@joshuaswarren/openclaw-engram 9.1.21 → 9.1.23

package/README.md

@@ -71,12 +71,12 @@ After installation, add Engram to your `openclaw.json`:
           // Option 1: Use OpenAI for extraction:
           "openaiApiKey": "${OPENAI_API_KEY}"
 
-          // Option 2: Use a local LLM (no API key needed):
+          // Option 2: Use Engram's local LLM path (plugin mode only; no API key needed):
           // "localLlmEnabled": true,
           // "localLlmUrl": "http://localhost:1234/v1",
           // "localLlmModel": "qwen2.5-32b-instruct"
 
-          // Option 3: Use the gateway model chain (multi-provider fallback):
+          // Option 3: Use the gateway model chain (primary path in gateway mode):
           // "modelSource": "gateway",
           // "gatewayAgentId": "engram-llm",
           // "fastGatewayAgentId": "engram-llm-fast"
@@ -87,7 +87,7 @@ After installation, add Engram to your `openclaw.json`:
 }
 ```
 
-> **Gateway model source:** When `modelSource` is `"gateway"`, Engram routes all LLM calls (extraction, consolidation, reranking) through an OpenClaw agent persona's model chain instead of its own config. Define agent personas in `openclaw.json → agents.list[]` with a `primary` model and `fallbacks[]` array — Engram tries each in order until one succeeds. This lets you build multi-provider fallback chains like Fireworks → local LLM → cloud OpenAI. See the [Gateway Model Source](docs/config-reference.md#gateway-model-source) guide for full setup.
+> **Gateway model source:** When `modelSource` is `"gateway"`, Engram routes all LLM calls (extraction, consolidation, reranking) through an OpenClaw agent persona's model chain instead of its own config. Extraction starts on the `gatewayAgentId` chain directly in this mode; `localLlm*` settings do not control primary extraction order. Define agent personas in `openclaw.json → agents.list[]` with a `primary` model and `fallbacks[]` array — Engram tries each in order until one succeeds. This lets you build multi-provider fallback chains like Fireworks → local LLM → cloud OpenAI. See the [Gateway Model Source](docs/config-reference.md#gateway-model-source) guide for full setup.
 
 Restart the gateway:
 
@@ -484,10 +484,10 @@ All settings live in `openclaw.json` under `plugins.entries.openclaw-engram.conf
 | Setting | Default | Description |
 |---------|---------|-------------|
 | `openaiApiKey` | `(env)` | OpenAI API key (optional when using a local LLM) |
-| `localLlmEnabled` | `false` | Use a local LLM instead of OpenAI for extraction |
+| `localLlmEnabled` | `false` | Enable Engram's local LLM path when `modelSource` is `plugin` |
 | `localLlmUrl` | unset | Local LLM endpoint (e.g., `http://localhost:1234/v1`) |
 | `localLlmModel` | unset | Local model name (e.g., `qwen2.5-32b-instruct`) |
-| `model` | `gpt-5.2` | OpenAI model for extraction (when not using local LLM) |
+| `model` | `gpt-5.2` | OpenAI model for extraction when `modelSource` is `plugin` and local LLM is disabled |
 | `searchBackend` | `"qmd"` | Search engine: `qmd`, `orama`, `lancedb`, `meilisearch`, `remote`, `noop` |
 | `captureMode` | `implicit` | Memory write policy: `implicit`, `explicit`, `hybrid` |
 | `recallBudgetChars` | `maxMemoryTokens * 4` | Recall budget (default ~8K chars; set 64K+ for large-context models) |

package/dist/access-cli.js

@@ -3,12 +3,12 @@ import {
   EngramAccessService,
   Orchestrator,
   parseConfig
-} from "./chunk-3ILF2OUT.js";
-import "./chunk-EM2BU45W.js";
+} from "./chunk-S6X3YWYF.js";
+import "./chunk-PIQWOQEX.js";
 import "./chunk-IMMYYNXG.js";
 import "./chunk-Y6DMH5LN.js";
 import "./chunk-6KX4XLQJ.js";
-import "./chunk-FP3MUEPD.js";
+import "./chunk-4722S3A6.js";
 import "./chunk-MKM2BCQH.js";
 import "./chunk-DEIBZP3O.js";
 import "./chunk-SSIIJJKA.js";

package/dist/{calibration-UY5SQU5K.js → calibration-WH4XEHIA.js}

@@ -1,7 +1,7 @@
 // openclaw-engram: Local-first memory plugin
 import {
   FallbackLlmClient
-} from "./chunk-FP3MUEPD.js";
+} from "./chunk-4722S3A6.js";
 import {
   listJsonFiles
 } from "./chunk-DEIBZP3O.js";
@@ -233,4 +233,4 @@ export {
   runCalibrationIfEnabled,
   synthesizeCalibrationRules
 };
-//# sourceMappingURL=calibration-UY5SQU5K.js.map
+//# sourceMappingURL=calibration-WH4XEHIA.js.map

package/dist/{causal-consolidation-5RFIDR74.js → causal-consolidation-62M3TOOX.js}

@@ -1,7 +1,7 @@
 // openclaw-engram: Local-first memory plugin
 import {
   FallbackLlmClient
-} from "./chunk-FP3MUEPD.js";
+} from "./chunk-4722S3A6.js";
 import {
   readChainIndex,
   resolveChainsDir
@@ -203,4 +203,4 @@ export {
   deriveCausalPromotionCandidates,
   synthesizeCausalPreferencesViaLlm
 };
-//# sourceMappingURL=causal-consolidation-5RFIDR74.js.map
+//# sourceMappingURL=causal-consolidation-62M3TOOX.js.map

package/dist/{chunk-FP3MUEPD.js → chunk-4722S3A6.js}

@@ -260,7 +260,7 @@ var FallbackLlmClient = class {
           const result = await this.tryModel(model, messages, options);
           if (result) {
             if (isFallback) {
