains/observability/index.js";
+import type { PromptsContract } from "../domains/prompts/contract.js";
+import { createPromptsDomainModule } from "../domains/prompts/index.js";
+import type { ProvidersContract, TargetDescriptor, ThinkingLevel } from "../domains/providers/index.js";
+import {
+	AGENT_ROLE_TOOLS_REQUIRED_REASON,
+	applyModelCapabilityPatch,
+	firstRuntimeResolutionError,
+	isOrchestratorEligibleRuntime,
+	ProvidersDomainModule,
+	refineRuntimeTargetWithModelHints,
+	resolveModelCapabilities,
+	resolveModelRuntimeCapabilitiesForProviders,
+	resolveRuntimeTarget,
+	supportsAgentRoleTools,
+	targetRequiresAuth,
+	VALID_THINKING_LEVELS,
+} from "../domains/providers/index.js";
+import { getRuntimeRegistry } from "../domains/providers/registry.js";
+import { registerBuiltinRuntimes } from "../domains/providers/runtimes/builtins.js";
+import {
+	createResourcesDomainModule,
+	discoverMarketplaceSkills,
+	modelVisibleSkills,
+	type ResourcesContract,
+} from "../domains/resources/index.js";
+import { DEFAULT_RECENT_ENTRY_LIMIT } from "../domains/safety/finish-contract.js";
+import { createFinishContractRegistration } from "../domains/safety/finish-contract-registration.js";
+import type { AutonomyLevel, SafetyContract } from "../domains/safety/index.js";
+import { parseRigorOverride, resolveRigor, SafetyDomainModule } from "../domains/safety/index.js";
+import type { ProtectedArtifactState } from "../domains/safety/protected-artifacts.js";
+import {
+	createProtectedArtifactsRegistration,
+	type ProtectedArtifactProtectEvent,
+} from "../domains/safety/protected-artifacts-registration.js";
+import type { SchedulingContract } from "../domains/scheduling/contract.js";
+import { SchedulingDomainModule } from "../domains/scheduling/index.js";
+import { type CompactResult, compact } from "../domains/session/compaction/compact.js";
+import { collectSessionEntries } from "../domains/session/compaction/session-entries.js";
+import { estimateTokens } from "../domains/session/compaction/tokens.js";
+import { ceilChars } from "../domains/session/context-accounting.js";
+import type { SessionContract, SessionMeta } from "../domains/session/contract.js";
+import { createDecisionBoardStore } from "../domains/session/decision-board.js";
+import type { CompactionSummaryEntry, CompactionTrigger, SessionEntry } from "../domains/session/entries.js";
+import { SessionDomainModule } from "../domains/session/index.js";
+import {
+	clearPendingProtectedArtifact,
+	reconcilePendingProtectedArtifacts,
+	stagePendingProtectedArtifact,
+} from "../domains/session/protected-artifact-journal.js";
+import {
+	protectedArtifactEntryFromArtifact,
+	protectedArtifactStateFromSessionEntries,
+} from "../domains/session/protected-artifacts.js";
+import { createTaskBoardStore } from "../domains/session/task-board.js";
+import { filterEntriesToActivePath } from "../domains/session/tree/active-path.js";
+import { type ShareContract, ShareDomainModule } from "../domains/share/index.js";
+import { createUserTasksStore } from "../domains/user-tasks/store.js";
+import { type AcpSafeSettingsPatch, type AcpSafeSettingsSnapshot, serveClioAcpAgent } from "../engine/acp/server.js";
+import { createStdioServerTransport } from "../engine/acp/transport.js";
+import { completeEngineText } from "../engine/ai.js";
+import { setProtectedModelsProvider } from "../engine/apis/residency.js";
+import {
+	createLoopGuardRegistration,
+	INTERACTIVE_LOOP_BLOCK_BUDGET,
+	readOrchTurnToolCallBudget,
+} from "../engine/loop-guard.js";
+import { openSession, readSessionTailTurns, sessionCurrentPath, sessionPaths } from "../engine/session.js";
+import type { EngineModel } from "../engine/types.js";
+import { createChatLoop } from "../interactive/chat-loop.js";
+import { buildReplayAgentMessagesFromTurns } from "../interactive/chat-renderer.js";
+import { type RunIo, startInteractive } from "../interactive/index.js";
+import type { BootOptions } from "./boot-options.js";
+
+export type { BootOptions, HeadlessSamplingOverrides } from "./boot-options.js";
+
+import {
+	detectPlatformKeybindingWarnings,
+	detectTerminalKeySupport,
+	formatInvalidKeybindingNotice,
+	formatPlatformKeybindingNotice,
+	validateKeybindings,
+} from "../interactive/keybinding-manager.js";
+import { subscribeLoopGuardStop } from "../interactive/loop-guard-interrupt.js";
+import { BUILTIN_SLASH_COMMANDS } from "../interactive/slash-commands.js";
+import { createToolProseRegistration } from "../interactive/tool-prose-registration.js";
+import { type AskUserHandler, cancelledAskUserResult } from "../tools/ask-user.js";
+import { registerAllTools } from "../tools/bootstrap.js";
+import { isGitRepository, recoverCleanupReadyCompeteGroups } from "../tools/compete-worktrees.js";
+import { createDispatchBackgroundRegistry } from "../tools/dispatch-background.js";
+import { createFileMutationObserver, createSkillActivationObserver } from "../tools/observers.js";
+import { createRegistry } from "../tools/registry.js";
+
+export interface BootResult {
+	exitCode: number;
+	bootTimeMs: number;
+}
+
+/**
+ * The bannered boot is the whole of what a piped or CI invocation of bare
+ * `clio` shows, so it is the entire first impression for a stranger who is not
+ * on a TTY. It used to end in a hardcoded `ready`, a word with no relationship
+ * to anything: a machine with no target configured at all printed it and
+ * exited 0, which is the one state where the installation can do nothing.
+ *
+ * What the line reports now is the orchestrator target the settings actually
+ * declare, and when none is declared it says so and names the command that
+ * fixes it. The exit status stays 0 either way, because this path answers
+ * "did Clio boot", which it did, and CI scripts already depend on that
+ * answer; the readiness of the configuration is reported in words instead.
+ */
+function bannerConfigurationLine(): string {
+	let settings: ClioSettings;
+	try {
+		settings = readSettings();
+	} catch {
+		return chalk.yellow("settings.yaml is not valid. Run `clio-coder doctor` for the exact keys.");
+	}
+	// A dangling chat target normalizes to null in the schema, so a deleted
+	// target arrives here as no target at all rather than as a name to report.
+	const targetId = settings.orchestrator?.target;
+	if (!targetId) {
+		return chalk.yellow("no model target configured. Run `clio-coder configure` to add one.");
+	}
+	const model = settings.orchestrator?.model;
+	return chalk.dim(`target ${targetId}${model ? ` · model ${model}` : " · no default model"}`);
+}
+
+function buildBanner(): string {
+	const { clio } = getVersionInfo();
+	return `
+  ${chalk.cyan("Clio Coder")}
+  ${chalk.dim(`v${clio} · CLIO: Context Layer for I/O · HPC & scientific software`)}
+  ${bannerConfigurationLine()}
+`;
+}
+
+function printJsonSessionHeader(meta: SessionMeta | null): Record<string, unknown> | null {
+	if (!meta) return null;
+	return {
+		type: "session",
+		version: meta.sessionFormatVersion ?? 1,
+		id: meta.id,
+		timestamp: meta.createdAt,
+		cwd: meta.cwd,
+		target: meta.target,
+		model: meta.model,
+		clioVersion: meta.clioVersion,
+	};
+}
+
+function applyHeadlessSettingsOverlay(
+	settings: ClioSettings,
+	overrides: BootOptions["headless"] | undefined,
+): ClioSettings {
+	const next = structuredClone(settings);
+	if (!overrides) return next;
+	const previousTarget = next.orchestrator.target;
+	if (overrides.target !== undefined) {
+		next.orchestrator.target = overrides.target;
+		if (overrides.model === undefined && (previousTarget !== overrides.target || !next.orchestrator.model)) {
+			const target = next.targets.find((entry) => entry.id === overrides.target);
+			if (target) next.orchestrator.model = target.defaultModel ?? null;
+		}
+	}
+	if (overrides.model !== undefined) next.orchestrator.model = overrides.model;
+	if (overrides.thinking !== undefined) next.orchestrator.thinkingLevel = overrides.thinking;
+	if (overrides.autonomy !== undefined) next.autonomy = overrides.autonomy;
+	return next;
+}
+
+interface CompactionResolution {
+	model: EngineModel;
+	targetId: string;
+	apiKey?: string;
+}
+
+function resolveTarget(providers: ProvidersContract, targetId: string | null | undefined): TargetDescriptor | null {
+	if (!targetId) return null;
+	return providers.getTarget(targetId);
+}
+
+function advanceThinkingLevel(current: ThinkingLevel, available: ReadonlyArray<ThinkingLevel>): ThinkingLevel {
+	const levels = available.length > 0 ? available : VALID_THINKING_LEVELS;
+	if (!levels.includes(current)) return levels[0] ?? "off";
+	const normalized = current;
+	const idx = levels.indexOf(normalized);
+	return levels[(idx + 1) % levels.length] ?? "off";
+}
+
+/**
+ * pi-ai's openai-completions provider refuses to stream without an apiKey even
+ * when the target is a local server that ignores the Authorization header
+ * entirely. The chat loop, the dispatch workers, and the background memory
+ * model all send this placeholder so a local llama.cpp or LM Studio endpoint
+ * works without the user inventing a credential.
+ */
+const LOCAL_API_KEY_FALLBACK = "clio-local-target";
+
+export async function resolveApiKeyForTarget(
+	target: TargetDescriptor,
+	providers: ProvidersContract,
+	signal?: AbortSignal,
+): Promise<string | undefined> {
+	const runtime = providers.getRuntime(target.runtime);
+	if (!runtime) return undefined;
+	// A target that needs no credential still needs the placeholder. Returning
+	// nothing here left compaction as the one path that did not send it, so
+	// `/context compact` and every automatic compaction at the window threshold
+	// failed on exactly the local runtimes Clio is built for, while ordinary
+	// turns against the same target succeeded.
+	if (!targetRequiresAuth(target, runtime)) return LOCAL_API_KEY_FALLBACK;
+	const resolved = await providers.auth.resolveForTarget(target, runtime, signal ? { signal } : undefined);
+	return resolved.apiKey;
+}
+
+function createBackgroundMemoryModelClient(
+	providers: ProvidersContract,
+	settings: Readonly<ClioSettings>,
+	timeoutMs: number,
+): TaskMemoryModelClient | null {
+	const targetId = settings.background.target?.trim();
+	const wireModelId = settings.background.model?.trim();
+	if (!targetId || !wireModelId) return null;
+	// One model cannot both drive the action agent and think its way through a
+	// memory step: the memory call would contend with the agent's own decoding on
+	// the same server, and a reasoning model spends most of a memory step
+	// deliberating over a task it is not executing. Memory falls back to its free
+	// deterministic tier rather than degrading the work the operator asked for.
