dsh-model-switch 0.4.24 → 0.4.26

package/lib/session-title.js

@@ -1,4 +1,4 @@
-import { n as sessionProviderOf, t as resolveCapabilityRoute } from "./session-route-h9SAqajc.js";
+import { i as sessionProviderOf, n as groupRoute, t as deploymentRoute } from "./session-route-BjNPlliu.js";
 //#region lib/types/session-title.js
 /**
 * Model Switch-owned session-naming row.
@@ -10,16 +10,32 @@ import { n as sessionProviderOf, t as resolveCapabilityRoute } from "./session-r
 * saved group is named through that group's pair while every other session keeps
 * following its own model.
 *
-* The call itself stays inside the official implementation: the exported
-* `generateSessionTitleWithLlm` receives the session-resolved config and records
-* exactly that route in `session/title-llm-request`, so the durable record and
-* the actual request can never disagree. Nothing is remounted when settings
-* change — the route is read through on every naming request — so a settings
-* save can neither interrupt an in-flight naming request nor leave the profile
-* without a registered provider.
+* The call itself stays inside the official implementation: the auxiliary-LLM
+* package records exactly the route that was called in `session/title-llm-request`,
+* so the durable record and the actual request can never disagree. Nothing is
+* remounted when settings change — the route is read through on every naming
+* request — so a settings save can neither interrupt an in-flight naming request
+* nor leave the profile without a registered provider.
 *
-* `cordis.patch.yml` disables the official row (the seam allows one registered
-* provider per context), and this row mirrors the cadence the disabled row
+* Two official auxiliary-LLM lines are supported, selected at runtime from the
+* module surface, because the 0.2.1 release replaced the shared title call:
+* - Up to DSH 0.2.0 (`generateSessionTitleWithLlm`): the module owns the system
+*   prompt, input framing, and output validation; the shared policy carries the
+*   title-length targets.
+* - DSH 0.2.1+ (`executeSessionTitleLlm`): the provider owns the system prompt,
+*   the input, the reasoning-effort selection, and the output interpretation;
+*   the shared policy carries execution controls only. This row mirrors the
+*   official provider's wording and interpretation so a configured route keeps
+*   the official behavior.
+*
+* A naming request that fails is retried once with a widened output budget
+* (max-token failures only) and then falls through the remaining route
+* candidates — the session group's pair, the deployment-wide pair, the row's own
+* pair, then the session's logged route — so one unavailable model or provider
+* cannot leave the session on its deterministic fallback title forever.
+*
+* `cordis.patch.yml` disables the official rows (the seam allows one registered
+* provider per context), and this row mirrors the cadence a disabled row
 * declared, so a profile that restated the row as all-prompts keeps its cadence.
 *
 * @module dsh-model-switch/session-title
@@ -54,6 +70,43 @@ const TITLE_CADENCES = Object.freeze([Object.freeze({
 	automatic: "all-prompts",
 	firstPromptOnly: false
 })]);
+/** Execution controls both official auxiliary-LLM lines validate; defaults mirror the shipped official rows. */
+const EXECUTION_DEFAULTS = Object.freeze({
+	maxInputBytes: 4096,
+	maxOutputTokens: 64,
+	timeoutMs: 6e4
+});
+/** Provider-owned title-length targets: required by the pre-0.2.1 line and moved into providers by 0.2.1. */
+const TARGET_DEFAULTS = Object.freeze({
+	targetWords: 5,
+	targetCjkCharacters: 10
+});
+/** Execution keys accepted by both lines; provider-owned keys are excluded on purpose. */
+const EXECUTION_KEYS = Object.freeze([
+	"maxInputBytes",
+	"maxOutputTokens",
+	"timeoutMs",
+	"provider",
+	"model"
+]);
+/**
+* Read one call entry point without assuming it exists.
+*
+* The module namespace of a host line that does not export the name can throw on
+* access (mocked namespaces do), so a missing entry point reads as undefined
+* instead of failing the row.
+* @param module - the loaded auxiliary-LLM package.
+* @param key - the call entry point to read.
+* @returns The callable, or undefined when this line does not expose it.
+*/
+function titleCallOf(module, key) {
+	try {
+		const value = module[key];
+		return typeof value === "function" ? value : void 0;
+	} catch {
+		return;
+	}
+}
 /**
 * The cadence one official package name declares.
 * @param pkg - the disabled row's package name.
