dsh-model-switch 0.4.23 → 0.4.25

package/lib/compaction.js

@@ -1,3 +1,4 @@
+import { n as sessionProviderOf, t as resolveCapabilityRoute } from "./session-route-BCQFzqRP.js";
 //#region lib/types/compaction.js
 /** Official compaction row this module supersedes. */
 const OFFICIAL_COMPACTION_ROW = "compaction-basic";
@@ -81,6 +82,118 @@ function currentCompactionRoute(ctx) {
 	return compactionRouteOf(settings);
 }
 /**
+* Resolve the compaction summarization route for one session id.
+*
+* Looks up the session from `ctx.sessions` when available, derives its current
+* provider (`pending` selection preferred over `lastUsed`, then logged routes),
+* and resolves the provider group's compaction pair before falling back to the
+* deployment-wide compaction pair.
+* @param ctx - this row's context.
+* @param sessionId - the session id carried on `GenerateOptions.sessionId`.
+* @returns The derived compaction route, or undefined for the engine's own route.
+*/
+function resolveSessionCompactionRoute(ctx, sessionId) {
+	if (typeof sessionId !== "string" || sessionId.trim() === "") return void 0;
+	let settings;
+	try {
+		settings = ctx.get("modelSwitch", false)?.currentSettings();
+	} catch {
+		settings = void 0;
+	}
+	if (!settings) return void 0;
+	let session;
+	try {
+		session = ctx.get("sessions", false)?.get?.(sessionId);
+	} catch {
+		session = void 0;
+	}
+	const provider = sessionProviderOf(ctx, session);
+	return resolveCapabilityRoute(settings, provider, "compaction");
+}
+/**
+* Wrap the real llm service so `purpose === 'compaction'` calls resolve their
+* summarization route per session while keeping `compaction/summary` truthful.
+*
+* `summarizeWithLlm` builds a mutable `options` literal with `purpose: 'compaction'`
+* and `sessionId: agent.session.id`, calls `ctx.llm.stream(options)`, and then
+* returns `{ provider: options.provider, model: options.model }` to `commitCompactionBody`.
+* Mutating `options.provider` and `options.model` in place before delegating to
+* `target.stream(options)` ensures the durable `compaction/summary` event
+* records the exact provider/model that wrote the summary. All other members
+* (such as `resolveModelInfo`) pass through bound to `real`.
+* @param real - the host's `ctx.llm`.
+* @param resolve - per-session route resolver; undefined keeps `options` untouched.
+* @param onError - optional error sink when the resolver throws.
+* @returns A transparent Proxy over `real`.
+*/
+function compactionLlm(real, resolve, onError) {
+	return new Proxy(real, { get(target, property, _receiver) {
+		if (property === "stream") return (options) => {
+			if (options.purpose === "compaction" && typeof options.sessionId === "string" && options.sessionId.trim() !== "") {
+				let route;
+				try {
+					route = resolve(options.sessionId);
+				} catch (cause) {
+					onError?.(cause);
+					route = void 0;
+				}
+				if (route !== void 0) {
+					if (Object.isFrozen(options)) return target.stream({
+						...options,
+						provider: route.provider,
+						model: route.model
+					});
+					options.provider = route.provider;
+					options.model = route.model;
+				}
+			}
+			return target.stream(options);
+		};
+		const value = Reflect.get(target, property, target);
+		return typeof value === "function" ? value.bind(target) : value;
+	} });
+}
+/**
+* Build the child context the compaction engine mounts under.
+*
+* When `ctx.llm` is present, shadows it with {@link compactionLlm} on an
+* extended context so the child engine fiber resolves the per-session wrapper
+* while registering `ctx.compaction` on the shared context tree.
+* @param ctx - this row's context.
+* @returns A context shadowing `llm` when available, else `ctx`.