+	if (backgroundSharesReasoningModelWithOrchestrator(providers, settings)) return null;
+	const resolved = resolveRuntimeTarget(providers, {
+		targetId,
+		wireModelId,
+		// Memory reads a trajectory and writes a fixed envelope. Reasoning adds
+		// latency and, on a small local model, routinely consumes the entire
+		// output budget before the envelope is ever written.
+		requestedThinkingLevel: "off",
+		use: "orchestrator",
+		requireTools: false,
+		requireOutputBudget: true,
+	});
+	if (!resolved.ok) {
+		const detail = firstRuntimeResolutionError(resolved.diagnostics) ?? "background target resolution failed";
+		throw new Error(detail);
+	}
+	const kbHit = providers.knowledgeBase?.lookup(resolved.target.wireModelId) ?? null;
+	const model = resolved.target.runtime.synthesizeModel(resolved.target.target, resolved.target.wireModelId, kbHit);
+	const refined = refineRuntimeTargetWithModelHints(resolved.target, model, providers.knowledgeBase);
+	applyModelCapabilityPatch(model, refined.capabilities);
+	return {
+		async complete(request) {
+			const apiKey = targetRequiresAuth(refined.target, refined.runtime)
+				? (await providers.auth.resolveForTarget(refined.target, refined.runtime, { signal: request.signal })).apiKey
+				: LOCAL_API_KEY_FALLBACK;
+			return completeEngineText({
+				model,
+				systemPrompt: request.systemPrompt,
+				userPrompt: request.userPrompt,
+				maxTokens: request.maxTokens,
+				// Always off, never the operator's chat thinking level. A model that
+				// reasons anyway still works: `completeEngineText` keeps only text
+				// blocks, and the memory output budget leaves room for the preamble.
+				thinkingLevel: "off",
+				signal: request.signal,
+				timeoutMs,
+				...(apiKey === undefined ? {} : { apiKey }),
+			});
+		},
+	};
+}
+
+/**
+ * Warn about a configured agent-role model the provider reports cannot call
+ * tools. Selection already refuses these, so reaching here means the config
+ * predates the check or was hand-edited; a run would otherwise fail later at
+ * dispatch admission with a message that names a missing tool rather than the
+ * model that cannot use any.
+ */
+function agentRoleToolWarnings(providers: ProvidersContract, settings: Readonly<ClioSettings>): string[] {
+	const roles: ReadonlyArray<{ label: string; target: string | null; model: string | null }> = [
+		{ label: "orchestrator", target: settings.orchestrator.target, model: settings.orchestrator.model },
+		{ label: "workers.default", target: settings.workers.default.target, model: settings.workers.default.model },
+	];
+	const warnings: string[] = [];
+	for (const role of roles) {
+		const targetId = role.target?.trim();
+		const wireModelId = role.model?.trim();
+		if (!targetId || !wireModelId) continue;
+		try {
+			const status = providers.list().find((entry) => entry.target.id === targetId);
+			if (!status) continue;
+			const capabilities = resolveModelCapabilities(status, wireModelId, providers.knowledgeBase);
+			if (supportsAgentRoleTools(capabilities)) continue;
+			warnings.push(
+				`${role.label} model '${wireModelId}' on target '${targetId}' ${AGENT_ROLE_TOOLS_REQUIRED_REASON}. ` +
+					`Pick another model, or state the correction in a model-catalog.d entry if the provider's flag is wrong.`,
+			);
+		} catch {
+			// An unresolvable capability is not evidence of a missing one.
+		}
+	}
+	return warnings;
+}
+
+/**
+ * True when the background role names the same model the orchestrator runs and
+ * that model reasons. Distinct models, or a non-reasoning shared model, are both
+ * fine; this only catches the one-model-does-everything configuration.
+ */
+function backgroundSharesReasoningModelWithOrchestrator(
+	providers: ProvidersContract,
+	settings: Readonly<ClioSettings>,
+): boolean {
+	const backgroundModel = settings.background.model?.trim();
+	const orchestratorModel = settings.orchestrator.model?.trim();
+	if (!backgroundModel || backgroundModel !== orchestratorModel) return false;
+	if (settings.background.target?.trim() !== settings.orchestrator.target?.trim()) return false;
+	try {
+		const status = providers.list().find((entry) => entry.target.id === settings.background.target?.trim());
+		if (!status) return false;
+		return resolveModelCapabilities(status, backgroundModel, providers.knowledgeBase).reasoning === true;
+	} catch {
+		// An unresolvable capability is not evidence of a reasoning model; the
+		// operator's explicit configuration stands.
+		return false;
+	}
+}
+
+export function synthesizeOrchestratorModel(
+	providers: ProvidersContract,
+	target: TargetDescriptor,
+	wireModelId: string,
+): EngineModel | null {
+	const runtime = providers.getRuntime(target.runtime);
+	if (!runtime) return null;
+	let model: EngineModel;
+	try {
+		const kbHit = providers.knowledgeBase?.lookup(wireModelId) ?? null;
+		model = runtime.synthesizeModel(target, wireModelId, kbHit);
+	} catch {
+		return null;
+	}
+	try {
+		const status = providers.list().find((entry) => entry.target.id === target.id);
+		if (status) {
+			const caps = resolveModelCapabilities(status, wireModelId, providers.knowledgeBase, {
+				detectedReasoning: providers.getDetectedReasoning(target.id, wireModelId),
+			});
+			applyModelCapabilityPatch(model, caps);
+		}
+	} catch {
+		// Older test doubles and degraded provider bundles may not expose live
+		// status. The synthesized model still carries runtime and catalog caps.
+	}
+	return model;
+}
+
+async function resolveCompactionModel(
+	settings: ClioSettings,
+	providers: ProvidersContract,
+): Promise<CompactionResolution | null> {
+	const targetId = settings.orchestrator?.target ?? null;
+	const wireModelId = settings.orchestrator?.model ?? null;
+	if (!targetId || !wireModelId) return null;
+	const target = resolveTarget(providers, targetId);
+	if (!target) return null;
+	const model = synthesizeOrchestratorModel(providers, target, wireModelId);
+	if (!model) return null;
+	const apiKey = await resolveApiKeyForTarget(target, providers);
+	const resolution: CompactionResolution = { model, targetId };
+	if (apiKey !== undefined) resolution.apiKey = apiKey;
+	return resolution;
+}
+
+function readSessionEntriesForCompact(sessionId: string): SessionEntry[] {
+	const reader = openSession(sessionId);
+	return collectSessionEntries(reader.turns(), sessionPaths(reader.meta()).current);
+}
+
+/**
+ * The finish-contract only inspects the window since the last user message
+ * (`recentEntries`, capped at 80), so it reads a bounded tail of the ledger
+ * instead of parsing the whole file every turn_end. 160 = twice the 80-entry cap
+ * leaves ample margin above any entries appended after the assistant turn, so the
+ * assessed window is byte-identical to the whole-file read (see the
+ * behaviour-equivalence contract test) while cost stays bounded by session
+ * *shape*, not session *length*.
+ */
+const FINISH_CONTRACT_TAIL_ENTRIES = DEFAULT_RECENT_ENTRY_LIMIT * 2;
+
+function readRecentSessionEntriesForContract(sessionId: string): SessionEntry[] {
+	return collectSessionEntries(
+		readSessionTailTurns(sessionId, FINISH_CONTRACT_TAIL_ENTRIES).entries,
+		sessionCurrentPath(sessionId),
+	);
+}
+
+function protectedArtifactStateForCurrentSession(
+	session: SessionContract,
+): ReturnType<typeof protectedArtifactStateFromSessionEntries> {
+	const meta = session.current();
+	if (!meta) return { artifacts: [] };
+	reconcilePendingProtectedArtifacts(session);
+	return protectedArtifactStateFromSessionEntries(readSessionEntriesForCompact(meta.id));
+}
+
+function appendProtectedArtifactRegistryEvent(
+	session: SessionContract | undefined,
+	event: ProtectedArtifactProtectEvent,
+): void {
+	const current = session?.current();
+	if (session === undefined || current === null || current === undefined) {
+		throw new Error("no active session is available for protected artifact persistence");
+	}
+	const pending = stagePendingProtectedArtifact(current.id, event);
+	session.appendEntry(
+		protectedArtifactEntryFromArtifact(event.artifact, {
+			parentTurnId: event.turnId ?? null,
+			toolName: event.toolName,
+			...(event.toolCallId !== undefined ? { toolCallId: event.toolCallId } : {}),
+			...(event.runId !== undefined ? { runId: event.runId } : {}),
+			...(event.correlationId !== undefined ? { correlationId: event.correlationId } : {}),
+		}),
+	);
+	if (session.flushAppends === undefined) {
+		throw new Error("session does not expose the durable append flush required by protected artifact persistence");
+	}
+	session.flushAppends();
+	clearPendingProtectedArtifact(pending);
+}
+
+/**
+ * Fold a terminal dispatch payload's worker skill activations into the session
+ * ledger, tagged with the runId, so worker skill provenance sits next to
+ * main-agent activations.
+ *
+ * Both terminal channels carry them and the run finalizer emits exactly one of
+ * the two, so this cannot double-record. The orchestrator never observes worker
+ * tool calls directly, because a worker runs its own registry in its own
+ * subprocess, which makes this the only recording path there is. Returns how
+ * many were folded.
+ */
+export function foldDispatchSkillActivations(
+	session: SessionContract | undefined,
+	payload: { runId?: unknown; skillActivations?: ReadonlyArray<unknown> } | undefined,
+): number {
+	const runId = payload?.runId;
+	if (typeof runId !== "string") return 0;
+	let folded = 0;
+	for (const activation of payload?.skillActivations ?? []) {
+		if (!isSkillActivation(activation)) continue;
+		appendSkillActivationRegistryEvent(session, { ...activation, runId });
+		folded += 1;
+	}
+	return folded;
+}
+
+function appendSkillActivationRegistryEvent(
+	session: SessionContract | undefined,
+	activation: Parameters<SessionContract["recordSkillActivation"]>[0],
+): void {
+	if (!session?.current()) return;
+	try {
+		session.recordSkillActivation(activation);
+	} catch {
+		// Activation metadata should never alter the result of a completed
+		// context(scope=skills) call. Missing ledger data is visible in diagnostics.
+	}
+}
+
+async function runCompactionFlow(
+	session: SessionContract,
+	settings: ClioSettings,
+	providers: ProvidersContract,
+	instructions?: string,
+	trigger?: CompactionTrigger,
+): Promise<CompactResult | null> {
+	const meta = session.current();
+	if (!meta) {
+		throw new Error("no current session to compact; start one with /new or /resume first");
+	}
+	const resolved = await resolveCompactionModel(settings, providers);
+	if (!resolved) {
+		throw new Error("no model configured; set orchestrator.target + orchestrator.model");
+	}
+	// Summarize only the active branch: after a /tree switch the raw file
+	// still holds abandoned sibling turns, and a summary that folds them in
+	// would persist abandoned content back into the active context. The full
+	// file read stays in place for protected artifacts and the masking
+	// rewrite, which are session-global. The task board is not: it used to
+	// read the full file too (last taskLedger entry in file order, with no
+	// branch filter at all), which was a second, independent instance of this
+	// same bug. See the taskBoard wiring below for the fix.