@@ -128,42 +181,341 @@ function forwardedConfig(entry) {
 */
 function officialTitlePackage(entry) {
 	const pkg = entry.options?.name;
-	return typeof pkg === "string" && pkg.trim() !== "" ? pkg : void 0;
+	if (typeof pkg !== "string" || pkg.trim() === "") return void 0;
+	return pkg;
+}
+/** A positive integer, or undefined for anything else (blank, fractional, negative, non-number). */
+function positiveInteger(value) {
+	if (typeof value !== "number" || !Number.isInteger(value) || value <= 0) return void 0;
+	return value;
 }
 /**
-* Read live Model Switch settings without requiring the host row.
+* Split one untrusted stored policy into the halves both official lines accept.
 *
-* The host is read structurally instead of injected so this row keeps mirroring
-* the official naming behavior when the `model-switch` row is disabled: an
-* unreachable host resolves to the row's own policy rather than to a pinned
-* route.
-* @param ctx - this row's context.
-* @returns The settings, or undefined when the host is unavailable.
+* Execution controls pass through untouched (an invalid value still fails the
+* official validator). Absent execution fields fall back to the shipped
+* official defaults, because a profile that mounts no official row has no
+* policy to forward and this row must still run. Title-length targets are read
+* here for the system instruction: the pre-0.2.1 line wants them in its shared
+* config, while 0.2.1 moved them into each provider.
+* @param forwarded - the disabled official row's stored config.
+* @returns The execution subset and the title-length targets.
 */
-function currentSettings(ctx) {
-	try {
-		return ctx.get("modelSwitch", false)?.currentSettings();
-	} catch {
-		return;
+function splitTitlePolicy(forwarded) {
+	const execution = {};
+	for (const key of EXECUTION_KEYS) {
+		const value = forwarded[key];
+		if (value !== void 0) execution[key] = value;
 	}
+	for (const [key, fallback] of Object.entries(EXECUTION_DEFAULTS)) if (execution[key] === void 0) execution[key] = fallback;
+	return {
+		execution,
+		targets: {
+			targetWords: positiveInteger(forwarded.targetWords) ?? TARGET_DEFAULTS.targetWords,
+			targetCjkCharacters: positiveInteger(forwarded.targetCjkCharacters) ?? TARGET_DEFAULTS.targetCjkCharacters
+		}
+	};
+}
+/**
+* Detect the 0.2.1+ auxiliary-LLM line.
+* @param module - the loaded auxiliary-LLM package.
+* @returns Whether the module exposes the provider-prepared call.
+*/
+function isModernTitleModule(module) {
+	return titleCallOf(module, "executeSessionTitleLlm") !== void 0;
+}
+/**
+* Validate the stored policy through the official validator of the installed line.
+* @param module - the loaded auxiliary-LLM package.
+* @param forwarded - the disabled official row's stored config.
+* @returns The validated execution config and title-length targets.
+* @throws when the official validator rejects the stored policy.
+*/
+function resolveNamingPolicy(module, forwarded) {
+	const split = splitTitlePolicy(forwarded);
+	const resolve = module.resolveSessionTitleLlmConfig;
+	if (isModernTitleModule(module)) return {
+		config: resolve(split.execution),
+		targets: split.targets
+	};
+	const config = resolve({
+		...split.execution,
+		...split.targets
+	});
+	return {
+		config,
+		targets: {
+			targetWords: config.targetWords ?? split.targets.targetWords,
+			targetCjkCharacters: config.targetCjkCharacters ?? split.targets.targetCjkCharacters
+		}
+	};
+}
+/**
+* Candidate routes for one naming request, most specific first and deduplicated.
+*
+* The stored session group's pair wins, then the deployment-wide pair, then the
+* row's own configured pair, then the session's logged route. The list is also
+* the fallback order: a failing candidate falls through to the next one.
+* @param settings - live Model Switch settings; undefined skips the stored pairs.
+* @param provider - the delegating session's provider, when one is observable.
+* @param policy - the validated row policy (its own pair is the third candidate).
+* @param requestRoute - the session's logged main-request route, when one exists.
+* @returns The ordered, deduplicated candidates; empty when only the row policy applies.