+*/
+function compactionScope(ctx) {
+	let real;
+	try {
+		real = ctx.get("llm", false);
+	} catch {
+		real = void 0;
+	}
+	if (real === void 0) return ctx;
+	return ctx.extend({ llm: compactionLlm(real, (sessionId) => resolveSessionCompactionRoute(ctx, sessionId), (cause) => {
+		ctx.logger.error("model-switch: failed to resolve per-session compaction route; keeping the engine route");
+		ctx.logger.error(cause);
+	}) });
+}
+/**
+* Whether the compaction capability is actually served from this context.
+*
+* Awaiting a plugin fiber settles even when the plugin's injections were never
+* satisfied, so the mount is confirmed through the public service instead of by
+* the absence of an error: an unactivated engine must never be recorded as the
+* live engine, or the row would stop retrying while `/compact` had no service.
+* @param ctx - this row's context.
+* @returns true when `ctx.compaction` exposes the engine's public entry points.
+*/
+function engineActive(ctx) {
+	try {
+		const service = ctx.get("compaction", false);
+		return service !== void 0 && typeof service.compactIfNeeded === "function";
+	} catch {
+		return false;
+	}
+}
+/**
 * Load and mount the official basic compaction engine under `config`.
 * @param ctx - this row's context; the engine mounts unisolated so `/compact` sees it.
 * @param config - composed engine config.
@@ -98,9 +211,8 @@ async function mountEngine(ctx, config, label) {
 	}
 	let fiber;
 	try {
-		fiber = ctx.plugin(loaded.default, config);
+		fiber = compactionScope(ctx).plugin(loaded.default, config);
 		await fiber;
-		return fiber;
 	} catch (cause) {
 		if (fiber) try {
 			await fiber.dispose();
@@ -109,6 +221,14 @@ async function mountEngine(ctx, config, label) {
 		ctx.logger.error(cause);
 		return;
 	}
+	if (engineActive(ctx)) return fiber;
+	ctx.logger.error(`model-switch: the compaction engine mounted with ${label} but never activated; unloading it so the next settings change retries`);
+	try {
+		await fiber.dispose();
+	} catch (cause) {
+		ctx.logger.error("model-switch: failed to unload the unactivated compaction engine");
+		ctx.logger.error(cause);
+	}
 }
 /**
 * Dispose the live engine, keeping the ledger clean for the next mount.
@@ -129,8 +249,11 @@ async function disposeEngine(ctx, state) {
 /**
 * Reconcile one compaction mount: remount only when the composed config changed.
 *
-* Failure ladder: configured config → official defaults (`{}`) → unmounted with
-* explicit errors, so the next settings event retries the whole ladder.
+* A settings change that leaves the composed config identical — every session
+* naming, subagent, or provider-group save — returns after the signature check,
+* so the engine is never unloaded for settings it does not read. A mount that
+* fails to activate stays unmounted with an explicit error and the next settings
+* change retries with a fresh read.
 * @param ctx - this row's context.
 * @param state - the row's reconcile ledger.
 */
@@ -154,22 +277,25 @@ async function runReconcile(ctx, state) {
 		state.engine = fiber;
 		return;
 	}
-	ctx.logger.error("model-switch: retrying the compaction engine on the official defaults");
-	const fallback = await mountEngine(ctx, {}, "the official defaults");
-	if (fallback) {
-		state.engine = fallback;
-		return;
-	}
 	ctx.logger.error("model-switch: compaction engine is unmounted; /compact and automatic compaction are unavailable until settings change");
 }
 /**
-* Queue a reconcile behind the row's serialized chain, absorbing failures.
+* Queue one reconcile behind the row's serialized chain, absorbing failures.
 * @param ctx - this row's context.
 * @param state - the row's reconcile ledger.
+* @param coalesce - collapse a burst of triggers into the single pass that reads the settled settings.
 * @returns The chain tail; never rejects.
