@remnic/core 1.0.0

package/dist/chunk-MDDAA2AO.js

@@ -0,0 +1,925 @@
+import {
+  launchProcessSync
+} from "./chunk-LK6SGL53.js";
+import {
+  mergeEnv,
+  readEnvVar
+} from "./chunk-MARWOCVP.js";
+import {
+  log
+} from "./chunk-KWBU5S5U.js";
+
+// src/local-llm.ts
+import { existsSync, readFileSync } from "fs";
+import os from "os";
+function trimTrailingSlashes(s) {
+  let end = s.length;
+  while (end > 0 && s[end - 1] === "/") end--;
+  return s.substring(0, end);
+}
+var LOCAL_SERVERS = [
+  {
+    type: "ollama",
+    defaultPort: 11434,
+    healthEndpoint: "/",
+    modelsEndpoint: "/api/tags",
+    detectFn: (resp) => typeof resp === "string" && resp.includes("Ollama")
+  },
+  {
+    type: "mlx",
+    defaultPort: 8080,
+    healthEndpoint: "/v1/models",
+    modelsEndpoint: "/v1/models",
+    detectFn: (resp) => typeof resp === "object" && resp !== null && "data" in resp && Array.isArray(resp.data)
+  },
+  {
+    type: "lmstudio",
+    defaultPort: 1234,
+    healthEndpoint: "/v1/models",
+    modelsEndpoint: "/v1/models",
+    detectFn: (resp) => typeof resp === "object" && resp !== null && "data" in resp && Array.isArray(resp.data)
+  },
+  {
+    type: "vllm",
+    defaultPort: 8e3,
+    healthEndpoint: "/health",
+    modelsEndpoint: "/v1/models",
+    detectFn: (resp) => resp === "" || typeof resp === "object" && resp !== null
+  }
+];
+var LOCAL_LLM_GLOBAL_BACKEND_STATE = "__openclawEngramLocalLlmBackendState";
+var LocalLlmClient = class _LocalLlmClient {
+  config;
+  isAvailable = null;
+  lastHealthCheck = 0;
+  detectedType = null;
+  cachedModelInfo = null;
+  cachedLmsContext = null;
+  lastLmsCheck = 0;
+  consecutive400s = 0;
+  cooldownUntilMs = 0;
+  modelRegistry;
+  _disableThinking = false;
+  requestQueues = {
+    "recall-critical": [],
+    background: []
+  };
+  queueProcessing = /* @__PURE__ */ new Set();
+  queueDrainScheduled = false;
+  static HEALTH_CHECK_INTERVAL_MS = 6e4;
+  // 1 minute
+  static LMS_CACHE_INTERVAL_MS = 3e4;
+  // 30 seconds
+  constructor(config, modelRegistry) {
+    this.config = config;
+    this.modelRegistry = modelRegistry;
+  }
+  /**
+   * Disable thinking/reasoning mode for models that support it (e.g. Qwen 3.5).
+   * When enabled, adds chat_template_kwargs to suppress chain-of-thought,
+   * reducing latency for fast-tier operations.
+   */
+  set disableThinking(value) {
+    this._disableThinking = value;
+  }
+  resolveHomeDir() {
+    return this.config.localLlmHomeDir || readEnvVar("HOME") || os.homedir();
+  }
+  buildRequestHeaders(base = {}) {
+    const headers = {
+      ...base,
+      ...this.config.localLlmHeaders ?? {}
+    };
+    if (this.config.localLlmApiKey && this.config.localLlmAuthHeader !== false) {
+      headers.Authorization = `Bearer ${this.config.localLlmApiKey}`;
+    }
+    return headers;
+  }
+  isAbortError(err) {
+    if (!err || typeof err !== "object") return false;
+    const maybe = err;
+    return maybe.name === "AbortError" || maybe.message === "This operation was aborted" || maybe.message === "The operation was aborted";
+  }
+  /**
+   * Set the ModelRegistry for caching detected capabilities
+   */
+  setModelRegistry(registry) {
+    this.modelRegistry = registry;
+  }
+  /**
+   * Get the detected server type (null if not detected)
+   */
+  getDetectedType() {
+    return this.detectedType;
+  }
+  getBackendKey() {
+    return trimTrailingSlashes(
+      this.config.localLlmUrl.replace("localhost", "127.0.0.1")
+    ).replace(/\/v1$/, "");
+  }
+  getGlobalBackendState() {
+    const globalAny = globalThis;
+    if (!globalAny[LOCAL_LLM_GLOBAL_BACKEND_STATE]) {
+      globalAny[LOCAL_LLM_GLOBAL_BACKEND_STATE] = /* @__PURE__ */ new Map();
+    }
+    return globalAny[LOCAL_LLM_GLOBAL_BACKEND_STATE];
+  }
+  getTrippedBackendState(now) {
+    const state = this.getGlobalBackendState().get(this.getBackendKey()) ?? null;
+    if (!state) return null;
+    if (state.untilMs <= now) {
+      this.getGlobalBackendState().delete(this.getBackendKey());
+      this.lastHealthCheck = 0;
+      return null;
+    }
+    return state;
+  }
+  markBackendUnavailable(reason, durationMs) {
+    const normalizedReason = this.normalizeBackendTripReason(reason);
+    if (durationMs > 0) {
+      const untilMs = Date.now() + durationMs;
+      this.getGlobalBackendState().set(this.getBackendKey(), { untilMs, reason: normalizedReason });
+    } else {
+      this.getGlobalBackendState().delete(this.getBackendKey());
+    }
