@mlx-node/agent 0.0.7 → 0.0.8

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mlx-node/agent",
-  "version": "0.0.7",
+  "version": "0.0.8",
   "homepage": "https://github.com/mlx-node/mlx-node",
   "bugs": {
     "url": "https://github.com/mlx-node/mlx-node/issues"
@@ -29,8 +29,8 @@
   "dependencies": {
     "@earendil-works/pi-ai": "0.80.6",
     "@earendil-works/pi-coding-agent": "0.80.6",
-    "@mlx-node/lm": "0.0.7",
-    "@mlx-node/server": "0.0.7",
+    "@mlx-node/lm": "0.0.8",
+    "@mlx-node/server": "0.0.8",
     "typebox": "1.3.6"
   },
   "devDependencies": {

package/dist/provider/inference-trace.d.ts

@@ -1,58 +0,0 @@
-/**
- * Privacy-safe, best-effort inference tracing for the in-process mlx agent.
- *
- * The trace is deliberately metadata-only. Callers can only supply fields from
- * the allowlist below; prompt/output text, tool payloads, paths, token ids, and
- * error messages have no serialization path. Tracing is enabled only when
- * `MLX_INFERENCE_TRACE` is explicitly truthy and
- * `MLX_INFERENCE_TRACE_FILE` is non-empty.
- *
- * Writes are synchronous because this is a bounded, high-level lifecycle trace
- * (roughly a dozen records per request), not a token/layer trace. Every write is
- * swallowed on failure so observability can never change inference behavior.
- */
-export type InferenceTraceEvent = 'request.created' | 'host.queue.enter' | 'host.start' | 'host.resident' | 'host.load.start' | 'host.load.end' | 'host.callback.end' | 'session.reset' | 'session.prime' | 'session.config' | 'native.stream.start' | 'native.stream.first_delta' | 'request.terminal';
-export type InferenceTraceStatus = 'ok' | 'error' | 'aborted';
-export interface InferenceTraceFields {
-    status?: InferenceTraceStatus;
-    resident?: 'hit' | 'load' | 'swap';
-    resetMode?: 'preserve' | 'full';
-    finishReason?: 'stop' | 'length' | 'cancelled' | 'error' | 'other';
-    reasoningEffort?: 'none' | 'low' | 'medium' | 'high';
-    modelType?: 'gemma4' | 'qwen3' | 'qwen3_5' | 'qwen3_5_moe' | 'lfm2' | 'lfm2_moe' | 'other';
-    streaming?: boolean;
-    reportPerformance?: boolean;
-    isReasoning?: boolean;
-    durationMs?: number;
-    queueMs?: number;
-    messageCount?: number;
-    toolCount?: number;
-    maxNewTokens?: number;
-    promptTokens?: number;
-    cachedTokens?: number;
-    generatedTokens?: number;
-    reasoningTokens?: number;
-    ttftMs?: number;
-    prefillTokensPerSecond?: number;
-    decodeTokensPerSecond?: number;
-    mtpMeanAcceptedTokens?: number;
-    mtpMeanAcceptedTokensTotal?: number;
-    mtpCycles?: number;
-    mtpMeanDepth?: number;
-    mtpAcceptanceByPosition?: readonly number[];
-}
-export interface InferenceTraceSink {
-    write(line: string): void;
-}
-export declare class InferenceTrace {
-    private readonly sink;
-    readonly requestId: string;
-    private terminalWritten;
-    constructor(sink?: InferenceTraceSink | undefined);
-    emit(event: InferenceTraceEvent, fields?: InferenceTraceFields): void;
-    terminal(status: InferenceTraceStatus, fields?: InferenceTraceFields): void;
-}
-export declare function createInferenceTrace(): InferenceTrace;
-export declare function classifyFinishReason(value: string): NonNullable<InferenceTraceFields['finishReason']>;
-export declare function classifyModelType(value: string): NonNullable<InferenceTraceFields['modelType']>;
-//# sourceMappingURL=inference-trace.d.ts.map

