@camstack/types 1.2.137 → 1.2.138

package/dist/capabilities/ai-cap-timeouts.d.ts

@@ -0,0 +1,29 @@
+/**
+ * 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.
+ */
+export declare const AI_CAP_TRANSPORT_TIMEOUT_MS = 600000;
+/** A multi-gigabyte download. An hour is not generous, it is honest. */
+export declare const AI_MODEL_INSTALL_TIMEOUT_MS: number;

package/dist/capabilities/llm-runtime.cap.d.ts

@@ -1,14 +1,3 @@
-/**
- * `llm-runtime` — node-side managed llama.cpp executor (spec §4). Registered
- * on EVERY node where `addon-ai` is installed; the hub `llm` provider reaches
- * a specific node's runtime with `nodePin(profile.runtime.nodeId)` — normal
- * cap routing, zero bespoke plumbing. `internal: true`: the operator reaches
- * this only through the `llm` cap's methods.
- *
- * One running llama-server child per node in v1 (models are RAM-heavy).
- * Resource ceiling = llama-server flags + idleStopMinutes ONLY (no RSS
- * watchdog — operator decision #3).
- */
 import { z } from 'zod';
 import { type InferProvider } from './capability-definition.js';
 /**

package/dist/index.js

@@ -11414,6 +11414,37 @@ var filesystemBrowseCapability = {
 	}
 };
 //#endregion
+//#region src/capabilities/ai-cap-timeouts.ts
+/**
+* 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;
+//#endregion
 //#region src/capabilities/llm-shared.ts
 /**
 * Shared LLM generate contracts — imported by BOTH `llm.cap.ts` (consumer
@@ -11729,11 +11760,13 @@ var llmRuntimeCapability = {
 		*/
 		complete: require_sleep.method(LlmRuntimeCompleteInputSchema, LlmGenerateResultSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		ensureStarted: require_sleep.method(zod.z.object({ runtime: ManagedRuntimeConfigSchema }), LlmRuntimeStatusSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		stop: require_sleep.method(zod.z.object({}), zod.z.void(), {
 			kind: "mutation",
@@ -11742,7 +11775,8 @@ var llmRuntimeCapability = {
 		status: require_sleep.method(zod.z.object({}), LlmRuntimeStatusSchema, { auth: "admin" }),
 		installModel: require_sleep.method(zod.z.object({ model: ManagedModelRefSchema }), zod.z.void(), {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_MODEL_INSTALL_TIMEOUT_MS
 		}),
 		deleteModel: require_sleep.method(zod.z.object({ file: zod.z.string() }), zod.z.void(), {
 			kind: "mutation",
@@ -11926,8 +11960,14 @@ var llmCapability = {
 	/** `nodeId` inputs below are DATA (the hub provider pins itself), never routing. */
 	nodeIdMode: "data",
 	methods: {
-		generate: require_sleep.method(LlmGenerateBaseInputSchema, LlmGenerateResultSchema, { kind: "mutation" }),
-		generateVision: require_sleep.method(GenerateVisionInputSchema, LlmGenerateResultSchema, { kind: "mutation" }),
+		generate: require_sleep.method(LlmGenerateBaseInputSchema, LlmGenerateResultSchema, {
+			kind: "mutation",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
+		}),
+		generateVision: require_sleep.method(GenerateVisionInputSchema, LlmGenerateResultSchema, {
+			kind: "mutation",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
+		}),
 		/**
 		* Stop a generation started with a `requestId`.
 		*
@@ -11951,7 +11991,8 @@ var llmCapability = {
 		}),
 		testProfile: require_sleep.method(ProfileRefInputSchema, LlmGenerateResultSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		/** Live vendor enumeration (GET /models etc.). */
 		listModels: require_sleep.method(ProfileRefInputSchema, zod.z.array(zod.z.string())),

package/dist/index.mjs

@@ -11413,6 +11413,37 @@ var filesystemBrowseCapability = {
 	}
 };
 //#endregion
+//#region src/capabilities/ai-cap-timeouts.ts
+/**
+* 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;
+//#endregion
 //#region src/capabilities/llm-shared.ts
 /**
 * Shared LLM generate contracts — imported by BOTH `llm.cap.ts` (consumer
@@ -11728,11 +11759,13 @@ var llmRuntimeCapability = {
 		*/
 		complete: method(LlmRuntimeCompleteInputSchema, LlmGenerateResultSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		ensureStarted: method(z.object({ runtime: ManagedRuntimeConfigSchema }), LlmRuntimeStatusSchema, {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_CAP_TRANSPORT_TIMEOUT_MS
 		}),
 		stop: method(z.object({}), z.void(), {
 			kind: "mutation",
@@ -11741,7 +11774,8 @@ var llmRuntimeCapability = {
 		status: method(z.object({}), LlmRuntimeStatusSchema, { auth: "admin" }),
 		installModel: method(z.object({ model: ManagedModelRefSchema }), z.void(), {
 			kind: "mutation",
-			auth: "admin"
+			auth: "admin",
+			timeoutMs: AI_MODEL_INSTALL_TIMEOUT_MS
 		}),
 		deleteModel: method(z.object({ file: z.string() }), z.void(), {
 			kind: "mutation",
@@ -11925,8 +11959,14 @@ 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`.
 		*
@@ -11950,7 +11990,8 @@ 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, z.array(z.string())),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@camstack/types",
-  "version": "1.2.137",
+  "version": "1.2.138",
   "description": "Shared types, interfaces, and model catalogs for the CamStack detection ecosystem",
   "keywords": [
     "camstack",