@iowarp/clio-coder 0.3.0 → 0.3.1

package/src/engine/apis/residency.ts

@@ -0,0 +1,640 @@
+/**
+ * One capacity-aware model-residency reconciler, shared by every local runtime
+ * that pins weights in VRAM. Both the interactive chat loop and the headless
+ * worker drive the provider stream path, so calling the reconciler at the top
+ * of each manageable runtime's stream gives exactly one place that decides
+ * load and evict for both paths.
+ *
+ * Co-residency is the default. Local servers are multi-model hosts with finite
+ * VRAM, not single-model slots: when the runtime advertises capacity (the
+ * llama.cpp router's `max_instances`) and a slot is free, Clio loads without
+ * evicting anything; when the runtime loads just-in-time and fails an
+ * oversized load cleanly (LM Studio), Clio attempts the co-resident load first
+ * and swaps only after that failure; when the server schedules fits itself
+ * (Ollama), Clio releases only its own unprotected stragglers. Eviction is the
+ * exception, taken only when a slot must be freed, and it never selects a
+ * protected resident while an unprotected one is available.
+ *
+ * Protection is symmetric and role-aware. Residents tagged `pinned:true` or
+ * `role:scout` by the server operator are never evicted. Residents referenced
+ * by the operator's own Clio configuration (see
+ * {@link setProtectedModelsProvider}) carry the role they serve (chat, memory,
+ * worker, target default) and are never evicted by another Clio stream while an
+ * unprotected candidate exists, and only with a loud warning naming the role
+ * when capacity is exhausted. No Clio profile silently evicts another profile's
+ * model, and no chat switch silently unloads the memory plane.
+ *
+ * The reconciler is best-effort and non-blocking. A slow or unreachable
+ * server, or a malformed resident listing, degrades to observe-only and never
+ * crashes a turn. `CLIO_CODER_RESIDENCY=observe` and an explicit target
+ * `lifecycle: user-managed` both force observe-only on every runtime path.
+ * Every collision or stress case emits a notice over the event bus instead of
+ * a thrown error. One reconcile decision holds per (target, model) within a
+ * TTL, and mutations for a target are serialized across processes through a
+ * state-dir lock file so a worker and the orchestrator cannot interleave
+ * unload/load against the same server.
+ */
+
+import { BusChannels, type RuntimeNoticeKind, type RuntimeNoticePayload } from "../../core/bus-events.js";
+import type { ProtectedModelRef, ResidencyRole } from "../../core/residency-protection.js";
+import { getSharedBus } from "../../core/shared-bus.js";
+import { withResidencyLock } from "./residency-lock.js";
+import { type ResidentModelInfo, residentMatchesKeep } from "./resident-models.js";
+
+export type ResidencyNotice = RuntimeNoticePayload;
+export type { RuntimeNoticeKind };
+
+// --- notice sink -----------------------------------------------------------
+
+export type ResidencyNoticeSink = (notice: ResidencyNotice) => void;
+
+function busNoticeSink(notice: ResidencyNotice): void {
+	getSharedBus().emit(BusChannels.RuntimeNotice, notice);
+}
+
+let noticeSink: ResidencyNoticeSink = busNoticeSink;
+
+/**
+ * Override where residency notices go. The main process keeps the default,
+ * which emits on the shared bus that the interactive layer renders. The worker
+ * subprocess installs a stderr sink so headless runs still surface the reason.
+ * Passing null restores the bus sink.
+ */
+export function setResidencyNoticeSink(sink: ResidencyNoticeSink | null): void {
+	noticeSink = sink ?? busNoticeSink;
+}
+
+/** Emit one notice through the active sink. Never throws into a turn. */
+export function emitResidencyNotice(notice: ResidencyNotice): void {
+	try {
+		noticeSink(notice);
+	} catch {
+		// A notice is informational; a sink failure must never escape into a turn.
+	}
+}
+
+// --- configured-model protection ---------------------------------------------
+
+// Wire model ids the operator's own configuration references, with the role
+// each one serves (chat, memory, worker, target default). The provider is
+// installed by the composition root: the orchestrator entry derives it from the
+// live effective settings, and the worker entry from the protectedModels list
+// carried on its WorkerSpec, so every process protects the same set. A bare
+// string is accepted for an id whose role the caller does not know.