+    this.isAvailable = false;
+    this.lastHealthCheck = 0;
+    log.warn(
+      `local LLM backend unavailable for ${durationMs}ms: model=${this.config.localLlmModel} reason=${normalizedReason}`
+    );
+  }
+  extractNonRecoverableBackendReason(reason) {
+    const match = reason.match(
+      /Failed to load model|Library not loaded|different Team IDs|code signature|llm_engine_mlx_amphibian/i
+    );
+    return match?.[0] ?? null;
+  }
+  extractNonRecoverableBackendReasonFromErrorText(errorText) {
+    const directReason = this.extractNonRecoverableBackendReason(errorText);
+    if (directReason) return directReason;
+    try {
+      const parsed = JSON.parse(errorText);
+      return this.extractNonRecoverableBackendReason(parsed?.error?.message ?? "");
+    } catch {
+      return null;
+    }
+  }
+  normalizeBackendTripReason(reason) {
+    const cleaned = reason.replace(/\s+/g, " ").replace(/^[-:–—\s]+/, "").trim();
+    if (!cleaned) return "unknown local backend failure";
+    return cleaned.length > 160 ? `${cleaned.slice(0, 157)}...` : cleaned;
+  }
+  /**
+   * Fetch with timeout for health checks
+   */
+  async fetchWithTimeout(url, timeoutMs = 2e3, headers) {
+    const controller = new AbortController();
+    const timeout = setTimeout(() => controller.abort(), timeoutMs);
+    try {
+      const response = await fetch(url, {
+        signal: controller.signal,
+        headers: this.buildRequestHeaders({ Accept: "application/json", ...headers ?? {} })
+      });
+      clearTimeout(timeout);
+      if (!response.ok) {
+        return { ok: false, data: null, status: response.status };
+      }
+      const contentType = response.headers.get("content-type");
+      if (contentType?.includes("application/json")) {
+        return { ok: true, data: await response.json(), status: response.status };
+      } else {
+        return { ok: true, data: await response.text(), status: response.status };
+      }
+    } catch (err) {
+      clearTimeout(timeout);
+      return { ok: false, data: null, status: null };
+    }
+  }
+  /**
+   * Check if local LLM is available
+   * Uses 127.0.0.1 instead of localhost to avoid DNS issues (consistent with tactician)
+   */
+  async checkAvailability() {
+    const now = Date.now();
+    const trippedState = this.getTrippedBackendState(now);
+    if (trippedState) {
+      this.isAvailable = false;
+      this.lastHealthCheck = 0;
+      log.info(
+        `local LLM availability: backend circuit open for ${Math.max(0, trippedState.untilMs - now)}ms (${trippedState.reason})`
+      );
+      return false;
+    }
+    if (this.isAvailable !== null && now - this.lastHealthCheck < _LocalLlmClient.HEALTH_CHECK_INTERVAL_MS) {
+      return this.isAvailable;
+    }
+    const baseUrl = trimTrailingSlashes(
+      this.config.localLlmUrl.replace("localhost", "127.0.0.1")
+    );
+    let sawUnauthorizedProbe = false;
+    for (const serverConfig of LOCAL_SERVERS) {
+      const healthUrl = `${baseUrl}${serverConfig.healthEndpoint}`;
+      log.debug(`checking ${serverConfig.type} at ${healthUrl}`);
+      const result = await this.fetchWithTimeout(healthUrl);
+      if (result.ok && serverConfig.detectFn(result.data)) {
+        this.isAvailable = true;
+        this.detectedType = serverConfig.type;
+        this.lastHealthCheck = now;
+        log.info(`detected ${serverConfig.type} at ${baseUrl}`);
+        return true;
+      }
+      if (result.status === 401 || result.status === 403) {
+        sawUnauthorizedProbe = true;
+      }
+    }
+    try {
+      const modelsUrl = `${baseUrl}/v1/models`;
+      const result = await this.fetchWithTimeout(modelsUrl);
+      if (result.ok) {
+        this.isAvailable = true;
+        this.detectedType = "generic";
+        this.lastHealthCheck = now;
+        log.info(`detected generic OpenAI-compatible server at ${baseUrl}`);
+        return true;
+      }
+      if (result.status === 401 || result.status === 403) {
+        sawUnauthorizedProbe = true;
+      }
+    } catch {
+    }
+    this.isAvailable = false;
+    this.detectedType = null;
+    this.lastHealthCheck = now;
+    if (sawUnauthorizedProbe) {
+      log.warn(
+        `local LLM availability probe was unauthorized at ${baseUrl}; verify localLlmApiKey and localLlmAuthHeader settings`
+      );
+    }
+    log.debug("local LLM not available at", baseUrl);
+    return false;
+  }
+  /**
+   * Try to get context window from LM Studio settings.json as fallback.