-              log.info(`fallback LLM: succeeded using ${model.modelString} (fallback ${i})`);
+              log.debug(`fallback LLM: succeeded using ${model.modelString} (fallback ${i})`);
             }
             return {
               content: result.content,
@@ -531,4 +531,4 @@ export {
   buildChatCompletionTokenLimit,
   FallbackLlmClient
 };
-//# sourceMappingURL=chunk-FP3MUEPD.js.map
+//# sourceMappingURL=chunk-4722S3A6.js.map

package/dist/chunk-4722S3A6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/json-extract.ts","../src/openai-chat-compat.ts","../src/resolve-provider-secret.ts","../src/fallback-llm.ts"],"sourcesContent":["/**\n * Utilities for extracting JSON payloads from LLM outputs.\n *\n * We see common failure modes:\n * - \"Here's an example: {..}\\nHere's the real answer: {..}\" (multiple JSON blocks)\n * - fenced ```json blocks\n * - leading/trailing prose around JSON\n *\n * These helpers attempt multiple candidates and let callers validate with schemas.\n */\n\nexport function stripCodeFences(text: string): string {\n  return text.replace(/```(?:json)?\\s*([\\s\\S]*?)```/gi, (_m, inner) => String(inner).trim());\n}\n\nexport function extractJsonCandidates(text: string): string[] {\n  const trimmed = text.trim();\n  const cleaned = stripCodeFences(trimmed);\n  const candidates: string[] = [];\n\n  if (cleaned.length > 0) candidates.push(cleaned);\n  candidates.push(...scanBalancedJsonBlocks(cleaned));\n\n  // Legacy regex fallback (single object)\n  const objMatch = cleaned.match(/\\{[\\s\\S]*\\}/);\n  if (objMatch) candidates.push(objMatch[0]);\n\n  const seen = new Set<string>();\n  return candidates\n    .map((c) => c.trim())\n    .filter((c) => c.length > 0)\n    .filter((c) => {\n      if (seen.has(c)) return false;\n      seen.add(c);\n      return true;\n    });\n}\n\nfunction scanBalancedJsonBlocks(text: string): string[] {\n  const out: string[] = [];\n  const opens = new Set([\"{\", \"[\"]);\n  const closes: Record<string, string> = { \"{\": \"}\", \"[\": \"]\" };\n\n  for (let i = 0; i < text.length; i++) {\n    const start = text[i];\n    if (!opens.has(start)) continue;\n\n    const expectedClose = closes[start];\n    let depth = 0;\n    let inString = false;\n    let escape = false;\n\n    for (let j = i; j < text.length; j++) {\n      const ch = text[j];\n\n      if (inString) {\n        if (escape) {\n          escape = false;\n        } else if (ch === \"\\\\\") {\n          escape = true;\n        } else if (ch === \"\\\"\") {\n          inString = false;\n        }\n        continue;\n      }\n\n      if (ch === \"\\\"\") {\n        inString = true;\n        continue;\n      }\n\n      if (ch === start) depth++;\n      if (ch === expectedClose) depth--;\n\n      if (depth === 0) {\n        out.push(text.slice(i, j + 1).trim());\n        i = j;\n        break;\n      }\n    }\n  }\n\n  return out;\n}\n\n","function normalizedModel(model: string): string {\n  return model.trim().toLowerCase();\n}\n\nfunction matchesModelFamily(normalized: string, familyPattern: RegExp): boolean {\n  return familyPattern.test(normalized);\n}\n\nexport function shouldAssumeOpenAiChatCompletions(baseUrl?: string): boolean {\n  if (!baseUrl) return true;\n  try {\n    return new URL(baseUrl).hostname.toLowerCase() === \"api.openai.com\";\n  } catch {\n    return false;\n  }\n}\n\nexport function usesMaxCompletionTokens(model: string, options?: { assumeOpenAI?: boolean }): boolean {\n  const normalized = normalizedModel(model);\n  if (options?.assumeOpenAI !== true) return false;\n  if (matchesModelFamily(normalized, /^gpt-5(?:$|[-.])/)) return true;\n  if (matchesModelFamily(normalized, /^gpt-4o(?:$|[-.])/)) return true;\n  if (matchesModelFamily(normalized, /^gpt-4\\.1(?:$|[-.])/)) return true;\n  if (matchesModelFamily(normalized, /^o1(?:$|[-.])/)) return true;\n  if (matchesModelFamily(normalized, /^o3(?:$|[-.])/)) return true;\n  return matchesModelFamily(normalized, /^o4-mini(?:$|[-.])/);\n}\n\nexport function buildChatCompletionTokenLimit(\n  model: string,\n  maxTokens: number,\n  options?: { assumeOpenAI?: boolean },\n): { max_tokens: number } | { max_completion_tokens: number } {\n  const safeMaxTokens = Math.max(0, Math.floor(maxTokens));\n  if (usesMaxCompletionTokens(model, options)) {\n    return { max_completion_tokens: safeMaxTokens };\n  }\n  return { max_tokens: safeMaxTokens };\n}\n","import { log } from \"./logger.js\";\nimport path from \"node:path\";\nimport os from \"node:os\";\n\n/**\n * Resolve a provider API key using OpenClaw's own auth resolution system.\n *\n * This module delegates to the gateway's `resolveApiKeyForProvider()` function,\n * which handles all secret reference formats (SecretRef objects, auth profiles,\n * \"secretref-managed\" markers, environment variables, etc.) using the same\n * codepath the gateway uses for its own agent sessions.\n *\n * For plain-text API keys, a fast path returns them directly without\n * involving the gateway auth system.