@remnic/core 10.0.35 → 10.0.36

package/dist/orchestrator.js

@@ -4,13 +4,13 @@ import {
   graphPathRelativeToStorage,
   mergeGraphExpandedResults,
   qmdStartupCollectionChecksWithTimeout
-} from "./chunk-G2PC6IZB.js";
+} from "./chunk-VKB2FGNW.js";
 import "./chunk-XFDHL252.js";
-import "./chunk-JJWOKRZA.js";
+import "./chunk-IKCGNTGS.js";
 import {
   formatCompressionGuidelinesForRecall
 } from "./chunk-5PHOU5QI.js";
-import "./chunk-ZRU7636R.js";
+import "./chunk-V3JZZJ2E.js";
 import "./chunk-I74SUMNI.js";
 import "./chunk-ILNPU5W5.js";
 import "./chunk-VZAKYDMJ.js";
@@ -31,7 +31,7 @@ import "./chunk-7N4KAIGN.js";
 import "./chunk-V5OCT34X.js";
 import "./chunk-TECVW3JP.js";
 import "./chunk-C5JF6EZ7.js";
-import "./chunk-TVM47JQB.js";
+import "./chunk-RYUSWAAU.js";
 import "./chunk-5KBIOI4G.js";
 import "./chunk-3Q3XD3NO.js";
 import "./chunk-YLF42JQ7.js";
@@ -74,7 +74,7 @@ import "./chunk-PHRLJFWZ.js";
 import "./chunk-64NJRYU2.js";
 import "./chunk-LQ5SS4OP.js";
 import "./chunk-XXSOVESA.js";
-import "./chunk-GK56GNNS.js";
+import "./chunk-65JLKIVA.js";
 import "./chunk-K567EIW3.js";
 import "./chunk-54V4BZWP.js";
 import "./chunk-2CGLBUW3.js";
@@ -121,7 +121,7 @@ import "./chunk-GAARPWMN.js";
 import "./chunk-7G5NTS7F.js";
 import "./chunk-4NZQ7IID.js";
 import "./chunk-3I4FUAVJ.js";
-import "./chunk-GRBQO56D.js";
+import "./chunk-NBH7IRME.js";
 import "./chunk-H4CKS5B2.js";
 import "./chunk-MMACM5G2.js";
 import "./chunk-54BGIBCR.js";
@@ -235,7 +235,7 @@ import "./chunk-27NO2KVB.js";
 import "./chunk-KFZY2OFH.js";
 import "./chunk-B32WYRXX.js";
 import "./chunk-QW27DOHG.js";
-import "./chunk-CM6JDFCN.js";
+import "./chunk-KQUN4RSX.js";
 import "./chunk-V7IZKJI3.js";
 import "./chunk-XW2T4QAW.js";
 import "./chunk-G3T6YMBO.js";
@@ -249,10 +249,10 @@ import "./chunk-ZSEFGDCI.js";
 import "./chunk-P6P35FY4.js";
 import "./chunk-3HSRWJTC.js";
 import "./chunk-AHS3QEEA.js";
-import "./chunk-GRZX3RHY.js";
-import "./chunk-KVNOHMVE.js";
+import "./chunk-XO4HOBM2.js";
+import "./chunk-X544TW54.js";
 import "./chunk-SRT6SHVZ.js";
-import "./chunk-6NAN4WFY.js";
+import "./chunk-NDJ42GKB.js";
 import "./chunk-TE6S7KDF.js";
 import "./chunk-MZUQVWVC.js";
 import "./chunk-EIB4FJVR.js";
@@ -260,16 +260,16 @@ import "./chunk-A5KNZSNA.js";
 import "./chunk-HUSTM6PU.js";
 import "./chunk-D24OXEPB.js";
 import "./chunk-64IO6ZEE.js";
-import "./chunk-OLAQWYNF.js";
+import "./chunk-5XTSIZ3L.js";
 import "./chunk-QQ4T622E.js";
-import "./chunk-HT6IEMQ2.js";
+import "./chunk-NIANATGL.js";
 import "./chunk-M3GKSCFN.js";
-import "./chunk-MGXLLL5L.js";
-import "./chunk-GQVTALUJ.js";
-import "./chunk-P2RCFQA6.js";
+import "./chunk-AD3ZB72W.js";
+import "./chunk-OHMGR6QN.js";
+import "./chunk-ZSAQSLKL.js";
 import "./chunk-7T4TOGMS.js";
 import "./chunk-7YQ6ESEA.js";
-import "./chunk-FW7KWZBI.js";
+import "./chunk-F2JLBXMG.js";
 import "./chunk-GDYLZ6FR.js";
 import "./chunk-PQM3B54K.js";
 import "./chunk-5LWODP5R.js";
@@ -284,14 +284,14 @@ import "./chunk-DFP37GUN.js";
 import "./chunk-4ISHIPOC.js";
 import "./chunk-AB3A7EFH.js";
 import "./chunk-KIUVEMBI.js";
-import "./chunk-NLQBTHBV.js";
+import "./chunk-LJSBF6PW.js";
 import "./chunk-2BPPGNAQ.js";
 import "./chunk-DJSJ4ULL.js";
-import "./chunk-4LOMATQA.js";
+import "./chunk-PDYP7SUC.js";
 import {
   hasIdentityRecoveryIntent,
   resolveEffectiveIdentityInjectionMode
-} from "./chunk-SXRX6K35.js";
+} from "./chunk-4YYIZ475.js";
 import "./chunk-W4F3C7N7.js";
 import "./chunk-MPC7TYTZ.js";
 import "./chunk-STM4T335.js";
@@ -304,7 +304,7 @@ import "./chunk-BRZRP5WD.js";
 import "./chunk-GNVZQ5JN.js";
 import "./chunk-YN2UN7N3.js";
 import "./chunk-JWX2Q6HS.js";
-import "./chunk-D75ET36O.js";
+import "./chunk-FL6SLVAJ.js";
 import "./chunk-53OWLEAD.js";
 import "./chunk-WWDO7ZWM.js";
 import "./chunk-R4SATHCV.js";
@@ -368,11 +368,11 @@ import "./chunk-2LSZVONP.js";
 import "./chunk-DEUNUKTD.js";
 import "./chunk-LXY5VVGM.js";
 import "./chunk-GKCXTSFF.js";
-import "./chunk-SQ554O3K.js";
-import "./chunk-MLED254J.js";
+import "./chunk-74AY3S7U.js";
+import "./chunk-T6OZW2WR.js";
 import "./chunk-QEDLJDHP.js";
 import "./chunk-U4LUIE6Y.js";
-import "./chunk-JUXOTY6J.js";
+import "./chunk-5YQXWVCP.js";
 import "./chunk-ZY2MNJR6.js";
 import "./chunk-7WTTE5FW.js";
 import "./chunk-3UXOZBHV.js";