+	const activeLeafTurnId = session.tree(meta.id).leafId ?? undefined;
+	const entries = filterEntriesToActivePath(readSessionEntriesForCompact(meta.id), activeLeafTurnId);
+	if (entries.length === 0) return null;
+
+	const result = await compact({
+		entries,
+		model: resolved.model,
+		...(resolved.apiKey !== undefined ? { apiKey: resolved.apiKey } : {}),
+		...(instructions !== undefined ? { instructions } : {}),
+	});
+	if (result.messagesSummarized === 0 || result.summary.length === 0) return null;
+
+	const entry: Omit<CompactionSummaryEntry, "turnId" | "timestamp"> = {
+		kind: "compactionSummary",
+		parentTurnId: result.firstKeptTurnId ?? null,
+		summary: result.summary,
+		tokensBefore: result.tokensBefore,
+		firstKeptTurnId: result.firstKeptTurnId ?? "",
+		messagesSummarized: result.messagesSummarized,
+		isSplitTurn: result.isSplitTurn,
+		tokensAfter: estimateTokensAfterCompaction(entries, result),
+		// The summarization call is a real model call. Persisting its provider
+		// usage on the entry is what puts it in front of `/cost` and `clio-coder usage
+		// report`, which folded the ledger and so counted every call but this one.
+		...(result.usage !== undefined ? { usage: result.usage } : {}),
+	};
+	if (trigger !== undefined) entry.trigger = trigger;
+	session.appendEntry(entry);
+	return result;
+}
+
+/**
+ * Compose the production chat-loop compaction callback. Errors intentionally
+ * propagate: the chat loop owns activity failure reporting and distinguishes
+ * a thrown read/model/persistence failure from the legitimate null no-op that
+ * `runCompactionFlow` returns for an empty session or an unavailable cut.
+ */
+export function createProductionAutoCompact(
+	session: SessionContract,
+	getSettings: () => ClioSettings,
+	providers: ProvidersContract,
+): (instructions?: string, trigger?: CompactionTrigger) => Promise<CompactResult | null> {
+	return (instructions, trigger) => runCompactionFlow(session, getSettings(), providers, instructions, trigger);
+}
+
+function estimateTokensFromSummary(summary: string): number {
+	// Mirrors the rough byte/4 heuristic the rest of the compaction stack
+	// uses for unmeasured payloads. Kept inline because this is the only
+	// caller; pi-mono's token estimator is provider-specific and we do not
+	// have a model handle at the persistence layer.
+	return Math.max(1, ceilChars(summary.length));
+}
+
+/**
+ * The post-compaction size, on the same scale as `tokensBefore`.
+ *
+ * `tokensBefore` is a whole-prompt figure: `calculateContextTokens` anchors it
+ * on the last assistant call's measured usage, so it carries the system prompt
+ * and tool schemas that compaction never touches. The after figure has to stay
+ * on that scale to be comparable, and the persistence layer has no model handle
+ * to re-measure with, so it is arithmetic on that same scale: drop what stops
+ * being replayed and add the summary that replaces it.
+ *
+ * Estimating it from the rebuilt message list instead reported `tokensBefore`
+ * back unchanged, which is what made /tree render "~16276 -> ~16276 tokens"
+ * beside a footer that said 16276 -> 11008 for the same compaction. That
+ * estimator anchors on the newest assistant usage, and the retained suffix
+ * still holds the assistant message whose usage describes the pre-compaction
+ * prompt. Anchoring on it reports precisely the number compaction removed.
+ */
+function estimateTokensAfterCompaction(entries: ReadonlyArray<SessionEntry>, result: CompactResult): number {
+	let droppedTokens = 0;
+	for (const entry of entries.slice(0, result.firstKeptEntryIndex)) droppedTokens += estimateTokens(entry);
+	const summaryTokens = estimateTokensFromSummary(result.summary);
+	// The summary alone is the floor: a session whose dropped estimate exceeds
+	// the measured anchor must not report a negative or sub-summary context.
+	return Math.max(summaryTokens, result.tokensBefore - droppedTokens + summaryTokens);
+}
+
+/**
+ * Alt+J / Alt+K step the orchestrator through the `scope` list of target
+ * ids or target/model refs. Absent scope is a no-op so unconfigured users
+ * feel nothing.
+ */
+export function advanceScopedTarget(
+	settings: Readonly<ClioSettings>,
+	direction: "forward" | "backward",
+): { target: string; model: string | null } | null {
+	const scope = settings.scope ?? [];
+	if (scope.length === 0) return null;
+	const registry = getRuntimeRegistry();
+	if (registry.list().length === 0) registerBuiltinRuntimes(registry);
+	const filteredScope = scope.filter((entry) => {
+		const [targetId] = entry.split("/");
+		const target = settings.targets.find((e) => e.id === targetId);
+		if (!target) return false;
+		const runtime = registry.get(target.runtime);
+		return runtime !== null && isOrchestratorEligibleRuntime(runtime);
+	});
+	if (filteredScope.length === 0) return null;
+	const activeTarget = settings.orchestrator.target ?? "";
+	const activeModel = settings.orchestrator.model ?? "";
+	const activeCombinedRef = activeTarget.length > 0 && activeModel.length > 0 ? `${activeTarget}/${activeModel}` : "";
+	const idx = filteredScope.findIndex((entry) => entry === activeCombinedRef || entry === activeTarget);
+	const base = idx === -1 ? 0 : idx + (direction === "forward" ? 1 : filteredScope.length - 1);
+	const next = filteredScope[base % filteredScope.length];
+	if (!next) return null;
+	const [targetId, ...modelParts] = next.split("/");
+	if (!targetId) return null;
+	if (modelParts.length > 0) {
+		return { target: targetId, model: modelParts.join("/") };
+	}
+	if (activeTarget === targetId) {
+		return { target: targetId, model: activeModel || null };
+	}
+	const descriptor = settings.targets.find((entry) => entry.id === targetId);
+	return { target: targetId, model: descriptor?.defaultModel ?? null };
+}
+
+export async function bootOrchestrator(options: BootOptions = {}): Promise<BootResult> {
+	const bootStdout = (text: string): void => {
+		if (options.terminalLease) options.terminalLease.writeDiagnostic("stdout", text);
+		else process.stdout.write(text);
+	};
+	const bootStderr = (text: string): void => {
+		if (options.terminalLease) options.terminalLease.writeDiagnostic("stderr", text);
+		else process.stderr.write(text);
+	};
+	const timer = new StartupTimer(
+		options.terminalLease
+			? (phase) => {
+					const line = formatBootTrace(phase);
+					if (line) options.terminalLease?.deferDiagnostic("stderr", line);
+				}
+			: undefined,
+	);
+	const bus = getSharedBus();
+	const termination = getTerminationCoordinator();
+	installBusTracer();
+	termination.installSignalHandlers();
+
+	ensureClioState();
+	timer.mark("install check");
+
+	// A hard-killed coordinator cannot run domain drains. Reconcile its compete
+	// process leases before dispatch startup scans the ledger, so abandoned rows
+	// observe dead workers and no stale worktree can be reused by this boot.
+	if (isGitRepository(process.cwd())) {
+		try {
+			const pendingGate = readPendingGateDecisions();
+			const pendingCompeteGroups = new Set<string>();
+			for (const handle of pendingGate.records) {
+				if (handle.record.kind === "output" && handle.record.topology === "compete") {
+					pendingCompeteGroups.add(handle.record.group);
+				} else if (handle.record.kind === "decision" && handle.record.decision.topology === "compete") {
+					pendingCompeteGroups.add(handle.record.decision.group);
+				}
+			}
+			const durableDecisions = readGateDecisionArtifacts();
+			const confirmedGroups = new Set(
+				durableDecisions
+					.filter(
+						({ artifact }) =>
+							artifact.topology === "compete" &&
+							(artifact.outcome === "operator-confirmed" || artifact.outcome === "full-auto-applied"),
+					)
+					.map(({ artifact }) => artifact.group),
+			);
+			for (const { artifact } of durableDecisions) {
+				if (artifact.topology === "compete" && artifact.outcome === "winner" && !confirmedGroups.has(artifact.group)) {
+					pendingCompeteGroups.add(artifact.group);
+				}
+			}
+			const recovery = recoverCleanupReadyCompeteGroups(process.cwd(), {
+				preserveActiveGroups: pendingCompeteGroups,
+				preserveAllActive: pendingGate.errors.length > 0,
+			});
+			for (const failure of recovery.failed) {
+				bootStderr(`[dispatch] compete recovery preserved ${failure.group}: ${failure.message}\n`);
+			}
+		} catch (err) {
+			bootStderr(`[dispatch] compete recovery failed closed: ${err instanceof Error ? err.message : String(err)}\n`);
+		}
+	}
+
+	let effectiveSettingsForDispatch: (() => Readonly<ClioSettings>) | null = null;
+	let protectedArtifactStateForDispatch: (() => ProtectedArtifactState) | null = null;
+
+	const result = await loadDomains(
+		[
+			options.startupSettings ? createConfigDomainModule(options.startupSettings) : ConfigDomainModule,
+			ExtensionsDomainModule,
+			InteropDomainModule,
+			createResourcesDomainModule({
+				reservedPromptNames: new Set(BUILTIN_SLASH_COMMANDS.map((entry) => entry.name)),
+				skills: () => ({
+					disableDiscovery: options.noSkills === true || options.headless?.noSkills === true,
+					...(options.skillPaths && options.skillPaths.length > 0
+						? { explicitSkillPaths: options.skillPaths }
+						: options.headless?.skillPaths && options.headless.skillPaths.length > 0
+							? { explicitSkillPaths: options.headless.skillPaths }
+							: {}),
+				}),
+			}),
+			ShareDomainModule,
+			createContextDomainModule({ noContextFiles: options.noContextFiles === true }),
+			ProvidersDomainModule,
+			SafetyDomainModule,
+			createPromptsDomainModule({
+				noContextFiles: options.noContextFiles === true,
+			}),
+			AgentsDomainModule,
+			MiddlewareDomainModule,
+			SessionDomainModule,
+			ObservabilityDomainModule,
+			SchedulingDomainModule,
+			// Dispatch resolves worker targets through the session's effective
+			// settings view once it exists (assigned below, after the config
+			// contract loads); until then it falls back to the shared snapshot.
