@remnic/core 10.0.91 → 10.0.93

package/dist/chunk-T6OZW2WR.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/local-llm-helpers.ts"],"sourcesContent":["import type { ResponseFormatJSONSchema } from \"openai/resources/shared\";\nimport type { Agent } from \"undici\";\nimport { log } from \"./logger.js\";\nimport { directProbeDispatcher } from \"./local-llm-transport.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 init: RequestInit & { dispatcher?: Agent } = { signal: requestSignal, headers: options.headers };\n    const dispatcher = directProbeDispatcher();\n    if (dispatcher) init.dispatcher = dispatcher;\n    const response = await fetch(url, init);\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":";;;;;;;;AAgCO,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,OAA6C,EAAE,QAAQ,eAAe,SAAS,QAAQ,QAAQ;AACrG,UAAM,aAAa,sBAAsB;AACzC,QAAI,WAAY,MAAK,aAAa;AAClC,UAAM,WAAW,MAAM,MAAM,KAAK,IAAI;AACtC,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":[]}