\n *\n * Results are cached per provider for the gateway process lifetime.\n */\n\ntype ResolveApiKeyFn = (params: {\n  provider: string;\n  cfg?: unknown;\n  agentDir?: string;\n}) => Promise<{ apiKey?: string; source?: string; mode?: string } | null>;\n\nlet _resolveApiKeyForProvider: ResolveApiKeyFn | null = null;\nlet _resolverLoaded = false;\nlet _resolverNextRetryAt = 0;\nconst RESOLVER_RETRY_BACKOFF_MS = 60_000; // 1 minute between retries after first failure\nconst resolvedCache = new Map<string, string | undefined>();\n\n/**\n * Lazily load the gateway's resolveApiKeyForProvider function.\n * Returns null if not available (e.g., running outside the gateway process).\n */\nasync function getGatewayResolver(): Promise<ResolveApiKeyFn | null> {\n  if (_resolverLoaded) {\n    return _resolveApiKeyForProvider;\n  }\n  // Backoff: don't re-scan filesystem on every call when module wasn't found.\n  // After a failure, wait RESOLVER_RETRY_BACKOFF_MS before trying again.\n  if (_resolverNextRetryAt > 0 && Date.now() < _resolverNextRetryAt) {\n    return null;\n  }\n\n  try {\n    // The gateway bundles this in a runtime chunk — import it dynamically.\n    // This import path is stable across gateway versions since it's a named runtime export.\n    const candidates = [\n      // Try glob-matching the runtime module name (hash varies per build)\n      ...await findRuntimeModules(),\n    ];\n\n    const { pathToFileURL } = await import(\"node:url\");\n    for (const candidate of candidates) {\n      try {\n        // Convert native path to file:// URL for cross-platform ESM import compatibility\n        const importUrl = pathToFileURL(candidate).href;\n        const mod = await import(importUrl);\n        if (typeof mod.resolveApiKeyForProvider === \"function\") {\n          _resolveApiKeyForProvider = mod.resolveApiKeyForProvider;\n          _resolverLoaded = true;\n          log.debug(\"loaded gateway resolveApiKeyForProvider from runtime module\");\n          return _resolveApiKeyForProvider;\n        }\n      } catch {\n        // Try next candidate\n      }\n    }\n  } catch {\n    // Silent\n  }\n\n  // Backoff before retrying — avoid repeated fs scanning.\n  // Retries after RESOLVER_RETRY_BACKOFF_MS so the resolver can\n  // recover if the gateway restarts or the module becomes available.\n  _resolverNextRetryAt = Date.now() + RESOLVER_RETRY_BACKOFF_MS;\n  log.debug(`gateway resolveApiKeyForProvider not available — will retry after ${RESOLVER_RETRY_BACKOFF_MS / 1000}s`);\n  return null;\n}\n\n/**\n * Find the gateway's model-auth runtime module by scanning the dist directory.\n * Uses require.resolve to find the openclaw package regardless of install method.\n */\nasync function findRuntimeModules(): Promise<string[]> {\n  const { readdirSync } = await import(\"node:fs\");\n  const { createRequire } = await import(\"node:module\");\n  const candidates: string[] = [];\n\n  // Discover the openclaw dist directory from the installed package,\n  // regardless of how it was installed (Homebrew, npm global, local, etc.)\n  const distDirs: string[] = [];\n\n  try {\n    // require.resolve finds the package from the current process context\n    const req = createRequire(import.meta.url);\n    const openclawMain = req.resolve(\"openclaw\");\n    const openclawDist = path.join(path.dirname(openclawMain), \"..\", \"dist\");\n    if (openclawDist) distDirs.push(path.resolve(openclawDist));\n  } catch {\n    // openclaw not resolvable from plugin context — try alternate paths\n  }\n\n  // Fallback: infer from the running process (gateway runs from its own dist/)\n  // Use fs.realpathSync to resolve symlinks (e.g., /usr/local/bin/openclaw → actual path)\n  try {\n    const { realpathSync } = await import(\"node:fs\");\n    const mainScript = process.argv[1];\n    if (mainScript) {\n      const realScript = realpathSync(mainScript);\n      if (realScript.includes(\"openclaw\")) {\n        const distDir = path.dirname(realScript);\n        if (!distDirs.includes(distDir)) distDirs.push(distDir);\n      }\n    }\n  } catch {\n    // Silent\n  }\n\n  for (const dir of distDirs) {\n    try {\n      const files = readdirSync(dir);\n      for (const f of files) {\n        if (f.startsWith(\"runtime-model-auth.runtime-\") && f.endsWith(\".js\")) {\n          candidates.push(path.join(dir, f));\n        }\n      }\n    } catch {\n      // Directory doesn't exist — skip\n    }\n  }\n\n  return candidates;\n}\n\n/**\n * Resolve a provider API key from various OpenClaw formats.