package/dist/provider/inference-trace.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"inference-trace.d.ts","sourceRoot":"","sources":["../../src/provider/inference-trace.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAIH,MAAM,MAAM,mBAAmB,GAC3B,iBAAiB,GACjB,kBAAkB,GAClB,YAAY,GACZ,eAAe,GACf,iBAAiB,GACjB,eAAe,GACf,mBAAmB,GACnB,eAAe,GACf,eAAe,GACf,gBAAgB,GAChB,qBAAqB,GACrB,2BAA2B,GAC3B,kBAAkB,CAAC;AAEvB,MAAM,MAAM,oBAAoB,GAAG,IAAI,GAAG,OAAO,GAAG,SAAS,CAAC;AAE9D,MAAM,WAAW,oBAAoB;IACnC,MAAM,CAAC,EAAE,oBAAoB,CAAC;IAC9B,QAAQ,CAAC,EAAE,KAAK,GAAG,MAAM,GAAG,MAAM,CAAC;IACnC,SAAS,CAAC,EAAE,UAAU,GAAG,MAAM,CAAC;IAChC,YAAY,CAAC,EAAE,MAAM,GAAG,QAAQ,GAAG,WAAW,GAAG,OAAO,GAAG,OAAO,CAAC;IACnE,eAAe,CAAC,EAAE,MAAM,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;IACrD,SAAS,CAAC,EACN,QAAQ,GACR,OAAO,GACP,SAAS,GACT,aAAa,GACb,MAAM,GACN,UAAU,GACV,OAAO,CAAC;IACZ,SAAS,CAAC,EAAE,OAAO,CAAC;IACpB,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B,WAAW,CAAC,EAAE,OAAO,CAAC;IACtB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,sBAAsB,CAAC,EAAE,MAAM,CAAC;IAChC,qBAAqB,CAAC,EAAE,MAAM,CAAC;IAC/B,qBAAqB,CAAC,EAAE,MAAM,CAAC;IAC/B,0BAA0B,CAAC,EAAE,MAAM,CAAC;IACpC,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,uBAAuB,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;CAC7C;AAED,MAAM,WAAW,kBAAkB;IACjC,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;CAC3B;AA+ID,qBAAa,cAAc;IAIb,OAAO,CAAC,QAAQ,CAAC,IAAI;IAHjC,QAAQ,CAAC,SAAS,SAAmB;IACrC,OAAO,CAAC,eAAe,CAAS;IAEhC,YAA6B,IAAI,GAAE,kBAAkB,GAAG,SAAyB,EAAI;IAErF,IAAI,CAAC,KAAK,EAAE,mBAAmB,EAAE,MAAM,GAAE,oBAAyB,GAAG,IAAI,CAiBxE;IAED,QAAQ,CAAC,MAAM,EAAE,oBAAoB,EAAE,MAAM,GAAE,oBAAyB,GAAG,IAAI,CAI9E;CACF;AAED,wBAAgB,oBAAoB,IAAI,cAAc,CAErD;AAED,wBAAgB,oBAAoB,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,CAAC,oBAAoB,CAAC,cAAc,CAAC,CAAC,CAUrG;AAED,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,CAAC,oBAAoB,CAAC,WAAW,CAAC,CAAC,CAY/F"}