+*/
+function titleRouteCandidates(settings, provider, policy, requestRoute) {
+	const candidates = [];
+	const push = (route) => {
+		if (route === void 0) return;
+		if (candidates.some((candidate) => candidate.provider === route.provider && candidate.model === route.model)) return;
+		candidates.push(route);
+	};
+	push(groupRoute(settings, provider, "title"));
+	push(deploymentRoute(settings, "title"));
+	if (policy.provider !== void 0 && policy.model !== void 0) push({
+		provider: policy.provider,
+		model: policy.model
+	});
+	push(requestRoute);
+	return candidates;
+}
+/**
+* Whether a failure may be retried on another candidate.
+* A superseded or disposed request keeps no budget, and a deadline that fired is
+* the caller's bound, not the route's fault: both end the attempt loop.
+* @param error - the failure thrown by the auxiliary call.
+* @param request - the service-owned request whose signal bounds the work.
+* @returns Whether another candidate may be tried.
+*/
+function shouldTryNextCandidate(error, request) {
+	if (request.signal.aborted) return false;
+	return error?.code !== "SESSION_TITLE_TIMEOUT";
+}
+/**
+* Whether the failure is the model filling its output budget before finishing.
+* Official lines report this as `title output reached maxOutputTokens`; a widened
+* budget is the only retry that can succeed for it.
+* @param error - the failure thrown by the auxiliary call.
+* @returns Whether one widened-budget retry is worth trying.
+*/
+function isMaxTokensFailure(error) {
+	if (!(error instanceof Error)) return false;
+	return /maxoutputtokens|max[-_ ]tokens/iu.test(error.message);
+}
+/**
+* A widened output budget for the same route.
+* @param maxOutputTokens - the policy's configured budget.
+* @returns Four times the budget, capped at 4096 tokens.
+*/
+function widerTitleBudget(maxOutputTokens) {
+	return Math.min(maxOutputTokens * 4, 4096);
+}
+/** Asterisk emphasis wrapping a whole line, such as `**Title**`. */
+const EMPHASIS_WRAPPER = /^(?<marker>\*{1,3})(?<inner>\S(?:.*\S)?)\k<marker>$/u;
+/** OSC control sequences, such as a title hyperlink. */
+const OSC_SEQUENCE = /(?:\u001B\]|\u009D)(?:(?!\u0007|\u001B\\)[\s\S])*(?:\u0007|\u001B\\|$)/gu;
+/** Control-sequence-introducer escapes such as SGR color codes. */
+const CSI_SEQUENCE = /(?:\u001B\[|\u009B)[0-?]*[ -/]*[@-~]/gu;
+/** Remaining two-byte ESC control sequences. */
+const ESC_SEQUENCE = /\u001B[@-_]/gu;
+/** Non-whitespace C0/C1 control characters. */
+const CONTROL_CHARACTER = /[\u0000-\u0008\u000B\u000C\u000E-\u001F\u007F-\u009F]/gu;
+/** Directional and invisible controls that can make a displayed title deceptive. */
+const DIRECTIONAL_CONTROL = /[\u200B\u200E\u200F\u202A-\u202E\u2060-\u2064\u2066-\u206F\uFEFF]/gu;
+/**
+* Normalize one accepted title exactly like the official service.
+* @param input - untrusted title text.
+* @returns A terminal-safe, whitespace-normalized one-line title, possibly empty.
+*/
+function normalizeTitleText(input) {
+	return input.replace(OSC_SEQUENCE, "").replace(CSI_SEQUENCE, "").replace(ESC_SEQUENCE, "").replace(CONTROL_CHARACTER, "").replace(DIRECTIONAL_CONTROL, "").replace(/\s+/gu, " ").trim();
+}
+/**
+* Take the title from model output: the first non-empty line, without asterisk
+* emphasis wrapping that whole line. Models that disregard the one-line
+* instruction put the title first and commentary after it.
+* @param text - joined text blocks of the auxiliary response.
+* @returns The candidate title line.
+*/
+function titleFromOutput(text) {
+	const line = text.split(/\r?\n/u).map((item) => item.trim()).find((item) => item.length > 0) ?? "";
+	const groups = EMPHASIS_WRAPPER.exec(line)?.groups;
+	const marker = groups?.["marker"];
+	const inner = groups?.["inner"];
+	return marker === void 0 || inner === void 0 || inner.includes(marker) ? line : inner;
+}
+/**
+* Interpret one assembled 0.2.1+ response as a title.
+* Rejects terminal tool-call and max-token outcomes exactly like the official
+* provider, so the row keeps the official acceptance rules.
+* @param response - the assembled auxiliary response.
+* @returns A non-empty normalized title.