package/dist/recall-navigation-operations.js

@@ -1,5 +1,5 @@
-import "./chunk-MGXLLL5L.js";
-import "./chunk-SXRX6K35.js";
+import "./chunk-AD3ZB72W.js";
+import "./chunk-4YYIZ475.js";
 import "./chunk-W4F3C7N7.js";
 import "./chunk-MPC7TYTZ.js";
 import "./chunk-STM4T335.js";
@@ -12,7 +12,7 @@ import "./chunk-BRZRP5WD.js";
 import "./chunk-GNVZQ5JN.js";
 import "./chunk-YN2UN7N3.js";
 import "./chunk-JWX2Q6HS.js";
-import "./chunk-D75ET36O.js";
+import "./chunk-FL6SLVAJ.js";
 import "./chunk-53OWLEAD.js";
 import "./chunk-WWDO7ZWM.js";
 import "./chunk-R4SATHCV.js";
@@ -74,11 +74,11 @@ import "./chunk-2LSZVONP.js";
 import "./chunk-DEUNUKTD.js";
 import "./chunk-LXY5VVGM.js";
 import "./chunk-GKCXTSFF.js";
-import "./chunk-SQ554O3K.js";
-import "./chunk-MLED254J.js";
+import "./chunk-74AY3S7U.js";
+import "./chunk-T6OZW2WR.js";
 import "./chunk-QEDLJDHP.js";
 import "./chunk-U4LUIE6Y.js";
-import "./chunk-JUXOTY6J.js";
+import "./chunk-5YQXWVCP.js";
 import "./chunk-ZY2MNJR6.js";
 import "./chunk-7WTTE5FW.js";
 import "./chunk-3UXOZBHV.js";

package/dist/recall-navigation.js

@@ -1,6 +1,6 @@
 import {
   runRecallNavigation
-} from "./chunk-SXRX6K35.js";
+} from "./chunk-4YYIZ475.js";
 import "./chunk-W4F3C7N7.js";
 import "./chunk-MPC7TYTZ.js";
 import "./chunk-STM4T335.js";
@@ -13,7 +13,7 @@ import "./chunk-BRZRP5WD.js";
 import "./chunk-GNVZQ5JN.js";
 import "./chunk-YN2UN7N3.js";
 import "./chunk-JWX2Q6HS.js";
-import "./chunk-D75ET36O.js";
+import "./chunk-FL6SLVAJ.js";
 import "./chunk-53OWLEAD.js";
 import "./chunk-WWDO7ZWM.js";
 import "./chunk-R4SATHCV.js";
@@ -75,11 +75,11 @@ import "./chunk-2LSZVONP.js";
 import "./chunk-DEUNUKTD.js";
 import "./chunk-LXY5VVGM.js";
 import "./chunk-GKCXTSFF.js";
-import "./chunk-SQ554O3K.js";
-import "./chunk-MLED254J.js";
+import "./chunk-74AY3S7U.js";
+import "./chunk-T6OZW2WR.js";
 import "./chunk-QEDLJDHP.js";
 import "./chunk-U4LUIE6Y.js";
-import "./chunk-JUXOTY6J.js";
+import "./chunk-5YQXWVCP.js";
 import "./chunk-ZY2MNJR6.js";
 import "./chunk-7WTTE5FW.js";
 import "./chunk-3UXOZBHV.js";

package/dist/summarizer.js

@@ -1,15 +1,15 @@
 import {
   HourlySummarizer,
   summarizeHourlyStatus
-} from "./chunk-D75ET36O.js";
+} from "./chunk-FL6SLVAJ.js";
 import "./chunk-53OWLEAD.js";
 import "./chunk-5Y43VGYB.js";
 import "./chunk-5MA7ZLPD.js";
-import "./chunk-SQ554O3K.js";
-import "./chunk-MLED254J.js";
+import "./chunk-74AY3S7U.js";
+import "./chunk-T6OZW2WR.js";
 import "./chunk-QEDLJDHP.js";
 import "./chunk-U4LUIE6Y.js";
-import "./chunk-JUXOTY6J.js";
+import "./chunk-5YQXWVCP.js";
 import "./chunk-KF7PQ7G2.js";
 import "./chunk-TOTT6S2V.js";
 import "./chunk-GMMZ45C4.js";