\n *\n * Resolution order:\n * 1. Plain-text string → returned immediately\n * 2. Gateway's resolveApiKeyForProvider → handles all secret ref formats\n * 3. Environment variable fallback (PROVIDER_NAME_API_KEY)\n * 4. undefined → provider is skipped in the fallback chain\n */\nexport async function resolveProviderApiKey(\n  providerId: string,\n  apiKeyValue: unknown,\n  gatewayConfig?: unknown,\n  agentDir?: string,\n): Promise<string | undefined> {\n  // Check cache first\n  const cacheKey = `provider:${providerId}`;\n  if (resolvedCache.has(cacheKey)) {\n    return resolvedCache.get(cacheKey);\n  }\n\n  let resolved: string | undefined;\n\n  // Fast path: plain-text string that looks like an actual API key\n  if (typeof apiKeyValue === \"string\" && apiKeyValue.trim().length > 0) {\n    // Skip known non-API-key markers used by the gateway for auth modes\n    // that don't use bearer tokens (OAuth, local endpoints, GCP credentials)\n    if (\n      apiKeyValue === \"secretref-managed\" ||\n      apiKeyValue.endsWith(\"-oauth\") ||\n      apiKeyValue.endsWith(\"-local\") ||\n      apiKeyValue === \"lm-studio\" ||\n      apiKeyValue.startsWith(\"gcp-\")\n    ) {\n      // Fall through to gateway resolver / env var fallback\n    } else {\n      resolved = apiKeyValue;\n      resolvedCache.set(cacheKey, resolved);\n      return resolved;\n    }\n  }\n\n  // The API key is either a SecretRef object, \"secretref-managed\", or empty.\n  // Try the gateway's own auth resolution system first.\n  const resolver = await getGatewayResolver();\n  if (resolver) {\n    try {\n      const resolvedAgentDir = agentDir ?? path.join(os.homedir(), \".openclaw\", \"agents\", \"main\", \"agent\");\n      const auth = await resolver({ provider: providerId, cfg: gatewayConfig, agentDir: resolvedAgentDir });\n      if (auth?.apiKey) {\n        resolved = auth.apiKey;\n        log.debug(`resolved API key for provider \"${providerId}\" via gateway auth (source: ${auth.source ?? \"unknown\"}, mode: ${auth.mode ?? \"unknown\"})`);\n        resolvedCache.set(cacheKey, resolved);\n        return resolved;\n      }\n    } catch (err) {\n      log.debug(\n        `gateway auth resolution failed for provider \"${providerId}\": ${err instanceof Error ? err.message : String(err)}`,\n      );\n    }\n  }\n\n  // Environment variable fallback\n  resolved = resolveFromEnv(providerId);\n  if (resolved) {\n    log.debug(`resolved API key for provider \"${providerId}\" from environment variable`);\n  } else {\n    log.debug(`could not resolve API key for provider \"${providerId}\" — skipping`);\n  }\n\n  // Only cache successful resolutions — failures are retried on next call\n  // so providers can recover after transient issues (e.g., 1Password agent restart)\n  if (resolved) {\n    resolvedCache.set(cacheKey, resolved);\n  }\n  return resolved;\n}\n\n/**\n * Try to resolve an API key from environment variables.\n */\nfunction resolveFromEnv(providerId: string): string | undefined {\n  const normalized = providerId.toUpperCase().replace(/[^A-Z0-9]/g, \"_\");\n  const candidates = [\n    `${normalized}_API_KEY`,\n    `${normalized}_TOKEN`,\n  ];\n  for (const envVar of candidates) {\n    const value = process.env[envVar];\n    if (value && value.trim().length > 0) {\n      return value.trim();\n    }\n  }\n  return undefined;\n}\n\n/**\n * Clear the resolution cache (useful for testing or key rotation).\n */\nexport function clearSecretCache(): void {\n  resolvedCache.clear();\n  _resolveApiKeyForProvider = null;\n  _resolverLoaded = false;\n  _resolverNextRetryAt = 0;\n}\n","import { log } from \"./logger.js\";\nimport type { GatewayConfig, ModelProviderConfig, AgentPersona } from \"./types.js\";\nimport { extractJsonCandidates } from \"./json-extract.js\";\nimport {\n  buildChatCompletionTokenLimit,\n  shouldAssumeOpenAiChatCompletions,\n} from \"./openai-chat-compat.js\";\nimport { resolveProviderApiKey } from \"./resolve-provider-secret.js\";\n\nexport interface FallbackLlmOptions {\n  temperature?: number;\n  maxTokens?: number;\n  timeoutMs?: number;\n  /** Override which agent persona's model chain to use (by ID from agents.list[]). */\n  agentId?: string;\n}\n\nexport interface FallbackLlmResponse {\n  content: string;\n  modelUsed: string;\n  usage?: {\n    inputTokens?: number;\n    outputTokens?: number;\n    totalTokens?: number;\n  };\n}\n\ninterface ModelRef {\n  providerId: string;\n  modelId: string;\n  providerConfig: ModelProviderConfig;\n  modelString: string;\n}\n\n/**\n * Generic fallback LLM client that uses the gateway's default AI configuration\n * and walks through the full fallback chain (primary + fallbacks).\n * Supports OpenAI and Anthropic API formats.\n */\nexport class FallbackLlmClient {\n  private gatewayConfig: GatewayConfig | undefined;\n\n  constructor(gatewayConfig?: GatewayConfig) {\n    this.gatewayConfig = gatewayConfig;\n  }\n\n  /**\n   * Check if fallback is available (gateway config has at least one model).\n   */\n  isAvailable(agentId?: string): boolean {\n    const models = this.getModelChain(agentId);\n    return models.length > 0;\n  }\n\n  /**\n   * Make a chat completion request using the gateway's default AI chain.\n   * Tries primary first, then each fallback in order.\n   * When agentId is provided, uses that agent persona's model chain instead of defaults.