+   * This reads the defaultContextLength setting which is what LM Studio uses
+   * when loading models without explicit context configuration.
+   */
+  getContextFromLmStudioSettings() {
+    try {
+      const homeDir = this.resolveHomeDir();
+      const settingsPath = `${homeDir}/.cache/lm-studio/settings.json`;
+      if (!existsSync(settingsPath)) {
+        log.debug(`LM Studio settings: file not found at ${settingsPath}`);
+        return null;
+      }
+      const content = readFileSync(settingsPath, "utf-8");
+      const settings = JSON.parse(content);
+      if (settings.defaultContextLength?.value) {
+        const contextWindow = settings.defaultContextLength.value;
+        log.debug(`LM Studio settings: found default context length: ${contextWindow}`);
+        return contextWindow;
+      }
+      return null;
+    } catch (err) {
+      const errorMsg = err instanceof Error ? err.message : String(err);
+      log.debug(`LM Studio settings: failed to read - ${errorMsg}`);
+      return null;
+    }
+  }
+  /**
+   * Try to get context window from LMS CLI (LM Studio specific).
+   * Uses --json flag for reliable parsing.
+   * Returns null if LMS CLI is not available or model not found.
+   */
+  getContextFromLmsCli(modelId) {
+    try {
+      const homeDir = this.resolveHomeDir();
+      const lmsPaths = [
+        this.config.localLmsCliPath || "",
+        `${homeDir}/.cache/lm-studio/bin/lms`,
+        "/usr/local/bin/lms",
+        "/opt/homebrew/bin/lms"
+      ];
+      const lmsPath = lmsPaths.find((p) => p.length > 0 && existsSync(p));
+      if (!lmsPath) {
+        log.debug(`LMS CLI: not found in standard locations (checked: ${lmsPaths.join(", ")})`);
+        return null;
+      }
+      log.debug(`LMS CLI: running: ${lmsPath} ps --json`);
+      const existingPath = readEnvVar("PATH") || "";
+      const result = launchProcessSync(lmsPath, ["ps", "--json"], {
+        encoding: "utf-8",
+        timeout: 5e3,
+        shell: false,
+        // Don't use shell for JSON output - more reliable
+        env: mergeEnv({
+          PATH: `${this.config.localLmsBinDir || `${homeDir}/.cache/lm-studio/bin`}:/usr/local/bin:/opt/homebrew/bin:/usr/bin:/bin:${existingPath}`,
+          HOME: homeDir
+        })
+      });
+      if (result.error) {
+        log.debug(`LMS CLI: spawn error - ${result.error.message}`);
+        return null;
+      }
+      if (result.stderr && result.stderr.trim()) {
+        log.debug(`LMS CLI: stderr - ${result.stderr.slice(0, 200)}`);
+      }
+      const output = result.stdout || "";
+      if (!output.trim()) {
+        log.debug("LMS CLI: empty output - LM Studio may not be running or no models loaded");
+        return null;
+      }
+      let models;
+      try {
+        models = JSON.parse(output);
+      } catch (parseErr) {
+        log.debug(`LMS CLI: JSON parse error - ${parseErr}`);
+        return null;
+      }
+      if (!Array.isArray(models) || models.length === 0) {
+        log.debug("LMS CLI: no models loaded");
+        return null;
+      }
+      const model = models.find(
+        (m) => m.identifier === modelId || m.modelKey === modelId || m.identifier?.includes(modelId.replace(/@\d+bit$/, ""))
+      );
+      if (!model) {
+        log.debug(`LMS CLI: model "${modelId}" not found in loaded models: ${models.map((m) => m.identifier).join(", ")}`);
+        return null;
+      }
+      const contextWindow = model.contextLength || model.maxContextLength;
+      if (contextWindow) {
+        log.info(`LMS CLI detected context window: ${contextWindow} for ${modelId} (max: ${model.maxContextLength})`);
+        return contextWindow;
+      }
+      return null;
+    } catch (err) {
+      const errorMsg = err instanceof Error ? err.message : String(err);
+      log.debug(`LMS CLI: failed - ${errorMsg}`);
+      return null;
+    }
+  }
+  /**
+   * Get full model info from LMS CLI including context length and max context length.
