@stackbone/cli 0.9.1 → 0.9.2

package/main.js

@@ -3229,7 +3229,7 @@ var init_studio_internal_prompts = __esmMin((() => {
 }));
 //#endregion
 //#region ../../libs/shared/validators/src/lib/studio/studio-model-routing.ts
-var MODEL_PROVIDER_SORTS, MODEL_PROVIDER_SORTS_TEXT, modelProviderSortSchema, isModelProviderSort, studioModelRoutingPathSchema, studioModelRoutingSchema, putStudioModelRoutingRequestSchema, studioModelRoutingDeleteResponseSchema;
+var MODEL_PROVIDER_SORTS, MODEL_PROVIDER_SORTS_TEXT, modelProviderSortSchema, MODEL_PROVIDER_QUANTIZATIONS, providerSlugSchema, providerSlugsSchema, percentileTargetSchema, dollarsSchema, modelProviderRoutingSchema, modelRoutingInputSchema, toModelProviderRouting, readModelProviderRouting, plainSortOf, studioModelRoutingPathSchema, studioModelRoutingSchema, putStudioModelRoutingRequestSchema, routingOfPutRequest, studioModelRoutingDeleteResponseSchema;
 var init_studio_model_routing = __esmMin((() => {
 	init_studio_model_selections();
 	MODEL_PROVIDER_SORTS = [
@@ -3239,14 +3239,69 @@ var init_studio_model_routing = __esmMin((() => {
 	];
 	MODEL_PROVIDER_SORTS_TEXT = MODEL_PROVIDER_SORTS.map((sort) => `'${sort}'`).join(", ");
 	modelProviderSortSchema = z.enum(MODEL_PROVIDER_SORTS);
-	isModelProviderSort = (value) => modelProviderSortSchema.safeParse(value).success;
+	MODEL_PROVIDER_QUANTIZATIONS = [
+		"int4",
+		"int8",
+		"fp4",
+		"mxfp4",
+		"nvfp4",
+		"fp6",
+		"fp8",
+		"mxfp8",
+		"fp16",
+		"bf16",
+		"fp32",
+		"unknown"
+	];
+	providerSlugSchema = z.string().regex(/^\S+$/, "a provider slug has no spaces").max(128);
+	providerSlugsSchema = z.array(providerSlugSchema).min(1).max(32);
+	percentileTargetSchema = z.union([z.number().positive(), z.strictObject({
+		p50: z.number().positive(),
+		p75: z.number().positive(),
+		p90: z.number().positive(),
+		p99: z.number().positive()
+	}).partial().refine((value) => Object.keys(value).length > 0, "name at least one percentile")]);
+	dollarsSchema = z.number().nonnegative();
+	modelProviderRoutingSchema = z.strictObject({
+		order: providerSlugsSchema,
+		only: providerSlugsSchema,
+		ignore: providerSlugsSchema,
+		allow_fallbacks: z.boolean(),
+		require_parameters: z.boolean(),
+		data_collection: z.enum(["allow", "deny"]),
+		zdr: z.boolean(),
+		enforce_distillable_text: z.boolean(),
+		quantizations: z.array(z.enum(MODEL_PROVIDER_QUANTIZATIONS)).min(1),
+		sort: z.union([modelProviderSortSchema, z.strictObject({
+			by: modelProviderSortSchema,
+			partition: z.enum(["model", "none"]).optional()
+		})]),
+		preferred_min_throughput: percentileTargetSchema,
+		preferred_max_latency: percentileTargetSchema,
+		max_price: z.strictObject({
+			prompt: dollarsSchema,
+			completion: dollarsSchema,
+			request: dollarsSchema,
+			image: dollarsSchema
+		}).partial()
+	}).partial().refine((value) => Object.keys(value).length > 0, "name at least one routing field");
+	modelRoutingInputSchema = z.union([modelProviderSortSchema, modelProviderRoutingSchema]);
+	toModelProviderRouting = (input) => typeof input === "string" ? { sort: input } : input;
+	readModelProviderRouting = (value) => {
+		const parsed = modelRoutingInputSchema.safeParse(value);
+		return parsed.success ? toModelProviderRouting(parsed.data) : null;
+	};
+	plainSortOf = (routing) => routing && Object.keys(routing).length === 1 && typeof routing.sort === "string" ? routing.sort : null;
 	studioModelRoutingPathSchema = studioModelSelectionPathSchema.extend({ ownerKind: z.literal("agent") });
 	studioModelRoutingSchema = studioModelRoutingPathSchema.extend({
-		sort: modelProviderSortSchema,
+		routing: modelProviderRoutingSchema,
+		/** `routing` as a plain sort, for a Studio older than the object; `null` otherwise. */
+		sort: modelProviderSortSchema.nullable(),
 		updatedBy: z.string().nullable(),
 		updatedAt: z.iso.datetime()
 	});
-	putStudioModelRoutingRequestSchema = z.strictObject({ sort: modelProviderSortSchema });
+	putStudioModelRoutingRequestSchema = z.union([z.strictObject({ routing: modelProviderRoutingSchema }), z.strictObject({ sort: modelProviderSortSchema })]);
+	routingOfPutRequest = (body) => "routing" in body ? body.routing : { sort: body.sort };
 	studioModelRoutingDeleteResponseSchema = studioModelRoutingPathSchema.extend({ status: z.literal("deleted") });
 }));
 //#endregion
@@ -3613,7 +3668,7 @@ var init_studio_rag = __esmMin((() => {
 		error: z.string().nullable()
 	});
 	ragRebuildStatusResponseSchema = z.object({ rebuild: ragRebuildStatusSchema.nullable() });
-})), VOICE_STT_MODELS, VOICE_TTS_MODELS, isCartesiaTtsModel, VOICE_BACKGROUND_CLIPS, VOICE_AMBIENT_SOUNDS, VOICE_THINKING_SOUNDS, thresholdSchema, millisecondsSchema, keytermsSchema, vadSchema, turnSchema, interruptionsSchema, ragSchema, volumeSchema, backgroundAudioSchema, VOICE_BACKGROUND_AUDIO_DEFAULTS, effectiveVoiceSettingsObjectSchema, checkCrossFieldRules, effectiveVoiceSettingsSchema, callVoiceSettingsObjectSchema, callVoiceSettingsSchema, toCallVoiceSettings, VOICE_SETTINGS_LEAF_PATHS, VOICE_TTS_PAIR, voiceSettingOriginSchema, effectiveVoiceSettingsViewSchema, checkTtsPair, voiceSettingsSchema, checkStudioTtsPair, studioVoiceSettingsPatchSchema, voiceSettingsResponseSchema, describeVoiceSettingsIssue, deepAgentNameSchema, voiceCallIdSchema, voiceTagsSchema, VOICE_TTS_MARKUP_CONTEXT, VOICE_TTS_MARKUP_GUIDES, isVoiceBoxUrl, VOICE_BOX_URL_MAX, voiceBoxUrlSchema, VOICE_TRACES_PATH, VOICE_WORKER_AGENT_NAME_ENV, VOICE_WORKER_PORT_ENV, VOICE_CALL_TOKEN_MAX, voiceDispatchSchema, voiceRoomRequestSchema, voiceRoomResponseSchema, VOICE_RECORDING_STATUSES, voiceRecordingQuerySchema, VOICE_RECORDING_URL_TTL_SECONDS, voiceRecordingResponseSchema, voiceStatusResponseSchema, HEARD_TEXT_MAX, aguiHeardSchema, VOICE_CALLS_PATH, voiceCallBootstrapSchema, voiceCallTranscriptSchema, VOICE_CALL_TURNS_MAX, VOICE_TURN_LATENCY_MAX_MS, turnLatencyMs, voiceCallTurnMetricsSchema, voiceCallTurnsSchema, readVoiceCallTurns, voiceCallCompleteSchema, voiceCallLifecycleResponseSchema;