package/dist/support-passport/index.js

@@ -4,7 +4,7 @@ import {
   SUPPORT_PASSPORT_MODEL_RESULT_PATH,
   SupportPassportModelBridge,
   parseSupportPassportModelJob
-} from "../chunk-JJWOKRZA.js";
+} from "../chunk-IKCGNTGS.js";
 import {
   DEFAULT_SUPPORT_PASSPORT_CONFIG,
   parseSupportPassportConfig
@@ -14,8 +14,8 @@ import {
   buildSupportPassportPublicRequestHandler,
   composeSupportPassportExternalRequestHandlers,
   normalizeSupportPassportNetworkAddress
-} from "../chunk-KVNOHMVE.js";
-import "../chunk-HT6IEMQ2.js";
+} from "../chunk-X544TW54.js";
+import "../chunk-NIANATGL.js";
 import {
   SUPPORT_PASSPORT_DRAFT_MAX_CONTENT_CHARACTERS,
   SUPPORT_PASSPORT_DRAFT_MAX_MEMORIES,
@@ -56,7 +56,7 @@ import {
   notifySupportPassportCommitted,
   resolveSupportPassportModelRoutePlan,
   syncDirectoryForDurability
-} from "../chunk-SXRX6K35.js";
+} from "../chunk-4YYIZ475.js";
 import "../chunk-W4F3C7N7.js";
 import "../chunk-MPC7TYTZ.js";
 import "../chunk-STM4T335.js";
@@ -69,7 +69,7 @@ import "../chunk-BRZRP5WD.js";
 import "../chunk-GNVZQ5JN.js";
 import "../chunk-YN2UN7N3.js";
 import "../chunk-JWX2Q6HS.js";
-import "../chunk-D75ET36O.js";
+import "../chunk-FL6SLVAJ.js";
 import "../chunk-53OWLEAD.js";
 import "../chunk-WWDO7ZWM.js";
 import "../chunk-R4SATHCV.js";
@@ -131,11 +131,11 @@ import "../chunk-2LSZVONP.js";
 import "../chunk-DEUNUKTD.js";
 import "../chunk-LXY5VVGM.js";
 import "../chunk-GKCXTSFF.js";
-import "../chunk-SQ554O3K.js";
-import "../chunk-MLED254J.js";
+import "../chunk-74AY3S7U.js";
+import "../chunk-T6OZW2WR.js";
 import "../chunk-QEDLJDHP.js";
 import "../chunk-U4LUIE6Y.js";
-import "../chunk-JUXOTY6J.js";
+import "../chunk-5YQXWVCP.js";
 import "../chunk-ZY2MNJR6.js";
 import "../chunk-7WTTE5FW.js";
 import "../chunk-3UXOZBHV.js";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@remnic/core",
-  "version": "10.0.35",
+  "version": "10.0.36",
   "description": "Framework-agnostic Remnic memory engine — orchestrator, storage, extraction, search, trust zones",
   "type": "module",
   "main": "dist/index.js",
@@ -3347,7 +3347,7 @@
     "core"
   ],
   "peerDependencies": {
-    "@remnic/coding-graph": "^10.0.35"
+    "@remnic/coding-graph": "^10.0.36"
   },
   "peerDependenciesMeta": {
     "@remnic/coding-graph": {

package/src/local-llm-headers-timeout.test.ts

@@ -12,6 +12,7 @@ import {
 } from "undici";
 
 import { LocalLlmClient } from "./local-llm.js";
+import { probeFetch } from "./local-llm-helpers.js";
 import { classifyAnalysisProviderError } from "./activity/timeline/analysis-provider.js";
 import { ChatTransport } from "./local-llm-transport.js";
 import { initLogger, resetLogger } from "./logger.js";
@@ -466,6 +467,69 @@ test("a process-wide custom dispatcher is left in place", async () => {
   await restored.close();
 });
 
+test("an explicit opt-in widens the pool even when a custom global dispatcher is installed", async () => {
+  // Some hosts install a process-wide dispatcher (for example an agent runtime's
+  // managed proxy) that is unrelated to the LAN model server. There the 300s
+  // header cap stood silently and every slow completion died with `fetch failed`
+  // at 301s. REMNIC_LOCAL_LLM_DIRECT=1 declares that the local endpoint is
+  // reached directly, so the transport uses its own pool. Without the flag the
+  // previous behavior (leave the custom dispatcher alone) is unchanged.
+  const original = getGlobalDispatcher();
+  const proxy = new ProxyAgent("http://127.0.0.1:9");
+  const priorFlag = process.env.REMNIC_LOCAL_LLM_DIRECT;
+  try {
+    setGlobalDispatcher(proxy);
+    delete process.env.REMNIC_LOCAL_LLM_DIRECT;
+    assert.equal(new ChatTransport().dispatcherFor(900_000), undefined, "no flag: custom dispatcher stays");
+
+    process.env.REMNIC_LOCAL_LLM_DIRECT = "1";
+    const transport = new ChatTransport();
+    const agent = transport.dispatcherFor(900_000);
+    assert.ok(agent instanceof Agent, "flag set: the widened pool must be used");
+    assert.ok(!(agent instanceof ProxyAgent), "flag set: the proxy dispatcher must not be used");
+    await agent.close();
+
+    process.env.REMNIC_LOCAL_LLM_DIRECT = "0";
+    assert.equal(new ChatTransport().dispatcherFor(900_000), undefined, "flag 0: custom dispatcher stays");
+  } finally {
+    if (priorFlag === undefined) delete process.env.REMNIC_LOCAL_LLM_DIRECT;
+    else process.env.REMNIC_LOCAL_LLM_DIRECT = priorFlag;
+    setGlobalDispatcher(original);
+    await proxy.close();
+  }
+});
+
+test("REMNIC_LOCAL_LLM_DIRECT=1 also sends availability probes around a global proxy dispatcher", async () => {
+  // A host whose global dispatcher cannot reach the LAN model server would fail
+  // checkAvailability() before any chat request runs, so direct mode must cover
+  // the probe as well as ChatTransport.post.
+  const server = http.createServer((_req, res) => {
+    res.setHeader("content-type", "application/json");
+    res.end('{"data":[]}');
+  });
+  await new Promise<void>((resolve) => server.listen(0, "127.0.0.1", resolve));
+  // SAFETY: listen(0, "127.0.0.1") above binds a TCP port, so address() is an AddressInfo, never a pipe string or null.
+  const { port } = server.address() as AddressInfo;
+  const url = `http://127.0.0.1:${port}/v1/models`;
+  const original = getGlobalDispatcher();
+  const priorFlag = process.env.REMNIC_LOCAL_LLM_DIRECT;
+  try {
+    setGlobalDispatcher(new ProxyAgent("http://127.0.0.1:9"));
+    delete process.env.REMNIC_LOCAL_LLM_DIRECT;
+    const viaProxy = await probeFetch(url, { timeoutMs: 2_000, headers: {} });
+    assert.equal(viaProxy.ok, false, "no flag: the probe goes through the (unreachable) proxy");
+
+    process.env.REMNIC_LOCAL_LLM_DIRECT = "1";
+    const direct = await probeFetch(url, { timeoutMs: 2_000, headers: {} });
+    assert.equal(direct.ok, true, "flag set: the probe reaches the endpoint directly");
+  } finally {
+    if (priorFlag === undefined) delete process.env.REMNIC_LOCAL_LLM_DIRECT;
+    else process.env.REMNIC_LOCAL_LLM_DIRECT = priorFlag;
+    setGlobalDispatcher(original);
+    await new Promise<void>((resolve) => server.close(() => resolve()));
+  }
+});
+
 test("aborting during a retry backoff stops the local request lane", async () => {
   const original = globalThis.fetch;
   const abortController = new AbortController();