 */
-function scheduleReconcile(ctx, state) {
-	state.chain = state.chain.then(() => runReconcile(ctx, state)).catch((cause) => {
+function scheduleReconcile(ctx, state, coalesce) {
+	if (coalesce) {
+		if (state.queued) return state.chain;
+		state.queued = true;
+	}
+	state.chain = state.chain.then(async () => {
+		state.queued = false;
+		await runReconcile(ctx, state);
+	}).catch((cause) => {
+		state.queued = false;
 		ctx.logger.error("model-switch: compaction engine reconcile failed");
 		ctx.logger.error(cause);
 	});
@@ -182,9 +308,10 @@ function scheduleReconcile(ctx, state) {
 * with the same package under the configurable summarization route. Host
 * settings and the loader are read structurally so a disabled `model-switch`
 * row leaves official compaction untouched. Reconciles run on
-* `model-switch/settings-update` (emitted by the host on volatile commits and
-* ordinary saves); the listener is global so instances of this row inside
-* agent-preset realms still hear the host row's emission. The initial
+* `model-switch/settings-update` (emitted by the host at construction, on
+* volatile commits, and on ordinary saves) and read the settled settings, so one
+* save costs at most one remount; the listener is global so instances of this
+* row inside agent-preset realms still hear the host row's emission. The initial
 * reconcile is awaited so `/compact` resolves the service as soon as this row
 * completes.
 * @param ctx - this row's context.
@@ -192,14 +319,15 @@ function scheduleReconcile(ctx, state) {
 async function apply(ctx) {
 	const state = {
 		chain: Promise.resolve(),
+		queued: false,
 		signature: null,
 		engine: void 0,
 		missingRow: false
 	};
 	ctx.on("model-switch/settings-update", () => {
-		scheduleReconcile(ctx, state);
+		scheduleReconcile(ctx, state, true);
 	}, { global: true });
-	await scheduleReconcile(ctx, state);
+	await scheduleReconcile(ctx, state, false);
 }
 //#endregion
-export { apply, compactionRouteOf, composeCompactionConfig, forwardedConfig, inject, name, officialCompactionEntry };
+export { apply, compactionLlm, compactionRouteOf, compactionScope, composeCompactionConfig, engineActive, forwardedConfig, inject, name, officialCompactionEntry, resolveSessionCompactionRoute };

package/lib/index.js

@@ -1,3 +1,4 @@
+import { n as normalizeProviderGroups } from "./provider-group-CwadImhH.js";
 import { ModelSwitchAdapterRegistry } from "./adapter-registry.js";
 import { createRequire } from "node:module";
 import z from "@deepseek-ai/schemastery";
@@ -9,80 +10,6 @@ import { launchEnvironmentOf } from "@deepseek-ai/dsh-launch-environment";
 import { Config as Config$1, DEEPSEEK_DEFAULT_API_VERSION, DEEPSEEK_DEFAULT_BASE_URL, DEEPSEEK_DEFAULT_MAX_TOKENS, DEEPSEEK_DEFAULT_MAX_USES, DEEPSEEK_DEFAULT_MODEL, DEEPSEEK_PROVIDER_ID, DeepSeekSearchProvider, WEB_SEARCH_DEEPSEEK_SETTINGS_NAMESPACE } from "@deepseek-ai/dsh-web-search-deepseek";
 import { readFileSync } from "node:fs";
 import { dirname, join } from "node:path";
-import { renderPrompt } from "@deepseek-ai/dsh-system-prompt";
-import { boundContextSummary, createMessage, createUserMessage } from "@deepseek-ai/dsh-llm";
-import { canonicalHeader } from "@deepseek-ai/dsh-session";
-//#region lib/types/provider-group.js
-/**
-* Provider configuration groups: per-provider presets for the compaction,
-* subagent, and auto-review routes.
-*
-* A group stores the exact flat settings fields it will write, so applying a
-* group is a plain settings-layer write: every capability consumer keeps
-* reading the live values unchanged. All functions here are pure and never
-* throw on malformed storage.