+export type ProtectedModelEntry = string | ProtectedModelRef;
+
+let protectedModelsProvider: (() => ReadonlyArray<ProtectedModelEntry>) | null = null;
+
+export function setProtectedModelsProvider(provider: (() => ReadonlyArray<ProtectedModelEntry>) | null): void {
+	protectedModelsProvider = provider;
+}
+
+/** Configured model ids mapped to the role that references them, when known. */
+function configProtectedRoles(): ReadonlyMap<string, ResidencyRole | undefined> {
+	const roles = new Map<string, ResidencyRole | undefined>();
+	try {
+		for (const entry of protectedModelsProvider?.() ?? []) {
+			const modelId = (typeof entry === "string" ? entry : entry.modelId).trim();
+			if (modelId.length === 0) continue;
+			const role = typeof entry === "string" ? undefined : entry.role;
+			if (!roles.has(modelId) || roles.get(modelId) === undefined) roles.set(modelId, role);
+		}
+	} catch {
+		return new Map();
+	}
+	return roles;
+}
+
+/** True when the runtime tagged this resident as operator-pinned (llama.cpp router tags). */
+export function residentTagProtected(tags: ReadonlyArray<string> | undefined): boolean {
+	if (!tags) return false;
+	return tags.includes("pinned:true") || tags.includes("role:scout");
+}
+
+// --- Clio-loaded registry --------------------------------------------------
+
+// Models Clio itself loaded, keyed by a stable per-server target key. The
+// registry is per-process; a resident model absent from it may still be
+// another Clio process's model, which is why eviction relies on the symmetric
+// protection above rather than on attribution alone.
+const clioLoaded = new Map<string, Set<string>>();
+
+/** Record that Clio loaded `modelId` on the target identified by `targetKey`. */
+export function markClioLoaded(targetKey: string, modelId: string): void {
+	let set = clioLoaded.get(targetKey);
+	if (!set) {
+		set = new Set();
+		clioLoaded.set(targetKey, set);
+	}
+	set.add(modelId);
+}
+
+function forgetClioLoaded(targetKey: string, modelId: string): void {
+	clioLoaded.get(targetKey)?.delete(modelId);
+}
+
+function isClioLoaded(targetKey: string, entry: ResidentModelInfo): boolean {
+	const set = clioLoaded.get(targetKey);
+	if (!set) return false;
+	if (set.has(entry.modelId)) return true;
+	return (entry.aliasIds ?? []).some((id) => set.has(id));
+}
+
+// --- TTL fast path ---------------------------------------------------------
+
+/** Skip the listResident round-trip for this long after a clean reconcile. */
+export const RECONCILE_TTL_MS = 60_000;
+const reconcileCache = new Map<string, { modelId: string; decision: "reconcile" | "observe"; at: number }>();
+
+/** Test-only: clear the cross-call Clio-loaded registry and TTL cache. */
+export function resetResidencyState(): void {
+	clioLoaded.clear();
+	reconcileCache.clear();
+}
+
+// --- residency decision (pure) ----------------------------------------------
+
+/**
+ * How the runtime makes a requested model resident, which decides when an
+ * eviction is even on the table:
+ *   - "router": Clio must POST an explicit load (llama.cpp router) and the
+ *     server advertises a slot capacity. A free slot loads without eviction;
+ *     a full server frees exactly the slots needed; unknown capacity falls
+ *     back to a conservative swap of unprotected residents.
+ *   - "jit": the runtime loads on open and fails an oversized load cleanly
+ *     (LM Studio with gpuStrictVramCap). Nothing is evicted up front; the
+ *     plan carries ranked `fallbackEvict` candidates for a retry after a
+ *     will-not-fit failure.
+ *   - "scheduler": the server places and fits models itself (Ollama). Only
+ *     Clio's own unprotected stragglers are released, because Clio pins them
+ *     with keep_alive -1 and nothing else ever reclaims them.
