dsh-model-switch 0.4.24 → 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 +82,100 @@ 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
@@ -116,7 +211,7 @@ async function mountEngine(ctx, config, label) {
 	}
 	let fiber;
 	try {
-		fiber = ctx.plugin(loaded.default, config);
+		fiber = compactionScope(ctx).plugin(loaded.default, config);
 		await fiber;
 	} catch (cause) {
 		if (fiber) try {
@@ -235,4 +330,4 @@ async function apply(ctx) {
 	await scheduleReconcile(ctx, state, false);
 }
 //#endregion
-export { apply, compactionRouteOf, composeCompactionConfig, engineActive, forwardedConfig, inject, name, officialCompactionEntry };
+export { apply, compactionLlm, compactionRouteOf, compactionScope, composeCompactionConfig, engineActive, forwardedConfig, inject, name, officialCompactionEntry, resolveSessionCompactionRoute };

package/lib/index.js

@@ -1,4 +1,4 @@
-import { n as normalizeProviderGroups } from "./provider-group-CsZi7H29.js";
+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";
@@ -10,9 +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/host-settings.js
 /**
 * Annotated rather than inferred: the nested entry schema's inferred type
@@ -28,7 +25,6 @@ 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(),
@@ -72,7 +68,6 @@ 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 },
@@ -617,300 +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);
-			const previous = stopGate;
-			let replacement;
-			try {
-				replacement = installGate();
-			} catch (cause) {
-				ctx.logger.warn("model-switch: could not refresh the switch-time compaction guard; the existing guard stays in place");
-				ctx.logger.warn(cause);
-				return assembled;
-			}
-			stopGate = replacement;
-			previous?.();
-		}
-		return assembled;
-	});
-	function installGate() {
-		return 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 {
@@ -933,7 +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);

package/lib/{provider-group-CsZi7H29.js → provider-group-CwadImhH.js}

@@ -1,20 +1,14 @@
 //#region lib/types/provider-group.js
 /**
-* Provider configuration groups: per-provider presets for the session-naming
-* and subagent routes.
+* Provider configuration groups: per-provider presets for the compaction,
+* auto-review, session-naming, and subagent routes.
 *
 * A group is a DERIVED policy, not a write: `src/session-route.ts` resolves the
 * routes one session should use from its own provider plus this stored table, so
 * switching a session's provider never rewrites deployment-wide settings and two
 * sessions on different providers cannot overwrite each other. The field names
 * below are the flat settings field names a group pins, which keeps the stored
-* shape identical to the earlier write-through design.
-*
-* Legacy entries may still carry `compactProvider`/`compactModel`/
-* `autoReviewProvider`/`autoReviewModel`: normalization preserves them so an
-* older stored document is never silently rewritten into a different meaning,
-* while the consumption side ignores them (a group save from the current UI
-* writes '' for every unsupported field, which normalization drops).
+* shape identical across releases.
 *
 * All functions here are pure and never throw on malformed storage.
 */

package/lib/{session-route-h9SAqajc.js → session-route-BCQFzqRP.js}

@@ -1,4 +1,4 @@
-import { t as groupOf } from "./provider-group-CsZi7H29.js";
+import { t as groupOf } from "./provider-group-CwadImhH.js";
 //#region lib/types/session-route.js
 /**
 * Per-session capability route derivation.

package/lib/session-title.js

@@ -1,4 +1,4 @@
-import { n as sessionProviderOf, t as resolveCapabilityRoute } from "./session-route-h9SAqajc.js";
+import { n as sessionProviderOf, t as resolveCapabilityRoute } from "./session-route-BCQFzqRP.js";
 //#region lib/types/session-title.js
 /**
 * Model Switch-owned session-naming row.