-*/
-const ROUTE_FIELDS = Object.freeze([
-	"compactProvider",
-	"compactModel",
-	"autoReviewProvider",
-	"autoReviewModel",
-	"subagentProvider",
-	"subagentModel",
-	"subagentReasoningEffort"
-]);
-/** Own data property, so a literal '__proto__' provider id can never touch a prototype. */
-function setKey(target, key, value) {
-	Object.defineProperty(target, key, {
-		value,
-		enumerable: true,
-		writable: true,
-		configurable: true
-	});
-}
-/** A non-blank string. Blank values behave exactly like an absent field. */
-function present(value) {
-	return typeof value === "string" && value.trim() !== "";
-}
-/** Normalize one raw entry to its owning provider and route view; undefined for non-object junk. */
-function normalizeEntry(value) {
-	if (value === null || typeof value !== "object" || Array.isArray(value)) return void 0;
-	const source = value;
-	const group = {};
-	for (const field of ROUTE_FIELDS) {
-		const item = source[field];
-		if (present(item)) group[field] = item;
-	}
-	return {
-		provider: present(source.provider) ? source.provider : "",
-		group
-	};
-}
-/**
-* Strip a raw stored value to well-formed groups; never throws. Both storage
-* shapes are accepted: the stored entry array (keys live in each entry's
-* `provider` field) and a hand-written provider → routes map (keys are the
-* providers). All-empty entries stay: they pin a provider to every default.
-*/
-function normalizeProviderGroups(raw) {
-	const groups = {};
-	if (Array.isArray(raw)) {
-		for (const entry of raw) {
-			const view = normalizeEntry(entry);
-			if (view === void 0 || view.provider === "") continue;
-			setKey(groups, view.provider, view.group);
-		}
-		return groups;
-	}
-	if (raw !== null && typeof raw === "object") for (const [key, value] of Object.entries(raw)) {
-		if (key.trim() === "") continue;
-		const view = normalizeEntry(value);
-		if (view === void 0) continue;
-		setKey(groups, key, view.group);
-	}
-	return groups;
-}
-//#endregion
 //#region lib/types/host-settings.js
 /**
 * Annotated rather than inferred: the nested entry schema's inferred type
@@ -98,11 +25,12 @@ const Config = z.object({
 	searchModel: z.string().volatile(),
 	imageProvider: z.string().volatile(),
 	imageModel: z.string().volatile(),
-	compactOnSwitch: z.boolean().default(true).volatile(),
 	autoReviewProvider: z.string().volatile(),
 	autoReviewModel: z.string().volatile(),
 	compactProvider: z.string().volatile(),
 	compactModel: z.string().volatile(),
+	titleProvider: z.string().volatile(),
+	titleModel: z.string().volatile(),
 	providerGroups: z.array(z.object({
 		provider: z.string().default(""),
 		compactProvider: z.string().default(""),
@@ -111,7 +39,9 @@ const Config = z.object({
 		autoReviewModel: z.string().default(""),
 		subagentProvider: z.string().default(""),
 		subagentModel: z.string().default(""),
-		subagentReasoningEffort: z.string().default("")
+		subagentReasoningEffort: z.string().default(""),
+		titleProvider: z.string().default(""),
+		titleModel: z.string().default("")
 	})).default([]).volatile()
 });
 function readConfig(config) {
@@ -126,6 +56,8 @@ function readConfig(config) {
 	const autoReviewModel = config.autoReviewModel.get();
 	const compactProvider = config.compactProvider.get();
 	const compactModel = config.compactModel.get();
+	const titleProvider = config.titleProvider.get();
+	const titleModel = config.titleModel.get();
 	const providerGroups = normalizeProviderGroups(config.providerGroups.get());
 	return {
 		subagentMode: config.subagentMode.get(),
@@ -136,11 +68,12 @@ function readConfig(config) {
 		...searchModel === void 0 ? {} : { searchModel },
 		...imageProvider === void 0 ? {} : { imageProvider },
 		...imageModel === void 0 ? {} : { imageModel },
-		compactOnSwitch: config.compactOnSwitch.get(),