+ */
+export type ResidencyStrategy = "router" | "jit" | "scheduler";
+
+/** Why a resident model may not be evicted, when it may not. */
+export type ResidencyProtection = "tag" | "config";
+
+export interface ResidentClassified extends ResidentModelInfo {
+	/** True when this process's Clio-loaded registry attributes the model to Clio. */
+	loadedByClio: boolean;
+	/** Set when the resident is protected from eviction (server tag or operator config). */
+	protection?: ResidencyProtection;
+	/** Which configured plane references this resident, when the protection came from config. */
+	role?: ResidencyRole;
+}
+
+export interface ResidencyFacts {
+	targetId: string;
+	runtimeId: string;
+	keepModelId: string;
+	resident: ReadonlyArray<ResidentClassified>;
+	/** False when the env or explicit target lifecycle opt-out asks Clio to only observe. */
+	managed: boolean;
+	strategy: ResidencyStrategy;
+	/** Max co-resident models when the runtime advertises it (llama.cpp router max_instances). */
+	capacity?: number;
+	/**
+	 * True when the runtime tags the keep model itself as operator-pinned, so it
+	 * will be excluded from every eviction tier once it is resident. Such a model
+	 * may only take an unprotected slot: evicting a config-protected resident for
+	 * it would be a one-way swap the configured role could never undo (#72).
+	 */
+	keepTagProtected?: boolean;
+	contextLength?: number;
+	modelMaxContext?: number;
+}
+
+export type ResidencyDecision = "reconcile" | "observe" | "decline";
+
+export interface ResidencyPlan {
+	decision: ResidencyDecision;
+	/** Residents the plan releases before making the keep model resident. */
+	evict: ResidentModelInfo[];
+	/** Ranked swap candidates for after a failed co-resident JIT load ("jit" strategy only). */
+	fallbackEvict: ResidentModelInfo[];
+	/** True when the keep model is already resident on the target. */
+	keepResident: boolean;
+	notices: ResidencyNotice[];
+}
+
+function gib(bytes: number): string {
+	return `${(bytes / 1024 ** 3).toFixed(1)} GiB`;
+}
+
+function makeNotice(
+	facts: ResidencyFacts,
+	kind: RuntimeNoticeKind,
+	level: ResidencyNotice["level"],
+	message: string,
+	detail?: ResidencyNotice["detail"],
+): ResidencyNotice {
+	const notice: ResidencyNotice = {
+		kind,
+		level,
+		targetId: facts.targetId,
+		runtimeId: facts.runtimeId,
+		model: facts.keepModelId,
+		message,
+	};
+	if (detail) notice.detail = detail;
+	return notice;
+}
+
+function clioLoadedFirst(entries: ReadonlyArray<ResidentClassified>): ResidentClassified[] {
+	return [...entries.filter((entry) => entry.loadedByClio), ...entries.filter((entry) => !entry.loadedByClio)];
+}
+
+/** `the memory model` / `the chat model` for a notice, or the bare id's article. */
+function describeRole(role: ResidencyRole | undefined): string {
+	if (role === "chat") return "the chat model";
+	if (role === "memory") return "the memory model";
+	if (role === "worker") return "a worker model";
+	if (role === "target-default") return "a target default model";
+	return "a configured model";
+}
+
+function evictionNotice(facts: ResidencyFacts, entry: ResidentClassified): ResidencyNotice {
+	if (entry.protection === "config") {
+		return makeNotice(
+			facts,
+			"swap",
+			"warning",
+			`evicting '${entry.modelId}' from '${facts.targetId}' to make room for '${facts.keepModelId}' even though settings still reference it as ${describeRole(entry.role)}: every slot is taken and no unprotected resident is left. Another profile pointed at it will reload it.`,
+			{ swappedOut: entry.modelId, configProtected: true, ...(entry.role ? { role: entry.role } : {}) },
+		);
+	}
+	if (entry.loadedByClio) {
+		return makeNotice(
+			facts,
+			"about-to-evict",
+			"info",
+			`evicting Clio-loaded model '${entry.modelId}' from '${facts.targetId}' to free a slot for '${facts.keepModelId}'.`,
+			entry.sizeVramBytes !== undefined ? { freedVramBytes: entry.sizeVramBytes } : undefined,
+		);
+	}
+	return makeNotice(
+		facts,
+		"swap",
+		"warning",
+		`swapping resident '${entry.modelId}' for requested '${facts.keepModelId}' on '${facts.targetId}' (Clio did not load it; recorded transition instead of a silent unload; set CLIO_CODER_RESIDENCY=observe or lifecycle: user-managed to forbid swaps).`,
+		{
+			...(entry.sizeVramBytes !== undefined ? { freedVramBytes: entry.sizeVramBytes } : {}),
+			swappedOut: entry.modelId,
+		},
+	);
+}
+
+/**
+ * Decide what to do with the target's resident set, purely from already
+ * gathered facts. The async {@link reconcileResidency} wraps this with the
+ * round-trips; keeping the decision pure makes the capacity math, protection
+ * tiers, eviction order, and every notice case directly testable without a
+ * server.