package/src/local-llm-helpers.ts

@@ -1,5 +1,7 @@
 import type { ResponseFormatJSONSchema } from "openai/resources/shared";
+import type { Agent } from "undici";
 import { log } from "./logger.js";
+import { directProbeDispatcher } from "./local-llm-transport.js";
 import type { LocalLlmRequestPriority } from "./local-llm.js";
 
 export interface LocalLlmChatCompletionOptions {
@@ -167,7 +169,10 @@ export async function probeFetch(
     : controller.signal;
   const timeout = setTimeout(() => controller.abort(), options.timeoutMs);
   try {
-    const response = await fetch(url, { signal: requestSignal, headers: options.headers });
+    const init: RequestInit & { dispatcher?: Agent } = { signal: requestSignal, headers: options.headers };
+    const dispatcher = directProbeDispatcher();
+    if (dispatcher) init.dispatcher = dispatcher;
+    const response = await fetch(url, init);
     clearTimeout(timeout);
     if (!response.ok) {
       log.debug(`local LLM probe: ${url} returned HTTP ${response.status}`);

package/src/local-llm-transport.ts

@@ -40,6 +40,18 @@ function isPlainPoolingAgent(dispatcher: unknown): boolean {
   return ctor.name === "Agent";
 }
 
+/**
+ * REMNIC_LOCAL_LLM_DIRECT=1 declares that the local endpoint is reached
+ * directly. Availability probes use this shared plain pool in that mode so they
+ * take the same route as chat requests; `undefined` keeps the global dispatcher.
+ */
+let directProbeAgent: Agent | null = null;
+export function directProbeDispatcher(): Agent | undefined {
+  if (process.env.REMNIC_LOCAL_LLM_DIRECT !== "1") return undefined;
+  directProbeAgent ??= new Agent();
+  return directProbeAgent;
+}
+
 /**
  * Connection pool whose header/body inactivity budgets track the configured
  * request budget.
@@ -72,7 +84,8 @@ export class ChatTransport {
    * because only the dispatcher's owner knows how to rebuild it.
    */
   dispatcherFor(budgetMs: number): Agent | undefined {
-    if (!isPlainPoolingAgent(getGlobalDispatcher())) {
+    // REMNIC_LOCAL_LLM_DIRECT=1: the operator declares the local endpoint is reached directly even though the process installed a global dispatcher for other traffic (an agent runtime's managed proxy). Without it the 300s header cap stands silently on such hosts and every completion slower than five minutes dies with `fetch failed`. Off by default so deployments that must use a proxy keep it.
+    if (!isPlainPoolingAgent(getGlobalDispatcher()) && process.env.REMNIC_LOCAL_LLM_DIRECT !== "1") {
       if (!this.warnedAboutCustomDispatcher) {
         this.warnedAboutCustomDispatcher = true;
         log.debug(