+})), VOICE_STT_MODELS, VOICE_TTS_MODELS, isCartesiaTtsModel, VOICE_BACKGROUND_CLIPS, VOICE_AMBIENT_SOUNDS, VOICE_THINKING_SOUNDS, thresholdSchema, millisecondsSchema, keytermsSchema, vadSchema, turnSchema, VOICE_EAGER_THRESHOLD_DEFAULT, interruptionsSchema, ragSchema, volumeSchema, backgroundAudioSchema, VOICE_BACKGROUND_AUDIO_DEFAULTS, effectiveVoiceSettingsObjectSchema, checkCrossFieldRules, effectiveVoiceSettingsSchema, callVoiceSettingsObjectSchema, callVoiceSettingsSchema, toCallVoiceSettings, VOICE_SETTINGS_LEAF_PATHS, VOICE_TTS_PAIR, voiceSettingOriginSchema, LEAVES_NEWER_THAN_SOME_BOXES, effectiveVoiceSettingsViewSchema, checkTtsPair, voiceSettingsSchema, checkStudioTtsPair, studioVoiceSettingsPatchSchema, voiceSettingsResponseSchema, describeVoiceSettingsIssue, deepAgentNameSchema, voiceCallIdSchema, voiceTagsSchema, VOICE_TTS_MARKUP_CONTEXT, VOICE_TTS_MARKUP_GUIDES, isVoiceBoxUrl, VOICE_BOX_URL_MAX, voiceBoxUrlSchema, VOICE_TRACES_PATH, VOICE_WORKER_AGENT_NAME_ENV, VOICE_WORKER_PORT_ENV, VOICE_CALL_TOKEN_MAX, voiceDispatchSchema, voiceRoomRequestSchema, voiceRoomResponseSchema, VOICE_RECORDING_STATUSES, voiceRecordingQuerySchema, VOICE_RECORDING_URL_TTL_SECONDS, voiceRecordingResponseSchema, voiceStatusResponseSchema, HEARD_TEXT_MAX, aguiHeardSchema, VOICE_CALLS_PATH, voiceCallBootstrapSchema, voiceCallKnowledgeSchema, voiceCallTranscriptSchema, VOICE_CALL_TURNS_MAX, VOICE_TURN_LATENCY_MAX_MS, turnLatencyMs, voiceCallTurnMetricsSchema, voiceCallTurnsSchema, readVoiceCallTurns, voiceCallCompleteSchema, voiceCallLifecycleResponseSchema;
 var init_studio_voice = __esmMin((() => {
 	init_memory_identities();
 	init_studio_rag();
@@ -3653,8 +3708,10 @@ var init_studio_voice = __esmMin((() => {
 		mode: z.enum(["stt", "livekit"]),
 		minEndpointingDelayMs: millisecondsSchema,
 		maxEndpointingDelayMs: millisecondsSchema,
-		detectorThreshold: thresholdSchema
+		detectorThreshold: thresholdSchema,
+		eagerThreshold: z.number().min(.3).max(.7)
 	});
+	VOICE_EAGER_THRESHOLD_DEFAULT = .3;
 	interruptionsSchema = z.strictObject({
 		enabled: z.boolean(),
 		minDurationMs: millisecondsSchema,
@@ -3695,7 +3752,7 @@ var init_studio_voice = __esmMin((() => {
 		sttModel: z.enum(VOICE_STT_MODELS),
 		keyterms: keytermsSchema.default(() => []),
 		vad: vadSchema,
-		turn: turnSchema,
+		turn: turnSchema.extend({ eagerThreshold: turnSchema.shape.eagerThreshold.default(VOICE_EAGER_THRESHOLD_DEFAULT) }),
 		interruptions: interruptionsSchema,
 		ttsModel: z.enum(VOICE_TTS_MODELS),
 		/**
@@ -3774,6 +3831,7 @@ var init_studio_voice = __esmMin((() => {
 		"turn.minEndpointingDelayMs",
 		"turn.maxEndpointingDelayMs",
 		"turn.detectorThreshold",
+		"turn.eagerThreshold",
 		"interruptions.enabled",
 		"interruptions.minDurationMs",
 		"interruptions.minWords",
@@ -3803,9 +3861,10 @@ var init_studio_voice = __esmMin((() => {
 		"code",
 		"default"
 	]);
+	LEAVES_NEWER_THAN_SOME_BOXES = new Set(["turn.eagerThreshold"]);
 	effectiveVoiceSettingsViewSchema = z.strictObject({
 		value: effectiveVoiceSettingsObjectSchema,
-		origins: z.strictObject(Object.fromEntries(VOICE_SETTINGS_LEAF_PATHS.map((path) => [path, voiceSettingOriginSchema])))
+		origins: z.strictObject(Object.fromEntries(VOICE_SETTINGS_LEAF_PATHS.map((path) => [path, LEAVES_NEWER_THAN_SOME_BOXES.has(path) ? voiceSettingOriginSchema.default("default") : voiceSettingOriginSchema])))
 	});
 	checkTtsPair = (value, ctx) => {
 		if (value.ttsModel !== void 0 && value.ttsVoice === void 0) ctx.addIssue({
@@ -3915,6 +3974,7 @@ var init_studio_voice = __esmMin((() => {
 	aguiHeardSchema = z.strictObject({ text: z.string().min(1).max(HEARD_TEXT_MAX) });
 	VOICE_CALLS_PATH = "/voice/v1/calls";
 	voiceCallBootstrapSchema = z.strictObject({ identities: integrationIdentitiesSchema.optional() });
+	voiceCallKnowledgeSchema = z.strictObject({ text: z.string().min(1).max(HEARD_TEXT_MAX) });
 	voiceCallTranscriptSchema = z.array(z.strictObject({
 		role: z.enum(["user", "assistant"]),
 		text: z.string().trim().min(1).max(16e3)
@@ -3929,7 +3989,9 @@ var init_studio_voice = __esmMin((() => {
 		transcriptionMs: turnLatencyMs,
 		llmTtftMs: turnLatencyMs,
 		ttsTtfbMs: turnLatencyMs,
-		e2eMs: turnLatencyMs
+		e2eMs: turnLatencyMs,
+		/** How long before the final transcript the turn's last eager end of turn left for the box (STA3-94). */
+		preflightLeadMs: turnLatencyMs
 	});
 	voiceCallTurnsSchema = z.array(voiceCallTurnMetricsSchema).max(VOICE_CALL_TURNS_MAX);
 	readVoiceCallTurns = (value) => {
@@ -3965,7 +4027,8 @@ var init_effective_voice_settings = __esmMin((() => {
 			mode: "stt",
 			minEndpointingDelayMs: 0,
 			maxEndpointingDelayMs: 1200,
-			detectorThreshold: .35
+			detectorThreshold: .35,
+			eagerThreshold: VOICE_EAGER_THRESHOLD_DEFAULT
 		}),
 		interruptions: Object.freeze({
 			enabled: true,
@@ -6018,6 +6081,15 @@ var init_studio_catalog = __esmMin((() => {
 		/** The provider routing an operator picked for this key, or `null` when nobody has. */
 		selectedProviderSort: modelProviderSortSchema.nullable().default(null),
 		/**
+		* The full provider routing the code writes for this key, or `null` when it writes
+		* none. `defaultProviderSort` is the same routing as a plain sort, for a Studio
+		* older than the object, and `null` whenever the routing is more than one sort.
