@camstack/addon-pipeline 1.2.181 → 1.2.182

package/dist/detection-pipeline/index.js

@@ -3019,6 +3019,42 @@ async function samplePool(pool, log, reportedDead) {
 }
 //#endregion
 //#region src/detection-pipeline/engine/inference-timeout-guard.ts
+/**
+* Inference-timeout regression guard — "did this update make the GPU time
+* out more than the last one did?", asked by the node itself.
+*
+* ## Why
+*
+* `@camstack/server` 1.2.218 reintroduced the inference timeouts 1.2.216 had
+* fixed: ~3 760 `inference request timed out` an hour for 4 h 19 min (16 215
+* lines, ~19 000 `runInference failed`), on a cluster whose previous hour had
+* seen ~5. Nothing said so. 1.2.219 fixed it again as a side effect of the
+* next deploy, and the regression was found a day later by an audit that
+* counted log lines per version by hand (2026-09-05). A green typecheck, a
+* green suite and a 200 on `/health` all held throughout.
+*
+* ## What it does
+*
+* Every five minutes the guard records how many inference requests timed out
+* on this node under the current FINGERPRINT (the closure version). The last
+* hour of buckets is persisted in the addon store, so a restart hands the
+* previous fingerprint's rate to the next one. Once the new fingerprint has a
+* quarter hour of its own, the two hourly rates are compared:
+*
+*  - ≥ `TIMEOUT_GUARD_RATIO`× worse AND above `TIMEOUT_GUARD_FLOOR_PER_HOUR`
+*    → `system.liveness-failed` (AlertCenter raises a persistent alert keyed
+*    on the finding id) and an ERROR line naming both versions and rates;
+*  - ≥ `TIMEOUT_GUARD_RATIO`× better → an INFO line, no alert;
+*  - the same fingerprint (a plain restart) → nothing to compare.
+*
+* The rate is per hour, not per frame: it is the unit the audit measured and
+* the unit an operator reads. A traffic doubling doubles the rate, which is
+* why the ratio is five and not two.
+*
+* Pure core (`recordTimeoutBucket`, `evaluateTimeoutRegression`) tested in
+* `__tests__/inference-timeout-guard.spec.ts`; the runtime wrapper owns the
+* timer, the store and the emit.
+*/
 var TIMEOUT_GUARD_BUCKET_MS = 5 * 6e4;
 /** Addon-store key the ledger persists under (the addon's own settings blob). */
 var INFERENCE_TIMEOUT_LEDGER_KEY = "_inferenceTimeoutLedger";
@@ -3062,6 +3098,40 @@ function evaluateTimeoutRegression(input) {
 		...base
 	};
 }
+/**
+* The closure version, read from the closure the runner was started from.
+*
+* The forked runner script is `<closure>/node_modules/@camstack/system/dist/
+* addon-runner.js` (hub: `/data/server-root/current/…`); the closure's own
+* `@camstack/server/package.json` sits two directories up. This is read from
+* disk rather than asked of `server-management` because that cap is
+* server-provided and the parent refuses to route it for a forked addon —
+* "no provider registered for cap server-management", measured on all
+* three nodes on 2026-09-05. A file the process was started from cannot
+* be refused.
+*/
+function closureVersionFromRunnerPath(runnerScript, fs) {
+	if (runnerScript === void 0 || runnerScript.length === 0) return null;
+	let dir = (0, node_path.dirname)(runnerScript);
+	for (let depth = 0; depth < 12; depth += 1) {
+		const candidate = (0, node_path.join)(dir, "node_modules", "@camstack", "server", "package.json");
+		const raw = fs.readFile(candidate);
+		if (raw !== null) {
+			try {
+				const parsed = JSON.parse(raw);
+				const version = typeof parsed === "object" && parsed !== null ? parsed.version : void 0;
+				if (typeof version === "string" && version.length > 0) return `server@${version}`;
+			} catch {
+				return null;
+			}
+			return null;
+		}
+		const parent = (0, node_path.dirname)(dir);
+		if (parent === dir) break;
+		dir = parent;
+	}
+	return null;
+}
 function readStoredLedgers(blob) {
 	const raw = blob[INFERENCE_TIMEOUT_LEDGER_KEY];
 	if (typeof raw !== "object" || raw === null) return {};
@@ -6891,12 +6961,13 @@ var DetectionPipelineProvider = class DetectionPipelineProvider {
 				readTotalTimeouts: () => this.listGuardedPools().reduce((sum, pool) => sum + pool.factory.getPoolBacklog().timedOut, 0),
 				readStore: () => this.readStore(),
 				writeStore: (patch) => this.writeStore(patch),
-				resolveFingerprint: async () => {
-					const ctx = this.addonCtx;
-					if (!ctx) return null;
-					const status = await ctx.api.serverManagement.getServerPackageStatus.query({ nodeId });
-					return status.runningVersion === null ? null : `server@${status.runningVersion}`;
-				},
+				resolveFingerprint: async () => closureVersionFromRunnerPath(process.argv[1], { readFile: (path) => {
+					try {
+						return node_fs.readFileSync(path, "utf8");
+					} catch {
+						return null;
+					}
+				} }),
 				emitFailed: (finding) => {
 					this.eventBus?.emit(require_dist.createEvent(require_dist.EventCategory.SystemLivenessFailed, TIMEOUT_GUARD_SOURCE, finding));
 				},