package/dist/chunk-JUXOTY6J.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/local-llm-transport.ts"],"sourcesContent":["import { Agent, getGlobalDispatcher } from \"undici\";\n\nimport { log } from \"./logger.js\";\n\n/**\n * Wire transport for local LLM chat completions.\n *\n * Split out of `local-llm.ts` so the client module keeps prompt and\n * response semantics while connection budgets live in one place.\n */\n\nexport interface ChatCompletionRequest {\n  url: string;\n  headers: Record<string, string>;\n  body: string;\n  signal: AbortSignal;\n  /**\n   * Ceiling for this endpoint, i.e. `localLlmTimeoutMs`. Sizes the\n   * connection's inactivity budgets. Per-request budgets below the ceiling\n   * stay enforced by `signal`.\n   */\n  budgetMs: number;\n}\n\n/**\n * Is this undici's stock pooling `Agent` — the transport we may widen?\n *\n * `instanceof` alone is not enough: Node's global fetch is backed by its\n * OWN bundled copy of undici, so the dispatcher parked on the well-known\n * global symbol is an `Agent` from a different copy than the one we import,\n * and cross-copy `instanceof` is false. Fall back to the constructor name,\n * which is stable across copies and still distinguishes `ProxyAgent` and\n * `MockAgent` from the stock pool.\n */\nfunction isPlainPoolingAgent(dispatcher: unknown): boolean {\n  if (dispatcher instanceof Agent) return true;\n  if (!dispatcher || typeof dispatcher !== \"object\") return false;\n  const ctor = (dispatcher as { constructor?: unknown }).constructor;\n  if (!ctor || typeof ctor !== \"function\") return false;\n  return ctor.name === \"Agent\";\n}\n\n/**\n * Connection pool whose header/body inactivity budgets track the configured\n * request budget.\n *\n * Node's global fetch runs on undici, which applies a default\n * `headersTimeout` and `bodyTimeout` of 300s. A non-streaming completion\n * sends its response headers only once generation finishes, so any\n * completion slower than 300s died with a bare `fetch failed` no matter how\n * high `localLlmTimeoutMs` was set — the configured budget was dead letter\n * past five minutes (issue #2148).\n *\n * One pool per client is enough: it is keyed on the ceiling and rebuilt when\n * that changes.\n */\nexport class ChatTransport {\n  private agent: Agent | null = null;\n  private agentBudgetMs: number | null = null;\n  private warnedAboutCustomDispatcher = false;\n\n  /**\n   * The widened pool, or `undefined` when the process installed its own\n   * transport that we must not displace.\n   *\n   * A deployment can route the local endpoint through a process-wide\n   * dispatcher — a `ProxyAgent` for an HTTP proxy, or a custom\n   * connect/TLS/DNS dispatcher. Swapping in our own pool would bypass the\n   * proxy entirely and break the request, so those are left alone. A\n   * `MockAgent` is excluded by the same check, so test transports keep\n   * working. The trade-off is explicit: on those setups the 300s cap stands,\n   * because only the dispatcher's owner knows how to rebuild it.\n   */\n  dispatcherFor(budgetMs: number): Agent | undefined {\n    if (!isPlainPoolingAgent(getGlobalDispatcher())) {\n      if (!this.warnedAboutCustomDispatcher) {\n        this.warnedAboutCustomDispatcher = true;\n        log.debug(\n          \"local LLM: a custom global dispatcher is installed; leaving it in place. \" +\n            \"localLlmTimeoutMs above that dispatcher's headersTimeout will not take effect.\",\n        );\n      }\n      return undefined;\n    }\n    const normalized = Number.isFinite(budgetMs) ? Math.max(0, budgetMs) : 0;\n    if (this.agent && this.agentBudgetMs === normalized) return this.agent;\n    void this.agent?.close().catch(() => {\n      /* pool teardown is best-effort */\n    });\n    this.agent = new Agent({\n      headersTimeout: normalized,\n      bodyTimeout: normalized,\n    });\n    this.agentBudgetMs = normalized;\n    return this.agent;\n  }\n\n  async post(request: ChatCompletionRequest): Promise<Response> {\n    const init: RequestInit & { dispatcher?: Agent } = {\n      method: \"POST\",\n      headers: request.headers,\n      body: request.body,\n      signal: request.signal,\n    };\n    const dispatcher = this.dispatcherFor(request.budgetMs);\n    if (dispatcher) init.dispatcher = dispatcher;\n    return await fetch(request.url, init);\n  }\n}\n"],"mappings":";;;;;AAAA,SAAS,OAAO,2BAA2B;AAkC3C,SAAS,oBAAoB,YAA8B;AACzD,MAAI,sBAAsB,MAAO,QAAO;AACxC,MAAI,CAAC,cAAc,OAAO,eAAe,SAAU,QAAO;AAC1D,QAAM,OAAQ,WAAyC;AACvD,MAAI,CAAC,QAAQ,OAAO,SAAS,WAAY,QAAO;AAChD,SAAO,KAAK,SAAS;AACvB;AAgBO,IAAM,gBAAN,MAAoB;AAAA,EACjB,QAAsB;AAAA,EACtB,gBAA+B;AAAA,EAC/B,8BAA8B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EActC,cAAc,UAAqC;AACjD,QAAI,CAAC,oBAAoB,oBAAoB,CAAC,GAAG;AAC/C,UAAI,CAAC,KAAK,6BAA6B;AACrC,aAAK,8BAA8B;AACnC,YAAI;AAAA,UACF;AAAA,QAEF;AAAA,MACF;AACA,aAAO;AAAA,IACT;AACA,UAAM,aAAa,OAAO,SAAS,QAAQ,IAAI,KAAK,IAAI,GAAG,QAAQ,IAAI;AACvE,QAAI,KAAK,SAAS,KAAK,kBAAkB,WAAY,QAAO,KAAK;AACjE,SAAK,KAAK,OAAO,MAAM,EAAE,MAAM,MAAM;AAAA,IAErC,CAAC;AACD,SAAK,QAAQ,IAAI,MAAM;AAAA,MACrB,gBAAgB;AAAA,MAChB,aAAa;AAAA,IACf,CAAC;AACD,SAAK,gBAAgB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,KAAK,SAAmD;AAC5D,UAAM,OAA6C;AAAA,MACjD,QAAQ;AAAA,MACR,SAAS,QAAQ;AAAA,MACjB,MAAM,QAAQ;AAAA,MACd,QAAQ,QAAQ;AAAA,IAClB;AACA,UAAM,aAAa,KAAK,cAAc,QAAQ,QAAQ;AACtD,QAAI,WAAY,MAAK,aAAa;AAClC,WAAO,MAAM,MAAM,QAAQ,KAAK,IAAI;AAAA,EACtC;AACF;","names":[]}