+   * Returns null if LMS CLI is unavailable or model not found.
+   */
+  getLmsModelInfo(modelId) {
+    try {
+      const result = launchProcessSync("lms", ["ps", "--json"], {
+        encoding: "utf-8",
+        timeout: 5e3,
+        shell: false
+      });
+      if (result.error) {
+        return null;
+      }
+      const output = result.stdout || "";
+      if (!output.trim()) {
+        return null;
+      }
+      let models;
+      try {
+        models = JSON.parse(output);
+      } catch {
+        return null;
+      }
+      if (!Array.isArray(models) || models.length === 0) {
+        return null;
+      }
+      const model = models.find(
+        (m) => m.identifier === modelId || m.modelKey === modelId || m.identifier?.includes(modelId.replace(/@\d+bit$/, ""))
+      );
+      if (!model || !model.contextLength) {
+        return null;
+      }
+      return {
+        contextLength: model.contextLength,
+        maxContextLength: model.maxContextLength || model.contextLength,
+        identifier: model.identifier || modelId
+      };
+    } catch {
+      return null;
+    }
+  }
+  /**
+   * Get context window for the configured model, using cache if available.
+   * This method caches the result to avoid repeated LMS CLI calls.
+   * Order: ModelRegistry (persistent) -> memory cache -> LMS CLI -> settings.json
+   */
+  getCachedContextWindow(modelId) {
+    const now = Date.now();
+    if (this.modelRegistry) {
+      const caps = this.modelRegistry.getCapabilities(modelId);
+      if (caps.source === "lmstudio" && caps.contextWindow) {
+        log.debug(`ModelRegistry: using persisted LM Studio context: ${caps.contextWindow}`);
+        this.cachedLmsContext = caps.contextWindow;
+        this.lastLmsCheck = now;
+        return caps.contextWindow;
+      }
+    }
+    if (this.cachedLmsContext && now - this.lastLmsCheck < _LocalLlmClient.LMS_CACHE_INTERVAL_MS) {
+      log.debug(`LMS CLI: returning in-memory cached context: ${this.cachedLmsContext}`);
+      return this.cachedLmsContext;
+    }
+    const lmsInfo = this.getLmsModelInfo(modelId);
+    if (lmsInfo?.contextLength) {
+      this.cachedLmsContext = lmsInfo.contextLength;
+      this.lastLmsCheck = now;
+      const calculatedOutputTokens = Math.min(Math.floor(lmsInfo.contextLength / 8), 16384);
+      const outputTokens = Math.max(calculatedOutputTokens, 4096);
+      if (this.modelRegistry) {
+        this.modelRegistry.setCapabilities(modelId, {
+          maxPositionEmbeddings: lmsInfo.maxContextLength || lmsInfo.contextLength,
+          contextWindow: lmsInfo.contextLength,
+          supportsExtendedContext: (lmsInfo.maxContextLength || lmsInfo.contextLength) > 65536,
+          typicalOutputTokens: outputTokens,
+          source: "lmstudio"
+        });
+        log.info(`LMS CLI: Stored capabilities for ${modelId}: ${lmsInfo.contextLength} context, ${outputTokens} output tokens`);
+      }
+      return lmsInfo.contextLength;
+    }
+    const legacyContext = this.getContextFromLmsCli(modelId);
+    if (legacyContext) {
+      this.cachedLmsContext = legacyContext;
+      this.lastLmsCheck = now;
+      if (this.modelRegistry) {
+        const calculatedOutputTokens = Math.min(Math.floor(legacyContext / 8), 16384);
+        const outputTokens = Math.max(calculatedOutputTokens, 4096);
+        this.modelRegistry.setCapabilities(modelId, {
+          maxPositionEmbeddings: legacyContext,
+          contextWindow: legacyContext,
+          supportsExtendedContext: false,
+          typicalOutputTokens: outputTokens,
+          source: "lmstudio"
+        });
+      }
+      return legacyContext;
+    }
+    const settingsContext = this.getContextFromLmStudioSettings();
+    if (settingsContext) {
+      log.info(`LM Studio settings: using default context: ${settingsContext}`);
+      this.cachedLmsContext = settingsContext;
+      this.lastLmsCheck = now;
+      return settingsContext;
+    }
+    return null;
+  }
+  /**
+   * Clear the LMS context cache. Call this when the model changes.