+* @throws when the finish reason, a tool-call block, or empty text makes the response unusable.
+*/
+function titleFromResponse(response) {
+	switch (response.finish.kind) {
+		case "stop": break;
+		case "tool-calls": throw new Error("session-title-llm: title output must contain text only");
+		case "max-tokens": throw new Error("session-title-llm: title output reached maxOutputTokens");
+		default: throw new Error(`session-title-llm: unsupported finish reason "${String(response.finish.kind)}"`);
+	}
+	if (response.blocks.some((block) => block.type === "tool-call")) throw new Error("session-title-llm: title output must contain text only");
+	const title = normalizeTitleText(titleFromOutput(response.blocks.filter((block) => block.type === "text").map((block) => block.text ?? "").join(" ")));
+	if (title.length === 0) throw new Error("session-title-llm: title model produced no text");
+	return title;
+}
+/**
+* Select the route's least supported reasoning effort for a title call, as the
+* official 0.2.1 provider does: naming is a short auxiliary task, and a full
+* reasoning budget can consume the whole output allowance.
+* @param model - the prepared model metadata.
+* @returns The first declared effort id, or undefined when the model has none.
+*/
+function selectTitleReasoningEffort(model) {
+	return model.reasoning?.efforts?.[0]?.id;
+}
+/**
+* The official 0.2.1+ system instruction, including the optional current-title anchor.
+* @param targets - title-length targets.
+* @param hasCurrentTitle - whether the request is anchored to an existing provider title.
+* @returns The system prompt text.
+*/
+function titleSystemPrompt(targets, hasCurrentTitle) {
+	return [
+		"Create a concise title for an AI coding-assistant session from the supplied human messages.",
+		"Return only the title on one line, **in plain text of natural language**, with no quotes, prefix, explanation, Markdown, XML, or terminal control codes. No code is allowed.",
+		"Use the language of the messages.",
+		`Aim for about ${targets.targetWords} words in non-CJK languages or ${targets.targetCjkCharacters} CJK characters.`,
+		"If the messages give little to name, still return a short best-effort title, such as Greeting, instead of explaining.",
+		...hasCurrentTitle ? [
+			"An existing title is supplied as currentTitle. If it still accurately describes the main topic or task, return it exactly unchanged.",
+			"Follow-up questions, additional details within the same topic, acknowledgements such as \"thanks\", and requests to continue do not by themselves justify a title change.",
+			"Do not reword, polish, shorten, or replace synonyms in an adequate title. Keeping its exact wording takes priority over the target length.",
+			"Change the title only when the messages materially change or expand the main topic or task so that the existing title is no longer accurate."
+		] : []
+	].join("\n");
+}
+/**
+* Frame the exact messages as JSON so user text cannot break structural delimiters.
+* @param messages - the selected human messages.
+* @param currentTitle - the accepted provider title to anchor on, when there is one.
+* @returns The framed user input.
+*/
+function frameTitleInput(messages, currentTitle) {
+	if (currentTitle === void 0) return `Generate the session title from this JSON array of human messages:\n${JSON.stringify(messages)}`;
+	return `Update the session title from this JSON object:\n${JSON.stringify({
+		currentTitle,
+		messages
+	})}`;
+}
+/**
+* The accepted provider title of one request, when the service supplied one.
+* Only a provider-sourced title anchors an update; a fallback title does not.
+* @param request - the service-owned naming request.
+* @returns The current title, or undefined when there is nothing to anchor on.
+*/
+function currentProviderTitle(request) {
+	const current = request.currentTitle;
+	if (current === void 0 || current.source?.kind !== "provider" || typeof current.title !== "string") return void 0;
+	return current.title;
+}
+/**
+* Run one candidate through the auxiliary-LLM line the profile installed.
+* @param ctx - this row's context, exposing `llm`.
+* @param module - the loaded auxiliary-LLM package.
+* @param plan - the profile's naming policy and cadence.
+* @param request - service-owned session, message revision, and cancellation.
+* @param attempt - one candidate route with its composed config.
+* @returns The official provider result.