package/dist/chunk-MLED254J.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/local-llm-helpers.ts"],"sourcesContent":["import type { ResponseFormatJSONSchema } from \"openai/resources/shared\";\nimport { log } from \"./logger.js\";\nimport type { LocalLlmRequestPriority } from \"./local-llm.js\";\n\nexport interface LocalLlmChatCompletionOptions {\n  temperature?: number;\n  maxTokens?: number;\n  responseFormat?: { type: string } | ResponseFormatJSONSchema;\n  timeoutMs?: number;\n  operation?: string;\n  forceDisableThinking?: boolean;\n  disableThinking?: boolean;\n  priority?: LocalLlmRequestPriority;\n  signal?: AbortSignal;\n  redactProviderErrors?: boolean;\n  /** Per-request model; defaults to config.localLlmModel. */\n  model?: string;\n  /**\n   * Optional out-box for the terminal failure cause (issue #2891): on a\n   * failed request the client records the last provider error here so\n   * callers can classify it (e.g. HTTP 429 → rate-limited) in memory\n   * instead of seeing only null. Never logged.\n   */\n  failureDiag?: { lastError?: unknown };\n}\n\nexport interface LocalLlmChatCompletionResult {\n  content: string;\n  usage?: { promptTokens: number; completionTokens: number; totalTokens: number };\n}\nexport function isAbortError(err: unknown): boolean {\n  if (!err || typeof err !== \"object\") return false;\n  const maybe = err as { name?: string; message?: string };\n  return (\n    maybe.name === \"AbortError\" ||\n    maybe.message === \"This operation was aborted\" ||\n    maybe.message === \"The operation was aborted\"\n  );\n}\n\nexport function waitForRetryBackoff(backoffMs: number, signal?: AbortSignal): Promise<boolean> {\n  return new Promise((resolve) => {\n    let settled = false;\n    let timer: NodeJS.Timeout | undefined;\n    const onAbort = (): void => {\n      if (settled) return;\n      settled = true;\n      if (timer) clearTimeout(timer);\n      signal?.removeEventListener(\"abort\", onAbort);\n      resolve(false);\n    };\n    const onTimer = (): void => {\n      if (settled) return;\n      settled = true;\n      signal?.removeEventListener(\"abort\", onAbort);\n      resolve(true);\n    };\n    timer = setTimeout(onTimer, backoffMs);\n    if (signal) {\n      if (signal.aborted) onAbort();\n      else signal.addEventListener(\"abort\", onAbort, { once: true });\n    }\n  });\n}\n\nexport function normalizeBackendTripReason(reason: string): string {\n  const cleaned = reason.replace(/\\s+/g, \" \").replace(/^[-:–—\\s]+/, \"\").trim();\n  if (!cleaned) return \"unknown local backend failure\";\n  return cleaned.length > 160 ? `${cleaned.slice(0, 157)}...` : cleaned;\n}\n\nexport function extractNonRecoverableBackendReason(reason: string): string | null {\n  const match = reason.match(\n    /Failed to load model|Library not loaded|different Team IDs|code signature|llm_engine_mlx_amphibian/i,\n  );\n  return match?.[0] ?? null;\n}\n\nexport function extractNonRecoverableBackendReasonFromErrorText(errorText: string): string | null {\n  const directReason = extractNonRecoverableBackendReason(errorText);\n  if (directReason) return directReason;\n  try {\n    const parsed = JSON.parse(errorText) as { error?: { message?: string } };\n    return extractNonRecoverableBackendReason(parsed?.error?.message ?? \"\");\n  } catch {\n    return null;\n  }\n}\n\n/**\n * Flatten a fetch rejection into one line, including the `cause` undici hides\n * the real reason behind (`fetch failed` alone says nothing).\n */\nexport function describeFetchFailure(err: unknown): string {\n  const name = err instanceof Error ? err.name : \"\";\n  const message = err instanceof Error ? err.message : String(err);\n  const cause: unknown = err instanceof Error ? (err as Error & { cause?: unknown }).cause : undefined;\n  const causeText = cause instanceof Error\n    ? ` (cause: ${cause.name}: ${cause.message}${\n      typeof (cause as Error & { code?: unknown }).code === \"string\"\n        ? ` [${String((cause as Error & { code?: unknown }).code)}]`\n        : \"\"\n    })`\n    : cause === undefined\n      ? \"\"\n      : ` (cause: ${String(cause)})`;\n  return `${name ? `${name}: ` : \"\"}${message}${causeText}`;\n}\n\n/**\n * What to record when every availability probe failed.\n *\n * A probe that TIMED OUT is not evidence the backend is down: the budget is a\n * fixed 2s and a busy event loop can burn it before the socket is scheduled,\n * so the backend looks dead while it answers other callers in single-digit ms.