package/lib/subagent-runtime.js

@@ -1,4 +1,4 @@
-import { r as sessionSubagentRouteOf } from "./session-route-h9SAqajc.js";
+import { r as sessionSubagentRouteOf } from "./session-route-BCQFzqRP.js";
 import { ReasoningEffortId } from "@deepseek-ai/dsh-llm";
 import OfficialSubagentRuntime from "@deepseek-ai/dsh-subagent";
 //#region lib/types/subagent-runtime.js

package/lib/types/adapter-registry.d.ts

@@ -41,7 +41,7 @@ export interface ModelSwitchImageAdapter {
 }
 export interface ModelSwitchProviderAdapters {
     readonly provider: string;
-    /** Native Agent providers own their context and are excluded from DSH switch compaction. */
+    /** Optional runtime role declared by the provider integration. */
     readonly role?: 'llm' | 'agent';
     readonly search?: ModelSwitchSearchAdapter;
     readonly image?: ModelSwitchImageAdapter;

package/lib/types/auto-review.d.ts

@@ -5,17 +5,20 @@
  * reviewer to the session's own request route (`header.config`). This row runs
  * the official `apply` under a scoped child context whose `llm` service is
  * shadowed by a wrapper that rewrites only `stream()`'s provider/model — and
- * only while the profile has a complete reviewer route configured. With no
- * route configured the wrapper passes options through by identity, so the
- * reviewer follows the session model exactly.
+ * only when the reviewing session's derived route (its provider group's
+ * reviewer pair, else the deployment-wide reviewer card) pins one. With no
+ * route pinned the wrapper passes options through by identity, so the reviewer
+ * follows the session model exactly.
  *
- * The reviewer route is deliberately DEPLOYMENT-WIDE. The official gate builds
- * its request with no session id and no purpose and closes over the single
- * context it was applied with, so this plugin has no public way to tell which
- * session a review serves; guessing one would attribute a review to the wrong
- * session and make the recorded reviewer route false. The unsupported
- * per-session reviewer route is recorded in `docs/seam-gap.md` with the public
- * seam it would need.
+ * Although the official gate omits `sessionId` and `purpose` from its frozen
+ * `GenerateOptions`, its `tools/pre-execute` waterfall listener receives
+ * `exec.agent` (`exec.agent.session`). After running `official.apply(scope)`,
+ * this module registers a `{ prepend: true }` `tools/pre-execute` listener on
+ * the same scope — which Cordis places ahead of the official gate's listener in
+ * the waterfall — and binds `exec.agent?.session` in an `AsyncLocalStorage`
+ * scope around `next()`. Every `scope.llm.stream()` call from `classifyRisk`
+ * therefore resolves the exact session being reviewed, with zero cross-session
+ * interference when multiple sessions review tool calls concurrently.
  *
  * The official gate stays an opt-in capability: the row mounts it only when
  * the profile actually started with the experimental-auto-review bundle.
@@ -24,8 +27,10 @@
  *
  * @module dsh-model-switch/auto-review
  */
+import { AsyncLocalStorage } from 'node:async_hooks';
 import type { Context } from '@deepseek-ai/cordis';
 import type { GenerateOptions, StreamChunk } from '@deepseek-ai/dsh-llm';
+import type { Session } from '@deepseek-ai/dsh-session';
 /** Cordis plugin name used by loader diagnostics. */
 export declare const name = "model-switch-auto-review";
 /** The official gate's host services plus Model Switch's settings owner. */
@@ -60,12 +65,34 @@ export declare function autoReviewRouteOf(settings: {
  * @returns A structurally compatible llm face for the scoped context.
  */
 export declare function reviewLlm(real: ReviewLlm, route: () => AutoReviewRoute | undefined): ReviewLlm;
+/**
+ * Resolve the reviewer route for one session (or deployment-wide when no session is bound).
+ *
+ * @param ctx - host context whose modelSwitch settings and sessionProjections derive the route.
+ * @param session - the session whose tool call is being reviewed, when known.
+ * @returns The derived reviewer route, or undefined to follow the session model.
+ */
+export declare function resolveSessionAutoReviewRoute(ctx: Context, session?: Session): AutoReviewRoute | undefined;
+/**
+ * Install the outer `tools/pre-execute` correlation listener ahead of the official gate.
+ *
+ * Because Cordis `on(..., { prepend: true })` unshifts onto the event hook list,
+ * registering this listener AFTER `official.apply(scope)` places it at index 0
+ * — directly ahead of the official gate's own prepended listener — so
+ * `activeSession.run(exec.agent?.session, next)` wraps the entire official
+ * review execution.
+ *
+ * @param scope - the review scope context where the gate is mounted.
+ * @param activeSession - AsyncLocalStorage holding the session under review.
+ */
+export declare function bindAutoReviewSessionCorrelation(scope: Context, activeSession: AsyncLocalStorage<Session | undefined>): void;
 /**
  * Build the scoped context the official gate runs under.
  * @param ctx - host context whose modelSwitch settings pin the reviewer route.
+ * @param activeSession - optional AsyncLocalStorage correlating each `tools/pre-execute` call to its session.
  * @returns A child context shadowing only the llm service.
  */
-export declare function reviewScope(ctx: Context): Context;
+export declare function reviewScope(ctx: Context, activeSession?: AsyncLocalStorage<Session | undefined>): Context;
 /**
  * Mount the official gate under the reviewer-route scope.
  *