+		* Absent from a box that predates it.
+		*/
+		defaultProviderRouting: modelProviderRoutingSchema.nullable().optional(),
+		/** The full provider routing an operator picked, as `defaultProviderRouting` is to the code's. */
+		selectedProviderRouting: modelProviderRoutingSchema.nullable().optional(),
+		/**
 		* `declared` — the code declares this key. `unused` — only a pick exists, for a
 		* key no declaration mentions any more; resetting it is the one thing left to do.
 		* An entry NOTHING scanned never carries an `unused` row: see the box's
@@ -9479,7 +9551,8 @@ var init_model_routing = __esmMin((() => {
 		ownerKind: text("owner_kind").notNull(),
 		ownerName: text("owner_name").notNull(),
 		key: text("key").notNull(),
-		sort: text("sort").notNull(),
+		sort: text("sort"),
+		routing: jsonb("routing"),
 		/** The acting user's id, or null on the un-authenticated `stackbone dev` datapath. */
 		updatedBy: text("updated_by"),
 		updatedAt: timestamp("updated_at", { withTimezone: true }).notNull().defaultNow()
@@ -11406,7 +11479,7 @@ var init_json = __esmMin((() => {
 //#region ../../libs/shared/model-provider-wire/src/lib/common/provider-routing.ts
 var providerRoutingFor, withProviderRouting;
 var init_provider_routing = __esmMin((() => {
-	providerRoutingFor = (sort) => Object.freeze({ sort });
+	providerRoutingFor = (routing) => Object.freeze({ ...routing });
 	withProviderRouting = (body, kept) => {
 		if (kept || !("provider" in body)) return body;
 		const next = { ...body };
@@ -19613,6 +19686,7 @@ var init_learn_control_plane_identity = __esmMin((() => {
 }));
 //#endregion
 //#region ../../libs/runtime/workspace-runtime/src/lib/catalog/catalog-projector.ts
+init_src$26();
 /**
 * `${kind}:${name}` — the identity a catalog row, a declaration, a prompt owner and
 * a selection's owner all agree on. The kind is half of it, never decoration: a deep
@@ -19774,16 +19848,26 @@ var asDeclaredModel = (ref) => ({
 	overridable: ref.overridable,
 	file: ref.file,
 	line: ref.line,
-	defaultProviderSort: ref.defaultProviderSort ?? null,
+	...defaultRoutingFields(declaredRouting(ref)),
 	selectedModelId: null,
 	selectedProviderId: null,
-	selectedProviderSort: null,
+	...selectedRoutingFields(null),
 	state: "declared"
 });
+/** The routing the code writes; a manifest older than the object carries only its sort. */
+var declaredRouting = (ref) => readModelProviderRouting(ref.defaultProviderRouting ?? ref.defaultProviderSort);
+var defaultRoutingFields = (routing) => ({
+	defaultProviderRouting: routing,
+	defaultProviderSort: plainSortOf(routing)
+});
+var selectedRoutingFields = (routing) => ({
+	selectedProviderRouting: routing,
+	selectedProviderSort: plainSortOf(routing)
+});
 var selectedFields = (pick) => ({
 	selectedModelId: pick?.modelId ?? null,
 	selectedProviderId: pick?.providerId ?? null,
-	selectedProviderSort: pick?.providerSort ?? null
+	...selectedRoutingFields(pick?.providerRouting ?? null)
 });
 /**
 * Cross what the code DECLARES against what an operator PICKED, for every row at
@@ -19817,7 +19901,7 @@ var attachModelSelections = (entries, selections) => entries.map((entry) => {
 		overridable: true,
 		file: null,
 		line: null,
-		defaultProviderSort: null,
+		...defaultRoutingFields(null),
 		...selectedFields(pick),
 		state: "unused"
 	}));
@@ -22674,7 +22758,7 @@ var init_role = __esmMin((() => {
 var STACKBONE_CLI_BUILD_VERSION, EMULATOR_CAPABILITY_ROUTES, EMULATOR_CAPABILITIES, buildContractResponse;
 var init_contract = __esmMin((() => {
 	init_src$26();
-	STACKBONE_CLI_BUILD_VERSION = "0.9.1";
+	STACKBONE_CLI_BUILD_VERSION = "0.9.2";
 	EMULATOR_CAPABILITY_ROUTES = {
 		"database.postgres_direct": "/api/db",
 		"rag.basic": "/api/rag",
@@ -28263,7 +28347,7 @@ var init_api_docs = __esmMin((() => {
 		},
 		{
 			name: VOICE_WORKER_TAG,
-			description: "What the LiveKit voice worker calls on the box besides AG-UI: identity bootstrap before output, heard-call completion after awaited closure, and the relay that takes a call's spans to the box's OTLP collector. It takes the call token the room route put in the call's dispatch, and nothing else. Never called by Studio or by the SDK."
+			description: "What the LiveKit voice worker calls on the box besides AG-UI: identity bootstrap before output, the knowledge search started on the caller's eager end of turn, heard-call completion after awaited closure, and the relay that takes a call's spans to the box's OTLP collector. It takes the call token the room route put in the call's dispatch, and nothing else. Never called by Studio or by the SDK."
 		}
 	];
 	SURFACE_TAG_DESCRIPTIONS = {
@@ -41640,6 +41724,11 @@ var init_mutations_outside_catalog = __esmMin((() => {
 			path: "/voice/v1/calls/{sessionKey}/bootstrap",
 			reason: "The accepted call freezes structured identity and recalls memory using its own call token, without an operator seat."
 		},
+		{
+			method: "POST",
+			path: "/voice/v1/calls/{sessionKey}/knowledge",
+			reason: "The call worker starts the knowledge search on the caller's eager end of turn using its own call token, without an operator seat. It persists nothing; its only caller is the LiveKit worker."