\n * Caching that verdict took extraction offline for a whole health-check\n * interval at a time (issue #2210), so a timeout leaves availability UNKNOWN —\n * this call fails, the next one re-probes instead of reading a stale false.\n *\n * A backend that ANSWERED and said no is a real verdict and stays cached, or\n * the daemon re-probes a known-bad endpoint on every request.\n */\nexport function resolveUnavailableVerdict(args: {\n  baseUrl: string;\n  sawAbortedProbe: boolean;\n  wasAvailable: boolean | null;\n}): { cacheVerdict: boolean; warning: string | null } {\n  if (args.sawAbortedProbe) {\n    return {\n      cacheVerdict: false,\n      warning:\n        `local LLM availability probe timed out at ${args.baseUrl} (event loop busy?); treating availability as ` +\n        \"unknown and re-probing on the next request rather than marking the backend down\",\n    };\n  }\n  return {\n    cacheVerdict: true,\n    // Surfaced at warn, not debug: extraction stops silently when this flips,\n    // and the daemon emits no debug at all unless configured for it.\n    warning: args.wasAvailable !== false ? `local LLM became unavailable at ${args.baseUrl}` : null,\n  };\n}\n\nexport interface ProbeFetchResult {\n  ok: boolean;\n  data: unknown;\n  status: number | null;\n  /** The probe hit its own budget rather than getting an answer. */\n  aborted?: boolean;\n}\n\n/**\n * One health-probe request.\n *\n * Every failure is logged with its cause: a transport error and a 404 both\n * returned `ok: false` with nothing recorded, so an unreachable backend was\n * indistinguishable from a wrong path and an availability failure had no\n * diagnosable reason at all (§22, issue #2210).\n */\nexport async function probeFetch(\n  url: string,\n  options: { timeoutMs: number; headers: Record<string, string>; signal?: AbortSignal },\n): Promise<ProbeFetchResult> {\n  const controller = new AbortController();\n  const requestSignal = options.signal\n    ? AbortSignal.any([options.signal, controller.signal])\n    : controller.signal;\n  const timeout = setTimeout(() => controller.abort(), options.timeoutMs);\n  try {\n    const response = await fetch(url, { signal: requestSignal, headers: options.headers });\n    clearTimeout(timeout);\n    if (!response.ok) {\n      log.debug(`local LLM probe: ${url} returned HTTP ${response.status}`);\n      return { ok: false, data: null, status: response.status };\n    }\n    const contentType = response.headers.get(\"content-type\");\n    const data = contentType?.includes(\"application/json\") ? await response.json() : await response.text();\n    return { ok: true, data, status: response.status };\n  } catch (err) {\n    clearTimeout(timeout);\n    const aborted = isAbortError(err);\n    log.debug(\n      `local LLM probe: ${url} ${aborted ? `timed out after ${options.timeoutMs}ms` : \"failed\"}: ${describeFetchFailure(err)}`,\n    );\n    return { ok: false, data: null, status: null, aborted };\n  }\n}\n\ninterface PendingProbe {\n  promise: Promise<boolean>;\n  controller: AbortController;\n  /** Callers still awaiting this sequence; at zero the sequence is aborted. */\n  waiters: number;\n}\n\n/**\n * Collapses concurrent probe sequences onto one in-flight run.\n *\n * A timed-out availability probe deliberately caches nothing (issue #2210),\n * which removed the false verdict that used to absorb concurrent callers:\n * every queued request would then run its own full probe sequence, on a host\n * already slow enough to blow the probe budget.\n *\n * Cancellation is refcounted rather than shared: one caller walking away must\n * not abort a sequence the others are still waiting on, so the underlying task\n * is aborted only once every waiter has gone.\n */\nexport class SingleFlightProbe {\n  private pending: PendingProbe | null = null;\n\n  run(task: (signal: AbortSignal) => Promise<boolean>, signal?: AbortSignal): Promise<boolean> {\n    let active = this.pending;\n    if (!active) {\n      const controller = new AbortController();\n      const created: PendingProbe = { promise: Promise.resolve(false), controller, waiters: 0 };\n      created.promise = task(controller.signal).finally(() => {\n        if (this.pending === created) this.pending = null;\n      });\n      this.pending = created;\n      active = created;\n    }\n    const entry = active;\n    entry.waiters += 1;\n    let released = false;\n    const release = () => {\n      if (released) return;\n      released = true;\n      entry.waiters -= 1;\n      if (entry.waiters > 0) return;\n      // Detach BEFORE aborting. `pending` would otherwise still point at this\n      // doomed entry until the task's own `finally` runs a microtask later, and\n      // a caller arriving in that window would join an aborted sequence and\n      // read its `false` as a verdict.