\n   */\n  async chatCompletion(\n    messages: Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }>,\n    options: FallbackLlmOptions = {},\n  ): Promise<FallbackLlmResponse | null> {\n    const models = this.getModelChain(options.agentId);\n    if (models.length === 0) {\n      log.warn(\"fallback LLM: no models configured in gateway\");\n      return null;\n    }\n\n    const runChain = async (): Promise<FallbackLlmResponse | null> => {\n      // Try each model in the chain\n      for (let i = 0; i < models.length; i++) {\n        const model = models[i];\n        const isFallback = i > 0;\n\n        try {\n          const result = await this.tryModel(model, messages, options);\n          if (result) {\n            if (isFallback) {\n              log.debug(`fallback LLM: succeeded using ${model.modelString} (fallback ${i})`);\n            }\n            return {\n              content: result.content,\n              modelUsed: model.modelString,\n              usage: result.usage,\n            };\n          }\n        } catch (err) {\n          const errorMsg = err instanceof Error ? err.message : String(err);\n          log.debug(`fallback LLM: ${model.modelString} failed (${errorMsg}), trying next...`);\n          // Continue to next model in chain\n        }\n      }\n\n      log.warn(`fallback LLM: all ${models.length} models in chain failed`);\n      return null;\n    };\n\n    if (typeof options.timeoutMs === \"number\") {\n      if (options.timeoutMs <= 0) {\n        log.warn(\"fallback LLM: timed out before request started\");\n        return null;\n      }\n      let timeoutHandle: ReturnType<typeof setTimeout> | undefined;\n      try {\n        return await Promise.race([\n          runChain(),\n          new Promise<null>((resolve) => {\n            timeoutHandle = setTimeout(() => {\n              log.warn(`fallback LLM: timed out after ${options.timeoutMs}ms`);\n              resolve(null);\n            }, options.timeoutMs);\n          }),\n        ]);\n      } finally {\n        if (timeoutHandle) clearTimeout(timeoutHandle);\n      }\n    }\n\n    return await runChain();\n  }\n\n  /**\n   * Make a request with structured output (Zod schema).\n   * Returns parsed JSON or null on failure.\n   */\n  async parseWithSchema<T>(\n    messages: Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }>,\n    schema: { parse: (data: unknown) => T },\n    options: FallbackLlmOptions = {},\n  ): Promise<T | null> {\n    const detailed = await this.parseWithSchemaDetailed(messages, schema, options);\n    return detailed?.result ?? null;\n  }\n\n  /**\n   * Like parseWithSchema but also returns the model that was used,\n   * so callers can emit accurate trace events.\n   */\n  async parseWithSchemaDetailed<T>(\n    messages: Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }>,\n    schema: { parse: (data: unknown) => T },\n    options: FallbackLlmOptions = {},\n  ): Promise<{ result: T; modelUsed: string } | null> {\n    const response = await this.chatCompletion(messages, options);\n    if (!response?.content) return null;\n\n    try {\n      const candidates = extractJsonCandidates(response.content);\n      for (const c of candidates) {\n        try {\n          const parsed = JSON.parse(c);\n          return { result: schema.parse(parsed), modelUsed: response.modelUsed };\n        } catch {\n          // keep trying other candidates\n        }\n      }\n      return null;\n    } catch (err) {\n      log.warn(\"fallback LLM: failed to parse structured output:\", err);\n      return null;\n    }\n  }\n\n  /**\n   * Get the full model chain from gateway config.\n   * Returns array of models in order: [primary, fallback1, fallback2, ...]\n   *\n   * When agentId is provided, looks up the matching entry in agents.list[]\n   * and uses that persona's model chain. Falls back to agents.defaults.model\n   * if agentId is not found or not provided.\n   */\n  private getModelChain(agentId?: string): ModelRef[] {\n    const chain: ModelRef[] = [];\n    const providers = this.gatewayConfig?.models?.providers;\n\n    if (!providers) return chain;\n\n    // Resolve the model config: agent persona chain or global defaults\n    let modelConfig: { primary?: string; fallbacks?: string[] } | undefined;\n\n    if (agentId) {\n      const persona = this.gatewayConfig?.agents?.list?.find(\n        (a) => a.id === agentId,\n      );\n      if (persona?.model) {\n        modelConfig = persona.model;\n        log.debug(`fallback LLM: using agent persona \"${agentId}\" model chain`);\n      } else {\n        log.warn(\n          `fallback LLM: agent persona \"${agentId}\" not found or has no model config, falling back to defaults`,\n        );\n      }\n    }\n\n    if (!modelConfig) {\n      modelConfig = this.gatewayConfig?.agents?.defaults?.model;\n    }\n\n    // Build list of model strings: primary + fallbacks\n    const modelStrings: string[] = [];\n\n    if (modelConfig?.primary) {\n      modelStrings.push(modelConfig.primary);\n    }\n\n    if (Array.isArray(modelConfig?.fallbacks)) {\n      for (const fb of modelConfig.fallbacks) {\n        if (typeof fb === \"string\" && !modelStrings.includes(fb)) {\n          modelStrings.push(fb);\n        }\n      }\n    }\n\n    // Parse each model string and look up provider config\n    for (const modelString of modelStrings) {\n      const modelRef = this.parseModelString(modelString, providers);\n      if (modelRef) {\n        chain.push(modelRef);\n      }\n    }\n\n    return chain;\n  }\n\n  /**\n   * Parse a \"provider/model\" string and look up its config.