 		...autoReviewProvider === void 0 || autoReviewProvider === "" ? {} : { autoReviewProvider },
 		...autoReviewModel === void 0 || autoReviewModel === "" ? {} : { autoReviewModel },
 		...compactProvider === void 0 || compactProvider === "" ? {} : { compactProvider },
 		...compactModel === void 0 || compactModel === "" ? {} : { compactModel },
+		...titleProvider === void 0 || titleProvider === "" ? {} : { titleProvider },
+		...titleModel === void 0 || titleModel === "" ? {} : { titleModel },
 		...Object.keys(providerGroups).length === 0 ? {} : { providerGroups }
 	};
 }
@@ -679,291 +612,6 @@ function allowDshRuntime(logger, pluginName, candidates) {
 	return shouldMountDshRuntime(logger, pluginName, version, new Set(Object.entries(compatibility?.dshReleases ?? {}).filter(([, status]) => status === "compatible" || status === "verified").map(([release]) => release)), compatibility?.blocklist);
 }
 //#endregion
-//#region lib/types/switch-compact.js
-/** Send-time context check and optional old-model compaction before a switched request. */
-/** Fraction of the target window reserved for output plus estimator error. */
-const OUTPUT_RESERVE = .2;
-/** Extra tokens on top of the ratio; the meter heuristic is not exact. */
-const ESTIMATE_MARGIN = 64;
-/**
-* Compare two fully specified model routes.
-* @param left - first route, if one was resolved.
-* @param right - second route, if one was resolved.
-* @returns Whether both routes identify the same provider and model.
-*/
-function sameSwitchRoute(left, right) {
-	return left !== void 0 && right !== void 0 && left.provider === right.provider && left.model === right.model;
-}
-/**
-* Decide whether a changed, complete route needs send-time context protection.
-* @param enabled - whether the user enabled send-time protection.
-* @param previous - route used by the preceding request.
-* @param selected - route assembled for the pending request.
-* @returns Whether the pending request changes its provider or model.
-*/
-function shouldProtectSwitch(enabled, previous, selected) {
-	if (!enabled) return false;
-	if (previous === void 0 || selected === void 0) return false;
-	return !sameSwitchRoute(previous, selected);
-}
-/**
-* Input budget for the target window. Reserves output (maxTokens or 20%) plus a
-* fixed estimator margin. `undefined` means the window cannot be used.
-* @param contextWindow - adapter-owned combined request/response capacity.
-* @param maxTokens - optional generation cap from the live agent options.
-* @returns Tokens available to the request input, or undefined for an invalid window.
-*/
-function switchInputBudget(contextWindow, maxTokens) {
-	if (!(contextWindow > 0) || !Number.isFinite(contextWindow)) return void 0;
-	const reserve = Math.max(Math.ceil(contextWindow * OUTPUT_RESERVE), maxTokens ?? 0) + ESTIMATE_MARGIN;
-	return Math.max(0, Math.floor(contextWindow - reserve));
-}
-/**
-* Determine whether an estimated input exceeds its target route budget.
-* @param tokens - estimated input tokens.
-* @param budget - input budget for the target window.
-* @returns Whether the budget is unusable or the estimate is too large.
-*/
-function overSwitchBudget(tokens, budget) {
-	return !(budget > 0) || tokens > budget;
-}
-/**
-* Build one transcript-visible Model Switch progress notice.
-* @param text - localized notice text shown in the transcript.
-* @param summary - bounded event summary for durable session metadata.
-* @returns A plugin-authored user message for the session log.
-*/
-function switchNotice(text, summary) {
-	return createUserMessage({
-		content: [{
-			type: "text",
-			text
-		}],
-		source: {
-			kind: "model-switch",
-			form: "notice",
-			summary: boundContextSummary(summary)
-		}
-	});
-}
-/**
-* Extract a complete route from partial request configuration.