+ */
+export function decideResidency(facts: ResidencyFacts): ResidencyPlan {
+	const notices: ResidencyNotice[] = [];
+	const keepResident = facts.resident.some((entry) => residentMatchesKeep(entry, facts.keepModelId));
+	const others = facts.resident.filter((entry) => !residentMatchesKeep(entry, facts.keepModelId));
+
+	// Stress: requested context window above the model's advertised maximum.
+	if (
+		facts.contextLength !== undefined &&
+		facts.modelMaxContext !== undefined &&
+		facts.modelMaxContext > 0 &&
+		facts.contextLength > facts.modelMaxContext
+	) {
+		notices.push(
+			makeNotice(
+				facts,
+				"stress",
+				"warning",
+				`context length ${facts.contextLength} exceeds ${facts.keepModelId}'s ${facts.modelMaxContext}-token limit on '${facts.targetId}'; lower the context window to avoid a truncated or split load.`,
+				{ requestedContext: facts.contextLength, modelMaxContext: facts.modelMaxContext },
+			),
+		);
+	}
+
+	// Stress: a resident model is partly on CPU (its total weight footprint is
+	// larger than its GPU-resident bytes), so the GPU is already oversubscribed.
+	for (const entry of facts.resident) {
+		if (entry.sizeBytes !== undefined && entry.sizeVramBytes !== undefined && entry.sizeBytes > entry.sizeVramBytes) {
+			notices.push(
+				makeNotice(
+					facts,
+					"stress",
+					"warning",
+					`model '${entry.modelId}' on '${facts.targetId}' is split across CPU and GPU (${gib(entry.sizeVramBytes)} of ${gib(entry.sizeBytes)} on GPU); expect slow generation.`,
+					{ residentVramBytes: entry.sizeVramBytes, residentTotalBytes: entry.sizeBytes },
+				),
+			);
+		}
+	}
+
+	// Observe-only: an explicit opt-out. Never evict; only report.
+	if (!facts.managed) {
+		return { decision: "observe", evict: [], fallbackEvict: [], keepResident, notices };
+	}
+
+	// The keep model is already served: co-residents stay, whoever loaded them.
+	// Report the sharing once per TTL so a genuinely oversubscribed box is
+	// never silent, but a fitting co-residency is normal, not a defect.
+	if (keepResident) {
+		if (others.length > 0) {
+			const names = others.map((entry) => (entry.role ? `${entry.modelId} (${entry.role})` : entry.modelId)).join(", ");
+			if (facts.capacity !== undefined && facts.resident.length > facts.capacity) {
+				notices.push(
+					makeNotice(
+						facts,
+						"stress",
+						"warning",
+						`'${facts.targetId}' holds ${facts.resident.length} models (${names} alongside '${facts.keepModelId}') above its ${facts.capacity}-instance capacity; unload one or raise the server's limit.`,
+						{ residentCount: facts.resident.length, maxInstances: facts.capacity },
+					),
+				);
+			} else {
+				notices.push(
+					makeNotice(
+						facts,
+						"co-resident",
+						"info",
+						`'${facts.targetId}' serves '${facts.keepModelId}' alongside ${names}. Clio leaves co-resident models loaded; keep the setup only while weights and KV caches fit in GPU memory.`,
+						facts.capacity !== undefined
+							? { residentCount: facts.resident.length, maxInstances: facts.capacity }
+							: { residentCount: facts.resident.length },
+					),
+				);
+			}
+		}
+		return { decision: "reconcile", evict: [], fallbackEvict: [], keepResident, notices };
+	}
+
+	// The keep model is not resident. Rank potential evictions by protection
+	// tier: unprotected residents first (Clio-attributed before foreign), then
+	// config-protected ones as a loud last resort. Tag-pinned residents are
+	// never candidates. A keep model that will itself come back tag-pinned may
+	// not displace the config tier: once resident it is never a candidate, so
+	// the configured role could not reclaim its slot.
+	const tierUnprotected = clioLoadedFirst(others.filter((entry) => entry.protection === undefined));
+	const tierConfig = clioLoadedFirst(others.filter((entry) => entry.protection === "config"));
+	const configEvictable = facts.keepTagProtected !== true;
+
+	let evict: ResidentClassified[] = [];
+	let fallbackEvict: ResidentClassified[] = [];
+
+	if (facts.strategy === "jit") {
+		// Attempt the co-resident load first; the runtime turns an oversized
+		// load into a clean failure, and only that failure justifies a swap.
