@infersec/conduit 1.112.0 → 1.113.0

package/dist/modelManagement/ModelManager.d.ts

@@ -24,6 +24,9 @@ export declare class ModelManager extends EventEmitter<ModelManagerEvents> {
     private lifecycleState;
     private downloadLockHandle;
     private stopRequested;
+    private lastEngineExitCode;
+    private lastEngineExitSignal;
+    private reachedRunningState;
     protected readonly modelsDirectory: string;
     constructor({ contextLength, engineConfig, enginePort, engineType, logger, model, root }: {
         contextLength?: number | null;
@@ -43,7 +46,10 @@ export declare class ModelManager extends EventEmitter<ModelManagerEvents> {
     stop(): Promise<void>;
     get canStart(): boolean;
     get canStop(): boolean;
+    get lastExitCode(): number | null;
+    get lastExitSignal(): NodeJS.Signals | null;
     get state(): EngineLifecycleState;
+    get wasRunning(): boolean;
     private checkEngineReadiness;
     private checkGenericHealthReadiness;
     private checkLlamacppReadiness;

package/dist/modelManagement/classifyEngineError.d.ts

@@ -0,0 +1,12 @@
+import { EngineExecutionErrorType } from "@infersec/definitions";
+/**
+ * Coarse engine-failure classification used only to fill `engine_execution.error_type`. Exit code and
+ * terminating signal are consulted FIRST (SIGKILL/SIGSEGV are reliable OOM signals regardless of
+ * how much stderr the engine produced); the message text then refines the category. Anything
+ * unrecognized is "unknown".
+ */
+export declare function classifyEngineFailure({ error, exitCode, signal }: {
+    error: Error;
+    exitCode: number | null;
+    signal?: NodeJS.Signals | null;
+}): EngineExecutionErrorType;

package/dist/reporting/engineExecutionReporter.d.ts

@@ -0,0 +1,54 @@
+import { EngineExecutionErrorType, EngineExecutionReportPayload, LLMEngine } from "@infersec/definitions";
+import { Logger } from "@infersec/logger";
+export interface EngineExecutionReportSource {
+    durationMs: number;
+    errorDetail: string | null;
+    errorType: EngineExecutionErrorType | null;
+    ttftMs: number;
+    throughput: {
+        avgTps: number;
+        peakTps: number | null;
+    };
+    usage: {
+        completionTokens: number;
+        promptTokens: number;
+        totalTokens: number;
+    };
+}
+/**
+ * Flattens flat CLI extra-arg tokens into [arg, value] pairs, sorted by ARG NAME (ascending, ties by
+ * value). `--flag=value` pairs split on the first `=`; a bare `--flag` consumes the following token
+ * as its value when that token does not start with "-" (classic CLI convention); anything else
+ * (flags, non-strings) is dropped.
+ */
+export declare function pairExtraArgs(tokens: unknown): Array<[string, string]>;
+/**
+ * Files AT MOST ONE engine_execution report per engine startup. `beginStartup(runAt)` re-arms the
+ * latch on every fresh model start (initial boot or cycle) and stamps the startup epoch used as the
+ * row's `run_at`. Whichever of the three report* paths fires first wins: startup failure, in-flight
+ * crash, or first successful full prompt completion.
+ */
+export declare class EngineExecutionReporter {
+    private readonly options;
+    private currentStartupAt;
+    private reportedForCurrentStartup;
+    constructor(options: {
+        buildContext: () => {
+            engineType: LLMEngine;
+            engineVersion: string | null;
+            extraArgsPairs: Array<[string, string]>;
+        };
+        logger: Logger;
+        report: (payload: EngineExecutionReportPayload) => Promise<void>;
+        sourceLabel: string;
+    });
+    /** Re-arms the latch; the stamp becomes the row's `run_at` (moment startup began). */
+    beginStartup(runAt: Date): void;
+    /** Reports a startup failure (rejected prepare/start, readiness timeout, pre-ready death). */
+    reportStartupFailure(report: EngineExecutionReportSource): Promise<void>;
+    /** Reports a spontaneous crash of an engine that had reached the running state. */
+    reportRuntimeCrash(report: EngineExecutionReportSource): Promise<void>;
+    /** Reports the first fully-responded, token-bearing prompt completion since startup. */
+    reportSuccess(report: EngineExecutionReportSource): Promise<void>;
+    private file;
+}

package/dist/reporting/index.d.ts

@@ -0,0 +1 @@
+export { EngineExecutionReporter, type EngineExecutionReportSource, pairExtraArgs } from "./engineExecutionReporter.js";

package/dist/requestHandlers/createConduitAnthropicAPIReferenceHandlers.d.ts

@@ -86,7 +86,7 @@ export declare function createPostMessagesHandler(options: {
             budget_tokens: number;
             type: "enabled";
         } | undefined;
-        tool_choice?: "none" | "any" | "auto" | {
+        tool_choice?: "any" | "none" | "auto" | {
             type: "auto";
             disable_parallel_tool_use?: boolean | undefined;
         } | {

package/dist/utils/machineInfo.d.ts

@@ -1,8 +1,26 @@
 import type { InferenceAgentMachineMetadata } from "@infersec/definitions";
 import type { Logger } from "@infersec/logger";
 export declare function normalizeBusAddress(bus: string | null | undefined): string | null;
+export declare function parseRocmSmiMemory({ keyPrefix, stdout }: {
+    keyPrefix: "GTT" | "VRAM";
+    stdout: string;
+}): Array<{
+    bus: string;
+    memoryTotalBytes: number | null;
+    memoryUsedBytes: number | null;
+}>;
 export declare function parseNvidiaSmiNumber(value: string | undefined): number | null;
 export declare function parseNvidiaSmiMemoryBytes(value: string | undefined): number | null;
+export declare function mergeGTTMemory({ gttTotalBytes, gttUsedBytes, systemTotalBytes, vramTotalBytes, vramUsedBytes }: {
+    gttTotalBytes: number | null;
+    gttUsedBytes: number | null;
+    systemTotalBytes: number | null;
+    vramTotalBytes: number | null;
+    vramUsedBytes: number | null;
+}): {
+    memoryTotalBytes: number | null;
+    memoryUsedBytes: number | null;
+};
 export declare function collectMachineMetadata({ logger }: {
     logger: Logger;
 }): Promise<InferenceAgentMachineMetadata>;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@infersec/conduit",
   "description": "End user conduit agent for connecting local LLMs to the cloud.",
-  "version": "1.112.0",
+  "version": "1.113.0",
   "bin": {
     "infersec-conduit": "./dist/cli.js"
   },