package/dist/detection-pipeline/index.mjs

@@ -7,6 +7,7 @@ import { n as startEventLoopStallMonitor } from "../event-loop-stall-monitor-DXL
 import { n as pickNodePlatformArch, t as readProcessMemory } from "../process-memory-zLBrzuFc.mjs";
 import * as fs from "node:fs";
 import * as path$1 from "node:path";
+import { dirname, join } from "node:path";
 import { spawn } from "node:child_process";
 import sharp from "sharp";
 import * as os from "node:os";
@@ -3012,6 +3013,42 @@ async function samplePool(pool, log, reportedDead) {
 }
 //#endregion
 //#region src/detection-pipeline/engine/inference-timeout-guard.ts
+/**
+* Inference-timeout regression guard — "did this update make the GPU time
+* out more than the last one did?", asked by the node itself.
+*
+* ## Why
+*
+* `@camstack/server` 1.2.218 reintroduced the inference timeouts 1.2.216 had
+* fixed: ~3 760 `inference request timed out` an hour for 4 h 19 min (16 215
+* lines, ~19 000 `runInference failed`), on a cluster whose previous hour had
+* seen ~5. Nothing said so. 1.2.219 fixed it again as a side effect of the
+* next deploy, and the regression was found a day later by an audit that
+* counted log lines per version by hand (2026-09-05). A green typecheck, a
+* green suite and a 200 on `/health` all held throughout.
+*
+* ## What it does
+*
+* Every five minutes the guard records how many inference requests timed out
+* on this node under the current FINGERPRINT (the closure version). The last
+* hour of buckets is persisted in the addon store, so a restart hands the
+* previous fingerprint's rate to the next one. Once the new fingerprint has a
+* quarter hour of its own, the two hourly rates are compared:
+*
+*  - ≥ `TIMEOUT_GUARD_RATIO`× worse AND above `TIMEOUT_GUARD_FLOOR_PER_HOUR`
+*    → `system.liveness-failed` (AlertCenter raises a persistent alert keyed
+*    on the finding id) and an ERROR line naming both versions and rates;
+*  - ≥ `TIMEOUT_GUARD_RATIO`× better → an INFO line, no alert;
+*  - the same fingerprint (a plain restart) → nothing to compare.
+*
+* The rate is per hour, not per frame: it is the unit the audit measured and
+* the unit an operator reads. A traffic doubling doubles the rate, which is
+* why the ratio is five and not two.
+*
+* Pure core (`recordTimeoutBucket`, `evaluateTimeoutRegression`) tested in
+* `__tests__/inference-timeout-guard.spec.ts`; the runtime wrapper owns the
+* timer, the store and the emit.
+*/
 var TIMEOUT_GUARD_BUCKET_MS = 5 * 6e4;
 /** Addon-store key the ledger persists under (the addon's own settings blob). */
 var INFERENCE_TIMEOUT_LEDGER_KEY = "_inferenceTimeoutLedger";
@@ -3055,6 +3092,40 @@ function evaluateTimeoutRegression(input) {
 		...base
 	};
 }
+/**
+* The closure version, read from the closure the runner was started from.
+*
+* The forked runner script is `<closure>/node_modules/@camstack/system/dist/
+* addon-runner.js` (hub: `/data/server-root/current/…`); the closure's own
+* `@camstack/server/package.json` sits two directories up. This is read from
+* disk rather than asked of `server-management` because that cap is
+* server-provided and the parent refuses to route it for a forked addon —
+* "no provider registered for cap server-management", measured on all
+* three nodes on 2026-09-05. A file the process was started from cannot
+* be refused.
+*/
+function closureVersionFromRunnerPath(runnerScript, fs) {
+	if (runnerScript === void 0 || runnerScript.length === 0) return null;
+	let dir = dirname(runnerScript);
+	for (let depth = 0; depth < 12; depth += 1) {
+		const candidate = join(dir, "node_modules", "@camstack", "server", "package.json");
+		const raw = fs.readFile(candidate);
+		if (raw !== null) {
+			try {
+				const parsed = JSON.parse(raw);
+				const version = typeof parsed === "object" && parsed !== null ? parsed.version : void 0;
+				if (typeof version === "string" && version.length > 0) return `server@${version}`;
+			} catch {
+				return null;
+			}
+			return null;
+		}
+		const parent = dirname(dir);
+		if (parent === dir) break;
+		dir = parent;
+	}
+	return null;
+}
 function readStoredLedgers(blob) {
 	const raw = blob[INFERENCE_TIMEOUT_LEDGER_KEY];
 	if (typeof raw !== "object" || raw === null) return {};
@@ -6884,12 +6955,13 @@ var DetectionPipelineProvider = class DetectionPipelineProvider {
 				readTotalTimeouts: () => this.listGuardedPools().reduce((sum, pool) => sum + pool.factory.getPoolBacklog().timedOut, 0),
 				readStore: () => this.readStore(),
 				writeStore: (patch) => this.writeStore(patch),
-				resolveFingerprint: async () => {
-					const ctx = this.addonCtx;
-					if (!ctx) return null;
-					const status = await ctx.api.serverManagement.getServerPackageStatus.query({ nodeId });
-					return status.runningVersion === null ? null : `server@${status.runningVersion}`;
-				},
+				resolveFingerprint: async () => closureVersionFromRunnerPath(process.argv[1], { readFile: (path) => {
+					try {
+						return fs.readFileSync(path, "utf8");
+					} catch {
+						return null;
+					}
+				} }),
 				emitFailed: (finding) => {
 					this.eventBus?.emit(createEvent(EventCategory.SystemLivenessFailed, TIMEOUT_GUARD_SOURCE, finding));
 				},

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@camstack/addon-pipeline",
-  "version": "1.2.181",
+  "version": "1.2.182",
   "description": "Pipeline bundle — runner, detection, motion, audio + stream broker. Multi-entry npm package shipping pipeline addons under a single bundle.",
   "keywords": [
     "camstack",