+   */
+  clearContextCache() {
+    this.cachedLmsContext = null;
+    this.lastLmsCheck = 0;
+    log.debug("LMS CLI: context cache cleared");
+  }
+  remainingCooldownMs(now = Date.now()) {
+    return Math.max(0, this.cooldownUntilMs - now);
+  }
+  scheduleQueueDrain() {
+    if (this.queueDrainScheduled) return;
+    this.queueDrainScheduled = true;
+    queueMicrotask(() => {
+      this.queueDrainScheduled = false;
+      this.startAvailableQueuedRequests();
+    });
+  }
+  hasQueuedRequests() {
+    return this.requestQueues["recall-critical"].length > 0 || this.requestQueues.background.length > 0;
+  }
+  dequeueQueuedRequest(priority) {
+    const next = this.requestQueues[priority].shift();
+    return next ?? null;
+  }
+  failOpenQueuedRequestsForCooldown() {
+    let dropped = 0;
+    for (const priority of ["recall-critical", "background"]) {
+      while (this.requestQueues[priority].length > 0) {
+        const queued = this.requestQueues[priority].shift();
+        queued?.resolve(null);
+        dropped += 1;
+      }
+    }
+    return dropped;
+  }
+  startAvailableQueuedRequests() {
+    if (!this.queueProcessing.has("recall-critical")) {
+      const nextCritical = this.dequeueQueuedRequest("recall-critical");
+      if (nextCritical) {
+        this.queueProcessing.add("recall-critical");
+        void this.runQueuedRequest(nextCritical);
+      }
+    }
+    if (!this.queueProcessing.has("background")) {
+      const nextBackground = this.dequeueQueuedRequest("background");
+      if (nextBackground) {
+        this.queueProcessing.add("background");
+        void this.runQueuedRequest(nextBackground);
+      }
+    }
+  }
+  async runQueuedRequest(next) {
+    try {
+      const remainingCooldownMs = this.remainingCooldownMs();
+      if (remainingCooldownMs > 0) {
+        const additionalDropped = this.failOpenQueuedRequestsForCooldown();
+        log.warn(
+          `local LLM: cooldown active (${remainingCooldownMs}ms remaining), dropping ${additionalDropped + 1} queued request(s) fail-open`
+        );
+        next.resolve(null);
+        return;
+      }
+      let result = null;
+      try {
+        result = await this.runChatCompletionRequest(next.messages, next.options, {
+          priority: next.priority,
+          enqueuedAtMs: next.enqueuedAtMs
+        });
+      } catch (err) {
+        log.warn(`local LLM queue drain failed open: ${err instanceof Error ? err.message : String(err)}`);
+      }
+      next.resolve(result);
+    } finally {
+      this.queueProcessing.delete(next.priority);
+      if (this.hasQueuedRequests()) {
+        this.scheduleQueueDrain();
+      }
+    }
+  }
+  async runChatCompletionRequest(messages, options, queueMeta) {
+    log.debug(
+      `local LLM chatCompletion: localLlmEnabled=${this.config.localLlmEnabled}, model=${this.config.localLlmModel}`
+    );
+    const operation = options.operation ?? "unspecified";
+    const startedAtMs = Date.now();
+    if (queueMeta) {
+      log.debug(
+        `local LLM queue start: priority=${queueMeta.priority} waitMs=${startedAtMs - queueMeta.enqueuedAtMs} op=${operation}`
+      );
+    }
+    try {
+      const isAvailable = await this.checkAvailability();
+      if (!isAvailable) {
+        log.debug(
+          `local LLM: checkAvailability returned false for ${this.config.localLlmUrl}`
+        );
+        return null;
+      }
+      const promptChars = messages.reduce((sum, m) => sum + (m.content?.length ?? 0), 0);
+      const requestBody = {
+        model: this.config.localLlmModel,
+        messages,
+        temperature: options.temperature ?? 0.7,
+        // Use max_tokens consistent with cloud models
+        max_tokens: options.maxTokens ?? 4096
+      };
+      if (options.responseFormat?.type === "json_schema") {
+        requestBody.response_format = options.responseFormat;
+      }
+      if (this._disableThinking) {
+        requestBody.chat_template_kwargs = { enable_thinking: false };
+      }
+      const baseUrl = trimTrailingSlashes(
+        this.config.localLlmUrl.replace("localhost", "127.0.0.1")
+      );
+      const chatUrl = baseUrl.endsWith("/v1") ? `${baseUrl}/chat/completions` : `${baseUrl}/v1/chat/completions`;
+      const requestBodyJson = JSON.stringify(requestBody);
+      log.debug(
+        `local LLM: sending request to ${chatUrl} with model ${this.config.localLlmModel}`
+      );
+      log.debug(`local LLM: request body length=${requestBodyJson.length}`);
+      if (this.config.debug) {
+        try {
+          const { writeFileSync } = await import("fs");