+			createDispatchDomainModule({
+				getSettings: () => effectiveSettingsForDispatch?.(),
+				getProtectedArtifactState: () => protectedArtifactStateForDispatch?.() ?? { artifacts: [] },
+				autonomyOverride: options.headless?.autonomy !== undefined,
+			}),
+			LifecycleDomainModule,
+		],
+		{ diagnostic: bootStderr },
+	);
+	timer.mark(`domains loaded (${result.loaded.length})`);
+
+	const dispatch = result.getContract<DispatchContract>("dispatch");
+	if (dispatch) {
+		termination.onDrain(async () => {
+			await dispatch.drain();
+		});
+	}
+	termination.onPersist(async () => {
+		await result.stop();
+	});
+
+	bus.emit(BusChannels.SessionStart, { at: Date.now() });
+	timer.mark("session_start fired");
+
+	const acpMode = options.acp !== undefined;
+	const interactive = !options.headless && !acpMode && process.env.CLIO_CODER_INTERACTIVE === "1";
+	if (!interactive && !options.headless && !acpMode) {
+		bootStdout(buildBanner());
+		if (process.env.CLIO_CODER_TIMING === "1") bootStdout(`${timer.report()}\n`);
+	}
+
+	const config = result.getContract<ConfigContract>("config");
+	const providers = result.getContract<ProvidersContract>("providers");
+	timer.mark("providers resolved");
+
+	if (options.apiKey) {
+		if (!providers) {
+			bootStderr("Clio Coder: --api-key supplied but providers domain unavailable; ignoring.\n");
+		} else {
+			const settingsNow = applyHeadlessSettingsOverlay(config?.get() ?? readSettings(), options.headless);
+			const activeTargetId = settingsNow.orchestrator?.target;
+			const target = resolveTarget(providers, activeTargetId);
+			const runtime = target ? providers.getRuntime(target.runtime) : null;
+			if (target && runtime) {
+				providers.auth.setRuntimeOverrideForTarget(target, runtime, options.apiKey);
+			} else {
+				bootStderr("Clio Coder: --api-key supplied but no active orchestrator target is configured; ignoring.\n");
+			}
+		}
+	}
+
+	if (!interactive && !options.headless && !acpMode) {
+		bootStdout(`${chalk.dim("  (non-interactive boot. pass CLIO_CODER_INTERACTIVE=1 to launch the TUI.)")}\n`);
+		await termination.shutdown(0);
+		return { exitCode: 0, bootTimeMs: timer.snapshot().totalMs };
+	}
+
+	const middleware = result.getContract<MiddlewareContract>("middleware");
+	const observability = result.getContract<ObservabilityContract>("observability");
+	const safety = result.getContract<SafetyContract>("safety");
+	const session = result.getContract<SessionContract>("session");
+	const prompts = result.getContract<PromptsContract>("prompts");
+	const agents = result.getContract<AgentsContract>("agents");
+	const resources = result.getContract<ResourcesContract>("resources");
+	const extensions = result.getContract<ExtensionsContract>("extensions");
+	const share = result.getContract<ShareContract>("share");
+	const contextDomain = result.getContract<ContextContract>("context");
+	const interop = result.getContract<InteropContract>("interop");
+	// Boot detection resolves paths only: no `--version` subprocess and no skill
+	// walk on the boot path. The hint is gated on `interactive`, which is already
+	// false under headless and ACP.
+	const interopReport = interactive && interop ? await interop.detect({ cwd: process.cwd() }) : null;
+	const initialNotices = interactive ? [...(contextDomain?.startupHints() ?? [])] : [];
+	// Once per version, interactive only: headless and ACP have no operator at
+	// the keyboard to tell, and the record is left unclaimed for the boot that does.
+	const upgrade = interactive ? takeUpgradeNotice() : null;
+	if (upgrade !== null) initialNotices.push(describeUpgradeNotice(upgrade));
+	if (interop && interopReport) {
+		const interopHint = interop.bootHint(interopReport);
+		if (interopHint !== null) initialNotices.push(interopHint);
+	}
+	if (!providers || !dispatch || !observability || !safety || !middleware) {
+		bootStderr(
+			"Clio Coder: chat mode requires safety + middleware + providers + dispatch + observability contracts; aborting.\n",
+		);
+		await termination.shutdown(1);
+		return { exitCode: 1, bootTimeMs: timer.snapshot().totalMs };
+	}
+
+	// A headless `--session`/`--continue` resolves here, where the session
+	// domain exists. It is a hard requirement rather than a hint: a caller that
+	// asked to continue a conversation must not receive an answer written
+	// without that conversation's history.
+	const requestedResume = options.headless?.resumeSession;
+	let headlessResumeFailure: string | null = null;
+	let resolvedResumeId: string | undefined;
+	if (requestedResume !== undefined) {
+		if (!session) {
+			headlessResumeFailure = "session continuation requires the session domain, which is not loaded";
+		} else if (requestedResume.kind === "id") {
+			resolvedResumeId = requestedResume.id;
+		} else {
+			const latest = session.history()[0];
+			if (latest === undefined) headlessResumeFailure = `no previous session recorded for ${process.cwd()}`;
+			else resolvedResumeId = latest.id;
+		}
+	}
+	const resumeId = resolvedResumeId ?? process.env.CLIO_CODER_RESUME_SESSION_ID?.trim();
+	let resumedSessionAtBoot = false;
+	if (resumeId && session && headlessResumeFailure === null) {
+		try {
+			session.resume(resumeId);
+			resumedSessionAtBoot = true;
+		} catch (err) {
+			const detail = err instanceof Error ? err.message : String(err);
+			if (requestedResume !== undefined) headlessResumeFailure = `failed to resume session ${resumeId}: ${detail}`;
+			else bootStderr(`Clio Coder: failed to resume session ${resumeId}: ${detail}\n`);
+		}
+	}
+	Reflect.deleteProperty(process.env, "CLIO_CODER_RESUME_SESSION_ID");
+	if (headlessResumeFailure !== null) {
+		bootStderr(`clio-coder run: ${headlessResumeFailure}\n`);
+		await termination.shutdown(2);
+		return { exitCode: 2, bootTimeMs: timer.snapshot().totalMs };
+	}
+
+	// Hook diagnostics ride the typed bus. The domain loader constructed the
+	// bundle with the stderr default; swap in a sink that publishes
+	// middleware.hookFailed (the interactive warn notice consumes it) and keep
+	// stderr for non-interactive runs, which have no notice subscriber.
+	middleware.setDiagnosticSink((diagnostic) => {
+		bus.emit(BusChannels.MiddlewareHookFailed, {
+			kind: diagnostic.kind,
+			registrationId: diagnostic.registrationId,
+			hook: diagnostic.hook,
+			at: Date.now(),
+			...(diagnostic.kind === "hook_failed"
+				? { message: diagnostic.message }
+				: {
+						elapsedMs: diagnostic.elapsedMs,
+						budgetMs: diagnostic.budgetMs,
+						steadyStateWarn: diagnostic.steadyStateWarn,
+						p50Ms: diagnostic.stats.p50Ms,
+						p95Ms: diagnostic.stats.p95Ms,
+						overCount: diagnostic.stats.overCount,
+						windowSamples: diagnostic.stats.window,
+					}),
+		});
+		if (!interactive) writeMiddlewareDiagnosticToStderr(diagnostic);
+	});
+
+	// Residency and runtime notices (model swaps, double residency, VRAM
+	// stress) reach the operator through the interactive notice renderer; a
+	// headless run has no subscriber, so mirror them to stderr there. Silent
+	// model swaps are exactly the failure mode the residency policy forbids.
+	if (!interactive) {
+		bus.on(BusChannels.RuntimeNotice, (payload: unknown) => {
+			const notice = payload as { level?: string; message?: string } | undefined;
+			if (typeof notice?.message !== "string") return;
+			process.stderr.write(`[clio:runtime] ${notice.level ?? "info"}: ${notice.message}\n`);
+		});
+	}
+
+	// Guardrail policy (tool-call budgets, tool byte caps, dispatch ledger cap)
+	// resolves settings-first with env as emergency override; install the
+	// settings section before any guard registration or tool reads it.
+	configureGuardrails((config?.get() ?? readSettings()).guardrails);
+
+	// Guard registrations on the middleware contract, in order: loop guard,
+	// protected artifacts (last among guards so it absorbs protect_path effects
+	// from everything before it), dispatch dedup. Workers register their own
+	// loop guard and protected-artifacts instances inside their subprocess in
+	// worker-runtime.ts; the orchestrator instances carry the bus and the
+	// session persistence sink.
+	middleware.registerHook(
+		createLoopGuardRegistration({
+			safety,
+			bus,
+			turnBlockBudget: INTERACTIVE_LOOP_BLOCK_BUDGET,
+			turnToolCallBudget: readOrchTurnToolCallBudget(),
+			// Interactive/headless/ACP all share this orchestrator guard: at the
+			// block budget, lock tools for the rest of the turn so the model
+			// answers from what it gathered instead of hard-cancelling a turn that
+			// may already hold the answer. The bounded backstop still cancels a
+			// model that keeps calling tools.
+			turnSynthesisLockout: true,
+		}),
+	);
+	let initialProtectedArtifactState: ProtectedArtifactState | undefined;
+	let initialProtectionReadError: string | null = null;
+	if (session) {
+		try {
+			initialProtectedArtifactState = protectedArtifactStateForCurrentSession(session);
+		} catch (error) {
+			initialProtectionReadError = error instanceof Error ? error.message : String(error);
+		}
+	}
+	const protectedArtifactsGuard = createProtectedArtifactsRegistration({
+		...(initialProtectedArtifactState !== undefined ? { initialState: initialProtectedArtifactState } : {}),
+		onProtect: (event) => appendProtectedArtifactRegistryEvent(session, event),
+		onDurabilityFailure: (health) => {
+			bus.emit(BusChannels.MiddlewareHookFailed, {
+				kind: "hook_failed",
+				registrationId: "guard.protected-artifacts",
+				hook: "before_tool",
+				at: Date.now(),
+				message: health.reason,
+			});
+		},
+	});
+	if (initialProtectionReadError !== null) {
+		protectedArtifactsGuard.markDegraded(
+			`initial session protection history could not be read: ${initialProtectionReadError}`,
+		);
+	}
+	protectedArtifactStateForDispatch = () => {
+		const health = protectedArtifactsGuard.health();
+		if (health.kind === "degraded") {
+			throw new Error(`dispatch: protected artifact durability degraded: ${health.reason}`);
+		}
+		return protectedArtifactsGuard.state();
+	};
+	middleware.registerHook(protectedArtifactsGuard);
+	middleware.registerHook(createDispatchDedupRegistration());
+	// Observers run after the guards; they emit no effects and their sinks are
+	// best-effort (session ledger, codewiki refresh).
+	middleware.registerHook(
+		createSkillActivationObserver((activation) => appendSkillActivationRegistryEvent(session, activation)),
+	);
+	// First-turn skills reminder: user-message-visible text is the one channel
+	// the battery-tested local models act on; suggestion protocol only, the
+	// operator load gate is untouched.
+	if (resources) {
+		middleware.registerHook(
+			createSkillsReminderRegistration({
+				countModelVisibleSkills: () => modelVisibleSkills(resources.skills(process.cwd()).items).length,
+				// Same lookup context(scope="skills") lists under its Marketplace
+				// heading, minus what is already installed, so the count the
+				// reminder quotes is the count the listing will show.
+				countInstallableSkills: () => {
+					const installed = new Set(resources.skills(process.cwd()).items.map((skill) => skill.name));
+					return discoverMarketplaceSkills({ cwd: process.cwd() }).skills.filter((skill) => !installed.has(skill.name))
+						.length;
+				},
+			}),
+		);
+	}
+	// Task-board reminder: same user-message-visible channel, fired once per
+	// session when a request literally enumerates three or more steps. The
+	// static routing line and tasks hint ask for the same board; battery-tested
+	// local models only comply when the instruction rides the user message.