+		fallbackEvict = configEvictable ? [...tierUnprotected, ...tierConfig] : tierUnprotected;
+	} else if (facts.strategy === "scheduler") {
+		// The server fits and places models itself. Release only Clio's own
+		// unprotected stragglers; foreign and protected residents stay.
+		evict = tierUnprotected.filter((entry) => entry.loadedByClio);
+	} else if (facts.capacity === undefined) {
+		// Router without readable capacity: no way to prove a free slot, so
+		// fall back to swapping the unprotected residents before the load.
+		evict = tierUnprotected;
+	} else {
+		const slotsNeeded = facts.resident.length + 1 - facts.capacity;
+		if (slotsNeeded > 0) {
+			const candidates = configEvictable ? [...tierUnprotected, ...tierConfig] : tierUnprotected;
+			if (
+				candidates.length < slotsNeeded &&
+				!configEvictable &&
+				tierUnprotected.length + tierConfig.length >= slotsNeeded
+			) {
+				// The load would only fit by evicting a configured model in favour of
+				// a pinned one. Decline up front instead of stranding the config role.
+				const blocked = tierConfig.slice(0, slotsNeeded - tierUnprotected.length);
+				const names = blocked.map((entry) => `'${entry.modelId}' (${describeRole(entry.role)})`).join(", ");
+				notices.push(
+					makeNotice(
+						facts,
+						"will-not-fit",
+						"error",
+						`cannot load pinned '${facts.keepModelId}' on '${facts.targetId}': it would evict ${names} that settings still reference, and a pinned model is never evicted, so the configured model could not return. Unload one manually, unpin '${facts.keepModelId}' on the server, or raise the server's max instances.`,
+						{
+							residentCount: facts.resident.length,
+							maxInstances: facts.capacity,
+							configProtected: true,
+							...(blocked[0]?.role ? { role: blocked[0].role } : {}),
+						},
+					),
+				);
+				return { decision: "decline", evict: [], fallbackEvict: [], keepResident, notices };
+			}
+			if (candidates.length < slotsNeeded) {
+				notices.push(
+					makeNotice(
+						facts,
+						"will-not-fit",
+						"error",
+						`cannot load '${facts.keepModelId}' on '${facts.targetId}': all ${facts.capacity} instance slots hold pinned models (${others.map((entry) => entry.modelId).join(", ")}). Unload one manually or raise the server's max instances.`,
+						{ residentCount: facts.resident.length, maxInstances: facts.capacity },
+					),
+				);
+				return { decision: "decline", evict: [], fallbackEvict: [], keepResident, notices };
+			}
+			evict = candidates.slice(0, slotsNeeded);
+		} else if (others.length > 0) {
+			notices.push(
+				makeNotice(
+					facts,
+					"co-resident",
+					"info",
+					`loading '${facts.keepModelId}' alongside ${others.map((entry) => entry.modelId).join(", ")} on '${facts.targetId}' (${facts.resident.length + 1}/${facts.capacity} instances).`,
+					{ residentCount: facts.resident.length + 1, maxInstances: facts.capacity },
+				),
+			);
+		}
+	}
+
+	for (const entry of evict) notices.push(evictionNotice(facts, entry));
+	return { decision: "reconcile", evict, fallbackEvict, keepResident, notices };
+}
+
+// --- async reconciler ------------------------------------------------------
+
+/**
+ * Per-runtime hooks the reconciler drives. Each manageable runtime builds an
+ * adapter that closes over its own client (the LM Studio SDK socket, the
+ * Ollama HTTP client, the llama.cpp router's HTTP surface) so the reconciler
+ * itself stays runtime-agnostic.
+ */
+export interface ResidencyAdapter {
+	/** Stable per-server key for the Clio-loaded registry, TTL cache, and lock file. */
+	targetKey: string;
+	targetId: string;
+	runtimeId: string;
+	keepModelId: string;
+	/** False when the env or explicit target lifecycle opt-out asks Clio to only observe. */
+	managed: boolean;
+	strategy: ResidencyStrategy;
+	contextLength?: number;
+	modelMaxContext?: number;
+	listResident(): Promise<ResidentModelInfo[]>;
+	/**
+	 * Runtime tags of the keep model itself, when the runtime lists them before
+	 * the load (the llama.cpp router lists tags for unloaded models too). Called
+	 * after {@link listResident}, so an adapter may serve it from that snapshot.
