@camstack/addon-ai 0.4.53 → 0.4.55

package/dist/addon.js

@@ -13460,6 +13460,35 @@ method(_void(), array(string()).readonly(), { auth: "admin" }), method(object({
 	auth: "admin"
 });
 /**
+* Transport bounds for AI capability methods.
+*
+* Its own module because `llm.cap.ts` and `llm-runtime.cap.ts` already import
+* each other: declaring the constant in either one hands the other `undefined`
+* at module-init time, and an undefined `timeoutMs` silently falls back to the
+* kernel's 60 s default — the exact failure this constant exists to prevent,
+* reintroduced invisibly. Found by the guard, not by review.
+*/
+/**
+* Ceiling for a cap method that runs inference or loads a model.
+*
+* Deliberately far above anything the AI layer itself may wait for (a summary
+* asks 180 s by default, a rule may ask 600 s), because the TRANSPORT must
+* never be the layer that gives up first: when it does, the failure is
+* reported as a transport error for a model that was still thinking, naming
+* the wrong layer and hiding the real one.
+*
+* Measured 2026-09-03: a 6-image vision request needs 44.8 s warm and ~20 s
+* more on a cold model load. With no declaration every digest failed at
+* exactly 60 s and shipped without its text.
+*
+* A ceiling, not a budget. The AI layer's own timeouts (the rule's `timeoutMs`,
+* the profile's `timeoutMs` / `firstTokenTimeoutMs`) remain the things that
+* decide when to stop, because only they can say WHY.
+*/
+var AI_CAP_TRANSPORT_TIMEOUT_MS = 6e5;
+/** A multi-gigabyte download. An hour is not generous, it is honest. */
+var AI_MODEL_INSTALL_TIMEOUT_MS = 60 * 6e4;
+/**
 * Shared LLM generate contracts — imported by BOTH `llm.cap.ts` (consumer
 * surface) and `llm-runtime.cap.ts` (node-side managed executor) so the two
 * caps stay wire-compatible without a circular cap→cap import.
@@ -13768,11 +13797,13 @@ var llmRuntimeCapability = {
 			timeoutMs: number$1().int().positive().optional()
 		}), LlmGenerateResultSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		ensureStarted: method(object({ runtime: ManagedRuntimeConfigSchema }), LlmRuntimeStatusSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		stop: method(object({}), _void(), {
 			kind: "mutation",
@@ -13781,7 +13812,8 @@ var llmRuntimeCapability = {
 		status: method(object({}), LlmRuntimeStatusSchema, { auth: "admin" }),
 		installModel: method(object({ model: ManagedModelRefSchema }), _void(), {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_MODEL_INSTALL_TIMEOUT_MS
 		}),
 		deleteModel: method(object({ file: string() }), _void(), {
 			kind: "mutation",
@@ -13963,8 +13995,14 @@ var llmCapability = {
 	/** `nodeId` inputs below are DATA (the hub provider pins itself), never routing. */
 	nodeIdMode: "data",
 	methods: {
-		generate: method(LlmGenerateBaseInputSchema, LlmGenerateResultSchema, { kind: "mutation" }),
-		generateVision: method(GenerateVisionInputSchema, LlmGenerateResultSchema, { kind: "mutation" }),
+		generate: method(LlmGenerateBaseInputSchema, LlmGenerateResultSchema, {
+			kind: "mutation",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
+		}),
+		generateVision: method(GenerateVisionInputSchema, LlmGenerateResultSchema, {
+			kind: "mutation",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
+		}),
 		/**
 		* Stop a generation started with a `requestId`.
 		*
@@ -13988,7 +14026,8 @@ var llmCapability = {
 		}),
 		testProfile: method(ProfileRefInputSchema, LlmGenerateResultSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		/** Live vendor enumeration (GET /models etc.). */
 		listModels: method(ProfileRefInputSchema, array(string())),
@@ -73861,6 +73900,10 @@ var CONSUMER_RETRY_POLICY = {
 	"ai-summary": RETRYING,
 	/** NcConfirmGate — 8 s total budget, fail-OPEN. */
 	"notifier-rules": NOT_RETRYING,
+	/** NcRuleAiAnalyst — the sentence on ONE notification, same 8 s budget and
+	*  the same reason: somebody is holding a phone, and a retry produces a
+	*  later notification rather than a better one. */
+	"notifier-narration": NOT_RETRYING,
 	/** SceneConfirmGate + checkSceneLlm — 8 s total budget, fail-CLOSED. */
 	"scene-monitor": NOT_RETRYING
 };