+	middleware.registerHook(createTaskBoardReminderRegistration());
+	const taskMemoryBank = new TaskMemoryBank();
+	let taskMemorySessionId = session?.current()?.id ?? null;
+	const ensureTaskMemorySession = (): void => {
+		const currentSessionId = session?.current()?.id ?? null;
+		if (currentSessionId === taskMemorySessionId) return;
+		taskMemoryBank.clear();
+		taskMemorySessionId = currentSessionId;
+	};
+	const memorySettings = (config?.get() ?? readSettings()).memory.intervention;
+	// Bound late: the registration is built here, but the buffer a deferred
+	// reminder lands in belongs to the chat loop that has not been composed yet.
+	let deferredMemoryReminderSink: ((message: string) => void) | null = null;
+	// Content-bearing, so it exists only when the operator named a file. The
+	// telemetry row says which silence happened; this says what the model wrote.
+	const memoryTracePath = taskMemoryTracePath();
+	const memoryTrace = memoryTracePath === null ? null : createTaskMemoryTrace(memoryTracePath);
+	const memoryIntervention = createMemoryInterventionRegistration({
+		bank: taskMemoryBank,
+		telemetry: createTaskMemoryTelemetrySink(),
+		...(memoryTrace === null ? {} : { onEnvelope: (envelope) => memoryTrace.record(envelope) }),
+		// A headless run submits no further turn, so a detached step could only
+		// finish after the process that would have read it has exited.
+		deliversDeferredReminders: options.headless === undefined,
+		onDeferredReminder: (message) => deferredMemoryReminderSink?.(message),
+		getSettings: () => {
+			ensureTaskMemorySession();
+			return effectiveSettingsForDispatch?.().memory.intervention ?? memorySettings;
+		},
+		getModelClient: () => {
+			const settings = effectiveSettingsForDispatch?.();
+			return settings === undefined
+				? null
+				: createBackgroundMemoryModelClient(providers, settings, settings.memory.intervention.timeoutMs);
+		},
+	});
+	middleware.registerHook(memoryIntervention);
+	const unsubscribeMemoryLoop = bus.on(BusChannels.LoopBlocked, () => memoryIntervention.signalLoop());
+	termination.onDrain(() => unsubscribeMemoryLoop());
+	if (contextDomain) {
+		middleware.registerHook(createFileMutationObserver(({ paths }) => contextDomain.noteFileChanges(paths)));
+	}
+	// User-defined hooks: extensions and the project (.clio-coder/hooks.yaml,
+	// .clio-coder/hooks.local.yaml) declare a conservative, receipted hook set on the
+	// same effect machinery. They register after the guards, so safety stays
+	// authoritative: a hook may add effects (including request block_tool) but
+	// cannot grant a permission safety would deny. Loading is best-effort.
+	const hookReceiptLog = createHookReceiptLog({ persistPath: join(clioStateDir(), "hook-receipts.json") });
+	const extensionHookRoots: ExtensionHookRoot[] = (extensions?.list(process.cwd()) ?? [])
+		.filter((ext) => ext.enabled && ext.effective)
+		.map((ext) => ({ id: ext.id, rootPath: ext.rootPath }));
+	const userHooks = installUserHooks({
+		cwd: process.cwd(),
+		extensions: extensionHookRoots,
+		registerHook: (registration) => middleware.registerHook(registration),
+		recordReceipt: (receipt) => hookReceiptLog.record(receipt),
+	});
+	if (!interactive) {
+		for (const issue of userHooks.fileIssues) {
+			process.stderr.write(`[clio:hooks] ${issue.message}\n`);
+		}
+		for (const issue of userHooks.issues) {
+			process.stderr.write(`[clio:hooks] ${issue.source.sourcePath}#${issue.index}: ${issue.issues.join("; ")}\n`);
+		}
+	}
+	termination.onDrain(() => hookReceiptLog.flush());
+	// Autonomy is hot-reloaded for interactive and headless admissions. ACP
+	// server prompts use the snapshot captured at session/new.
+	let activeAcpSessionAutonomy: AutonomyLevel | null = null;
+	// The one effective-autonomy resolution. Every admission surface (registry
+	// admission, dispatch plan provenance, ACP session snapshot) resolves
+	// through these two functions so a fallback added to one surface cannot
+	// silently skip another.
+	const resolveBaselineAutonomy = (): AutonomyLevel =>
+		effectiveSettingsForDispatch?.().autonomy ??
+		options.headless?.autonomy ??
+		(config?.get() ?? readSettings()).autonomy ??
+		"auto-edit";
+	const resolveEffectiveAutonomy = (): AutonomyLevel => activeAcpSessionAutonomy ?? resolveBaselineAutonomy();
+	const middlewareToolChoice = createMiddlewareToolChoiceControl();
+	const toolRegistry = createRegistry({
+		safety,
+		middleware,
+		onMiddlewareEffects: (effects) => middlewareToolChoice.apply(effects),
+		autonomy: resolveEffectiveAutonomy,
+	});
+	const mainPermissionOrigin = acpMode ? "acp-server" : "main";
+	toolRegistry.onPermissionRequired((call, decision, meta) => {
+		bus.emit(BusChannels.PermissionRequested, {
+			tool: call.tool,
+			actionClass: decision.classification.actionClass,
+			requestId: meta.requestId,
+			origin: mainPermissionOrigin,
+			axis: meta.axis,
+			...(decision.kind === "ask" ? { rejection: decision.rejection } : {}),
+			...(decision.policy?.ruleId !== undefined ? { ruleId: decision.policy.ruleId } : {}),
+			...(decision.policy?.policySource !== undefined ? { policySource: decision.policy.policySource } : {}),
+			...(decision.policy?.reasonCode !== undefined ? { reasonCode: decision.policy.reasonCode } : {}),
+		});
+	});
+	let askUserHandler: AskUserHandler | null = null;
+	// ask_user is interactive-only by design: it is a human interview tool and
+	// headless/ACP surfaces have no operator to interview, so the tool is not
+	// registered there at all (documented in `clio-coder run --help`). Skills that
+	// interview fall back to their stated defaults when the tool is absent.
+	const askUserBridge: AskUserHandler = async (questions, invokeOptions) =>
+		askUserHandler ? await askUserHandler(questions, invokeOptions) : cancelledAskUserResult();
+	const userTasks = createUserTasksStore({ cwd: process.cwd() });
+	// One task board per orchestrator: the tasks tool mutates it, the turn-end
+	// open-tasks nudge reads it, and the footer/overlay render it. Keyed on the
+	// current session id so resume/fork/new refolds it from taskLedger entries.
+	// Folding through filterEntriesToActivePath (not the raw file) is what
+	// keeps a /resume from picking up whichever branch happened to write its
+	// taskLedger entry last in file order; readEntries here used to skip that
+	// filter entirely (issue #94).
+	const taskBoard = createTaskBoardStore({
+		getSessionId: () => session?.current()?.id ?? null,
+		readEntries: () => {
+			const meta = session?.current();
+			if (!meta) return [];
+			const leafTurnId = session?.tree(meta.id).leafId ?? undefined;
+			return filterEntriesToActivePath(readSessionEntriesForCompact(meta.id), leafTurnId);
+		},
+		appendEntry: (entry) => {
+			session?.appendEntry(entry);
+		},
+	});
+	// Prime the projection once at composition. Interactive repaint paths use
+	// cachedSnapshot() below, so a first paint can never become a ledger read.
+	taskBoard.snapshot();
+	const decisionBoard = createDecisionBoardStore({
+		getSessionId: () => session?.current()?.id ?? null,
+		readEntries: () => {
+			const meta = session?.current();
+			if (!meta) return [];
+			const leafTurnId = session?.tree(meta.id).leafId ?? undefined;
+			return filterEntriesToActivePath(readSessionEntriesForCompact(meta.id), leafTurnId);
+		},
+		getActiveLeafTurnId: () => {
+			const meta = session?.current();
+			return meta ? (session?.tree(meta.id).leafId ?? null) : null;
+		},
+		appendEntry: (entry) => {
+			if (!session) throw new Error("decision board: no session ledger is available");
+			session.appendEntry(entry);
+		},
+	});
+	// getSessionId alone never notices a /tree switch: it moves the active
+	// append point inside the same session, so the id-keyed cache above kept
+	// showing the abandoned branch's board (issue #94). SessionTurnSwitched is
+	// the signal that switch actually happened; invalidate() forces the next
+	// read to refold from the now-current leaf.
+	bus.on(BusChannels.SessionTurnSwitched, () => {
+		taskBoard.invalidate();
+		// The tree switch is the I/O boundary: refold eagerly here so the 250-ms
+		// island ticker remains a cache-only consumer after changing branches.
+		taskBoard.snapshot();
+		decisionBoard.invalidate();
+	});
+	// Link in-flight dispatch runs to the live board via the ledger's
+	// activeRunIds field: a run is tracked from the moment its child process is
+	// live until it finalizes either way. attach/detach are no-ops when no board
+	// is declared, so an ambient dispatch never forces a board into existence.
+	bus.on(BusChannels.DispatchStarted, (payload) => {
+		if (typeof payload?.runId === "string") taskBoard.attachRun(payload.runId);
+	});
+	bus.on(BusChannels.DispatchCompleted, (payload) => {
+		if (typeof payload?.runId !== "string") return;
+		taskBoard.detachRun(payload.runId);
+		foldDispatchSkillActivations(session, payload);
+	});
+	bus.on(BusChannels.DispatchFailed, (payload) => {
+		if (typeof payload?.runId !== "string") return;
+		taskBoard.detachRun(payload.runId);
+		// A run that failed after loading a skill is the case the operator most
+		// needs the provenance for, and the receipt already carried it here.
+		foldDispatchSkillActivations(session, payload);
+	});
+	// Operator-initiated backgrounding is a TUI affordance: the registry is the
+	// one object the dispatch tool and the keypress both hold.
+	const dispatchBackground = createDispatchBackgroundRegistry();
+	registerAllTools(toolRegistry, {
+		...(session ? { session } : {}),
+		taskBoard,
+		userTasks,
+		dispatch,
+		bus,
+		...(interactive ? { askUser: askUserBridge } : {}),
+		...(agents ? { getAgentCatalog: () => renderAgentCatalogSectionsFromSpecs(agents.listSpecs()).stable } : {}),
+		...(agents ? { getAgentSpecs: () => agents.listSpecs() } : {}),
+		...(agents ? { getAgentRoleFacts: agentRoleFactsResolver((id: string) => agents.getSpec(id)) } : {}),
+		// Same effective-autonomy resolution the registry admission uses, so plan
+		// provenance and compete winner handling agree with the approval surface.
+		getAutonomy: resolveEffectiveAutonomy,
+		...(interactive ? { dispatchBackground } : {}),
+		getCostCeilingUsd: () => result.getContract<SchedulingContract>("scheduling")?.ceilingUsd() ?? 0,
+		getSkillLoaderOptions: () => ({
+			trustProjectCompatRoots: config?.get().skills.trustProjectCompatRoots === true,
+			disableDiscovery: options.noSkills === true || options.headless?.noSkills === true,
+			...(options.skillPaths && options.skillPaths.length > 0
+				? { explicitSkillPaths: options.skillPaths }
+				: options.headless?.skillPaths && options.headless.skillPaths.length > 0
+					? { explicitSkillPaths: options.headless.skillPaths }
+					: {}),
+		}),
+	});
+
+	// Live routing is owned by this process. Seed it once from saved settings
+	// (with any headless CLI overrides baked in); from here on every consumer
+	// reads the effective view — shared snapshot + session routing overlay — so
+	// another process writing settings.yaml can update defaults and the
+	// target catalog but never redirect this session's routing.