+          writeFileSync("/tmp/engram-last-request.json", requestBodyJson);
+        } catch {
+        }
+      }
+      const effectiveTimeoutMs = typeof options.timeoutMs === "number" ? Math.min(this.config.localLlmTimeoutMs, options.timeoutMs) : this.config.localLlmTimeoutMs;
+      const maxAttempts = 1 + Math.max(0, this.config.localLlmRetry5xxCount);
+      let response = null;
+      let lastAbortError = null;
+      for (let attempt = 1; attempt <= maxAttempts; attempt += 1) {
+        const attemptAbort = new AbortController();
+        const attemptTimeout = setTimeout(() => attemptAbort.abort(), effectiveTimeoutMs);
+        try {
+          response = await fetch(chatUrl, {
+            method: "POST",
+            headers: this.buildRequestHeaders({
+              "Content-Type": "application/json"
+            }),
+            body: JSON.stringify(requestBody),
+            signal: attemptAbort.signal
+          });
+        } catch (err) {
+          if (!this.isAbortError(err)) throw err;
+          lastAbortError = err instanceof Error ? err : new Error(String(err));
+          if (attempt < maxAttempts) {
+            const backoffMs2 = this.config.localLlmRetryBackoffMs * attempt;
+            log.warn(
+              `local LLM request aborted: op=${operation} attempt=${attempt}/${maxAttempts} timeoutMs=${effectiveTimeoutMs} model=${this.config.localLlmModel}; retrying after ${backoffMs2}ms`
+            );
+            await new Promise((resolve) => setTimeout(resolve, backoffMs2));
+            continue;
+          }
+          break;
+        } finally {
+          clearTimeout(attemptTimeout);
+        }
+        if (response.ok) break;
+        if (response.status >= 500 && attempt < maxAttempts) {
+          try {
+            const errorText = await response.clone().text();
+            const nonRecoverableReason = this.extractNonRecoverableBackendReasonFromErrorText(errorText);
+            if (nonRecoverableReason) {
+              this.markBackendUnavailable(
+                nonRecoverableReason,
+                this.config.localLlm400CooldownMs
+              );
+              this.consecutive400s = 0;
+              return null;
+            }
+          } catch (e) {
+            log.debug(`local LLM failed to inspect retryable error body: ${e}`);
+          }
+        }
+        if (response.status < 500 || attempt >= maxAttempts) break;
+        const backoffMs = this.config.localLlmRetryBackoffMs * attempt;
+        log.warn(
+          `local LLM request got ${response.status}; retrying (attempt ${attempt + 1}/${maxAttempts}) after ${backoffMs}ms`
+        );
+        await new Promise((resolve) => setTimeout(resolve, backoffMs));
+      }
+      log.debug(
+        `local LLM: received response, status=${response?.status}, ok=${response?.ok}`
+      );
+      if (!response) {
+        if (lastAbortError) {
+          log.warn(
+            `local LLM request aborted after ${maxAttempts} attempt(s): op=${operation} timeoutMs=${effectiveTimeoutMs} model=${this.config.localLlmModel} promptChars=${promptChars} durationMs=${Date.now() - startedAtMs}`
+          );
+        } else {
+          log.warn(
+            `local LLM request failed: no response object (op=${operation} model=${this.config.localLlmModel} durationMs=${Date.now() - startedAtMs})`
+          );
+        }
+        return null;
+      }
+      if (!response.ok) {
+        let reason = "";
+        let errorText = "";
+        try {
+          errorText = await response.text();
+          try {
+            const parsed = JSON.parse(errorText);
+            reason = parsed?.error?.message ? ` \u2014 ${parsed.error.message}` : "";
+          } catch {
+            log.debug(`local LLM error body: ${errorText.slice(0, 500)}`);
+          }
+        } catch (e) {
+          log.debug(`local LLM failed to read error body: ${e}`);
+        }
+        log.warn(
+          `local LLM request failed: ${response.status} ${response.statusText}${reason} (op=${operation}, model=${this.config.localLlmModel}, url=${chatUrl}, promptChars=${promptChars}, maxTokens=${requestBody.max_tokens})`
+        );
+        const nonRecoverableReason = this.extractNonRecoverableBackendReason(reason) ?? this.extractNonRecoverableBackendReasonFromErrorText(errorText);
+        if (nonRecoverableReason) {
+          this.markBackendUnavailable(
+            nonRecoverableReason,
+            this.config.localLlm400CooldownMs
+          );
+          this.consecutive400s = 0;
+          return null;
+        }
+        if (response.status === 400) {
+          this.consecutive400s += 1;