package/lib/types/client/ModelSwitchSettings.d.ts

@@ -26,7 +26,6 @@ export interface ModelSwitchSettingsFace {
      */
     saveSubagent: (ops: readonly ConfigViewOp<SubagentSettingsView>[]) => Promise<void>;
     saveCapability: (route: 'search' | 'image' | 'autoReview' | 'compaction' | 'title', ops: readonly ConfigViewOp<CapabilityRouteView>[]) => Promise<void>;
-    setCompactOnSwitch: (value: boolean) => Promise<void>;
     /** Replace or remove (undefined) one provider's saved capability group. */
     saveProviderGroup: (provider: string, group: ProviderGroupConfig | undefined) => Promise<void>;
     loadCatalog: () => Promise<readonly ModelProviderGroup[]>;

package/lib/types/client/locales.d.ts

@@ -4,12 +4,9 @@ export declare const zh: {
     readonly subtitle: "为子代理与能力路由设置固定模型；当前会话模型用输入框旁的选择器直接切换。";
     readonly conversationRoutes: "对话路由";
     readonly capabilityRoutes: "能力路由";
-    readonly sendProtection: "发送保护";
     readonly settingsSynced: "设置已同步";
     readonly remoteSettingsUnavailable: "官方 DSH 在非本机浏览器禁用配置读取与写入；请在 Host 的 127.0.0.1 页面打开设置。";
     readonly cancel: "取消";
-    readonly compactOnSwitch: "切换时压缩";
-    readonly compactOnSwitchHelp: "发送且换了模型时，先按目标窗口检查上下文；超了就用切换前的模型压一次历史，再继续请求。只选模型、不发送，不会压缩。官方自动压缩和 /compact 不受影响。压缩会产生额外模型用量，摘要可能省略细节。";
     readonly provider: "提供商";
     readonly model: "模型";
     readonly providerDefault: "使用提供商默认值";
@@ -46,8 +43,8 @@ export declare const zh: {
     readonly providerGroups: "供应商配置组";
     readonly providerGroupsCard: "按供应商预设路由";
     readonly providerGroupsSummary: "使用该供应商的会话自动采用对应组";
-    readonly providerGroupsHelp: "为会话命名与子代理按供应商分别保存路由。";
-    readonly providerGroupsBodyHelp: "每个供应商保存一组会话命名与子代理路由。会话自身的供应商与该组一致时，该会话的命名与子代理按组解析，否则回落到上方卡片里的全局设置。切换某个会话的模型不会写入任何全局设置，因此不同供应商的多个会话互不影响。压缩与自动审查仍是全局卡片：压缩事件要如实记录实际写入摘要的模型，审查的拦截点没有会话身份。";
+    readonly providerGroupsHelp: "为压缩、自动审查、会话命名与子代理按供应商分别保存路由。";
+    readonly providerGroupsBodyHelp: "每个供应商保存一组压缩、自动审查、会话命名与子代理路由。会话自身的供应商与该组一致时，这四项能力按组实时解析，未固定项回落到上方卡片里的全局设置。切换某个会话的模型不会写入任何全局设置，因此不同供应商的多个会话并行时互不影响，会话中切换供应商后下一次调用也会立即跟随新供应商的配置组。";
     readonly providerGroupPick: "选择供应商以编辑其配置组。";
     readonly providerGroupDelete: "删除该组";
     readonly saveTimeout: "保存超时，请重试";

package/lib/types/client-contract.d.ts

@@ -10,7 +10,6 @@ export interface ModelSwitchSettingsView {
     searchModel?: string;
     imageProvider?: string;
     imageModel?: string;
-    compactOnSwitch?: boolean;
     autoReviewProvider?: string;
     autoReviewModel?: string;
     compactProvider?: string;