+	const sessionRouting = seedSessionRouting(
+		applyHeadlessSettingsOverlay(config?.get() ?? readSettings(), options.headless),
+	);
+	// Non-routing settings a session changed "for this session only" via the
+	// /settings overlay. Layered under the routing overlay in the effective
+	// view, so the live session reflects them immediately while settings.yaml
+	// (the global default for new sessions) stays untouched until the operator
+	// chooses to save globally.
+	const sessionOverrides: SessionOverrides = new Map(
+		options.headless?.autonomy === undefined ? [] : [["autonomy", options.headless.autonomy]],
+	);
+	// The effective view is derived by deep-cloning the saved snapshot, and it
+	// is read on the tool-admission hot path (every call resolves autonomy
+	// through it). Memoize on the two things it depends on: the config
+	// domain's snapshot identity (swapped wholesale on every update and
+	// external hot reload) and a generation counter bumped by every session
+	// routing/override mutation below. Without this, one tool call cost a full
+	// settings structuredClone.
+	let sessionStateGeneration = 0;
+	let cachedSettingsBase: Readonly<ClioSettings> | null = null;
+	let cachedSettingsGeneration = -1;
+	let cachedSettingsView: ClioSettings | null = null;
+	const bumpSessionState = (): void => {
+		sessionStateGeneration += 1;
+		// Child-process seams read the effective setting from the environment,
+		// and the session view may override the saved value.
+		setGitCommitAttributionEnabled(getCurrentSettings().attribution.gitCommits);
+	};
+	const getCurrentSettings = (): ClioSettings => {
+		// Recents live in the data dir (core/recent-models.ts), never in
+		// settings.yaml; consumers that need them call listRecentModels
+		// directly, so an Alt+L pick in another session does not churn the
+		// config watcher here.
+		const base = config?.get();
+		// No config domain (unit tests, degraded boot): readSettings() returns a
+		// fresh object every call, so there is nothing stable to key a cache on.
+		if (base === undefined) return applySessionRouting(applyOverrides(readSettings(), sessionOverrides), sessionRouting);
+		if (
+			cachedSettingsView !== null &&
+			cachedSettingsBase === base &&
+			cachedSettingsGeneration === sessionStateGeneration
+		) {
+			return cachedSettingsView;
+		}
+		const view = applySessionRouting(applyOverrides(base, sessionOverrides), sessionRouting);
+		cachedSettingsBase = base;
+		cachedSettingsGeneration = sessionStateGeneration;
+		cachedSettingsView = view;
+		return view;
+	};
+	effectiveSettingsForDispatch = getCurrentSettings;
+	bumpSessionState();
+	// The config bundle publishes the saved value on reload; a session-scoped
+	// override must win, so re-derive from the session view after it.
+	const unsubscribeCommitAttributionSync = bus.on(BusChannels.ConfigHotReload, () => bumpSessionState());
+	termination.onDrain(() => unsubscribeCommitAttributionSync());
+	const getTaskMemorySeedOffer = (): { source: string; count: number } | null => {
+		return taskMemoryHandoffSeedOffer(process.cwd(), getCurrentSettings().memory.intervention.enabled);
+	};
+	const seedCurrentTaskMemoryFromHandoff = () => {
+		ensureTaskMemorySession();
+		return seedTaskMemoryFromNewestHandoff(
+			taskMemoryBank,
+			process.cwd(),
+			getCurrentSettings().memory.intervention.enabled,
+		);
+	};
+	if (resumedSessionAtBoot) {
+		const offer = getTaskMemorySeedOffer();
+		if (offer && offer.count > 0) {
+			initialNotices.push(
+				`task memory: ${offer.count} handoff entr${offer.count === 1 ? "y" : "ies"} available from ${offer.source}; run /memory seed to import`,
+			);
+		}
+	}
+	for (const warning of agentRoleToolWarnings(providers, getCurrentSettings())) {
+		if (interactive) initialNotices.push(warning);
+		else process.stderr.write(`${warning}\n`);
+	}
+	// Residency protection follows the live effective settings: the models the
+	// operator's config references (orchestrator, worker default/profiles,
+	// target defaults) may never be evicted by another Clio stream, and a
+	// routing change updates the set on the next read.
+	setProtectedModelsProvider(() => protectedResidencyModels(getCurrentSettings()));
+
+	const validatedKeybindings = validateKeybindings((config?.get() ?? readSettings()).keybindings ?? {});
+	const invalidBindings = validatedKeybindings.invalid;
+	if (invalidBindings.length > 0) {
+		const notice = formatInvalidKeybindingNotice(invalidBindings);
+		if (interactive) initialNotices.push(notice);
+		else process.stderr.write(notice);
+	}
+	const platformWarnings = process.stdin.isTTY
+		? detectPlatformKeybindingWarnings(validatedKeybindings.valid, detectTerminalKeySupport(process.env))
+		: [];
+	if (platformWarnings.length > 0) {
+		const notice = formatPlatformKeybindingNotice(platformWarnings);
+		if (interactive) initialNotices.push(notice);
+		else process.stderr.write(notice);
+	}
+	/**
+	 * Locked read-modify-write of saved settings. Routes through the config
+	 * contract (which refreshes its snapshot and dispatches change events) when
+	 * available, else straight through core updateSettings. Either way the
+	 * mutator runs against the freshest on-disk state under the advisory
+	 * settings lock, so two processes saving defaults at the same time cannot
+	 * interleave and drop each other's patches.
+	 */
+	const persistSavedMutation = (mutator: SettingsMutator): void => {
+		if (config?.update) config.update(mutator);
+		else updateSettings(mutator);
+		bumpSessionState();
+	};
+	/**
+	 * Apply a routing change with one consistent scope: it takes effect in this
+	 * session immediately and writes through to saved settings as the default
+	 * for future sessions. Only the patched fields hit the file, so concurrent
+	 * sessions cannot clobber each other's saved defaults wholesale.
+	 */
+	const updateSessionRouting = (patch: RoutingPatch, mutateSaved?: (saved: ClioSettings) => void): void => {
+		applyRoutingPatch(sessionRouting, patch);
+		bumpSessionState();
+		persistSavedMutation((saved) => {
+			mergeRoutingPatchIntoSettings(saved, patch);
+			mutateSaved?.(saved);
+		});
+	};
+	const readAcpSafeSettings = (): AcpSafeSettingsSnapshot => {
+		const settings = getCurrentSettings();
+		return {
+			target: settings.orchestrator.target,
+			model: settings.orchestrator.model,
+			thinkingLevel: settings.orchestrator.thinkingLevel ?? "off",
+			autonomy: settings.autonomy,
+		};
+	};
+	/**
+	 * ACP safe settings are one atomic persisted mutation followed by infallible
+	 * in-process routing assignment. Persisting first avoids reporting a live
+	 * route that failed to become the future-session default.
+	 */
+	const commitAcpSafeSettings = (patch: AcpSafeSettingsPatch): AcpSafeSettingsSnapshot => {
+		const orchestrator: NonNullable<RoutingPatch["orchestrator"]> = {};
+		if (patch["orchestrator.target"] !== undefined) orchestrator.target = patch["orchestrator.target"];
+		if (patch["orchestrator.model"] !== undefined) orchestrator.model = patch["orchestrator.model"];
+		if (patch["orchestrator.thinkingLevel"] !== undefined) {
+			orchestrator.thinkingLevel = patch["orchestrator.thinkingLevel"];
+		}
+		const routingPatch: RoutingPatch | null = Object.keys(orchestrator).length > 0 ? { orchestrator } : null;
+		persistSavedMutation((saved) => {
+			if (routingPatch !== null) mergeRoutingPatchIntoSettings(saved, routingPatch);
+			if (patch.autonomy !== undefined) saved.autonomy = patch.autonomy;
+		});
+		if (routingPatch !== null) {
+			applyRoutingPatch(sessionRouting, routingPatch);
+			bumpSessionState();
+		}
+		return readAcpSafeSettings();
+	};
+	/**
+	 * Persist a whole-settings blob coming from the effective view (the
+	 * /settings overlay, favorites toggles). Routing edits in the blob are
+	 * absorbed into the session state and written through; everything else is
+	 * persisted without leaking this session's routing into the saved defaults.
+	 */
+	const applySettingsBlob = (next: ClioSettings): void => {
+		const patch = diffRouting(getCurrentSettings(), next);
+		if (patch) {
+			applyRoutingPatch(sessionRouting, patch);
+			bumpSessionState();
+		}
+		persistSavedMutation((fresh) => {
+			const persisted = structuredClone(next);
+			restoreRoutingFields(persisted, fresh);
+			// A whole-blob write (providers, favorites) must not globalize a
+			// session-only override: restore every overridden leaf from the
+			// fresh file so it stays session-local until explicitly saved.
+			for (const path of sessionOverrides.keys()) setAtPath(persisted, path, getAtPath(fresh, path));
+			if (patch) mergeRoutingPatchIntoSettings(persisted, patch);
+			return persisted;
+		});
+	};
+	/**
+	 * Commit a single /settings edit, keyed by its config-path id. `next` is the
+	 * effective view with the one leaf already changed.
+	 *   - scope "session": apply live only. Routing ids feed the routing state;
+	 *     every other id becomes a session override. settings.yaml is untouched.
+	 *   - scope "global": apply live and persist just that leaf as the new
+	 *     default, clearing any prior session override for it.
+	 * Restart-required ids (budget.concurrency, runtimePlugins,
+	 * terminal.tuiMode, terminal.fullscreenScrollbar) cannot apply
+	 * live, so the overlay only offers "global" for them; the file write is what
+	 * a later restart picks up.
+	 */
+	const commitSetting = (id: string, next: ClioSettings, scope: "session" | "global"): void => {
+		if (isRoutingPath(id)) {
+			// Build the patch from `next` keyed by the edited id, not by diffing
+			// against the live view: a prior session-only apply already moved the
+			// routing state, so a diff would be empty and the global save would
+			// silently no-op. Only the touched fields are persisted, so concurrent
+			// sessions never clobber each other's saved routing.
