@mlx-node/server 0.0.9 → 0.0.12

package/dist/chat-session-warm-reuse.d.ts

@@ -7,19 +7,17 @@
  * and nothing re-exports it. Callers are the two server endpoints —
  * `endpoints/responses.ts` (tier-1 / tier-2 HIT branch) and
  * `endpoints/messages.ts` (`getOrCreateWarmAny` HIT branch) — each
- * invoking the helper exclusively under a `SessionRegistry` HIT gate.
+ * invoking the helper only when native cache reuse is authorized: either a
+ * `SessionRegistry` hit or a block-paged model whose allocator validates
+ * reusable prefixes by content hash and isolates live cache owners.
  *
- * Why this helper exists at all: `ChatSession.reset()` is the safe
- * public wipe — it always calls `model.resetCaches()` because the
- * underlying `SessionCapableModel` is shared across every session
- * lifetime via the native `ModelRegistry`. A partial wipe that leaves
- * the shared native KV cache intact without the HIT gate would leak a
- * previous (unrelated) request's cached prefix into the next
- * `chat_session_start_sync` call. The server's warm-lease replay path
- * DOES have that HIT gate — the registry's own `getOrCreate` hit
- * signal is the authoritative proof that the native cache genuinely
- * belongs to this chain — so a JS-state-only reset that preserves the
- * native cache is correct there and only there.
+ * Why this helper exists at all: public `ChatSession.reset()` invalidates the
+ * current session's native state — owner-scoped on block-paged models and
+ * model-wide on exclusive/flat models. A warm registry HIT instead needs to
+ * preserve the already-authorized native prefix while clearing only the JS
+ * conversation wrapper. Warm leases provide the ownership gate; block-paged
+ * adapters provide the content-verification gate. A JS-state-only reset is
+ * valid in either case.
  *
  * Fields accessed: `inFlight`, `history`, `lastImagesKey`, `lastAudioKey`, `turnCount`,
  * `unresolvedOkToolCallCount`, `needsFullReplay`, `defaultConfig`, `activeTools`.

package/dist/chat-session-warm-reuse.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"chat-session-warm-reuse.d.ts","sourceRoot":"","sources":["../src/chat-session-warm-reuse.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AAEH,OAAO,KAAK,EAAc,WAAW,EAAE,mBAAmB,EAAE,MAAM,cAAc,CAAC;AAqBjF;;;;;;;;;;;;;;GAcG;AACH,wBAAsB,sCAAsC,CAAC,CAAC,SAAS,mBAAmB,EACxF,OAAO,EAAE,WAAW,CAAC,CAAC,CAAC,GACtB,OAAO,CAAC,IAAI,CAAC,CAyBf"}
+{"version":3,"file":"chat-session-warm-reuse.d.ts","sourceRoot":"","sources":["../src/chat-session-warm-reuse.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8BG;AAEH,OAAO,KAAK,EAAc,WAAW,EAAE,mBAAmB,EAAE,MAAM,cAAc,CAAC;AAqBjF;;;;;;;;;;;;;;GAcG;AACH,wBAAsB,sCAAsC,CAAC,CAAC,SAAS,mBAAmB,EACxF,OAAO,EAAE,WAAW,CAAC,CAAC,CAAC,GACtB,OAAO,CAAC,IAAI,CAAC,CAyBf"}

package/dist/chat-session-warm-reuse.js

@@ -7,19 +7,17 @@
  * and nothing re-exports it. Callers are the two server endpoints —
  * `endpoints/responses.ts` (tier-1 / tier-2 HIT branch) and
  * `endpoints/messages.ts` (`getOrCreateWarmAny` HIT branch) — each
- * invoking the helper exclusively under a `SessionRegistry` HIT gate.
+ * invoking the helper only when native cache reuse is authorized: either a
+ * `SessionRegistry` hit or a block-paged model whose allocator validates
+ * reusable prefixes by content hash and isolates live cache owners.
  *
- * Why this helper exists at all: `ChatSession.reset()` is the safe
- * public wipe — it always calls `model.resetCaches()` because the
- * underlying `SessionCapableModel` is shared across every session
- * lifetime via the native `ModelRegistry`. A partial wipe that leaves
- * the shared native KV cache intact without the HIT gate would leak a
- * previous (unrelated) request's cached prefix into the next
- * `chat_session_start_sync` call. The server's warm-lease replay path
- * DOES have that HIT gate — the registry's own `getOrCreate` hit
- * signal is the authoritative proof that the native cache genuinely
- * belongs to this chain — so a JS-state-only reset that preserves the
- * native cache is correct there and only there.
+ * Why this helper exists at all: public `ChatSession.reset()` invalidates the
+ * current session's native state — owner-scoped on block-paged models and
+ * model-wide on exclusive/flat models. A warm registry HIT instead needs to
+ * preserve the already-authorized native prefix while clearing only the JS
+ * conversation wrapper. Warm leases provide the ownership gate; block-paged
+ * adapters provide the content-verification gate. A JS-state-only reset is
+ * valid in either case.
  *
  * Fields accessed: `inFlight`, `history`, `lastImagesKey`, `lastAudioKey`, `turnCount`,
  * `unresolvedOkToolCallCount`, `needsFullReplay`, `defaultConfig`, `activeTools`.