-* @param value - candidate provider and model values.
-* @returns A complete route, or undefined when either identifier is absent.
-*/
-function routeOf(value) {
-	if (value === void 0 || value.provider === void 0 || value.provider.length === 0 || value.model === void 0 || value.model.length === 0) return void 0;
-	return {
-		provider: value.provider,
-		model: value.model
-	};
-}
-/**
-* Install send-time switch protection on the host context.
-* @param ctx - host context that owns agent-loop, llm, and tokenMeter.
-* @param settings - live Model Switch settings; `compactOnSwitch !== false` is on.
-* @param roleOf - provider-role lookup used to exclude native Agent runtimes.
-*/
-function installSwitchCompaction(ctx, settings, roleOf = () => void 0) {
-	const assemblies = /* @__PURE__ */ new WeakMap();
-	let stopGate;
-	ctx.on("system-prompt/assemble", async (_assembly, context, next) => {
-		const assembled = await next();
-		if (context.agent !== void 0) {
-			assemblies.set(context.agent, assembled);
-			stopGate?.();
-			installGate();
-		}
-		return assembled;
-	});
-	function installGate() {
-		stopGate = ctx.on("agent/pre-step", async ({ agent, messages, signal }, next) => {
-			const decision = await next();
-			if (decision.kind !== "enter" || signal.aborted) return decision;
-			const enabled = settings().compactOnSwitch !== false;
-			const previous = routeOf(agent.session.requestHeader()?.config);
-			const assembly = assemblies.get(agent);
-			const selected = routeOf(assembly === void 0 ? void 0 : {
-				...assembly.variables.provider === void 0 ? {} : { provider: assembly.variables.provider },
-				...assembly.variables.model === void 0 ? {} : { model: assembly.variables.model }
-			});
-			if (!shouldProtectSwitch(enabled, previous, selected) || previous === void 0 || selected === void 0) return decision;
-			if (roleOf(previous.provider) === "agent" || roleOf(selected.provider) === "agent") return decision;
-			return protectSwitch(ctx, agent, previous, selected, assembly, decision, messages, signal);
-		}, { prepend: true });
-	}
-}
-async function protectSwitch(ctx, agent, previous, selected, assembly, decision, claimed, signal) {
-	const session = agent.session;
-	const notice = (text, summary) => {
-		session.append("user/message", switchNotice(text, summary), { surfaceOp: "append" });
-	};
-	const reject = (reason, summary, compacted) => {
-		notice(compacted ? `无法继续本次模型切换：压缩已发生，但请求未继续（${reason}）。` : `无法继续本次模型切换：${reason}`, summary);
-		retainUserClaims(session, claimed);
-		return { kind: "reject" };
-	};
-	notice("正在检查上下文…", "checking context");
-	if (assembly === void 0) return reject("无法确定本次请求组装的目标模型。", "switch blocked: no assembled target", false);
-	const llm = ctx.get("llm", false);
-	if (llm?.resolveModelInfo === void 0) return reject("无法确定目标模型的上下文窗口。", "switch blocked: unknown window", false);
-	let window;
-	try {
-		const value = (await llm.resolveModelInfo(selected.provider, selected.model, signal)).context?.contextWindow;
-		window = typeof value === "number" && value > 0 ? value : void 0;
-	} catch {
-		window = void 0;
-	}
-	if (signal.aborted) return reject("已取消。", "switch blocked: cancelled", false);
-	if (window === void 0) return reject("无法确定目标模型的上下文窗口。", "switch blocked: unknown window", false);
-	const meter = ctx.get("tokenMeter", false);
-	if (meter === void 0 || typeof meter.measure !== "function" || typeof meter.estimateMessage !== "function") return reject("无法估算待发上下文（缺少 token 计量）。", "switch blocked: no token meter", false);
-	const header = targetHeader(selected, assembly);
-	const previousHeader = session.requestHeader();
-	let pressure;
-	try {