+			const patch = routingPatchForId(id, next);
+			if (!patch) return;
+			applyRoutingPatch(sessionRouting, patch);
+			bumpSessionState();
+			if (scope === "global") persistSavedMutation((saved) => mergeRoutingPatchIntoSettings(saved, patch));
+			return;
+		}
+		const value = getAtPath(next, id);
+		if (scope === "session") {
+			sessionOverrides.set(id, value);
+			bumpSessionState();
+			return;
+		}
+		sessionOverrides.delete(id);
+		bumpSessionState();
+		persistSavedMutation((saved) => setAtPath(saved, id, value));
+	};
+	/** Alt+J / Alt+K: step this session's orchestrator through the scope list. */
+	const cycleScopedSession = (direction: "forward" | "backward"): boolean => {
+		const next = advanceScopedTarget(getCurrentSettings(), direction);
+		if (!next) return false;
+		updateSessionRouting({ orchestrator: { target: next.target, model: next.model } });
+		return true;
+	};
+
+	const readCurrentSessionEntries = (): ReadonlyArray<SessionEntry> => {
+		if (session === undefined) return [];
+		const meta = session.current();
+		if (!meta) return [];
+		reconcilePendingProtectedArtifacts(session);
+		return readSessionEntriesForCompact(meta.id);
+	};
+
+	// turn_end assessors, fired by the chat-loop when the final assistant
+	// message of a run lands. Tool-prose first so its hard-block interruption
+	// precedes the finish-contract advisory in effect order.
+	middleware.registerHook(createToolProseRegistration());
+	middleware.registerHook(createTaskNudgeRegistration({ getBoard: () => taskBoard.snapshot() }));
+	middleware.registerHook(createReadOnlyExplorationNudgeRegistration());
+	middleware.registerHook(createUnbackedWorkerClaimRegistration());
+	middleware.registerHook(
+		createDetachedDispatchNudgeRegistration({ getOpenBatches: () => openDetachedBatchViews(dispatch) }),
+	);
+	if (session) {
+		middleware.registerHook(
+			createFinishContractRegistration({
+				// Tail-scoped: the contract only needs the last-user-message window, so
+				// it parses a bounded ledger tail per turn_end rather than the whole
+				// file (which grows unbounded with session length).
+				readSessionEntries: () => {
+					const meta = session.current();
+					return meta ? readRecentSessionEntriesForContract(meta.id) : null;
+				},
+				resolveRigor: () =>
+					resolveRigor({ cwd: process.cwd(), override: parseRigorOverride(process.env.CLIO_CODER_RIGOR) }),
+				recordDecision: (record) => safety.audit.recordCompletionContract?.(record),
+			}),
+		);
+	}
+
+	const chat = createChatLoop({
+		getSettings: getCurrentSettings,
+		providers,
+		middleware,
+		middlewareToolChoice,
+		protectedArtifacts: {
+			replace: (state) => protectedArtifactsGuard.replaceState(state),
+			markDegraded: (reason) => protectedArtifactsGuard.markDegraded(reason),
+		},
+		knownTargets: () => new Set(providers.list().map((entry) => entry.target.id)),
+		observability,
+		bus,
+		...(prompts ? { prompts } : {}),
+		...(session ? { session } : {}),
+		getMemorySection: () => {
+			try {
+				const records = loadMemoryRecordsSync(clioDataDir());
+				return buildMemoryPromptSection(records, {
+					activeRepository: canonicalMemoryRepositoryIdentity(process.cwd()),
+				}).section;
+			} catch {
+				return "";
+			}
+		},
+		getTaskMemoryHandoffSource: () => {
+			ensureTaskMemorySession();
+			return renderTaskMemoryHandoffSource(taskMemoryBank.snapshot());
+		},
+		registerDeferredReminderSink: (sink) => {
+			deferredMemoryReminderSink = sink;
+		},
+		onAskUserFinalized: (policy) => {
+			decisionBoard.recordFinalizedInterview(policy);
+		},
+		...(session
+			? {
+					readSessionEntries: readCurrentSessionEntries,
+					autoCompact: createProductionAutoCompact(session, getCurrentSettings, providers),
+				}
+			: {}),
+		toolRegistry,
+		hasAttachedDispatch: () => dispatchBackground.size() > 0,
+	});
+
+	// Coordinated shutdown (SIGINT/SIGTERM, TUI quit) must abort any in-flight
+	// turn before domains stop. The agent abort fans out to every running
+	// tool's AbortSignal, and bash-exec answers it by signalling the tool's
+	// detached process group. Without this, a headless SIGINT exited the CLI
+	// while a running tool's children survived as orphans of init.
+	termination.onDrain(async () => {
+		chat.dispose();
+		// The abort fans out to running tools, but their results still land and
+		// persist through the aborted run's subscribers. Domains (the session
+		// writer among them) stop in the persist phase, strictly after drain, so
+		// awaiting settlement here makes a session append after session stop
+		// impossible by ordering.
+		await chat.whenSettled();
+	});
+
+	// A boot-time resume (CLIO_CODER_RESUME_SESSION_ID) must replay the resumed
+	// session into the chat loop the same way the interactive /resume overlay
+	// does. Without this, the first submit runs with an empty provider context
+	// and parents its user turn at null, appending a second root that silently
+	// abandons the resumed session's active path. The leaf id is restored even
+	// when rebuilding replay messages fails, so parenting stays correct and
+	// only the provider context degrades.
+	if (resumedSessionAtBoot && session) {
+		const resumedMeta = session.current();
+		if (resumedMeta) {
+			let leafTurnId: string | null = null;
+			try {
+				leafTurnId = session.tree(resumedMeta.id).leafId;
+			} catch (err) {
+				bootStderr(
+					`Clio Coder: failed to read resumed session tree ${resumedMeta.id}: ${err instanceof Error ? err.message : String(err)}\n`,
+				);
+			}
+			try {
+				// Scoped to the leaf resume landed on for the same reason the /resume
+				// overlay is (issue #107): with a /tree pin persisted, the file still
+				// holds the abandoned branch after the pinned turn, and replaying it
+				// unfiltered seeds the provider with turns the next append does not
+				// parent onto.
+				chat.resetForSession(
+					leafTurnId,
+					buildReplayAgentMessagesFromTurns(readCurrentSessionEntries(), leafTurnId ? { activeLeafTurnId: leafTurnId } : {}),
+				);
+			} catch (err) {
+				chat.resetForSession(leafTurnId);
+				bootStderr(
+					`Clio Coder: failed to replay resumed session context ${resumedMeta.id}: ${err instanceof Error ? err.message : String(err)}\n`,
+				);
+			}
+		}
+	}
+
+	if (options.acp) {
+		// ACP-served sessions get the same routing isolation as interactive
+		// ones, but ACP v1 has no channel for agent-initiated advisory text:
+		// the session/update union (agent_message_chunk, agent_thought_chunk,
+		// tool_call*, plan, …) carries turn content, and notifications outside
+		// an active session/prompt would break strict clients (see the matching
+		// note in src/engine/acp/server.ts). The external-divergence and
+		// target-removed notices therefore go to the session ledger as `custom`
+		// entries, where /resume and session tooling can surface them.
+		// ACP is operatorless too: bound a runaway turn with the shared
+		// interrupt->stop subscriber, the same way the headless path does.
+		const unsubscribeAcpLoopGuardStop = subscribeLoopGuardStop(bus, chat);
+		const unsubscribeAcpRoutingNotices = bus.on(BusChannels.ConfigNextTurn, (payload) => {
+			const evt = payload as { diff?: { nextTurn?: string[] }; settings?: Readonly<ClioSettings> } | null | undefined;
+			if (!evt?.settings || !Array.isArray(evt.diff?.nextTurn)) return;
+			if (!session?.current()) return;
+			const notices = routingChangeNotices(evt.diff.nextTurn, evt.settings, getCurrentSettings());
+			for (const notice of notices) {
+				try {
+					session.appendEntry({
+						kind: "custom",
+						customType: "clio.routing-notice",
+						parentTurnId: null,
+						data: { kind: notice.kind, level: notice.level, text: notice.text },
+					});
+				} catch {
+					// Advisory only; a ledger write failure must not affect the
+					// ACP turn loop.
+				}
+			}
+		});
+		try {
+			const transport = options.acp.transport ?? createStdioServerTransport(options.acp.transportOptions);
+			const code = await serveClioAcpAgent({
+				transport,
+				chat,
+				...(session ? { session } : {}),
+				...(session
+					? {
+							readSessionEntries: readSessionEntriesForCompact,
+							buildReplayMessages: (entries: ReadonlyArray<SessionEntry>, leafTurnId: string | null) =>
+								buildReplayAgentMessagesFromTurns(entries, leafTurnId === null ? {} : { activeLeafTurnId: leafTurnId }),
+						}
+					: {}),
+				providers,
+				settings: {
+					read: readAcpSafeSettings,
+					commit: commitAcpSafeSettings,
+				},
+				toolRegistry,
+				bus,
+				autonomy: resolveBaselineAutonomy,
+				routing: () => {
+					const settings = getCurrentSettings();
+					return {
+						target: settings.orchestrator.target,
+						model: settings.orchestrator.model,
+					};
+				},
+				onActiveSessionAutonomyChange: (level) => {
+					activeAcpSessionAutonomy = level;
+				},
+				cwd: process.cwd(),
+				version: getVersionInfo().clio,
+				// Stdout belongs to JSON-RPC. Text this process did not author, such
+				// as a provider's failure body, is kept off the wire and written to
+				// the unstructured stderr tail instead.
+				diagnostics: (line) => {
+					process.stderr.write(`[clio:acp] ${line}\n`);
+				},
+				permissionTimeoutMs:
+					options.acp.permissionTimeoutMs ??
+					config?.get().delegation.defaults.permissionTimeoutMs ??
+					DEFAULT_DELEGATION_PERMISSION_TIMEOUT_MS,
+			});
+			// Same ordering the termination drain hook relies on: an aborted turn's
+			// tool results still land and persist through the run's subscribers,
+			// and the session writer stops in result.stop() below. Awaiting
+			// settlement here makes a session append after session stop impossible
+			// by ordering rather than by timing.
+			await chat.whenSettled();
+			chat.dispose();
+			await dispatch.drain();
+			await result.stop();
+			return { exitCode: code, bootTimeMs: timer.snapshot().totalMs };
+		} finally {
+			unsubscribeAcpRoutingNotices();
+			unsubscribeAcpLoopGuardStop();
+		}
+	}
+
+	if (options.headless) {
+		// Operatorless: there is no TUI subscriber to turn a loop-guard interrupt
+		// into a run stop, so a degenerate local model would spin until an
+		// external timeout (each call blocked, the agent loop never aborted). Wire
+		// the shared interrupt->stop subscriber so the run ends with the same
+		// durable closing turn the interactive surface produces.
+		const unsubscribeLoopGuardStop = subscribeLoopGuardStop(bus, chat);
+		const headlessPermissionReason = HEADLESS_PERMISSION_DENIED_REASON;
+		const unsubscribeHeadlessPermission = toolRegistry.onPermissionRequired((call, decision, meta) => {
+			bus.emit(BusChannels.PermissionResolved, {
+				status: "denied",
+				requestId: meta.requestId,
+				origin: "main",
+				decidedBy: "policy:no-operator",
+				tool: call.tool,
+				actionClass: decision.classification.actionClass,
+				reason: headlessPermissionReason,
+				requestedBy: "headless",
+			});
+			toolRegistry.cancelParkedCalls(headlessPermissionReason);
+		});
+		try {
+			const parsedSkillRequest = resources?.parsePendingSkillRequests(options.headless.prompt, process.cwd()) ?? {
+				text: options.headless.prompt,
+				pendingSkillRequests: [],
+			};
+			const promptExpansion = resources?.expandPromptTemplate(parsedSkillRequest.text, process.cwd());
+			// A prompt named as `/name` is a request for that template. When the
+			// template refuses, the run says why and stops; sending the literal
+			// `/name` on to the model spends a turn answering a command it cannot
+			// run, and the operator never sees the refusal.