+		},
 		{
 			method: "POST",
 			path: "/voice/v1/calls/{sessionKey}/complete",
@@ -65522,12 +65611,12 @@ var init_registry = __esmMin((() => {
 		};
 		const instructions = read(() => def.resolvedInstructions?.());
 		const model = read(() => def.resolvedModel?.());
-		const providerSort = read(() => def.resolvedProviderSort?.());
-		const { providerSort: _previous, ...rest } = metadata;
+		const providerRouting = readModelProviderRouting(read(() => def.resolvedProviderRouting?.()) ?? read(() => def.resolvedProviderSort?.()));
+		const { providerRouting: _previous, ...rest } = metadata;
 		return {
 			...rest,
 			...typeof model === "string" && model !== "" ? { model } : {},
-			...isModelProviderSort(providerSort) ? { providerSort } : {},
+			...providerRouting !== null ? { providerRouting } : {},
 			...typeof instructions === "string" && instructions !== "" ? { instructions } : {}
 		};
 	};
@@ -73127,7 +73216,7 @@ var selectionsSelect = (kinds, names) => sql`
   SELECT COALESCE(s.owner_kind, r.owner_kind) AS owner_kind,
          COALESCE(s.owner_name, r.owner_name) AS owner_name,
          COALESCE(s.key, r.key) AS key,
-         s.model_id, s.provider_id, r.sort
+         s.model_id, s.provider_id, r.sort, r.routing
     FROM ${modelSelections} s
     FULL OUTER JOIN ${modelRouting} r
       ON r.owner_kind = s.owner_kind AND r.owner_name = s.owner_name AND r.key = s.key
@@ -73148,7 +73237,7 @@ var createModelSelectionsReader = (opts) => {
 			keys.set(row.key, {
 				modelId: row.model_id,
 				providerId: row.provider_id,
-				providerSort: isModelProviderSort(row.sort) ? row.sort : null
+				providerRouting: readModelProviderRouting(row.routing ?? row.sort)
 			});
 			byOwner.set(id, keys);
 		}
@@ -86683,14 +86772,19 @@ var notFound$1 = {
 	code: "studio_model_selection_not_found",
 	message: "No model is picked for that key. Its code default already applies."
 };
-var rowToRouting = (row) => ({
-	ownerKind: "agent",
-	ownerName: row.ownerName,
-	key: row.key,
-	sort: row.sort,
-	updatedBy: row.updatedBy,
-	updatedAt: row.updatedAt.toISOString()
-});
+var rowToRouting = (row) => {
+	const routing = readModelProviderRouting(row.routing ?? row.sort);
+	if (!routing) throw new Error("model_routing_row_unreadable");
+	return {
+		ownerKind: "agent",
+		ownerName: row.ownerName,
+		key: row.key,
+		routing,
+		sort: plainSortOf(routing),
+		updatedBy: row.updatedBy,
+		updatedAt: row.updatedAt.toISOString()
+	};
+};
 var routingNotFound = {
 	code: "studio_model_routing_not_found",
 	message: "No provider routing is picked for that key. The code's `modelRouting` already applies."
@@ -86820,10 +86914,10 @@ var createModelSelectionsRouter = (opts) => {
 		tags: [MODEL_SELECTIONS_TAG],
 		middleware: [requireCapability("config:write")],
 		summary: "Pick the provider routing for one key",
-		description: "Sets which OpenRouter provider one model key of a deep agent prefers (`provider: { sort }` on its model calls): `latency`, `throughput` or `price`. It outranks the code's `modelRouting`, and a subagent key with none follows the agent's. The box drops the field on any gateway that is not OpenRouter. The agent is rebuilt to apply it.",
+		description: "Sets which OpenRouter providers one model key of a deep agent goes to: OpenRouter's `provider` object on its model calls, e.g. `{ \"routing\": { \"only\": [\"deepinfra/fp8\"], \"allow_fallbacks\": false } }`, or `{ \"sort\": \"latency\" }` as shorthand for `{ \"routing\": { \"sort\": \"latency\" } }`. It outranks the code's `modelRouting`, and a subagent key with none follows the agent's. The box drops the field on any gateway that is not OpenRouter. The agent is rebuilt to apply it.",
 		request: {
 			params: studioModelRoutingPathSchema,
-			body: jsonBody(named("PutStudioModelRoutingRequest", putStudioModelRoutingRequestSchema), "The provider sort.")
+			body: jsonBody(named("PutStudioModelRoutingRequest", putStudioModelRoutingRequestSchema), "The provider routing, or a plain sort.")
 		},
 		responses: {
 			200: jsonResponse$1(named("StudioModelRouting", studioModelRoutingSchema), "The routing as it now stands."),
@@ -86832,12 +86926,14 @@ var createModelSelectionsRouter = (opts) => {
 		}
 	}), async (c) => {
 		const path = c.req.valid("param");
-		const { sort } = c.req.valid("json");
+		const routing = routingOfPutRequest(c.req.valid("json"));
+		const sort = plainSortOf(routing);
 		const actor = readActorUserId(c);
 		const row = await opts.db.transaction(async (tx) => {
 			const [written] = await tx.insert(modelRouting).values({
 				...path,
 				sort,
+				routing,
 				updatedBy: actor
 			}).onConflictDoUpdate({
 				target: [
@@ -86847,6 +86943,7 @@ var createModelSelectionsRouter = (opts) => {
 				],
 				set: {
 					sort,
+					routing,
 					updatedBy: actor,
 					updatedAt: sql`now()`
 				}
@@ -86857,7 +86954,8 @@ var createModelSelectionsRouter = (opts) => {
 				subjectType: "model_selection",
 				payload: {
 					...selectionSubject(path),
-					sort
+					...sort !== null ? { sort } : {},
+					routing
 				}
 			}, tx);
 			return written;
@@ -93386,11 +93484,17 @@ var answerOf = async (storage, row) => {
 };
 //#endregion
 //#region ../../libs/runtime/workspace-runtime/src/lib/voice/voice-rag.ts
-var WORD_CHARACTER, spokenWordCount, knowledgeQuery, selectKnowledge;
+var WORD_CHARACTER, spokenWordCount, knowledgeQuery, PUNCTUATION_OR_SYMBOL_RUN, prefetchSentence, prefetchKey, selectKnowledge;
 var init_voice_rag = __esmMin((() => {
 	WORD_CHARACTER = /[\p{L}\p{N}]/u;
 	spokenWordCount = (text) => text.trim().split(/\s+/u).filter((token) => WORD_CHARACTER.test(token)).length;
 	knowledgeQuery = ({ exchanges, sentence }) => [...exchanges.flatMap(({ question, reply }) => [question, reply]), sentence].map((piece) => piece.trim()).filter((piece) => piece !== "").join("\n");
+	PUNCTUATION_OR_SYMBOL_RUN = /[\p{P}\p{S}]+/gu;
+	prefetchSentence = (text) => text.normalize("NFC").toLowerCase().replace(PUNCTUATION_OR_SYMBOL_RUN, " ").replace(/\s+/gu, " ").trim();
+	prefetchKey = ({ exchanges, sentence }) => knowledgeQuery({
+		exchanges,
+		sentence: prefetchSentence(sentence)
+	});
 	selectKnowledge = (hits, { minScore, topK }) => {
 		const passages = hits.filter((hit) => Number.isFinite(hit.score) && hit.score >= minScore).map((hit) => hit.content).filter((content) => typeof content === "string" && content.trim() !== "").slice(0, topK);
 		return passages.length === 0 ? void 0 : JSON.stringify(passages);
@@ -93436,7 +93540,7 @@ var init_voice_calls = __esmMin((() => {
 			const words = spokenWordCount(text);
 			return words !== 0 && words >= rag.minWords;
 		};
-		const searchKnowledge = async (call, text) => {
+		const searchKnowledge = async (call, text, origin = "turn") => {
 			const { rag } = call;
 			if (!searchable(rag, text)) return void 0;
 			const { collection } = rag;
@@ -93444,6 +93548,7 @@ var init_voice_calls = __esmMin((() => {
 				exchanges: call.exchanges,
 				sentence: text
 			});