-		pressure = estimateSwitchPressure(meter, session, header, assembly, decision.messages, previousHeader);
-	} catch (error) {
-		return reject(`无法估算待发上下文（${error instanceof Error ? error.message : String(error)}）。`, "switch blocked: estimate failed", false);
-	}
-	const budget = agent.options.maxTokens === void 0 ? switchInputBudget(window) : switchInputBudget(window, agent.options.maxTokens);
-	if (budget === void 0) return reject("无法确定目标模型的上下文窗口。", "switch blocked: unknown window", false);
-	const passed = switchNotice("上下文容量检查通过，无需压缩。", "context check passed");
-	const completed = switchNotice("压缩完成，复查通过。", "compacted");
-	if (!overSwitchBudget(pressure + meter.estimateMessage(passed), budget)) {
-		session.append("user/message", passed, { surfaceOp: "append" });
-		return decision;
-	}
-	let fixedPressure;
-	try {
-		fixedPressure = estimateFixedInput(meter, session, header, assembly, decision.messages);
-	} catch (error) {
-		return reject(`无法估算待发上下文（${error instanceof Error ? error.message : String(error)}）。`, "switch blocked: estimate failed", false);
-	}
-	if (overSwitchBudget(fixedPressure + meter.estimateMessage(completed), budget)) return reject("新消息或固定请求内容已超过目标模型可用容量。", "switch blocked: fixed input oversized", false);
-	notice("当前上下文超过目标模型可用容量，正在使用旧模型压缩…", "compacting with previous model");
-	const compactionGeneration = session.surface.replaceGeneration;
-	let compacted = false;
-	try {
-		const result = await compactWithPrevious(ctx, agent, previous, signal);
-		compacted = result !== "none" || session.surface.replaceGeneration > compactionGeneration;
-		if (result === "none") return reject("没有可压缩的历史。", "switch blocked: nothing to compact", compacted);
-	} catch (error) {
-		compacted = compacted || session.surface.replaceGeneration > compactionGeneration;
-		if (signal.aborted) return reject("已取消。", "switch blocked: cancelled", compacted);
-		return reject(`压缩失败（${error instanceof Error ? error.message : String(error)}）。`, "switch blocked: compact failed", compacted);
-	}
-	if (signal.aborted) return reject("已取消。", "switch blocked: cancelled", compacted);
-	let again;
-	try {
-		again = estimateSwitchPressure(meter, session, header, assembly, decision.messages);
-	} catch (error) {
-		return reject(`压缩后无法复查（${error instanceof Error ? error.message : String(error)}）。`, "switch blocked: remeasure failed", compacted);
-	}
-	if (overSwitchBudget(again + meter.estimateMessage(completed), budget)) return reject("压缩后仍超过目标模型可用容量。", "switch blocked: still over budget", compacted);
-	session.append("user/message", completed, { surfaceOp: "append" });
-	return decision;
-}
-function targetHeader(selected, assembly) {
-	return canonicalHeader({
-		config: {
-			provider: selected.provider,
-			model: selected.model
-		},
-		...assembly.tools.length > 0 ? { tools: [...assembly.tools] } : {}
-	});
-}
-/**
-* Next-request input under the target route. Uses the committed surface priced
-* for the target (including image tokens), then adds still-pending enter
-* messages and the prospective system prompt while dropping already-priced
-* system nodes so they are not counted twice.
-*
-* The target envelope has no provider usage of its own, so the meter can only
-* price it with the chars-per-token heuristic. That heuristic is unconservative
-* for dense text (base64 payloads, minified sources, CJK). The previous
-* envelope holds the provider's reported usage for the same surface, so the
-* switch takes the larger of the two prices: the anchor keeps the check
-* truthful exactly where a switch makes the heuristic weakest.
-*
-* @param meter - replay token meter.
-* @param session - session being measured.
-* @param header - target request envelope.
-* @param assembly - prospective prompt assembly.