package/dist/endpoints/messages.d.ts

@@ -31,7 +31,7 @@
  *   * **Non-paged path** (Qwen3.5 dense + MoE — default-off pending a
  *     perf decision; the Qianfan-OCR VLM — no adapter wired). Each
  *     request looks up the warm slot via
- *     `SessionRegistry.getOrCreateWarmAny(requestedSystem)`. On a
+ *     `SessionRegistry.getOrCreateWarmAny(requestedSystem, cacheSalt)`. On a
  *     HIT we keep the underlying native KV cache alive
  *     (`resetPreservingNativeCacheForWarmReuse` wipes only JS-side
  *     session state) so the native `verify_cache_prefix_direct` can
@@ -58,7 +58,7 @@
  */
 import type { IncomingMessage, ServerResponse } from 'node:http';
 import type { IdleSweeper } from '../idle-sweeper.js';
-import type { ModelWorkCoordinator } from '../model-work-coordinator.js';
+import { type ModelWorkCoordinator } from '../model-work-coordinator.js';
 import type { ModelRegistry } from '../registry.js';
 import type { AnthropicMessagesRequest } from '../types-anthropic.js';
 export declare function handleCreateMessage(res: ServerResponse, body: AnthropicMessagesRequest, registry: ModelRegistry, httpReq?: IncomingMessage, idleSweeper?: IdleSweeper | null, resolveModel?: (name: string) => Promise<void>, modelWorkCoordinator?: ModelWorkCoordinator): Promise<void>;

package/dist/endpoints/messages.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"messages.d.ts","sourceRoot":"","sources":["../../src/endpoints/messages.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAyDG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,cAAc,EAAE,MAAM,WAAW,CAAC;AAajE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AAgBtD,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,8BAA8B,CAAC;AACzE,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAepD,OAAO,KAAK,EAAE,wBAAwB,EAAE,MAAM,uBAAuB,CAAC;AAg3BtE,wBAAsB,mBAAmB,CACvC,GAAG,EAAE,cAAc,EACnB,IAAI,EAAE,wBAAwB,EAC9B,QAAQ,EAAE,aAAa,EACvB,OAAO,CAAC,EAAE,eAAe,EACzB,WAAW,CAAC,EAAE,WAAW,GAAG,IAAI,EAChC,YAAY,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,OAAO,CAAC,IAAI,CAAC,EAC9C,oBAAoB,CAAC,EAAE,oBAAoB,GAC1C,OAAO,CAAC,IAAI,CAAC,CAioBf"}
+{"version":3,"file":"messages.d.ts","sourceRoot":"","sources":["../../src/endpoints/messages.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAyDG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,cAAc,EAAE,MAAM,WAAW,CAAC;AAajE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AAgBtD,OAAO,EAGL,KAAK,oBAAoB,EAC1B,MAAM,8BAA8B,CAAC;AACtC,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAsBpD,OAAO,KAAK,EAAE,wBAAwB,EAAE,MAAM,uBAAuB,CAAC;AA64BtE,wBAAsB,mBAAmB,CACvC,GAAG,EAAE,cAAc,EACnB,IAAI,EAAE,wBAAwB,EAC9B,QAAQ,EAAE,aAAa,EACvB,OAAO,CAAC,EAAE,eAAe,EACzB,WAAW,CAAC,EAAE,WAAW,GAAG,IAAI,EAChC,YAAY,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,OAAO,CAAC,IAAI,CAAC,EAC9C,oBAAoB,CAAC,EAAE,oBAAoB,GAC1C,OAAO,CAAC,IAAI,CAAC,CAiwBf"}