+	 */
+	keepModelTags?(): Promise<ReadonlyArray<string> | undefined>;
+	unload(modelId: string): Promise<void>;
+	/** Router-style servers: read the max co-resident instance count. */
+	capacity?(): Promise<number | undefined>;
+	/**
+	 * Router-style servers: make the keep model resident. Must be idempotent
+	 * (a loaded model returns immediately) and bounded by the runtime's load
+	 * wait. A thrown error propagates to the caller and fails the turn with
+	 * its message.
+	 */
+	load?(modelId: string): Promise<void>;
+	/** Cross-process mutation serializer; defaults to the state-dir lock file. */
+	withLock?<T>(targetKey: string, fn: () => Promise<T>): Promise<T>;
+	now?: () => number;
+	ttlMs?: number;
+}
+
+export type ReconcileResult = ResidencyPlan;
+
+/**
+ * Gather the resident set and capacity facts, decide, emit notices, and
+ * perform the mutations under the cross-process lock. List and unload probes
+ * are best-effort and degrade to observe-only; an explicit `load` failure
+ * propagates so the turn fails with the router's reason instead of a bare
+ * connection error.
+ */
+export async function reconcileResidency(adapter: ResidencyAdapter): Promise<ReconcileResult> {
+	const now = adapter.now ?? Date.now;
+	const ttl = adapter.ttlMs ?? RECONCILE_TTL_MS;
+
+	const cached = reconcileCache.get(adapter.targetKey);
+	if (cached && cached.modelId === adapter.keepModelId && now() - cached.at < ttl) {
+		return {
+			decision: cached.decision,
+			evict: [],
+			fallbackEvict: [],
+			keepResident: cached.decision === "reconcile",
+			notices: [],
+		};
+	}
+
+	let resident: ResidentModelInfo[];
+	try {
+		resident = await adapter.listResident();
+	} catch {
+		// Unreachable or slow server: never block the turn, just observe.
+		return { decision: "observe", evict: [], fallbackEvict: [], keepResident: false, notices: [] };
+	}
+
+	let capacity: number | undefined;
+	if (adapter.capacity) {
+		try {
+			capacity = await adapter.capacity();
+		} catch {
+			capacity = undefined;
+		}
+	}
+
+	let keepTagProtected = false;
+	if (adapter.keepModelTags) {
+		try {
+			keepTagProtected = residentTagProtected(await adapter.keepModelTags());
+		} catch {
+			keepTagProtected = false;
+		}
+	}
+
+	const protectedRoles = configProtectedRoles();
+	const classified: ResidentClassified[] = resident.map((entry) => {
+		const configId = [entry.modelId, ...(entry.aliasIds ?? [])].find((id) => protectedRoles.has(id));
+		const protection: ResidencyProtection | undefined = residentTagProtected(entry.tags)
+			? "tag"
+			: configId !== undefined
+				? "config"
+				: undefined;
+		const role = configId === undefined ? undefined : protectedRoles.get(configId);
+		return {
+			...entry,
+			loadedByClio: isClioLoaded(adapter.targetKey, entry),
+			...(protection ? { protection } : {}),
+			...(role ? { role } : {}),
+		};
+	});
+
+	const facts: ResidencyFacts = {
+		targetId: adapter.targetId,
+		runtimeId: adapter.runtimeId,
+		keepModelId: adapter.keepModelId,
+		resident: classified,
+		managed: adapter.managed,
+		strategy: adapter.strategy,
+		...(capacity !== undefined ? { capacity } : {}),
+		...(keepTagProtected ? { keepTagProtected } : {}),
+		...(adapter.contextLength !== undefined ? { contextLength: adapter.contextLength } : {}),
+		...(adapter.modelMaxContext !== undefined ? { modelMaxContext: adapter.modelMaxContext } : {}),
+	};
+
+	const plan = decideResidency(facts);
+	for (const notice of plan.notices) emitResidencyNotice(notice);
+
+	if (plan.decision === "reconcile") {
+		const mutate = async (): Promise<void> => {
+			for (const entry of plan.evict) {
+				try {
+					await adapter.unload(entry.modelId);
+					forgetClioLoaded(adapter.targetKey, entry.modelId);
+				} catch {
+					// Best-effort: a failed unload self-heals on the next reconcile.