+          if (this.consecutive400s >= this.config.localLlm400TripThreshold) {
+            this.cooldownUntilMs = Date.now() + this.config.localLlm400CooldownMs;
+            log.warn(
+              `local LLM: entering cooldown for ${this.config.localLlm400CooldownMs}ms after ${this.consecutive400s} consecutive 400 responses`
+            );
+            this.consecutive400s = 0;
+          }
+        } else {
+          this.consecutive400s = 0;
+        }
+        return null;
+      }
+      this.consecutive400s = 0;
+      const data = await response.json();
+      log.debug(
+        `local LLM response: choices=${data.choices?.length}, usage=${JSON.stringify(data.usage)}`
+      );
+      const msg = data.choices?.[0]?.message;
+      const content = msg?.content || msg?.reasoning_content || "";
+      if (!content) {
+        log.warn(`local LLM returned empty content. choices=${JSON.stringify(data.choices)?.slice(0, 200)}`);
+        return null;
+      }
+      const usage = data.usage ? {
+        promptTokens: data.usage.prompt_tokens ?? 0,
+        completionTokens: data.usage.completion_tokens ?? 0,
+        totalTokens: data.usage.total_tokens ?? 0
+      } : this.estimateTokens(messages, content);
+      const durationMs = Date.now() - startedAtMs;
+      if (this.config.slowLogEnabled && durationMs >= this.config.slowLogThresholdMs) {
+        const promptChars2 = messages.reduce((sum, m) => sum + (m.content?.length ?? 0), 0);
+        const op = options.operation ? ` op=${options.operation}` : "";
+        log.warn(
+          `SLOW local LLM:${op} durationMs=${durationMs} model=${this.config.localLlmModel} url=${chatUrl} promptChars=${promptChars2} outputTokens=${usage.completionTokens} totalTokens=${usage.totalTokens}`
+        );
+      }
+      log.debug("local LLM: request succeeded, tokens:", usage.totalTokens);
+      return { content, usage };
+    } catch (err) {
+      const errMsg = err instanceof Error ? err.message : String(err);
+      const durationMs = Date.now() - startedAtMs;
+      if (this.isAbortError(err)) {
+        log.warn(
+          `local LLM request aborted: op=${operation} timeoutMs=${options.timeoutMs ?? this.config.localLlmTimeoutMs} model=${this.config.localLlmModel} durationMs=${durationMs} error=${errMsg}`
+        );
+        return null;
+      }
+      log.warn(`local LLM request error: op=${operation} error=${errMsg}`);
+      this.isAvailable = false;
+      const nonRecoverableReason = this.extractNonRecoverableBackendReason(errMsg);
+      if (nonRecoverableReason) {
+        this.markBackendUnavailable(
+          nonRecoverableReason,
+          this.config.localLlm400CooldownMs
+        );
+      }
+      return null;
+    } finally {
+      if (queueMeta) {
+        const finishedAtMs = Date.now();
+        const waitMs = startedAtMs - queueMeta.enqueuedAtMs;
+        log.debug(
+          `local LLM queue finish: priority=${queueMeta.priority} waitMs=${waitMs} runMs=${finishedAtMs - startedAtMs} totalMs=${finishedAtMs - queueMeta.enqueuedAtMs} op=${operation}`
+        );
+      }
+    }
+  }
+  /**
+   * Query the local LLM server for loaded model information.
+   * Returns null if unavailable or if the model is not found.
+   */
+  async getLoadedModelInfo() {
+    const baseUrl = trimTrailingSlashes(
+      this.config.localLlmUrl.replace("localhost", "127.0.0.1")
+    );
+    const modelsUrl = baseUrl.endsWith("/v1") ? `${baseUrl}/models` : `${baseUrl}/v1/models`;
+    log.debug(`Fetching model info from ${modelsUrl}`);
+    try {
+      const result = await this.fetchWithTimeout(modelsUrl, 3e3);
+      if (!result.ok) {
+        if (result.status === 401 || result.status === 403) {
+          log.warn(
+            `Local LLM: unauthorized while fetching models from ${modelsUrl}; verify localLlmApiKey and localLlmAuthHeader settings`
+          );
+        }
+        log.warn(`Local LLM: Failed to fetch models from ${modelsUrl} - server returned error`);
+        return null;
+      }
+      if (!result.data) {
+        log.warn(`Local LLM: No data returned from ${modelsUrl}`);
+        return null;
+      }
+      const data = result.data;
+      if (!Array.isArray(data.data) || data.data.length === 0) {
+        log.warn("Local LLM returned no models");
+        return null;
+      }
+      const modelIds = data.data.map((m) => m.id).filter(Boolean);
+      log.debug(
+        `Local LLM: Found ${modelIds.length} model(s). First 10: ${modelIds.slice(0, 10).join(", ")}`
+      );
+      const configuredModel = this.config.localLlmModel;
+      let model = data.data.find((m) => m.id === configuredModel);
+      if (!model) {