\n   */\n  private parseModelString(\n    modelString: string,\n    providers: Record<string, ModelProviderConfig>,\n  ): ModelRef | null {\n    // Parse \"provider/model\" format (e.g., \"openai/gpt-5.2\", \"anthropic/claude-opus-4-6\")\n    const parts = modelString.split(\"/\");\n    if (parts.length < 2) {\n      log.warn(`fallback LLM: invalid model format: ${modelString}`);\n      return null;\n    }\n\n    const providerId = parts[0];\n    const modelId = parts.slice(1).join(\"/\"); // Handle cases like \"openai/gpt-5.2-turbo\"\n\n    const providerConfig = providers[providerId];\n    if (!providerConfig) {\n      log.warn(`fallback LLM: provider not found: ${providerId}`);\n      return null;\n    }\n\n    return { providerId, modelId, providerConfig, modelString };\n  }\n\n  /**\n   * Resolve the API key for a provider, handling OpenClaw secret ref formats.\n   * Results are cached per provider so exec calls only happen once.\n   */\n  private async resolveApiKey(\n    providerId: string,\n    providerConfig: ModelProviderConfig,\n  ): Promise<string | undefined> {\n    return resolveProviderApiKey(providerId, providerConfig.apiKey, this.gatewayConfig);\n  }\n\n  /**\n   * Try to call a single model.\n   */\n  private async tryModel(\n    model: ModelRef,\n    messages: Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }>,\n    options: FallbackLlmOptions,\n  ): Promise<{ content: string; usage?: FallbackLlmResponse[\"usage\"] } | null> {\n    // Resolve the API key from secret refs before making the call.\n    // If the raw key looks like an unresolved secret ref and resolution fails,\n    // skip this provider entirely so the chain falls through to the next.\n    const rawKey = model.providerConfig.apiKey;\n    const needsResolution = rawKey === \"secretref-managed\"\n      || (typeof rawKey === \"object\" && rawKey !== null);\n    const resolvedApiKey = await this.resolveApiKey(model.providerId, model.providerConfig);\n\n    if (needsResolution && !resolvedApiKey) {\n      throw new Error(`API key for provider \"${model.providerId}\" could not be resolved from secret ref`);\n    }\n\n    const configWithResolvedKey = resolvedApiKey\n      ? { ...model.providerConfig, apiKey: resolvedApiKey }\n      : model.providerConfig;\n\n    switch (model.providerConfig.api) {\n      case \"anthropic-messages\":\n        return await this.callAnthropic(configWithResolvedKey, model.modelId, messages, options);\n      case \"openai-completions\":\n      default:\n        return await this.callOpenAI(\n          configWithResolvedKey,\n          model.modelId,\n          messages,\n          options,\n          shouldAssumeOpenAiChatCompletions(model.providerConfig.baseUrl),\n        );\n    }\n  }\n\n  /**\n   * Call OpenAI-compatible API.\n   */\n  private async callOpenAI(\n    config: ModelProviderConfig,\n    modelId: string,\n    messages: Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }>,\n    options: FallbackLlmOptions,\n    assumeOpenAI: boolean,\n  ): Promise<{ content: string; usage?: FallbackLlmResponse[\"usage\"] } | null> {\n    const base = config.baseUrl.replace(/\\/$/, \"\");\n    const url = base.endsWith(\"/v1\")\n      ? `${base}/chat/completions`\n      : `${base}/v1/chat/completions`;\n\n    const headers: Record<string, string> = {\n      \"Content-Type\": \"application/json\",\n      ...config.headers,\n    };\n\n    // Handle auth — apiKey is already resolved to a string by tryModel()\n    if (config.apiKey && typeof config.apiKey === \"string\") {\n      if (config.authHeader !== false) {\n        headers[\"Authorization\"] = `Bearer ${config.apiKey}`;\n      }\n    }\n\n    const body = {\n      model: modelId,\n      messages,\n      temperature: options.temperature ?? 0.3,\n      ...buildChatCompletionTokenLimit(modelId, options.maxTokens ?? 4096, {\n        assumeOpenAI,\n      }),\n    };\n\n    const response = await fetch(url, {\n      method: \"POST\",\n      headers,\n      body: JSON.stringify(body),\n    });\n\n    if (!response.ok) {\n      const error = await response.text();\n      throw new Error(`OpenAI API error: ${response.status} ${error}`);\n    }\n\n    const data = (await response.json()) as {\n      choices: Array<{\n        message: {\n          content: string;\n        };\n      }>;\n      usage?: {\n        prompt_tokens?: number;\n        completion_tokens?: number;\n        total_tokens?: number;\n      };\n    };\n\n    const content = data.choices?.