+*/
+async function runTitleCall(ctx, module, plan, request, attempt) {
+	const messages = selectTitleMessages(request.messages, plan.cadence.firstPromptOnly);
+	const execute = titleCallOf(module, "executeSessionTitleLlm");
+	if (execute !== void 0) {
+		const anchor = plan.cadence.firstPromptOnly ? void 0 : currentProviderTitle(request);
+		const prepared = {
+			system: titleSystemPrompt(plan.targets, anchor !== void 0),
+			input: frameTitleInput(messages, anchor),
+			messageSeqs: messages.map((message) => message.seq),
+			selectReasoningEffort: selectTitleReasoningEffort
+		};
+		const response = await execute(ctx, attempt.config, request, plan.cadence.id, prepared);
+		return {
+			title: titleFromResponse(response),
+			messageSeqs: prepared.messageSeqs,
+			model: response.model
+		};
+	}
+	const legacy = titleCallOf(module, "generateSessionTitleWithLlm");
+	if (legacy === void 0) throw new Error(`model-switch: ${TITLE_LLM_PACKAGE} exposes no title call this row understands`);
+	return legacy(ctx, {
+		...plan.targets,
+		...attempt.config
+	}, request, messages, plan.cadence.id);
 }
 /**
 * Execute one naming request through the official implementation.
 *
 * The route is resolved from the session being titled — its provider selects the
-* saved group, and the flat settings pair is the deployment default — and is
-* handed to `generateSessionTitleWithLlm`, which frames the request, records the
-* exact route in `session/title-llm-request`, and validates the result.
+* saved group, the flat settings pair is the deployment default — and every
+* candidate is tried in order until one succeeds. A max-token failure first gets
+* one widened-budget retry on the same route; every other failure moves on.
 * @param ctx - this row's context, exposing `llm`.
 * @param module - the loaded auxiliary-LLM module.
 * @param plan - the profile's fixed naming policy.
 * @param request - service-owned session, message revision, and cancellation.
 * @returns The official provider result.
+* @throws The last candidate failure when every candidate fails.
 */
 async function generateTitle(ctx, module, plan, request) {
-	const provider = sessionProviderOf(ctx, request.session);
-	const route = resolveCapabilityRoute(currentSettings(ctx), provider, "title");
-	return module.generateSessionTitleWithLlm(ctx, composeTitleRoute(plan.config, route), request, selectTitleMessages(request.messages, plan.cadence.firstPromptOnly), plan.cadence.id);
+	const candidates = titleRouteCandidates(currentSettings(ctx), sessionProviderOf(ctx, request.session), plan.config, request.route);
+	const attempts = candidates.length === 0 ? [{
+		route: void 0,
+		config: plan.config
+	}] : candidates.map((route) => ({
+		route,
+		config: composeTitleRoute(plan.config, route)
+	}));
+	let lastError;
+	for (const attempt of attempts) try {
+		return await runTitleCall(ctx, module, plan, request, attempt);
+	} catch (error) {
+		lastError = error;
+		if (!shouldTryNextCandidate(error, request)) break;
+		if (!isMaxTokensFailure(error)) continue;
+		const widened = {
+			route: attempt.route,
+			config: {
+				...attempt.config,
+				maxOutputTokens: widerTitleBudget(attempt.config.maxOutputTokens)
+			}
+		};
+		try {
+			return await runTitleCall(ctx, module, plan, request, widened);
+		} catch (retryError) {
+			lastError = retryError;
+			if (!shouldTryNextCandidate(retryError, request)) break;
+		}
+	}
+	const tried = attempts.map((attempt) => attempt.route === void 0 ? "the row policy" : `${attempt.route.provider}/${attempt.route.model}`).join(" → ");
+	if (!request.signal.aborted) ctx.logger.warn(`model-switch: session title generation failed on every candidate (${tried}): ${String(lastError)}`);
+	throw lastError;
+}
+/**
+* Read live Model Switch settings without requiring the host row.
+*
+* The host is read structurally instead of injected so this row keeps mirroring
+* the official naming behavior when the `model-switch` row is disabled: an
+* unreachable host resolves to the row's own policy rather than to a pinned
+* route.
+* @param ctx - this row's context.
+* @returns The settings, or undefined when the host is unavailable.
+*/
+function currentSettings(ctx) {
+	try {
+		return ctx.get("modelSwitch", false)?.currentSettings();
+	} catch {
+		return;
+	}
 }
 /**
 * Mount the Model Switch session-naming row for this profile.