+        model = data.data.find(
+          (m) => configuredModel.includes(m.id || "") || (m.id || "").includes(configuredModel.replace(/@\d+bit$/, ""))
+        );
+      }
+      if (!model) {
+        model = data.data[0];
+        const availablePreview = data.data.map((m) => m.id).filter(Boolean).slice(0, 10).join(", ");
+        log.warn(
+          `Configured model "${configuredModel}" not found in local LLM. Using "${model.id}" instead. Available (first 10): ${availablePreview}`
+        );
+      }
+      let contextWindow = model.max_context_length || model.max_tokens;
+      if (!contextWindow) {
+        log.info("Local LLM: API did not report context window, trying LMS CLI...");
+        const lmsContext = this.getCachedContextWindow(model.id || "");
+        if (lmsContext) {
+          contextWindow = lmsContext;
+        }
+      }
+      this.cachedModelInfo = {
+        id: model.id || "unknown",
+        contextWindow,
+        maxTokens: model.max_tokens
+      };
+      log.info(
+        `Local LLM model detected: ${this.cachedModelInfo.id}, context window: ${contextWindow?.toLocaleString() || "unknown (may use default)"}`
+      );
+      return this.cachedModelInfo;
+    } catch (err) {
+      log.warn(`Failed to fetch model info: ${err}`);
+      return null;
+    }
+  }
+  /**
+   * Check if the configured model is available and get its actual context window.
+   * Warns if there's a mismatch between expected and actual context.
+   */
+  async validateModelConfig(expectedContextWindow) {
+    const warnings = [];
+    const modelInfo = await this.getLoadedModelInfo();
+    if (!modelInfo) {
+      return { available: false, warnings: ["Could not query local LLM for model info"] };
+    }
+    if (expectedContextWindow && modelInfo.contextWindow) {
+      if (modelInfo.contextWindow < expectedContextWindow) {
+        warnings.push(
+          `Context window mismatch: Model ${modelInfo.id} supports ${modelInfo.contextWindow.toLocaleString()} tokens, but engram is configured for ${expectedContextWindow.toLocaleString()}. Set localLlmMaxContext: ${modelInfo.contextWindow} in config to avoid errors.`
+        );
+      }
+    }
+    if (!modelInfo.contextWindow) {
+      warnings.push(
+        `Local LLM server did not report context window for ${modelInfo.id}. If you get "context length exceeded" errors, set localLlmMaxContext in config.`
+      );
+    }
+    return {
+      available: true,
+      actualContextWindow: modelInfo.contextWindow,
+      warnings
+    };
+  }
+  /**
+   * Make a chat completion request to local LLM
+   */
+  async chatCompletion(messages, options = {}) {
+    if (!this.config.localLlmEnabled) {
+      log.debug("local LLM: disabled, returning null");
+      return null;
+    }
+    const remainingMs = this.remainingCooldownMs();
+    if (remainingMs > 0) {
+      log.debug(`local LLM: cooldown active (${remainingMs}ms remaining), skipping request`);
+      return null;
+    }
+    if (options.priority) {
+      const priority = options.priority;
+      return await new Promise((resolve) => {
+        this.requestQueues[priority].push({
+          messages,
+          options,
+          priority,
+          enqueuedAtMs: Date.now(),
+          resolve
+        });
+        this.scheduleQueueDrain();
+      });
+    }
+    return await this.runChatCompletionRequest(messages, options);
+  }
+  /**
+   * Estimate tokens when local LLM doesn't return usage stats
+   * Rough estimate: 1 token ≈ 4 characters
+   */
+  estimateTokens(messages, response) {
+    const promptChars = messages.reduce((sum, m) => sum + m.content.length, 0);
+    const promptTokens = Math.ceil(promptChars / 4);
+    const completionTokens = Math.ceil(response.length / 4);
+    return {
+      promptTokens,
+      completionTokens,
+      totalTokens: promptTokens + completionTokens
+    };
+  }
+  /**
+   * Try local LLM first, fallback to cloud provider if configured
+   */
+  async withFallback(localOperation, fallbackOperation, operationName) {
+    if (this.config.localLlmEnabled) {
+      const localResult = await localOperation();
+      if (localResult !== null) {
+        log.debug(`${operationName}: used local LLM`);
+        return localResult;
+      }
+      if (this.config.localLlmFallback) {
+        log.info(`${operationName}: local LLM unavailable, falling back to cloud`);
+      } else {
+        throw new Error(`${operationName}: local LLM unavailable and fallback disabled`);
+      }
+    }
+    return fallbackOperation();
+  }
+};
+
+export {
+  LocalLlmClient
+};
+//# sourceMappingURL=chunk-MDDAA2AO.js.map