+			const search = origin === "prefetch" ? "rag_prefetch" : "rag_search";
 			const warn = (code) => options.logger.warn(`voice ${call.input.agent}/${call.input.sessionKey}: ${code}.`);
 			try {
 				const result = await withTimeout(retrieve({
@@ -93452,7 +93557,7 @@ var init_voice_calls = __esmMin((() => {
 					topK: rag.topK
 				}), rag.timeoutMs);
 				if (result.error) {
-					warn(`rag_search_failed (${result.error.code})`);
+					warn(`${search}_failed (${result.error.code})`);
 					return;
 				}
 				if (result.data.length === 0 && !call.ragEmptyWarned) {
@@ -93461,11 +93566,31 @@ var init_voice_calls = __esmMin((() => {
 				}
 				return selectKnowledge(result.data, rag);
 			} catch (error) {
-				if (error instanceof ActionTimeoutError) warn("rag_search_timeout");
-				else warn(`rag_search_failed (${error instanceof Error ? error.name : typeof error})`);
+				if (error instanceof ActionTimeoutError) warn(`${search}_timeout`);
+				else warn(`${search}_failed (${error instanceof Error ? error.name : typeof error})`);
 			}
 		};
+		let inputGuard;
 		return {
+			useInputGuard: (guard) => {
+				inputGuard = guard;
+			},
+			prefetchKnowledge: async (agent, sessionKey, text) => {
+				const call = find(agent, sessionKey);
+				const sequence = ++call.prefetches;
+				const guard = inputGuard;
+				if (!guard || call.completion || !searchable(call.rag, text)) return;
+				const guarded = await Promise.resolve().then(() => guard(agent, text)).catch(() => "skip");
+				if (sequence !== call.prefetches || call.completion) return;
+				if (guarded === "skip" || guarded.blocked || !searchable(call.rag, guarded.content)) return;
+				call.prefetch = {
+					key: prefetchKey({
+						exchanges: call.exchanges,
+						sentence: guarded.content
+					}),
+					promise: searchKnowledge(call, guarded.content, "prefetch")
+				};
+			},
 			accept: async (input) => {
 				const registry = options.getRegistry();
 				if (!registry) throw new Error("Voice call memory service is unavailable.");
@@ -93479,6 +93604,7 @@ var init_voice_calls = __esmMin((() => {
 					rag: input.rag && Object.freeze({ ...input.rag }),
 					refusalPhrase: input.refusalPhrase ?? VOICE_SETTINGS_DEFAULTS.refusalPhrase,
 					exchanges: [],
+					prefetches: 0,
 					startedAt,
 					expiresAt: startedAt + CALL_TOKEN_TTL_SECONDS * 1e3,
 					work: /* @__PURE__ */ new Set()
@@ -93545,7 +93671,21 @@ var init_voice_calls = __esmMin((() => {
 								done();
 							};
 						},
-						knowledgeFor: (text) => searchKnowledge(call, text)
+						knowledgeFor: (text, notePrefetch) => {
+							const { prefetch } = call;
+							delete call.prefetch;
+							call.prefetches++;
+							if (!searchable(call.rag, text)) return Promise.resolve(void 0);
+							if (prefetch?.key === prefetchKey({
+								exchanges: call.exchanges,
+								sentence: text
+							})) {
+								notePrefetch?.("hit");
+								return prefetch.promise;
+							}
+							notePrefetch?.(prefetch ? "miss" : "none");
+							return searchKnowledge(call, text);
+						}
 					});
 					call.context = context;
 					if (call.input.startRecording) Promise.resolve().then(() => call.input.startRecording(sessionAddressFor(agent, sessionKey, void 0, sessionIdentityFor(identity)))).catch(() => options.logger.warn(`voice ${agent}/${sessionKey}: recording_start_failed.`));
@@ -93634,7 +93774,7 @@ var createVoiceRoomsRouter = (options) => {
 		middleware: [requireCapability(VOICE_CAPABILITY$1)],
 		"x-stackbone-catalog": excludeFromCatalog("Answers a LiveKit participant token. The operation catalog never offers an operation whose answer is a credential."),
 		summary: "Open a voice call with a deep agent",
-		description: "Opens a fresh room `call-<agent>-<8 hex>` and answers the LiveKit URL, the browser's token to join it (it joins within five minutes; a call already running is not cut) and the settings the call runs with. The token carries the room's agent dispatch: joining the room sends the voice worker in with the agent's name, its effective voice settings, the session key `voice-<room>` every turn lands in, the tags each turn's run carries (`channel: voice`, `room`, and `origin`, which the caller states and which defaults to `studio`), the box's public URL the worker posts its turns to, and a call token: the worker's bearer for this call, valid two hours, until the box restarts, on the AG-UI chat of this agent and this session key on the trace relay `POST /voice/v1/traces`, and on this call's identity bootstrap and heard completion at `/voice/v1/calls/{sessionKey}/bootstrap` and `/voice/v1/calls/{sessionKey}/complete`, nowhere else. The box freezes memory policy, the agent's knowledge settings (`rag`) and its refusal phrase (`refusalPhrase`, what the caller hears when a guardrail stops their sentence or the agent's reply), none of them in the dispatch, at room acceptance, so a change saved in Studio reaches the next call; the worker supplies the original SIP participant's phoneNumber before output, or the browser call uses explicit identities supplied on this request. The bootstrap response contains no recalled facts or policy. On a box that exports traces (`OTEL_EXPORTER_OTLP_*`), the dispatch also asks the worker to send the call's spans to that relay with that token. The call is recorded to the box bucket as audio-only OGG only when the agent's `recording` voice setting is on (it is off by default) and the box bucket is not on a loopback or private host; a recording that fails never fails the call. The LiveKit API key and secret are never answered. The call token is answered only inside the room token's dispatch, which anyone holding the room token can read.",
+		description: "Opens a fresh room `call-<agent>-<8 hex>` and answers the LiveKit URL, the browser's token to join it (it joins within five minutes; a call already running is not cut) and the settings the call runs with. The token carries the room's agent dispatch: joining the room sends the voice worker in with the agent's name, its effective voice settings, the session key `voice-<room>` every turn lands in, the tags each turn's run carries (`channel: voice`, `room`, and `origin`, which the caller states and which defaults to `studio`), the box's public URL the worker posts its turns to, and a call token: the worker's bearer for this call, valid two hours, until the box restarts, on the AG-UI chat of this agent and this session key on the trace relay `POST /voice/v1/traces`, and on this call's identity bootstrap, knowledge prefetch and heard completion at `/voice/v1/calls/{sessionKey}/bootstrap`, `/voice/v1/calls/{sessionKey}/knowledge` and `/voice/v1/calls/{sessionKey}/complete`, nowhere else. The box freezes memory policy, the agent's knowledge settings (`rag`) and its refusal phrase (`refusalPhrase`, what the caller hears when a guardrail stops their sentence or the agent's reply), none of them in the dispatch, at room acceptance, so a change saved in Studio reaches the next call; the worker supplies the original SIP participant's phoneNumber before output, or the browser call uses explicit identities supplied on this request. The bootstrap response contains no recalled facts or policy. On a box that exports traces (`OTEL_EXPORTER_OTLP_*`), the dispatch also asks the worker to send the call's spans to that relay with that token. The call is recorded to the box bucket as audio-only OGG only when the agent's `recording` voice setting is on (it is off by default) and the box bucket is not on a loopback or private host; a recording that fails never fails the call. The LiveKit API key and secret are never answered. The call token is answered only inside the room token's dispatch, which anyone holding the room token can read.",