+			if (promptExpansion?.expanded === false && promptExpansion.refusal) {
+				process.stderr.write(`clio-coder: ${promptExpansion.refusal.message}\n`);
+				await termination.shutdown(1);
+				return { exitCode: 1, bootTimeMs: timer.snapshot().totalMs };
+			}
+			const fileExpansion = await expandInlineFileReferencesAsync(
+				promptExpansion?.expanded ? promptExpansion.text : parsedSkillRequest.text,
+				{
+					cwd: process.cwd(),
+					includeImages: true,
+					missing: "leave",
+				},
+			);
+			const images = [...(options.headless.images ?? []), ...fileExpansion.images];
+			const workingContextPaths = [...(options.headless.workingContextPaths ?? []), ...fileExpansion.referencedPaths];
+			const code = await runHeadlessMainAgent(chat, {
+				prompt: fileExpansion.text,
+				...(images.length > 0 ? { images } : {}),
+				...(workingContextPaths.length > 0 ? { workingContextPaths } : {}),
+				...(options.headless.sampling ? { sampling: options.headless.sampling } : {}),
+				...(parsedSkillRequest.pendingSkillRequests.length > 0
+					? { pendingSkillRequests: parsedSkillRequest.pendingSkillRequests }
+					: {}),
+				mode: options.headless.mode ?? "text",
+				...(options.headless.jsonEvents ? { jsonEvents: options.headless.jsonEvents } : {}),
+				...(options.headless.steerChannel ? { steerChannel: options.headless.steerChannel } : {}),
+				getSessionHeader: () => printJsonSessionHeader(session?.current() ?? null),
+			});
+			await termination.shutdown(code);
+			return { exitCode: code, bootTimeMs: timer.snapshot().totalMs };
+		} finally {
+			unsubscribeHeadlessPermission();
+			unsubscribeLoopGuardStop();
+		}
+	}
+
+	if (options.terminalLease) {
+		for (const diagnostic of options.terminalLease.takeDiagnostics()) {
+			const text = diagnostic.text.trimEnd();
+			if (text.length > 0) initialNotices.push(text);
+		}
+	}
+
+	await startInteractive({
+		bus,
+		providers,
+		dispatch,
+		...(agents ? { agents } : {}),
+		observability,
+		chat,
+		...(options.terminalLease
+			? {
+					terminalLease: options.terminalLease,
+					onHydratedFrameCommit: () => timer.mark("Stage 1 hydration"),
+				}
+			: { onFirstFrameCommit: () => timer.mark("first TUI paint") }),
+		...(initialNotices.length > 0 ? { initialNotices } : {}),
+		...(resources ? { resources } : {}),
+		...(extensions ? { extensions } : {}),
+		...(interop ? { interop } : {}),
+		...(share ? { share } : {}),
+		toolRegistry,
+		...(session ? { session } : {}),
+		...(session ? { readSessionEntries: readCurrentSessionEntries } : {}),
+		getTaskBoard: () => taskBoard.cachedSnapshot(),
+		getDecisionBoard: () => decisionBoard.snapshot(),
+		supersedeDecision: (interviewId, key, correction) => decisionBoard.supersede(interviewId, key, correction),
+		userTasks,
+		getTaskMemoryStatus: () => {
+			ensureTaskMemorySession();
+			const settings = getCurrentSettings();
+			const bank = taskMemoryBank.snapshot();
+			return {
+				enabled: settings.memory.intervention.enabled,
+				tier: settings.background.target && settings.background.model ? "llm" : "rules",
+				size: taskMemoryBankSize(bank),
+				lastDecision: memoryIntervention.lastDecision(),
+				bank,
+				activity: memoryIntervention.recentActivity(),
+				stepInFlight: memoryIntervention.stepInFlight(),
+			};
+		},
+		getTaskMemorySeedOffer,
+		seedTaskMemory: seedCurrentTaskMemoryFromHandoff,
+		stateDir: clioStateDir(),
+		dataDir: clioDataDir(),
+		registerAskUserHandler: (handler) => {
+			askUserHandler = handler;
+			return () => {
+				if (askUserHandler === handler) askUserHandler = null;
+			};
+		},
+		getSettings: getCurrentSettings,
+		getFleetNodes: () => result.getContract<SchedulingContract>("scheduling")?.fleet?.list() ?? [],
+		onBackgroundDispatch: () => dispatchBackground.backgroundNewest(),
+		...(session ? { getSessionId: () => session.current()?.id ?? null } : {}),
+		...(contextDomain
+			? {
+					getContextState: (cwd?: string) => contextDomain.contextState(cwd),
+					onInit: async (
+						options: {
+							preview?: boolean;
+							adopt?: boolean;
+							applyClioMd?: boolean;
+							rewriteClioMd?: boolean;
+							proposeClioMd?: boolean;
+							includeGlobalImports?: boolean;
+							heuristic?: boolean;
+						},
+						runIo?: RunIo,
+					) => {
+						// Interactive context-init explores the repo with the configured target by
+						// default, grounded in the freshly built codewiki, and falls back to the
+						// deterministic heuristic when no target is reachable. --heuristic and
+						// --preview skip model generation.
+						const useModel = options.heuristic !== true && options.preview !== true;
+						const bootstrapOptions = bootstrapInputFromInitOptions(options);
+						await contextDomain.runBootstrap({
+							cwd: process.cwd(),
+							confirmGitignore: () => true,
+							adopt: options.adopt === true,
+							...bootstrapOptions,
+							...(useModel
+								? {
+										generate: modelBootstrapGenerate({
+											dispatch,
+											resolveRoute: () => {
+												if (!config) throw new Error("context-bootstrap configuration unavailable");
+												return resolveBootstrapRoute(config.get());
+											},
+											// Names the agent that actually ran and reports the throw as
+											// what it is. "Scout unavailable" was wrong twice over: the
+											// agent is context-bootstrap, and the same line was printed
+											// for a worker that failed, a worker whose answer the loop
+											// guard removed, and a worker that succeeded and whose
+											// payload the reader then refused.
+											onFallback: (err, mode) =>
+												runIo?.stderr(
+													`context init: context-bootstrap did not produce a handbook, using ${mode === "existing" ? "the existing CLIO-CODER.md" : "the heuristic writer"} (${err.message})\n`,
+												),
+										}),
+										modelId: "configured-clio-target",
+									}
+								: {}),
+						});
+					},
+					onContextClear: async (options: { all?: boolean; confirmed?: boolean; confirmedAll?: boolean }) => {
+						await contextDomain.runContextClear({
+							cwd: process.cwd(),
+							all: options.all === true,
+							io: {
+								stdout: (s) => process.stdout.write(s),
+								stderr: (s) => process.stderr.write(s),
+							},
+							confirmContext: () => options.confirmed === true,
+							confirmAll: () => options.confirmedAll === true,
+						});
+					},
+					onContextRefresh: async () => {
+						await contextDomain.runContextRefresh({
+							cwd: process.cwd(),
+							io: {
+								stdout: (s) => process.stdout.write(s),
+								stderr: (s) => process.stderr.write(s),
+							},
+						});
+					},
+				}
+			: {}),
+		onSetThinkingLevel: (level) => {
+			const current = getCurrentSettings();
+			const nextLevel =
+				resolveModelRuntimeCapabilitiesForProviders(
+					providers,
+					current.orchestrator.target,
+					current.orchestrator.model,
+					level,
+				)?.thinking.effectiveLevel ?? "off";
+			updateSessionRouting({ orchestrator: { thinkingLevel: nextLevel } });
+		},
+		onCycleThinking: () => {
+			const current = getCurrentSettings();
+			const thinking = resolveModelRuntimeCapabilitiesForProviders(
+				providers,
+				current.orchestrator.target,
+				current.orchestrator.model,
+				current.orchestrator.thinkingLevel ?? "off",
+			)?.thinking;
+			const effectiveAvailable = thinking?.supportedLevels ?? (["off"] as ThinkingLevel[]);
+			const nextLevel = advanceThinkingLevel(
+				thinking?.effectiveLevel ?? current.orchestrator.thinkingLevel ?? "off",
+				effectiveAvailable,
+			);
+			updateSessionRouting({ orchestrator: { thinkingLevel: nextLevel } });
+		},
+		onSelectModel: ({ target, model }) => {
+			const registry = getRuntimeRegistry();
+			const settings = getCurrentSettings();
+			const descriptor = settings.targets.find((e) => e.id === target);
+			if (descriptor) {
+				const runtime = registry.get(descriptor.runtime);
+				if (!runtime) {
+					throw new Error(
+						`cannot use target '${target}' as orchestrator target because runtime '${descriptor.runtime}' is not registered`,
+					);
+				}
+				if (!isOrchestratorEligibleRuntime(runtime)) {
+					throw new Error(
+						`cannot use target '${target}' as orchestrator target because runtime '${runtime.id}' is not an HTTP/native runtime`,
+					);
+				}
+			}
+			updateSessionRouting({ orchestrator: { target, model } });
+			rememberRecentModel(`${target}/${model}`, getCurrentSettings().modelSelector.recentLimit);
+		},
+		writeSettings: (next) => applySettingsBlob(next),
+		commitSetting: (id, next, scope) => commitSetting(id, next, scope),
+		...(session
+			? {
+					onResumeSession: (sessionId) => {
+						try {
+							const previousSessionId = session.current()?.id ?? null;
+							session.resume(sessionId);
+							taskBoard.snapshot();
+							if (previousSessionId !== sessionId) ensureTaskMemorySession();
+						} catch (err) {
+							process.stderr.write(
+								`[/resume] failed to resume ${sessionId}: ${err instanceof Error ? err.message : String(err)}\n`,
+							);
+						}
+					},
+					onNewSession: () => {
+						const settings = getCurrentSettings();
+						const input: { cwd: string; target?: string; model?: string } = { cwd: process.cwd() };
+						if (settings.orchestrator.target) input.target = settings.orchestrator.target;
+						if (settings.orchestrator.model) input.model = settings.orchestrator.model;
+						session.create(input);
+						taskBoard.snapshot();
+						ensureTaskMemorySession();
+					},
+					onForkSession: (parentTurnId) => {
+						try {
+							session.fork(parentTurnId);
+							taskBoard.snapshot();
+							ensureTaskMemorySession();
+						} catch (err) {
+							process.stderr.write(
+								`[/fork] failed at turn ${parentTurnId}: ${err instanceof Error ? err.message : String(err)}\n`,
+							);
+						}
+					},
+					onCompact: async (instructions) => {
+						await chat.compact(instructions);
+					},
+				}
+			: {}),
+		onCycleScopedModelForward: () => cycleScopedSession("forward"),
+		onCycleScopedModelBackward: () => cycleScopedSession("backward"),
+		onShutdown: async () => {
+			await termination.shutdown(0);
+		},
+	});
+	return { exitCode: 0, bootTimeMs: timer.snapshot().totalMs };
+}