[0]?.message?.content;\n    if (!content) {\n      throw new Error(\"Empty response from OpenAI API\");\n    }\n\n    return {\n      content,\n      usage: data.usage\n        ? {\n            inputTokens: data.usage.prompt_tokens,\n            outputTokens: data.usage.completion_tokens,\n            totalTokens: data.usage.total_tokens,\n          }\n        : undefined,\n    };\n  }\n\n  /**\n   * Call Anthropic Messages API.\n   */\n  private async callAnthropic(\n    config: ModelProviderConfig,\n    modelId: string,\n    messages: Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }>,\n    options: FallbackLlmOptions,\n  ): Promise<{ content: string; usage?: FallbackLlmResponse[\"usage\"] } | null> {\n    const url = `${config.baseUrl.replace(/\\/$/, \"\")}/messages`;\n\n    const headers: Record<string, string> = {\n      \"Content-Type\": \"application/json\",\n      \"anthropic-version\": \"2023-06-01\",\n      ...config.headers,\n    };\n\n    // Handle auth - Anthropic uses x-api-key header (apiKey resolved by tryModel)\n    if (config.apiKey && typeof config.apiKey === \"string\") {\n      headers[\"x-api-key\"] = config.apiKey;\n    }\n\n    // Extract system message (Anthropic handles it separately)\n    const systemMessage = messages.find((m) => m.role === \"system\")?.content;\n    const nonSystemMessages = messages.filter((m) => m.role !== \"system\");\n\n    // Convert messages to Anthropic format\n    const anthropicMessages = nonSystemMessages.map((m) => ({\n      role: m.role,\n      content: m.content,\n    }));\n\n    const body: Record<string, unknown> = {\n      model: modelId,\n      messages: anthropicMessages,\n      max_tokens: options.maxTokens ?? 4096,\n      temperature: options.temperature ?? 0.3,\n    };\n\n    if (systemMessage) {\n      body.system = systemMessage;\n    }\n\n    const response = await fetch(url, {\n      method: \"POST\",\n      headers,\n      body: JSON.stringify(body),\n    });\n\n    if (!response.ok) {\n      const error = await response.text();\n      throw new Error(`Anthropic API error: ${response.status} ${error}`);\n    }\n\n    const data = (await response.json()) as {\n      content: Array<{\n        type: string;\n        text: string;\n      }>;\n      usage?: {\n        input_tokens?: number;\n        output_tokens?: number;\n      };\n    };\n\n    const content = data.content?.[0]?.text;\n    if (!content) {\n      throw new Error(\"Empty response from Anthropic API\");\n    }\n\n    return {\n      content,\n      usage: data.usage\n        ? {\n            inputTokens: data.usage.input_tokens,\n            outputTokens: data.usage.output_tokens,\n            totalTokens: (data.usage.input_tokens ?? 0) + (data.usage.output_tokens ?? 0),\n          }\n        : undefined,\n    };\n  }\n}\n"],"mappings":";;;;;;AAWO,SAAS,gBAAgB,MAAsB;AACpD,SAAO,KAAK,QAAQ,kCAAkC,CAAC,IAAI,UAAU,OAAO,KAAK,EAAE,KAAK,CAAC;AAC3F;AAEO,SAAS,sBAAsB,MAAwB;AAC5D,QAAM,UAAU,KAAK,KAAK;AAC1B,QAAM,UAAU,gBAAgB,OAAO;AACvC,QAAM,aAAuB,CAAC;AAE9B,MAAI,QAAQ,SAAS,EAAG,YAAW,KAAK,OAAO;AAC/C,aAAW,KAAK,GAAG,uBAAuB,OAAO,CAAC;AAGlD,QAAM,WAAW,QAAQ,MAAM,aAAa;AAC5C,MAAI,SAAU,YAAW,KAAK,SAAS,CAAC,CAAC;AAEzC,QAAM,OAAO,oBAAI,IAAY;AAC7B,SAAO,WACJ,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC,EAC1B,OAAO,CAAC,MAAM;AACb,QAAI,KAAK,IAAI,CAAC,EAAG,QAAO;AACxB,SAAK,IAAI,CAAC;AACV,WAAO;AAAA,EACT,CAAC;AACL;AAEA,SAAS,uBAAuB,MAAwB;AACtD,QAAM,MAAgB,CAAC;AACvB,QAAM,QAAQ,oBAAI,IAAI,CAAC,KAAK,GAAG,CAAC;AAChC,QAAM,SAAiC,EAAE,KAAK,KAAK,KAAK,IAAI;AAE5D,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,QAAQ,KAAK,CAAC;AACpB,QAAI,CAAC,MAAM,IAAI,KAAK,EAAG;AAEvB,UAAM,gBAAgB,OAAO,KAAK;AAClC,QAAI,QAAQ;AACZ,QAAI,WAAW;AACf,QAAI,SAAS;AAEb,aAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,YAAM,KAAK,KAAK,CAAC;AAEjB,UAAI,UAAU;AACZ,YAAI,QAAQ;AACV,mBAAS;AAAA,QACX,WAAW,OAAO,MAAM;AACtB,mBAAS;AAAA,QACX,WAAW,OAAO,KAAM;AACtB,qBAAW;AAAA,QACb;AACA;AAAA,MACF;AAEA,UAAI,OAAO,KAAM;AACf,mBAAW;AACX;AAAA,MACF;AAEA,UAAI,OAAO,MAAO;AAClB,UAAI,OAAO,cAAe;AAE1B,UAAI,UAAU,GAAG;AACf,YAAI,KAAK,KAAK,MAAM,GAAG,IAAI,CAAC,EAAE,KAAK,CAAC;AACpC,YAAI;AACJ;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;;;ACnFA,SAAS,gBAAgB,OAAuB;AAC9C,SAAO,MAAM,KAAK,EAAE,YAAY;AAClC;AAEA,SAAS,mBAAmB,YAAoB,eAAgC;AAC9E,SAAO,cAAc,KAAK,UAAU;AACtC;AAEO,SAAS,kCAAkC,SAA2B;AAC3E,MAAI,CAAC,QAAS,QAAO;AACrB,MAAI;AACF,WAAO,IAAI,IAAI,OAAO,EAAE,SAAS,YAAY,MAAM;AAAA,EACrD,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAEO,SAAS,wBAAwB,OAAe,SAA+C;AACpG,QAAM,aAAa,gBAAgB,KAAK;AACxC,MAAI,SAAS,iBAAiB,KAAM,QAAO;AAC3C,MAAI,mBAAmB,YAAY,kBAAkB,EAAG,QAAO;AAC/D,MAAI,mBAAmB,YAAY,mBAAmB,EAAG,QAAO;AAChE,MAAI,mBAAmB,YAAY,qBAAqB,EAAG,QAAO;AAClE,MAAI,mBAAmB,YAAY,eAAe,EAAG,QAAO;AAC5D,MAAI,mBAAmB,YAAY,eAAe,EAAG,QAAO;AAC5D,SAAO,mBAAmB,YAAY,oBAAoB;AAC5D;AAEO,SAAS,8BACd,OACA,WACA,SAC4D;AAC5D,QAAM,gBAAgB,KAAK,IAAI,GAAG,KAAK,MAAM,SAAS,CAAC;AACvD,MAAI,wBAAwB,OAAO,OAAO,GAAG;AAC3C,WAAO,EAAE,uBAAuB,cAAc;AAAA,EAChD;AACA,SAAO,EAAE,YAAY,cAAc;AACrC;;;ACrCA,OAAO,UAAU;AACjB,OAAO,QAAQ;AAsBf,IAAI,4BAAoD;AACxD,IAAI,kBAAkB;AACtB,IAAI,uBAAuB;AAC3B,IAAM,4BAA4B;AAClC,IAAM,gBAAgB,oBAAI,IAAgC;AAM1D,eAAe,qBAAsD