 		request: { body: jsonBody(named("VoiceRoomRequest", voiceRoomRequestSchema), "The agent to call.") },
 		responses: {
 			200: jsonResponse$1(named("VoiceRoomResponse", voiceRoomResponseSchema), "The room, its participant token, and what the dispatch carries."),
@@ -94092,18 +94232,21 @@ var createVoiceCallsRouter = (options) => {
 		options.logger.warn(principal?.type === "call" ? `voice ${principal.agent}/${principal.sessionKey}: lifecycle_failed.` : "voice lifecycle request failed.");
 		return c.json(errorEnvelope("voice_lifecycle_failed", "The voice lifecycle operation failed."), 500);
 	});
+	const callResponses = {
+		200: jsonResponse$1(named("VoiceCallLifecycleResponse", voiceCallLifecycleResponseSchema), "The call operation completed."),
+		400: errorResponse("Invalid bounded call payload."),
+		401: errorResponse("Missing, invalid or expired call token."),
+		403: errorResponse("Only the call token for this session is admitted."),
+		404: errorResponse("No accepted call for this credential."),
+		413: errorResponse("The payload exceeds 1 MiB.")
+	};
 	const shared = {
 		method: "post",
 		tags: ["voice-worker"],
 		security: [{ voiceCallToken: [] }],
 		responses: {
-			200: jsonResponse$1(named("VoiceCallLifecycleResponse", voiceCallLifecycleResponseSchema), "The call operation completed."),
-			400: errorResponse("Invalid bounded call payload."),
-			401: errorResponse("Missing, invalid or expired call token."),
-			403: errorResponse("Only the call token for this session is admitted."),
-			404: errorResponse("No accepted call for this credential."),
+			...callResponses,
 			409: errorResponse("The call identity is frozen or the call is completing."),
-			413: errorResponse("The payload exceeds 1 MiB."),
 			500: errorResponse("Settings, recall or extraction failed.")
 		}
 	};
@@ -94121,6 +94264,22 @@ var createVoiceCallsRouter = (options) => {
 		await options.calls.bootstrap(principal.agent, principal.sessionKey, c.req.valid("json").identities);
 		return c.json({ ok: true }, 200);
 	});
+	app.openapi(createRoute({
+		...shared,
+		responses: callResponses,
+		path: "/{sessionKey}/knowledge",
+		summary: "Start the knowledge search on the caller's eager end of turn",
+		description: "Runs the agent's deterministic input guardrails over the sentence the speech model guessed ends the caller's turn, starts the knowledge search on the masked text and answers without waiting for it. The turn whose sentence matches it, ignoring case, spacing and punctuation, takes those passages instead of searching again. An agent without a knowledge collection or with a model-backed input guardrail, a sentence too short to search, a blocked sentence and a completing call answer the same and start nothing. Writes nothing to the conversation and calls no chat model.",
+		request: {
+			params: z$1.object({ sessionKey: z$1.string() }),
+			body: jsonBody(named("VoiceCallKnowledge", voiceCallKnowledgeSchema), "The sentence the speech model guessed ends the caller's turn.")
+		}
+	}), async (c) => {
+		const principal = readRequestPrincipal(c);
+		if (principal.type !== "call") throw new Error("Missing call principal");
+		await options.calls.prefetchKnowledge(principal.agent, principal.sessionKey, c.req.valid("json").text);
+		return c.json({ ok: true }, 200);
+	});
 	app.openapi(createRoute({
 		...shared,
 		path: "/{sessionKey}/complete",
@@ -94335,7 +94494,7 @@ var modelWarmupBody = (target) => ({
 		content: WARMUP_USER_TEXT
 	}],
 	reasoning: { enabled: true },
-	...target.providerSort ? { provider: providerRoutingFor(target.providerSort) } : {}
+	...target.providerRouting ? { provider: providerRoutingFor(target.providerRouting) } : {}
 });
 var createModelWarmup = (deps) => {
 	const timeoutMs = deps.timeoutMs ?? 2e4;
@@ -94346,7 +94505,7 @@ var createModelWarmup = (deps) => {
 		const body = modelWarmupBody({
 			model: entry.metadata.model,
 			instructions: entry.metadata.instructions,
-			providerSort: entry.metadata.providerSort
+			providerRouting: entry.metadata.providerRouting
 		});
 		const signal = AbortSignal.timeout(timeoutMs);
 		try {
@@ -95146,7 +95305,7 @@ async function probeDurableInterruptState(graph, threadId, logger) {
 		return { supported: false };
 	}
 }
-var nowIso, CLIENT_MOVED_ON, CLIENT_TOOL_ANSWER_ROUNDS, createDeepAgentChat, hintDurableSessions, voiceCallConfigurable, guardrailHoldReply, toDecisionStep, toRedactionStep, toParkStep;
+var nowIso, CLIENT_MOVED_ON, CLIENT_TOOL_ANSWER_ROUNDS, createDeepAgentChat, hintDurableSessions, voiceCallConfigurable, guardrailHoldReply, toDecisionStep, toRedactionStep, toParkStep, agentPolicy, evaluateInput;
 var init_chat = __esmMin((() => {
 	init_studio_continuation();
 	init_abortable();
@@ -95218,10 +95377,7 @@ var init_chat = __esmMin((() => {
 		async function readPolicy(agentName) {
 			if (!opts.guardrails) return [];
 			try {
-				return await opts.guardrails.effectiveGuardrails({
-					kind: "agent",
-					name: agentName
-				});
+				return await agentPolicy(opts.guardrails, agentName);
 			} catch (err) {
 				logger.warn(`deep-agent guardrail policy read failed for "${agentName}" — the turn runs unguarded: ${messageOf$6(err)}`);
 				return [];
@@ -95249,13 +95405,7 @@ var init_chat = __esmMin((() => {
 			if (guardrails.length === 0) return unguarded;
 			try {
 				const original = lastUserMessage(turn.messages);
-				const classify = classifierFor(signal);
-				const result = await evaluateGuardrails({
-					guardrails,
-					content: original,
-					target: "input",
-					...classify ? { classify } : {}
-				});
+				const result = await evaluateInput(guardrails, original, classifierFor(signal));
 				return {
 					turn: result.content === original ? turn : withLastUserMessage(turn, result.content),
 					blocked: result.allowed ? null : result.decision,
@@ -95268,6 +95418,27 @@ var init_chat = __esmMin((() => {
 			}
 		}
 		/**
+		* The input border's deterministic rules over one sentence that is not a turn yet:
+		* the guess a voice call's knowledge prefetch searches (STA3-94 D2). It never asks
+		* the classifier, which the guardrail ADR spends once per turn, and writes no run
+		* step: the turn that follows evaluates the same rules and records them.