+				}
+			}
+			if (adapter.load) await adapter.load(adapter.keepModelId);
+		};
+		// Serialize actual mutations across processes; a keep model that is
+		// already fully resident needs no lock (the load hook is a no-op) and
+		// waiting on a still-loading model is read-only polling.
+		if (plan.evict.length > 0 || (adapter.load && !plan.keepResident)) {
+			const lock = adapter.withLock ?? withResidencyLock;
+			await lock(adapter.targetKey, mutate);
+		} else {
+			await mutate();
+		}
+		// Attribute the keep model to Clio only when Clio is the one loading it.
+		if (!plan.keepResident) markClioLoaded(adapter.targetKey, adapter.keepModelId);
+		reconcileCache.set(adapter.targetKey, { modelId: adapter.keepModelId, decision: "reconcile", at: now() });
+	} else if (plan.decision === "observe" && !adapter.managed) {
+		// Cache the opt-out observation so its notices dedupe per TTL; a
+		// degraded observe (listResident failure) is never cached, so the next
+		// stream re-probes.
+		reconcileCache.set(adapter.targetKey, { modelId: adapter.keepModelId, decision: "observe", at: now() });
+	}
+
+	return plan;
+}
+
+/**
+ * The env-level residency opt-out. Clio manages residency for every
+ * manageable runtime by default; `CLIO_CODER_RESIDENCY=observe` (or off/0/false/
+ * user) flips the whole process to observe-only. Per-target opt-out goes
+ * through {@link residencyManagedFor}.
+ */
+export function residencyManaged(env: NodeJS.ProcessEnv = process.env): boolean {
+	const opt = (env.CLIO_CODER_RESIDENCY ?? "").trim().toLowerCase();
+	if (opt === "observe" || opt === "off" || opt === "0" || opt === "false" || opt === "user" || opt === "user-managed") {
+		return false;
+	}
+	return true;
+}
+
+/**
+ * Combined residency opt-out for one target: the process-wide env switch plus
+ * the target's explicit `lifecycle: user-managed`. An absent lifecycle means
+ * the operator made no choice, and Clio manages by default; runtime synthesis
+ * only sets the metadata field when settings carry an explicit value.
+ */
+export function residencyManagedFor(
+	lifecycle: "user-managed" | "clio-managed" | undefined,
+	env: NodeJS.ProcessEnv = process.env,
+): boolean {
+	if (lifecycle === "user-managed") return false;
+	return residencyManaged(env);
+}

package/src/engine/apis/resident-models.ts

@@ -0,0 +1,42 @@
+/**
+ * Runtime-agnostic resident-model shapes shared by the VRAM-aware reconciler
+ * (residency.ts) and the per-runtime adapters. The reconciler decides load and
+ * evict; this module only describes what a runtime reports as resident and how
+ * the keep target is matched against divergent id fields.
+ */
+
+/**
+ * One model currently resident (loaded) on a local runtime, with its footprint
+ * when the runtime reports it. `sizeVramBytes` is the GPU-resident portion;
+ * `sizeBytes` the total weight footprint. A `sizeBytes` larger than
+ * `sizeVramBytes` means the model is split onto CPU.
+ */
+export interface ResidentModelInfo {
+	modelId: string;
+	/**
+	 * Other ids the runtime reports for the same resident model. Ollama's
+	 * `/api/ps` returns both `model` and `name`, which can diverge, so the keep
+	 * target may match either field. The reconciler keeps a model resident when
+	 * the keep target equals `modelId` or any alias; collapsing to one id risks
+	 * evicting the model the chat loop just selected.
+	 */
+	aliasIds?: string[];
+	sizeVramBytes?: number;
+	sizeBytes?: number;
+	/**
+	 * Runtime-reported tags for the resident model (the llama.cpp router's
+	 * per-model `tags` array). The reconciler treats `pinned:true` and
+	 * `role:scout` as operator pins that are never evicted.
+	 */
+	tags?: string[];
+}
+
+/**
+ * A resident entry is the keep target when its canonical id or any alias
+ * matches. Matching on aliases too means a runtime that reports divergent id
+ * fields (Ollama's `model` vs `name`) never has the freshly selected model
+ * evicted out from under the chat loop.
+ */
+export function residentMatchesKeep(entry: ResidentModelInfo, keepModelId: string): boolean {
+	return entry.modelId === keepModelId || (entry.aliasIds?.includes(keepModelId) ?? false);
+}