@mlx-node/agent 0.0.7 → 0.0.9

package/dist/provider/warm-reuse.d.ts

@@ -18,18 +18,19 @@
  * the reused prefix and skip the corresponding re-prefill.
  *
  * Fields accessed: `inFlight`, `history`, `lastImagesKey`, `lastAudioKey`, `turnCount`,
- * `unresolvedOkToolCallCount`, `needsFullReplay`. These are TypeScript `private` fields on
- * `ChatSession` (compile-time only) — at runtime they are ordinary
- * properties. The cast through {@link ChatSessionWarmReuseInternals}
- * gives this helper a typed view of the instance without relaxing the
- * class's `private` declarations. The field names MUST stay in sync
- * with `packages/lm/src/chat-session.ts`; a mismatch would silently
- * skip the intended state wipe — the drift test in
+ * `unresolvedOkToolCallCount`, `needsFullReplay`, `defaultConfig`, `activeTools`.
+ * These are TypeScript `private` fields on `ChatSession` (compile-time
+ * only) — at runtime they are ordinary properties. The cast through
+ * {@link ChatSessionWarmReuseInternals} gives this helper a typed view
+ * of the instance without relaxing the class's `private` declarations.
+ * The field names MUST stay in sync with
+ * `packages/lm/src/chat-session.ts`; a mismatch would silently skip
+ * the intended state wipe — the drift test in
  * `packages/agent/__test__/warm-reuse.test.ts` checks every name in
  * {@link WARM_REUSE_TOUCHED_FIELDS} against a real `ChatSession`
  * instance.
  */
-import type { ChatSession, SessionCapableModel } from '@mlx-node/lm';
+import type { ChatConfig, ChatSession, SessionCapableModel } from '@mlx-node/lm';
 /**
  * Private structural view of the `ChatSession` JS-side state that the
  * warm-reuse helper needs to wipe. Mirrors the internal state
@@ -45,6 +46,8 @@ interface ChatSessionWarmReuseInternals {
     turnCount: number;
     unresolvedOkToolCallCount: number | null;
     needsFullReplay: boolean;
+    defaultConfig?: ChatConfig;
+    activeTools: ChatConfig['tools'];
 }
 /**
  * Runtime list of the `ChatSession` private field names this module

package/dist/provider/warm-reuse.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"warm-reuse.d.ts","sourceRoot":"","sources":["../../src/provider/warm-reuse.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAE,mBAAmB,EAAE,MAAM,cAAc,CAAC;AAErE;;;;;;GAMG;AACH,UAAU,6BAA6B;IACrC,QAAQ,EAAE,OAAO,CAAC;IAClB,OAAO,EAAE,OAAO,EAAE,CAAC;IACnB,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5B,SAAS,EAAE,MAAM,CAAC;IAClB,yBAAyB,EAAE,MAAM,GAAG,IAAI,CAAC;IACzC,eAAe,EAAE,OAAO,CAAC;CAC1B;AAkBD;;;;GAIG;AACH,eAAO,MAAM,yBAAyB,EAAgD,aAAa,CACjG,MAAM,6BAA6B,CACpC,CAAC;AAEF;;;;;;;;;;;;;;;GAeG;AACH,wBAAsB,sCAAsC,CAAC,CAAC,SAAS,mBAAmB,EACxF,OAAO,EAAE,WAAW,CAAC,CAAC,CAAC,GACtB,OAAO,CAAC,IAAI,CAAC,CAqBf"}
+{"version":3,"file":"warm-reuse.d.ts","sourceRoot":"","sources":["../../src/provider/warm-reuse.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AAEH,OAAO,KAAK,EAAE,UAAU,EAAE,WAAW,EAAE,mBAAmB,EAAE,MAAM,cAAc,CAAC;AAEjF;;;;;;GAMG;AACH,UAAU,6BAA6B;IACrC,QAAQ,EAAE,OAAO,CAAC;IAClB,OAAO,EAAE,OAAO,EAAE,CAAC;IACnB,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5B,SAAS,EAAE,MAAM,CAAC;IAClB,yBAAyB,EAAE,MAAM,GAAG,IAAI,CAAC;IACzC,eAAe,EAAE,OAAO,CAAC;IACzB,aAAa,CAAC,EAAE,UAAU,CAAC;IAC3B,WAAW,EAAE,UAAU,CAAC,OAAO,CAAC,CAAC;CAClC;AAoBD;;;;GAIG;AACH,eAAO,MAAM,yBAAyB,EAAgD,aAAa,CACjG,MAAM,6BAA6B,CACpC,CAAC;AAEF;;;;;;;;;;;;;;;GAeG;AACH,wBAAsB,sCAAsC,CAAC,CAAC,SAAS,mBAAmB,EACxF,OAAO,EAAE,WAAW,CAAC,CAAC,CAAC,GACtB,OAAO,CAAC,IAAI,CAAC,CAyBf"}