+		*/
+		const guardInput = async (agentName, text) => {
+			let guardrails = [];
+			if (opts.guardrails) try {
+				guardrails = await agentPolicy(opts.guardrails, agentName);
+			} catch {
+				return "skip";
+			}
+			const input = guardrails.filter((rule) => rule.enabled && rule.target === "input");
+			if (input.some((rule) => isModelBackedCheck(rule.checkType))) return "skip";
+			const result = await evaluateInput(input, text);
+			return {
+				blocked: !result.allowed,
+				content: result.content
+			};
+		};
+		/**
 		* Wrap a turn's stream with the OUTPUT border: the reply is held until the policy
 		* has judged it, then released, released masked, or dropped with a reason.
 		*
@@ -95573,7 +95744,10 @@ var init_chat = __esmMin((() => {
 			const guardedTurn = guarded.turn;
 			const blocked = guarded.blocked;
 			const userMessage = lastUserMessage(guardedTurn.messages);
-			const knowledge = registered && blocked === null ? voiceCall?.knowledgeFor(userMessage) : void 0;
+			let prefetch;
+			const knowledge = registered && blocked === null ? voiceCall?.knowledgeFor(userMessage, (outcome) => {
+				prefetch = outcome;
+			}) : void 0;
 			const runInput = {
 				trigger,
 				input: { message: userMessage },
@@ -95583,6 +95757,7 @@ var init_chat = __esmMin((() => {
 					memoryIdentity,
 					...voiceCall ? { voiceCall: true } : {},
 					...voiceCall?.searchesKnowledge ? { voiceKnowledgeRule: true } : {},
+					...prefetch ? { voiceKnowledgePrefetch: prefetch } : {},
 					...groupingKey !== void 0 ? { sessionKey: groupingKey } : {},
 					...turn.evalRunId !== void 0 ? evalRunStamp({
 						evalRunId: turn.evalRunId,
@@ -96317,6 +96492,7 @@ var init_chat = __esmMin((() => {
 			getMessageHistory,
 			seedState,
 			reconcileHeard,
+			guardInput,
 			runTurn: (turn) => ownedByAgent(turn.agent, runTurn(turn)),
 			resumeTurn: ownedResumeTurn
 		};
@@ -96343,6 +96519,16 @@ var init_chat = __esmMin((() => {
 		...guardrailParkRow(decision, approvalId),
 		status: "running"
 	});
+	agentPolicy = (reader, agentName) => reader.effectiveGuardrails({
+		kind: "agent",
+		name: agentName
+	});
+	evaluateInput = (guardrails, content, classify) => evaluateGuardrails({
+		guardrails,
+		content,
+		target: "input",
+		...classify ? { classify } : {}
+	});
 }));
 //#endregion
 //#region ../../libs/runtime/workspace-runtime/src/lib/deep-agents/http/run-tags-header.ts
@@ -97001,10 +97187,12 @@ var init_mount$1 = __esmMin((() => {
 			...options.classify ? { classify: options.classify } : {},
 			...options.plantAmbientBridges ? { plantAmbientBridges: options.plantAmbientBridges } : {}
 		});
+		const voiceCalls = voiceCallsForApp(app);
+		voiceCalls?.useInputGuard(chat.guardInput);
 		const wire = {
 			sessions: options.sessions,
 			invocations: options.invocations,
-			voiceCalls: voiceCallsForApp(app),
+			voiceCalls,
 			chat,
 			registry,
 			logger,
@@ -104335,7 +104523,7 @@ var calls_workflow_ts_default = "import { z } from \"zod\";\nimport { streamDeep
 var plain_workflow_ts_default = "import { z } from \"zod\";\n\n/**\n * Workflow \"{{name}}\" — a durable pipeline.\n *\n * Each `\"use step\"` is a durable checkpoint: kill the runtime mid-run and it\n * resumes from the last completed step instead of replaying everything.\n */\n\n// THE contract — sibling exports next to the bare workflow function. The build\n// harvests these (by convention) into the manifest + a live validator. The\n// parameter type is DERIVED from `inputSchema` and the RETURN type from\n// `outputSchema` (both `z.infer`), so neither side can drift from the contract\n// the build harvests. Without the return annotation the schema is published to\n// Studio and the generated types and never compared to what the body returns\n// (PLA-1272).\nexport const inputSchema = z.object({\n  message: z.string().describe(\"The input to process\"),\n});\n\nexport const outputSchema = z.object({\n  message: z.string(),\n  processed: z.boolean(),\n});\n\ntype {{Name}}Input = z.infer<typeof inputSchema>;\n\nexport async function {{export}}(input: {{Name}}Input): Promise<z.infer<typeof outputSchema>> {\n  \"use workflow\";\n\n  const result = await processMessage(input); // step 1 — durable checkpoint\n  return result;\n}\n\nasync function processMessage(input: {{Name}}Input) {\n  \"use step\";\n  // Replace this with the deterministic work of your step (DB write, HTTP call,\n  // an LLM call, …). Returning a value records it as the step's checkpoint.\n  return { message: input.message, processed: true };\n}\n";
 //#endregion
 //#region src/services/scaffold/template-loader.ts
-var SDK_TEMPLATE_VERSION = "0.10.1";
+var SDK_TEMPLATE_VERSION = "0.10.2";
 var PREFIX = "./templates/";
 var templates = new Map(Object.entries(/* @__PURE__ */ Object.assign({
 	"./templates/deep-agents/index.ts.tmpl": index_ts_default,
@@ -106849,6 +107037,57 @@ var literalText = (ts, node) => {
 	return null;
 };
 /**
+* A value written down as JSON would be: strings, numbers (a leading `-` included),
+* booleans, `null`, and arrays and object literals of those. `undefined` for
+* anything else ANYWHERE inside it — an identifier, a spread, a computed key, an
+* interpolation — because a value read in part is a value guessed at. `as const`
+* and `satisfies` around it are type-only and looked through.