\n      if (this.pending === entry) this.pending = null;\n      entry.controller.abort();\n    };\n    if (!signal) return entry.promise.finally(release);\n    return new Promise<boolean>((resolve, reject) => {\n      const onAbort = () => {\n        release();\n        resolve(false);\n      };\n      signal.addEventListener(\"abort\", onAbort, { once: true });\n      entry.promise.then(\n        (value) => {\n          signal.removeEventListener(\"abort\", onAbort);\n          release();\n          resolve(value);\n        },\n        (err: unknown) => {\n          signal.removeEventListener(\"abort\", onAbort);\n          release();\n          reject(err instanceof Error ? err : new Error(String(err)));\n        },\n      );\n    });\n  }\n}\n"],"mappings":";;;;;AA8BO,SAAS,aAAa,KAAuB;AAClD,MAAI,CAAC,OAAO,OAAO,QAAQ,SAAU,QAAO;AAC5C,QAAM,QAAQ;AACd,SACE,MAAM,SAAS,gBACf,MAAM,YAAY,gCAClB,MAAM,YAAY;AAEtB;AAEO,SAAS,oBAAoB,WAAmB,QAAwC;AAC7F,SAAO,IAAI,QAAQ,CAAC,YAAY;AAC9B,QAAI,UAAU;AACd,QAAI;AACJ,UAAM,UAAU,MAAY;AAC1B,UAAI,QAAS;AACb,gBAAU;AACV,UAAI,MAAO,cAAa,KAAK;AAC7B,cAAQ,oBAAoB,SAAS,OAAO;AAC5C,cAAQ,KAAK;AAAA,IACf;AACA,UAAM,UAAU,MAAY;AAC1B,UAAI,QAAS;AACb,gBAAU;AACV,cAAQ,oBAAoB,SAAS,OAAO;AAC5C,cAAQ,IAAI;AAAA,IACd;AACA,YAAQ,WAAW,SAAS,SAAS;AACrC,QAAI,QAAQ;AACV,UAAI,OAAO,QAAS,SAAQ;AAAA,UACvB,QAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AAAA,IAC/D;AAAA,EACF,CAAC;AACH;AAEO,SAAS,2BAA2B,QAAwB;AACjE,QAAM,UAAU,OAAO,QAAQ,QAAQ,GAAG,EAAE,QAAQ,cAAc,EAAE,EAAE,KAAK;AAC3E,MAAI,CAAC,QAAS,QAAO;AACrB,SAAO,QAAQ,SAAS,MAAM,GAAG,QAAQ,MAAM,GAAG,GAAG,CAAC,QAAQ;AAChE;AAEO,SAAS,mCAAmC,QAA+B;AAChF,QAAM,QAAQ,OAAO;AAAA,IACnB;AAAA,EACF;AACA,SAAO,QAAQ,CAAC,KAAK;AACvB;AAEO,SAAS,gDAAgD,WAAkC;AAChG,QAAM,eAAe,mCAAmC,SAAS;AACjE,MAAI,aAAc,QAAO;AACzB,MAAI;AACF,UAAM,SAAS,KAAK,MAAM,SAAS;AACnC,WAAO,mCAAmC,QAAQ,OAAO,WAAW,EAAE;AAAA,EACxE,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAMO,SAAS,qBAAqB,KAAsB;AACzD,QAAM,OAAO,eAAe,QAAQ,IAAI,OAAO;AAC/C,QAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,QAAM,QAAiB,eAAe,QAAS,IAAoC,QAAQ;AAC3F,QAAM,YAAY,iBAAiB,QAC/B,YAAY,MAAM,IAAI,KAAK,MAAM,OAAO,GACxC,OAAQ,MAAqC,SAAS,WAClD,KAAK,OAAQ,MAAqC,IAAI,CAAC,MACvD,EACN,MACE,UAAU,SACR,KACA,YAAY,OAAO,KAAK,CAAC;AAC/B,SAAO,GAAG,OAAO,GAAG,IAAI,OAAO,EAAE,GAAG,OAAO,GAAG,SAAS;AACzD;AAeO,SAAS,0BAA0B,MAIY;AACpD,MAAI,KAAK,iBAAiB;AACxB,WAAO;AAAA,MACL,cAAc;AAAA,MACd,SACE,6CAA6C,KAAK,OAAO;AAAA,IAE7D;AAAA,EACF;AACA,SAAO;AAAA,IACL,cAAc;AAAA;AAAA;AAAA,IAGd,SAAS,KAAK,iBAAiB,QAAQ,mCAAmC,KAAK,OAAO,KAAK;AAAA,EAC7F;AACF;AAkBA,eAAsB,WACpB,KACA,SAC2B;AAC3B,QAAM,aAAa,IAAI,gBAAgB;AACvC,QAAM,gBAAgB,QAAQ,SAC1B,YAAY,IAAI,CAAC,QAAQ,QAAQ,WAAW,MAAM,CAAC,IACnD,WAAW;AACf,QAAM,UAAU,WAAW,MAAM,WAAW,MAAM,GAAG,QAAQ,SAAS;AACtE,MAAI;AACF,UAAM,WAAW,MAAM,MAAM,KAAK,EAAE,QAAQ,eAAe,SAAS,QAAQ,QAAQ,CAAC;AACrF,iBAAa,OAAO;AACpB,QAAI,CAAC,SAAS,IAAI;AAChB,UAAI,MAAM,oBAAoB,GAAG,kBAAkB,SAAS,MAAM,EAAE;AACpE,aAAO,EAAE,IAAI,OAAO,MAAM,MAAM,QAAQ,SAAS,OAAO;AAAA,IAC1D;AACA,UAAM,cAAc,SAAS,QAAQ,IAAI,cAAc;AACvD,UAAM,OAAO,aAAa,SAAS,kBAAkB,IAAI,MAAM,SAAS,KAAK,IAAI,MAAM,SAAS,KAAK;AACrG,WAAO,EAAE,IAAI,MAAM,MAAM,QAAQ,SAAS,OAAO;AAAA,EACnD,SAAS,KAAK;AACZ,iBAAa,OAAO;AACpB,UAAM,UAAU,aAAa,GAAG;AAChC,QAAI;AAAA,MACF,oBAAoB,GAAG,IAAI,UAAU,mBAAmB,QAAQ,SAAS,OAAO,QAAQ,KAAK,qBAAqB,GAAG,CAAC;AAAA,IACxH;AACA,WAAO,EAAE,IAAI,OAAO,MAAM,MAAM,QAAQ,MAAM,QAAQ;AAAA,EACxD;AACF;AAqBO,IAAM,oBAAN,MAAwB;AAAA,EACrB,UAA+B;AAAA,EAEvC,IAAI,MAAiD,QAAwC;AAC3F,QAAI,SAAS,KAAK;AAClB,QAAI,CAAC,QAAQ;AACX,YAAM,aAAa,IAAI,gBAAgB;AACvC,YAAM,UAAwB,EAAE,SAAS,QAAQ,QAAQ,KAAK,GAAG,YAAY,SAAS,EAAE;AACxF,cAAQ,UAAU,KAAK,WAAW,MAAM,EAAE,QAAQ,MAAM;AACtD,YAAI,KAAK,YAAY,QAAS,MAAK,UAAU;AAAA,MAC/C,CAAC;AACD,WAAK,UAAU;AACf,eAAS;AAAA,IACX;AACA,UAAM,QAAQ;AACd,UAAM,WAAW;AACjB,QAAI,WAAW;AACf,UAAM,UAAU,MAAM;AACpB,UAAI,SAAU;AACd,iBAAW;AACX,YAAM,WAAW;AACjB,UAAI,MAAM,UAAU,EAAG;AAKvB,UAAI,KAAK,YAAY,MAAO,MAAK,UAAU;AAC3C,YAAM,WAAW,MAAM;AAAA,IACzB;AACA,QAAI,CAAC,OAAQ,QAAO,MAAM,QAAQ,QAAQ,OAAO;AACjD,WAAO,IAAI,QAAiB,CAAC,SAAS,WAAW;AAC/C,YAAM,UAAU,MAAM;AACpB,gBAAQ;AACR,gBAAQ,KAAK;AAAA,MACf;AACA,aAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AACxD,YAAM,QAAQ;AAAA,QACZ,CAAC,UAAU;AACT,iBAAO,oBAAoB,SAAS,OAAO;AAC3C,kBAAQ;AACR,kBAAQ,KAAK;AAAA,QACf;AAAA,QACA,CAAC,QAAiB;AAChB,iBAAO,oBAAoB,SAAS,OAAO;AAC3C,kBAAQ;AACR,iBAAO,eAAe,QAAQ,MAAM,IAAI,MAAM,OAAO,GAAG,CAAC,CAAC;AAAA,QAC5D;AAAA,MACF;AAAA,IACF,CAAC;AAAA,EACH;AACF;","names":[]}