-* @param pending - messages the pre-step is about to enter.
-* @param previousHeader - envelope of the last request, when the session has one.
-* @returns Estimated next-request input tokens.
-*/
-function estimateSwitchPressure(meter, session, header, assembly, pending, previousHeader) {
-	const estimate = routePressure(meter, session, header, assembly, pending);
-	if (previousHeader === void 0) return estimate;
-	return Math.max(estimate, routePressure(meter, session, previousHeader, assembly, pending));
-}
-/**
-* Price one request envelope against the committed surface.
-* @param meter - replay token meter.
-* @param session - session being measured.
-* @param header - request envelope to price.
-* @param assembly - prospective prompt assembly.
-* @param pending - messages the pre-step is about to enter.
-* @returns Estimated input tokens for that envelope.
-*/
-function routePressure(meter, session, header, assembly, pending) {
-	const measured = meter.measure(session, header);
-	let systemTokens = 0;
-	for (const node of measured.nodes) if (session.eventAt(node.seq)?.type === "system/message") systemTokens += node.tokens;
-	return measured.totalTokens - systemTokens + systemPromptTokens(meter, assembly) + pendingTokens(meter, pending);
-}
-function estimateFixedInput(meter, session, header, assembly, pending) {
-	const measured = meter.measure(session, header);
-	return measured.totalTokens - measured.surfaceTokens + systemPromptTokens(meter, assembly) + pendingTokens(meter, pending);
-}
-function pendingTokens(meter, messages) {
-	let total = 0;
-	for (const message of messages) {
-		const tokens = meter.estimateMessage(message);
-		if (!(tokens >= 0) || !Number.isFinite(tokens)) throw new Error("pending message token estimate is not a number");
-		total += tokens;
-	}
-	return total;
-}
-function systemPromptTokens(meter, assembly) {
-	const text = renderPrompt(assembly);
-	if (text.length === 0) return 0;
-	return meter.estimateMessage(createMessage({
-		role: "system",
-		content: [{
-			type: "text",
-			text
-		}],
-		source: { kind: "system-prompt" }
-	}));
-}
-function retainUserClaims(session, claimed) {
-	for (const message of claimed) {
-		if (message.source.kind !== "user") continue;
-		session.append("user/message", message, { surfaceOp: "append" });
-	}
-}
-async function compactWithPrevious(ctx, agent, previous, signal) {
-	const isolated = ctx.isolate("compaction");
-	const Engine = (await import("@deepseek-ai/dsh-compaction-basic")).default;
-	const fiber = isolated.plugin(Engine, {
-		auto: false,
-		summarizationProvider: previous.provider,
-		summarizationModel: previous.model,
-		compactionRetries: 0
-	});
-	try {
-		await fiber;
-		const engine = isolated.get("compaction", false);
-		if (engine === void 0) throw new Error("private compaction engine is unavailable");
-		return await engine.compactIfNeeded(agent, "context-overflow", signal) === null ? "none" : "compacted";
-	} finally {
-		await fiber.dispose();
-	}
-}
-//#endregion
 //#region lib/types/host-runtime.js
 /** Host owner for Model Switch settings. */
 var ModelSwitchRuntime = class extends Service {
@@ -986,7 +634,6 @@ var ModelSwitchRuntime = class extends Service {
 		this.source = () => readConfig(entry);
 		ctx.inject(["settings"], (scope) => scope.effect(() => scope.settings.configure({ auto: false }, ctx.fiber), "Model Switch: disable automatic settings form"));
 		if (!allowDshRuntime(ctx.logger, "dsh-model-switch", ["@deepseek-ai/dsh-agent"])) return;
-		installSwitchCompaction(ctx, () => this.source(), (provider) => this.adapters.get(provider)?.role);
 		installDeepSeekSearchAdapter(ctx);
 		installModelSwitchSearchProvider(ctx, this);
 		installCapabilitiesRpc(ctx, this.adapters, () => this.capabilities);