+*/
+var literalJson = (ts, node) => {
+	const K = ts.SyntaxKind;
+	if (!node) return void 0;
+	const text = literalText(ts, node);
+	if (text !== null) return text;
+	switch (node.kind) {
+		case K["AsExpression"]:
+		case K["SatisfiesExpression"]:
+		case K["ParenthesizedExpression"]: return literalJson(ts, node.expression);
+		case K["NumericLiteral"]: return Number(node.text);
+		case K["TrueKeyword"]: return true;
+		case K["FalseKeyword"]: return false;
+		case K["NullKeyword"]: return null;
+		case K["PrefixUnaryExpression"]: {
+			const unary = node;
+			if (unary.operator !== K["MinusToken"] || unary.operand.kind !== K["NumericLiteral"]) return;
+			return -Number(unary.operand.text);
+		}
+		case K["ArrayLiteralExpression"]: {
+			const items = [];
+			for (const element of node.elements) {
+				const item = literalJson(ts, element);
+				if (item === void 0) return void 0;
+				items.push(item);
+			}
+			return items;
+		}
+		case K["ObjectLiteralExpression"]: {
+			const object = {};
+			for (const property of node.properties) {
+				if (property.kind !== K["PropertyAssignment"]) return void 0;
+				const assignment = property;
+				const key = literalKey(ts, assignment.name);
+				const value = literalJson(ts, assignment.initializer);
+				if (key === null || value === void 0) return void 0;
+				object[key] = value;
+			}
+			return object;
+		}
+		default: return;
+	}
+};
+/** An object literal key written as a name or a string, never a computed one. */
+var literalKey = (ts, name) => name.kind === ts.SyntaxKind["Identifier"] || name.kind === ts.SyntaxKind["StringLiteral"] ? name.text ?? null : null;
+/**
 * A plain string literal's text, and nothing else: narrower than `literalText`,
 * which also takes a template with no substitution.
 */
@@ -107080,6 +107319,14 @@ var MODELS_USE_METHOD = "use";
 var MODEL_PROPERTY = "model";
 /** The property an agent and each of its subagents names its provider routing in. */
 var ROUTING_PROPERTY = "modelRouting";
+/**
+* A key's routing as the manifest records it: the object, and its plain-sort twin,
+* which is all a runtime older than the object reads.
+*/
+var routingFields = (routing) => ({
+	defaultProviderRouting: routing,
+	defaultProviderSort: plainSortOf(routing)
+});
 var MODEL_SCAN_MESSAGES = {
 	noCompiler: "[models] cannot resolve typescript from the workspace, so the model inventory was not collected — Studio will not know which models this workspace declares. Install it as a dev dependency of the project.",
 	unparsable: (file) => `[models] could not parse ${file}; the models it declares are missing from the inventory.`
@@ -107114,7 +107361,7 @@ var scanSource$1 = (source) => {
 			key,
 			defaultModelId: literalText(ts, call.arguments[1]),
 			overridable: true,
-			defaultProviderSort: null,
+			...routingFields(null),
 			file,
 			line: lineOf(call)
 		});
@@ -107138,7 +107385,7 @@ var recordModel = (existing, ref, warn) => {
 		existing.push(ref);
 		return;
 	}
-	if (already.defaultModelId !== ref.defaultModelId || already.overridable !== ref.overridable || already.defaultProviderSort !== ref.defaultProviderSort) warn(`[models] \`${ref.key}\` is declared again in ${ref.file} with a different default. One key is one pick, so the first declaration (${already.file}:${already.line}) is the one an operator sees — make the two agree, or give them different keys.`, {
+	if (already.defaultModelId !== ref.defaultModelId || already.overridable !== ref.overridable || JSON.stringify(already.defaultProviderRouting) !== JSON.stringify(ref.defaultProviderRouting)) warn(`[models] \`${ref.key}\` is declared again in ${ref.file} with a different default. One key is one pick, so the first declaration (${already.file}:${already.line}) is the one an operator sees — make the two agree, or give them different keys.`, {
 		file: ref.file,
 		line: ref.line
 	});
@@ -107156,16 +107403,19 @@ var collectAgentModels = (source) => {
 	const { ts, file, lineOf, at, warn } = source;
 	const found = [];
 	/**
-	* The `modelRouting` the config writes. `defineDeepAgent` refuses a value outside
-	* the sorts, so the agent will not load; one built at run time cannot be read
-	* here. Both are recorded as "no routing in the code", and the first is warned.
+	* The `modelRouting` the config writes: a sort, or OpenRouter's `provider` object
+	* written as a literal. `defineDeepAgent` refuses a value outside the schema, so
+	* the agent will not load; one built at run time cannot be read here. Both are
+	* recorded as "no routing in the code", and the first is warned.
 	*/
 	const routingOf = (label, config) => {
 		const value = propertyValue(ts, config, ROUTING_PROPERTY);
 		if (value === void 0) return null;
-		const written = literalText(ts, value);
-		if (written === null || isModelProviderSort(written)) return written;
-		warn(`[models] ${label} in ${file} declares \`modelRouting\` '${written}', which is not one of ${MODEL_PROVIDER_SORTS_TEXT}. The agent is refused when it is defined.`, at(value));
+		const written = literalJson(ts, value);
+		if (written === void 0) return null;
+		const parsed = modelRoutingInputSchema.safeParse(written);
+		if (parsed.success) return toModelProviderRouting(parsed.data);
+		warn(`[models] ${label} in ${file} declares \`modelRouting\` ${JSON.stringify(written)}, which is neither one of ${MODEL_PROVIDER_SORTS_TEXT} nor an OpenRouter \`provider\` object. The agent is refused when it is defined.`, at(value));
 		return null;
 	};
 	/**
@@ -107177,13 +107427,13 @@ var collectAgentModels = (source) => {
 	const declare = (args) => {
 		const { label, key, config, node } = args;
 		const declaration = propertyValue(ts, config, MODEL_PROPERTY);
-		const defaultProviderSort = routingOf(label, config);
+		const routing = routingFields(routingOf(label, config));
 		if (declaration === void 0) {
 			found.push({
 				key,
 				defaultModelId: null,
 				overridable: true,
-				defaultProviderSort,
+				...routing,
 				file,
 				line: lineOf(node)
 			});
@@ -107196,7 +107446,7 @@ var collectAgentModels = (source) => {
 			key,
 			defaultModelId: blank ? null : written,
 			overridable: written !== null,
-			defaultProviderSort,
+			...routing,
 			file,
 			line: lineOf(declaration)
 		});