package/dist/addon.mjs

@@ -13487,6 +13487,35 @@ method(_void(), array(string()).readonly(), { auth: "admin" }), method(object({
 	auth: "admin"
 });
 /**
+* Transport bounds for AI capability methods.
+*
+* Its own module because `llm.cap.ts` and `llm-runtime.cap.ts` already import
+* each other: declaring the constant in either one hands the other `undefined`
+* at module-init time, and an undefined `timeoutMs` silently falls back to the
+* kernel's 60 s default — the exact failure this constant exists to prevent,
+* reintroduced invisibly. Found by the guard, not by review.
+*/
+/**
+* Ceiling for a cap method that runs inference or loads a model.
+*
+* Deliberately far above anything the AI layer itself may wait for (a summary
+* asks 180 s by default, a rule may ask 600 s), because the TRANSPORT must
+* never be the layer that gives up first: when it does, the failure is
+* reported as a transport error for a model that was still thinking, naming
+* the wrong layer and hiding the real one.
+*
+* Measured 2026-09-03: a 6-image vision request needs 44.8 s warm and ~20 s
+* more on a cold model load. With no declaration every digest failed at
+* exactly 60 s and shipped without its text.
+*
+* A ceiling, not a budget. The AI layer's own timeouts (the rule's `timeoutMs`,
+* the profile's `timeoutMs` / `firstTokenTimeoutMs`) remain the things that
+* decide when to stop, because only they can say WHY.
+*/
+var AI_CAP_TRANSPORT_TIMEOUT_MS = 6e5;
+/** A multi-gigabyte download. An hour is not generous, it is honest. */
+var AI_MODEL_INSTALL_TIMEOUT_MS = 60 * 6e4;
+/**
 * Shared LLM generate contracts — imported by BOTH `llm.cap.ts` (consumer
 * surface) and `llm-runtime.cap.ts` (node-side managed executor) so the two
 * caps stay wire-compatible without a circular cap→cap import.
@@ -13795,11 +13824,13 @@ var llmRuntimeCapability = {
 			timeoutMs: number$1().int().positive().optional()
 		}), LlmGenerateResultSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		ensureStarted: method(object({ runtime: ManagedRuntimeConfigSchema }), LlmRuntimeStatusSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		stop: method(object({}), _void(), {
 			kind: "mutation",
@@ -13808,7 +13839,8 @@ var llmRuntimeCapability = {
 		status: method(object({}), LlmRuntimeStatusSchema, { auth: "admin" }),
 		installModel: method(object({ model: ManagedModelRefSchema }), _void(), {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_MODEL_INSTALL_TIMEOUT_MS
 		}),
 		deleteModel: method(object({ file: string() }), _void(), {
 			kind: "mutation",
@@ -13990,8 +14022,14 @@ var llmCapability = {
 	/** `nodeId` inputs below are DATA (the hub provider pins itself), never routing. */
 	nodeIdMode: "data",
 	methods: {
-		generate: method(LlmGenerateBaseInputSchema, LlmGenerateResultSchema, { kind: "mutation" }),
-		generateVision: method(GenerateVisionInputSchema, LlmGenerateResultSchema, { kind: "mutation" }),
+		generate: method(LlmGenerateBaseInputSchema, LlmGenerateResultSchema, {
+			kind: "mutation",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
+		}),
+		generateVision: method(GenerateVisionInputSchema, LlmGenerateResultSchema, {
+			kind: "mutation",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
+		}),
 		/**
 		* Stop a generation started with a `requestId`.
 		*
@@ -14015,7 +14053,8 @@ var llmCapability = {
 		}),
 		testProfile: method(ProfileRefInputSchema, LlmGenerateResultSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		/** Live vendor enumeration (GET /models etc.). */
 		listModels: method(ProfileRefInputSchema, array(string())),
@@ -73888,6 +73927,10 @@ var CONSUMER_RETRY_POLICY = {
 	"ai-summary": RETRYING,
 	/** NcConfirmGate — 8 s total budget, fail-OPEN. */
 	"notifier-rules": NOT_RETRYING,
+	/** NcRuleAiAnalyst — the sentence on ONE notification, same 8 s budget and
+	*  the same reason: somebody is holding a phone, and a retry produces a
+	*  later notification rather than a better one. */
+	"notifier-narration": NOT_RETRYING,
 	/** SceneConfirmGate + checkSceneLlm — 8 s total budget, fail-CLOSED. */
 	"scene-monitor": NOT_RETRYING
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@camstack/addon-ai",
-  "version": "0.4.53",
+  "version": "0.4.55",
   "description": "AI addon for CamStack — the `llm` collection provider (cloud, LAN, and camstack-managed local llama.cpp profiles) plus the per-node `llm-runtime` managed executor.",
   "keywords": [
     "camstack",