package/dist/provider/warm-reuse.js

@@ -18,13 +18,14 @@
  * the reused prefix and skip the corresponding re-prefill.
  *
  * Fields accessed: `inFlight`, `history`, `lastImagesKey`, `lastAudioKey`, `turnCount`,
- * `unresolvedOkToolCallCount`, `needsFullReplay`. These are TypeScript `private` fields on
- * `ChatSession` (compile-time only) — at runtime they are ordinary
- * properties. The cast through {@link ChatSessionWarmReuseInternals}
- * gives this helper a typed view of the instance without relaxing the
- * class's `private` declarations. The field names MUST stay in sync
- * with `packages/lm/src/chat-session.ts`; a mismatch would silently
- * skip the intended state wipe — the drift test in
+ * `unresolvedOkToolCallCount`, `needsFullReplay`, `defaultConfig`, `activeTools`.
+ * These are TypeScript `private` fields on `ChatSession` (compile-time
+ * only) — at runtime they are ordinary properties. The cast through
+ * {@link ChatSessionWarmReuseInternals} gives this helper a typed view
+ * of the instance without relaxing the class's `private` declarations.
+ * The field names MUST stay in sync with
+ * `packages/lm/src/chat-session.ts`; a mismatch would silently skip
+ * the intended state wipe — the drift test in
  * `packages/agent/__test__/warm-reuse.test.ts` checks every name in
  * {@link WARM_REUSE_TOUCHED_FIELDS} against a real `ChatSession`
  * instance.
@@ -43,6 +44,8 @@ const WARM_REUSE_TOUCHED_FIELD_SET = {
     turnCount: true,
     unresolvedOkToolCallCount: true,
     needsFullReplay: true,
+    defaultConfig: true,
+    activeTools: true,
 };
 /**
  * Runtime list of the `ChatSession` private field names this module
@@ -85,4 +88,8 @@ export async function resetPreservingNativeCacheForWarmReuse(session) {
     internals.turnCount = 0;
     internals.unresolvedOkToolCallCount = null;
     internals.needsFullReplay = false;
+    // Tools are conversation state. A provider replay can switch to an
+    // unrelated history, so restore constructor defaults exactly like
+    // ChatSession.reset() instead of leaking the prior committed overlay.
+    internals.activeTools = internals.defaultConfig?.tools;
 }

package/dist/run-agent.d.ts

@@ -17,7 +17,7 @@
  *   critical on pi's hard `process.exit()` paths.
  */
 import type { InlineExtension } from '@earendil-works/pi-coding-agent';
-import { type FilterableModelRegistryConstructor } from './provider/model-registry-filter.js';
+import { type FilterableModelRuntimeConstructor } from './provider/model-registry-filter.js';
 import type { MlxModelInfo } from './provider/models.js';
 /** Shape of pi's `main(argv, { extensionFactories })` — also the test seam. */
 export type RunAgentMain = (args: string[], opts: {
@@ -25,11 +25,11 @@ export type RunAgentMain = (args: string[], opts: {
 }) => Promise<void>;
 export interface RunAgentPi {
     main: RunAgentMain;
-    ModelRegistry: FilterableModelRegistryConstructor;
+    ModelRuntime: FilterableModelRuntimeConstructor;
 }
 /** @internal Narrow lifecycle seam for the agent's paged config overlays. */
 export interface AgentPagedConfigOverrides {
-    resolve(modelPath: string, modelType?: string): Promise<string>;
+    resolve(modelPath: string, modelType?: string, persistPagedCache?: boolean): Promise<string>;
     cleanup(): Promise<void>;
 }
 export interface RunAgentOptions {
@@ -41,6 +41,15 @@ export interface RunAgentOptions {
     argv: string[];
     /** Native inference-log path to surface after Pi takes over the TUI. */
     traceLogFile?: string;
+    /**
+     * Enable the SSD cold tier by default (the agent's default; the CLI sets it
+     * false for `--no-persist-cache`). Forwarded to {@link MlxModelHost}, which
+     * applies this ONE value to every load whose family is in
+     * `COLD_TIER_RESTORE_FAMILIES` — not to qwen3 alone. Families off that list
+     * are handed no policy because they can never persist, not because this flag
+     * spares them. `undefined` keeps the host's on-by-default behavior.
+     */
+    persistPagedCache?: boolean;
     /** @internal Test seam; when set, the pi dynamic import is skipped entirely. */
     piImpl?: RunAgentPi;
     /** @internal Test seam for paged model-path resolution and cleanup. */