package/dist/provider/inference-trace.js

@@ -1,205 +0,0 @@
-/**
- * Privacy-safe, best-effort inference tracing for the in-process mlx agent.
- *
- * The trace is deliberately metadata-only. Callers can only supply fields from
- * the allowlist below; prompt/output text, tool payloads, paths, token ids, and
- * error messages have no serialization path. Tracing is enabled only when
- * `MLX_INFERENCE_TRACE` is explicitly truthy and
- * `MLX_INFERENCE_TRACE_FILE` is non-empty.
- *
- * Writes are synchronous because this is a bounded, high-level lifecycle trace
- * (roughly a dozen records per request), not a token/layer trace. Every write is
- * swallowed on failure so observability can never change inference behavior.
- */
-import { appendFileSync } from 'node:fs';
-const EVENTS = new Set([
-    'request.created',
-    'host.queue.enter',
-    'host.start',
-    'host.resident',
-    'host.load.start',
-    'host.load.end',
-    'host.callback.end',
-    'session.reset',
-    'session.prime',
-    'session.config',
-    'native.stream.start',
-    'native.stream.first_delta',
-    'request.terminal',
-]);
-const STATUS = new Set(['ok', 'error', 'aborted']);
-const RESIDENT = new Set(['hit', 'load', 'swap']);
-const RESET_MODE = new Set(['preserve', 'full']);
-const FINISH_REASON = new Set([
-    'stop',
-    'length',
-    'cancelled',
-    'error',
-    'other',
-]);
-const REASONING_EFFORT = new Set([
-    'none',
-    'low',
-    'medium',
-    'high',
-]);
-const MODEL_TYPE = new Set([
-    'gemma4',
-    'qwen3',
-    'qwen3_5',
-    'qwen3_5_moe',
-    'lfm2',
-    'lfm2_moe',
-    'other',
-]);
-let requestSequence = 0;
-let eventSequence = 0;
-function envFlagEnabled(value) {
-    if (value == null)
-        return false;
-    return ['1', 'true', 'yes', 'on'].includes(value.trim().toLowerCase());
-}
-function defaultSink() {
-    if (!envFlagEnabled(process.env.MLX_INFERENCE_TRACE))
-        return undefined;
-    const path = process.env.MLX_INFERENCE_TRACE_FILE?.trim();
-    if (!path)
-        return undefined;
-    return {
-        write(line) {
-            appendFileSync(path, line, { encoding: 'utf8', flag: 'a', mode: 0o600 });
-        },
-    };
-}
-function nextRequestId() {
-    requestSequence += 1;
-    return `agent-${process.pid.toString(36)}-${requestSequence.toString(36)}`;
-}
-function nonNegativeFinite(value) {
-    return typeof value === 'number' && Number.isFinite(value) && value >= 0 ? value : undefined;
-}
-function nonNegativeInteger(value) {
-    const n = nonNegativeFinite(value);
-    return n !== undefined && Number.isSafeInteger(n) ? n : undefined;
-}
-function copyNumber(out, fields, key, integer) {
-    const value = integer ? nonNegativeInteger(fields[key]) : nonNegativeFinite(fields[key]);
-    if (value !== undefined)
-        out[key] = value;
-}
-/** Runtime allowlist: protects the file even if an untyped caller supplies extras. */
-function sanitizeFields(fields) {
-    const out = {};
-    if (fields.status !== undefined && STATUS.has(fields.status))
-        out.status = fields.status;
-    if (fields.resident !== undefined && RESIDENT.has(fields.resident))
-        out.resident = fields.resident;
-    if (fields.resetMode !== undefined && RESET_MODE.has(fields.resetMode))
-        out.resetMode = fields.resetMode;
-    if (fields.finishReason !== undefined && FINISH_REASON.has(fields.finishReason)) {
-        out.finishReason = fields.finishReason;
-    }
-    if (fields.reasoningEffort !== undefined && REASONING_EFFORT.has(fields.reasoningEffort)) {
-        out.reasoningEffort = fields.reasoningEffort;
-    }
-    if (fields.modelType !== undefined && MODEL_TYPE.has(fields.modelType))
-        out.modelType = fields.modelType;
-    for (const key of ['streaming', 'reportPerformance', 'isReasoning']) {
-        if (typeof fields[key] === 'boolean')
-            out[key] = fields[key];
-    }
-    for (const key of [
-        'durationMs',
-        'queueMs',
-        'ttftMs',
-        'prefillTokensPerSecond',
-        'decodeTokensPerSecond',
-        'mtpMeanAcceptedTokens',
-        'mtpMeanAcceptedTokensTotal',
-        'mtpMeanDepth',
-    ]) {
-        copyNumber(out, fields, key, false);
-    }
-    for (const key of [
-        'messageCount',
-        'toolCount',
-        'maxNewTokens',
-        'promptTokens',
-        'cachedTokens',
-        'generatedTokens',
-        'reasoningTokens',
-        'mtpCycles',
-    ]) {
-        copyNumber(out, fields, key, true);
-    }
-    if (Array.isArray(fields.mtpAcceptanceByPosition)) {
-        const values = fields.mtpAcceptanceByPosition
-            .slice(0, 16)
-            .map(nonNegativeFinite)
-            .filter((value) => value !== undefined);
-        if (values.length > 0)
-            out.mtpAcceptanceByPosition = values;
-    }
-    return out;
-}
-export class InferenceTrace {
-    sink;
-    requestId = nextRequestId();
-    terminalWritten = false;
-    constructor(sink = defaultSink()) {
-        this.sink = sink;
-    }
-    emit(event, fields = {}) {
-        if (!this.sink || !EVENTS.has(event))
-            return;
-        try {
-            eventSequence += 1;
-            const record = {
-                timestamp: new Date().toISOString(),
-                pid: process.pid,
-                seq: eventSequence,
-                requestId: this.requestId,
-                component: 'agent',
-                event,
-                ...sanitizeFields(fields),
-            };
-            this.sink.write(`${JSON.stringify(record)}\n`);
-        }
-        catch {
-            // Diagnostics are best-effort and must never affect inference.
-        }
-    }
-    terminal(status, fields = {}) {
-        if (this.terminalWritten)
-            return;
-        this.terminalWritten = true;
-        this.emit('request.terminal', { ...fields, status });
-    }
-}
-export function createInferenceTrace() {
-    return new InferenceTrace();
-}
-export function classifyFinishReason(value) {
-    switch (value) {
-        case 'stop':
-        case 'length':
-        case 'cancelled':
-        case 'error':
-            return value;
-        default:
-            return 'other';
-    }
-}
-export function classifyModelType(value) {
-    switch (value) {
-        case 'gemma4':
-        case 'qwen3':
-        case 'qwen3_5':
-        case 'qwen3_5_moe':
-        case 'lfm2':
-        case 'lfm2_moe':
-            return value;
-        default:
-            return 'other';
-    }
-}