@@ -180,7 +532,7 @@ async function apply(ctx) {
 	const cadence = titleCadence(entry === void 0 ? void 0 : officialTitlePackage(entry)) ?? TITLE_CADENCES[0];
 	if (cadence === void 0) return;
 	if (entry === void 0) ctx.logger.info("model-switch: no instantiated official session-title row to mirror; naming uses the first-prompt cadence");
-	/** The displaced row's naming policy, or the official defaults when it is absent. */
+	/** The displaced row's naming policy, or the shipped defaults when it is absent. */
 	const forwarded = entry === void 0 ? {} : forwardedConfig(entry);
 	let module;
 	try {
@@ -190,9 +542,9 @@ async function apply(ctx) {
 		ctx.logger.error(cause);
 		return;
 	}
-	let config;
+	let policy;
 	try {
-		config = module.resolveSessionTitleLlmConfig(forwarded);
+		policy = resolveNamingPolicy(module, forwarded);
 	} catch (cause) {
 		ctx.logger.error("model-switch: the official session-title row has no valid naming policy; naming keeps the deterministic fallback");
 		ctx.logger.error(cause);
@@ -200,7 +552,8 @@ async function apply(ctx) {
 	}
 	const plan = {
 		cadence,
-		config
+		config: policy.config,
+		targets: policy.targets
 	};
 	try {
 		ctx.sessionTitle.register({
@@ -214,4 +567,4 @@ async function apply(ctx) {
 	}
 }
 //#endregion
-export { TITLE_CADENCES, apply, composeTitleRoute, forwardedConfig, inject, name, officialTitleEntry, officialTitlePackage, selectTitleMessages, titleCadence };
+export { TITLE_CADENCES, apply, composeTitleRoute, forwardedConfig, frameTitleInput, inject, isMaxTokensFailure, isModernTitleModule, name, normalizeTitleText, officialTitleEntry, officialTitlePackage, resolveNamingPolicy, selectTitleMessages, selectTitleReasoningEffort, shouldTryNextCandidate, splitTitlePolicy, titleCadence, titleFromOutput, titleFromResponse, titleRouteCandidates, titleSystemPrompt, widerTitleBudget };

package/lib/subagent-runtime.js

@@ -1,4 +1,4 @@
-import { r as sessionSubagentRouteOf } from "./session-route-h9SAqajc.js";
+import { a as sessionSubagentRouteOf } from "./session-route-BjNPlliu.js";
 import { ReasoningEffortId } from "@deepseek-ai/dsh-llm";
 import OfficialSubagentRuntime from "@deepseek-ai/dsh-subagent";
 //#region lib/types/subagent-runtime.js
@@ -144,9 +144,32 @@ async function mountWithStartupFallback(candidate, fallback) {
 		startupFailure(error, await ledger.disposeAll());
 	}
 }
+/**
+* Operator-facing notice for a Host whose Subagent runtime cannot be routed.
+*
+* DSH 0.2.1 replaced `start()`/`startContinuable()` with the activation API, so
+* this adapter's override point is gone; the row then mounts the untouched
+* official runtime instead of guessing a new seam.
+* @param error - the typed startup incompatibility that rejected the candidate.
+* @returns The notice logged once at startup.
+*/
+function routedRuntimeUnavailableNotice(error) {
+	return `model-switch: the routed Subagent runtime is unavailable on this DSH Host (${error.message}); delegation follows the official route instead`;
+}
+/** Log one notice without letting a missing logger service fail the startup. */
+function reportNotice(ctx, message) {
+	try {
+		ctx.logger.warn(message);
+	} catch {}
+}
 async function mountProfileRuntime(ctx) {
 	return (await mountWithStartupFallback(async (track) => {
-		assertOfficialRuntimeSurface();
+		try {
+			assertOfficialRuntimeSurface();
+		} catch (error) {
+			if (error instanceof StartupIncompatibilityError) reportNotice(ctx, routedRuntimeUnavailableNotice(error));
+			throw error;
+		}
 		assertRoutingSurface(ctx);
 		const candidate = ctx.plugin(ModelSwitchSubagentRuntime);
 		track(() => candidate.dispose());
@@ -186,4 +209,4 @@ var ModelSwitchSubagentRuntime = class extends OfficialSubagentRuntime {
 	}
 };
 //#endregion
-export { ModelSwitchSubagentRuntime, StartupIncompatibilityError, profileSubagentRuntime as default, profileSubagentRuntime, mountWithStartupFallback, routeSubagentRequest };
+export { ModelSwitchSubagentRuntime, StartupIncompatibilityError, profileSubagentRuntime as default, profileSubagentRuntime, mountWithStartupFallback, routeSubagentRequest, routedRuntimeUnavailableNotice };

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";