package/dist/run-agent.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"run-agent.d.ts","sourceRoot":"","sources":["../src/run-agent.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAKH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,iCAAiC,CAAC;AASvE,OAAO,EACL,KAAK,kCAAkC,EAExC,MAAM,qCAAqC,CAAC;AAC7C,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,sBAAsB,CAAC;AAEzD,+EAA+E;AAC/E,MAAM,MAAM,YAAY,GAAG,CAAC,IAAI,EAAE,MAAM,EAAE,EAAE,IAAI,EAAE;IAAE,kBAAkB,EAAE,eAAe,EAAE,CAAA;CAAE,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;AAE9G,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,YAAY,CAAC;IACnB,aAAa,EAAE,kCAAkC,CAAC;CACnD;AAED,6EAA6E;AAC7E,MAAM,WAAW,yBAAyB;IACxC,OAAO,CAAC,SAAS,EAAE,MAAM,EAAE,SAAS,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IAChE,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAAC;CAC1B;AAED,MAAM,WAAW,eAAe;IAC9B,2FAA2F;IAC3F,SAAS,EAAE,MAAM,CAAC;IAClB,wEAAwE;IACxE,MAAM,EAAE,YAAY,EAAE,CAAC;IACvB,yDAAyD;IACzD,IAAI,EAAE,MAAM,EAAE,CAAC;IACf,wEAAwE;IACxE,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,gFAAgF;IAChF,MAAM,CAAC,EAAE,UAAU,CAAC;IACpB,uEAAuE;IACvE,oBAAoB,CAAC,EAAE,yBAAyB,CAAC;CAClD;AAED,mFAAmF;AACnF,wBAAgB,sBAAsB,CAAC,GAAG,GAAE,MAAM,CAAC,UAAwB,GAAG,OAAO,CAEpF;AAED;;;;;;;;;GASG;AACH,wBAAsB,QAAQ,CAAC,IAAI,EAAE,eAAe,GAAG,OAAO,CAAC,IAAI,CAAC,CAmDnE"}
+{"version":3,"file":"run-agent.d.ts","sourceRoot":"","sources":["../src/run-agent.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAKH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,iCAAiC,CAAC;AAWvE,OAAO,EACL,KAAK,iCAAiC,EAEvC,MAAM,qCAAqC,CAAC;AAC7C,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,sBAAsB,CAAC;AAEzD,+EAA+E;AAC/E,MAAM,MAAM,YAAY,GAAG,CAAC,IAAI,EAAE,MAAM,EAAE,EAAE,IAAI,EAAE;IAAE,kBAAkB,EAAE,eAAe,EAAE,CAAA;CAAE,KAAK,OAAO,CAAC,IAAI,CAAC,CAAC;AAE9G,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,YAAY,CAAC;IACnB,YAAY,EAAE,iCAAiC,CAAC;CACjD;AAED,6EAA6E;AAC7E,MAAM,WAAW,yBAAyB;IACxC,OAAO,CAAC,SAAS,EAAE,MAAM,EAAE,SAAS,CAAC,EAAE,MAAM,EAAE,iBAAiB,CAAC,EAAE,OAAO,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IAC7F,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC,CAAC;CAC1B;AAED,MAAM,WAAW,eAAe;IAC9B,2FAA2F;IAC3F,SAAS,EAAE,MAAM,CAAC;IAClB,wEAAwE;IACxE,MAAM,EAAE,YAAY,EAAE,CAAC;IACvB,yDAAyD;IACzD,IAAI,EAAE,MAAM,EAAE,CAAC;IACf,wEAAwE;IACxE,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB;;;;;;;OAOG;IACH,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B,gFAAgF;IAChF,MAAM,CAAC,EAAE,UAAU,CAAC;IACpB,uEAAuE;IACvE,oBAAoB,CAAC,EAAE,yBAAyB,CAAC;CAClD;AAED,mFAAmF;AACnF,wBAAgB,sBAAsB,CAAC,GAAG,GAAE,MAAM,CAAC,UAAwB,GAAG,OAAO,CAEpF;AAED;;;;;;;;;GASG;AACH,wBAAsB,QAAQ,CAAC,IAAI,EAAE,eAAe,GAAG,OAAO,CAAC,IAAI,CAAC,CA2EnE"}

package/dist/run-agent.js

@@ -18,7 +18,9 @@
  */
 import { homedir } from 'node:os';
 import { join } from 'node:path';
+import { coldCacheDrain } from '@mlx-node/core';
 import { PagedConfigOverrideManager } from '@mlx-node/lm';
+import { createLocalImageInputExtension } from './extensions/local-image-input.js';
 import { createPermissionGateExtension } from './extensions/permission-gate.js';
 import { createSubagentExtension } from './extensions/subagent.js';
 import { createTerminalTitleExtension } from './extensions/terminal-title.js';
@@ -46,6 +48,15 @@ export async function runAgent(opts) {
     // Mirrors `mlx launch claude`: chunked paged prefill keeps long-prompt
     // TTFT bounded on the default paged path.
     process.env.MLX_PAGED_PREFILL_CHUNK_SIZE ??= '2048';
+    // Hard offline invariant — NOT a user-overridable default, hence `=` not `??=`.
+    // `mlx agent` is local-only: no cloud provider may ever be contacted. pi 0.81.1's
+    // interactive and RPC startup call `ModelRuntime.refresh()`, which when PI_OFFLINE
+    // is unset fetches remote provider catalogs from pi.dev and can refresh a persisted
+    // cloud credential — a network path the mlx-only prototype filter does NOT cover
+    // (it patches the read methods, not `refresh`). Forcing PI_OFFLINE=1 here, before pi
+    // is imported below, also pins every ModelRuntime in this process to `allowNetwork`
+    // off, so no ambient/prior cloud credential can leak outbound traffic.
+    process.env.PI_OFFLINE = '1';
     // Force every paged-capable agent family through an isolated config clone.
     // This includes quantized LFM2 (whose standalone default is deliberately
     // flat) and Qwen3.5 dense/MoE (whose text-only defaults are flat). Gemma4's
@@ -55,19 +66,23 @@ export async function runAgent(opts) {
     const preserveEmbeddedGemmaDraft = agentGemmaDraftEnabled();
     const pagedConfigOverrides = opts.pagedConfigOverrides ?? new PagedConfigOverrideManager({ preserveEmbeddedGemmaDraft });
     const modelHost = new MlxModelHost(opts.models.map((model) => model.discovered), {
-        resolveModelPathFn: (model) => pagedConfigOverrides.resolve(model.path, model.modelType),
+        resolveModelPathFn: (model, policy) => pagedConfigOverrides.resolve(model.path, model.modelType, policy?.persistPagedCache),
         requirePagedCache: true,
+        persistPagedCache: opts.persistPagedCache,
     });
-    // Keep the pi import strictly behind the seam. The seam carries BOTH main
-    // and its registry class, so tests and production exercise the same policy
-    // installation/lifecycle instead of being able to bypass it accidentally.
+    // Keep the pi import strictly behind the seam. The seam carries BOTH main and
+    // the ModelRuntime class, so tests and production exercise the same policy
+    // installation/lifecycle instead of being able to bypass it accidentally. The
+    // filter patches the runtime prototype (not the extension-only ModelRegistry
+    // facade), which is where the selector / listing / resolution paths read.
     const pi = opts.piImpl ?? (await import('@earendil-works/pi-coding-agent'));
-    const restoreModelRegistry = installMlxOnlyModelRegistryFilter(pi.ModelRegistry, opts.models.map((model) => model.discovered.name));
+    const restoreModelRegistry = installMlxOnlyModelRegistryFilter(pi.ModelRuntime, opts.models.map((model) => model.discovered.name));
     const subagentsEnabled = opts.models.length > 0 && !opts.argv.includes('--no-extensions') && !opts.argv.includes('-ne');
     try {
         await pi.main(opts.argv, {
             extensionFactories: [
                 createMlxProviderExtension(opts.models, modelHost),
+                createLocalImageInputExtension(),
                 createPermissionGateExtension(),
                 ...(subagentsEnabled ? [createSubagentExtension()] : []),
                 ...(opts.traceLogFile !== undefined ? [createTraceNoticeExtension(opts.traceLogFile)] : []),
@@ -81,6 +96,17 @@ export async function runAgent(opts) {
         }
         finally {
             await pagedConfigOverrides.cleanup();
+            // `mlx agent -p` is one-shot: flush any accepted cold-tier prefix blocks
+            // to disk before the process exits, otherwise a prompt's just-persisted
+            // KV could still be queued/mid-write when we return. No-op when the tier
+            // was never opened, bounded so a stuck fsync can't hang exit, and never
+            // allowed to throw out of cleanup (best-effort durability).
+            try {
+                coldCacheDrain(5000);
+            }
+            catch {
+                // Best-effort: a drain failure must never mask the real exit path.
+            }
         }
     }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mlx-node/agent",
-  "version": "0.0.7",
+  "version": "0.0.9",
   "homepage": "https://github.com/mlx-node/mlx-node",
   "bugs": {
     "url": "https://github.com/mlx-node/mlx-node/issues"
@@ -21,16 +21,21 @@
     ".": {
       "types": "./dist/index.d.ts",
       "import": "./dist/index.js"
+    },
+    "./catalog": {
+      "types": "./dist/catalog.d.ts",
+      "default": "./dist/catalog.js"
     }
   },
   "scripts": {
     "build": "tsc -b"
   },
   "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",
+    "@earendil-works/pi-ai": "0.81.1",
+    "@earendil-works/pi-coding-agent": "0.81.1",
+    "@mlx-node/core": "0.0.9",
+    "@mlx-node/lm": "0.0.9",
+    "@mlx-node/server": "0.0.9",
     "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';
-    }
-}