;AACnE,MAAI,iBAAiB;AACnB,WAAO;AAAA,EACT;AAGA,MAAI,uBAAuB,KAAK,KAAK,IAAI,IAAI,sBAAsB;AACjE,WAAO;AAAA,EACT;AAEA,MAAI;AAGF,UAAM,aAAa;AAAA;AAAA,MAEjB,GAAG,MAAM,mBAAmB;AAAA,IAC9B;AAEA,UAAM,EAAE,cAAc,IAAI,MAAM,OAAO,KAAU;AACjD,eAAW,aAAa,YAAY;AAClC,UAAI;AAEF,cAAM,YAAY,cAAc,SAAS,EAAE;AAC3C,cAAM,MAAM,MAAM,OAAO;AACzB,YAAI,OAAO,IAAI,6BAA6B,YAAY;AACtD,sCAA4B,IAAI;AAChC,4BAAkB;AAClB,cAAI,MAAM,6DAA6D;AACvE,iBAAO;AAAA,QACT;AAAA,MACF,QAAQ;AAAA,MAER;AAAA,IACF;AAAA,EACF,QAAQ;AAAA,EAER;AAKA,yBAAuB,KAAK,IAAI,IAAI;AACpC,MAAI,MAAM,0EAAqE,4BAA4B,GAAI,GAAG;AAClH,SAAO;AACT;AAMA,eAAe,qBAAwC;AACrD,QAAM,EAAE,YAAY,IAAI,MAAM,OAAO,IAAS;AAC9C,QAAM,EAAE,cAAc,IAAI,MAAM,OAAO,QAAa;AACpD,QAAM,aAAuB,CAAC;AAI9B,QAAM,WAAqB,CAAC;AAE5B,MAAI;AAEF,UAAM,MAAM,cAAc,YAAY,GAAG;AACzC,UAAM,eAAe,IAAI,QAAQ,UAAU;AAC3C,UAAM,eAAe,KAAK,KAAK,KAAK,QAAQ,YAAY,GAAG,MAAM,MAAM;AACvE,QAAI,aAAc,UAAS,KAAK,KAAK,QAAQ,YAAY,CAAC;AAAA,EAC5D,QAAQ;AAAA,EAER;AAIA,MAAI;AACF,UAAM,EAAE,aAAa,IAAI,MAAM,OAAO,IAAS;AAC/C,UAAM,aAAa,QAAQ,KAAK,CAAC;AACjC,QAAI,YAAY;AACd,YAAM,aAAa,aAAa,UAAU;AAC1C,UAAI,WAAW,SAAS,UAAU,GAAG;AACnC,cAAM,UAAU,KAAK,QAAQ,UAAU;AACvC,YAAI,CAAC,SAAS,SAAS,OAAO,EAAG,UAAS,KAAK,OAAO;AAAA,MACxD;AAAA,IACF;AAAA,EACF,QAAQ;AAAA,EAER;AAEA,aAAW,OAAO,UAAU;AAC1B,QAAI;AACF,YAAM,QAAQ,YAAY,GAAG;AAC7B,iBAAW,KAAK,OAAO;AACrB,YAAI,EAAE,WAAW,6BAA6B,KAAK,EAAE,SAAS,KAAK,GAAG;AACpE,qBAAW,KAAK,KAAK,KAAK,KAAK,CAAC,CAAC;AAAA,QACnC;AAAA,MACF;AAAA,IACF,QAAQ;AAAA,IAER;AAAA,EACF;AAEA,SAAO;AACT;AAWA,eAAsB,sBACpB,YACA,aACA,eACA,UAC6B;AAE7B,QAAM,WAAW,YAAY,UAAU;AACvC,MAAI,cAAc,IAAI,QAAQ,GAAG;AAC/B,WAAO,cAAc,IAAI,QAAQ;AAAA,EACnC;AAEA,MAAI;AAGJ,MAAI,OAAO,gBAAgB,YAAY,YAAY,KAAK,EAAE,SAAS,GAAG;AAGpE,QACE,gBAAgB,uBAChB,YAAY,SAAS,QAAQ,KAC7B,YAAY,SAAS,QAAQ,KAC7B,gBAAgB,eAChB,YAAY,WAAW,MAAM,GAC7B;AAAA,IAEF,OAAO;AACL,iBAAW;AACX,oBAAc,IAAI,UAAU,QAAQ;AACpC,aAAO;AAAA,IACT;AAAA,EACF;AAIA,QAAM,WAAW,MAAM,mBAAmB;AAC1C,MAAI,UAAU;AACZ,QAAI;AACF,YAAM,mBAAmB,YAAY,KAAK,KAAK,GAAG,QAAQ,GAAG,aAAa,UAAU,QAAQ,OAAO;AACnG,YAAM,OAAO,MAAM,SAAS,EAAE,UAAU,YAAY,KAAK,eAAe,UAAU,iBAAiB,CAAC;AACpG,UAAI,MAAM,QAAQ;AAChB,mBAAW,KAAK;AAChB,YAAI,MAAM,kCAAkC,UAAU,+BAA+B,KAAK,UAAU,SAAS,WAAW,KAAK,QAAQ,SAAS,GAAG;AACjJ,sBAAc,IAAI,UAAU,QAAQ;AACpC,eAAO;AAAA,MACT;AAAA,IACF,SAAS,KAAK;AACZ,UAAI;AAAA,QACF,gDAAgD,UAAU,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,MAClH;AAAA,IACF;AAAA,EACF;AAGA,aAAW,eAAe,UAAU;AACpC,MAAI,UAAU;AACZ,QAAI,MAAM,kCAAkC,UAAU,6BAA6B;AAAA,EACrF,OAAO;AACL,QAAI,MAAM,2CAA2C,UAAU,mBAAc;AAAA,EAC/E;AAIA,MAAI,UAAU;AACZ,kBAAc,IAAI,UAAU,QAAQ;AAAA,EACtC;AACA,SAAO;AACT;AAKA,SAAS,eAAe,YAAwC;AAC9D,QAAM,aAAa,WAAW,YAAY,EAAE,QAAQ,cAAc,GAAG;AACrE,QAAM,aAAa;AAAA,IACjB,GAAG,UAAU;AAAA,IACb,GAAG,UAAU;AAAA,EACf;AACA,aAAW,UAAU,YAAY;AAC/B,UAAM,QAAQ,QAAQ,IAAI,MAAM;AAChC,QAAI,SAAS,MAAM,KAAK,EAAE,SAAS,GAAG;AACpC,aAAO,MAAM,KAAK;AAAA,IACpB;AAAA,EACF;AACA,SAAO;AACT;;;AC9LO,IAAM,oBAAN,MAAwB;AAAA,EACrB;AAAA,EAER,YAAY,eAA+B;AACzC,SAAK,gBAAgB;AAAA,EACvB;AAAA;AAAA;AAAA;AAAA,EAKA,YAAY,SAA2B;AACrC,UAAM,SAAS,KAAK,cAAc,OAAO;AACzC,WAAO,OAAO,SAAS;AAAA,EACzB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,eACJ,UACA,UAA8B,CAAC,GACM;AACrC,UAAM,SAAS,KAAK,cAAc,QAAQ,OAAO;AACjD,QAAI,OAAO,WAAW,GAAG;AACvB,UAAI,KAAK,+CAA+C;AACxD,aAAO;AAAA,IACT;AAEA,UAAM,WAAW,YAAiD;AAEhE,eAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK;AACtC,cAAM,QAAQ,OAAO,CAAC;AACtB,cAAM,aAAa,IAAI;AAEvB,YAAI;AACF,gBAAM,SAAS,MAAM,KAAK,SAAS,OAAO,UAAU,OAAO;AAC3D,cAAI,QAAQ;AACV,gBAAI,YAAY;AACd,kBAAI,MAAM,iCAAiC,MAAM,WAAW,cAAc,CAAC,GAAG;AAAA,YAChF;AACA,mBAAO;AAAA,cACL,SAAS,OAAO;AAAA,cAChB,WAAW,MAAM;AAAA,cACjB,OAAO,OAAO;AAAA,YAChB;AAAA,UACF;AAAA,QACF,SAAS,KAAK;AACZ,gBAAM,WAAW,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAChE,cAAI,MAAM,iBAAiB,MAAM,WAAW,YAAY,QAAQ,mBAAmB;AAAA,QAErF;AAAA,MACF;AAEA,UAAI,KAAK,qBAAqB,OAAO,MAAM,yBAAyB;AACpE,aAAO;AAAA,IACT;AAEA,QAAI,OAAO,QAAQ,cAAc,UAAU;AACzC,UAAI,QAAQ,aAAa,GAAG;AAC1B,YAAI,KAAK,gDAAgD;AACzD,eAAO;AAAA,MACT;AACA,UAAI;AACJ,UAAI;AACF,eAAO,MAAM,QAAQ,KAAK;AAAA,UACxB,SAAS;AAAA,UACT,IAAI,QAAc,CAAC,YAAY;AAC7B,4BAAgB,WAAW,MAAM;AAC/B,kBAAI,KAAK,iCAAiC,QAAQ,SAAS,IAAI;AAC/D,sBAAQ,IAAI;AAAA,YACd,GAAG,QAAQ,SAAS;AAAA,UACtB,CAAC;AAAA,QACH,CAAC;AAAA,MACH,UAAE;AACA,YAAI,cAAe,cAAa,aAAa;AAAA,MAC/C;AAAA,IACF;AAEA,WAAO,MAAM,SAAS;AAAA,EACxB;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,gBACJ,UACA,QACA,UAA8B,CAAC,GACZ;AACnB,UAAM,WAAW,MAAM,KAAK,wBAAwB,UAAU,QAAQ,OAAO;AAC7E,WAAO,UAAU,UAAU;AAAA,EAC7B;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,wBACJ,UACA,QACA,UAA8B,CAAC,GACmB;AAClD,UAAM,WAAW,MAAM,KAAK,eAAe,UAAU,OAAO;AAC5D,QAAI,CAAC,UAAU,QAAS,QAAO;AAE/B,QAAI;AACF,YAAM,aAAa,sBAAsB,SAAS,OAAO;AACzD,iBAAW,KAAK,YAAY;AAC1B,YAAI;AACF,gBAAM,SAAS,KAAK,MAAM,CAAC;AAC3B,iBAAO,EAAE,QAAQ,OAAO,MAAM,MAAM,GAAG,WAAW,SAAS,UAAU;AAAA,QACvE,QAAQ;AAAA,QAER;AAAA,MACF;AACA,aAAO;AAAA,IACT,SAAS,KAAK;AACZ,UAAI,KAAK,oDAAoD,GAAG;AAChE,aAAO;AAAA,IACT;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUQ,cAAc,SAA8B;AAClD,UAAM,QAAoB,CAAC;AAC3B,UAAM,YAAY,KAAK,eAAe,QAAQ;AAE9C,QAAI,CAAC,UAAW,QAAO;AAGvB,QAAI;AAEJ,QAAI,SAAS;AACX,YAAM,UAAU,KAAK,eAAe,QAAQ,MAAM;AAAA,QAChD,CAAC,MAAM,EAAE,OAAO;AAAA,MAClB;AACA,UAAI,SAAS,OAAO;AAClB,sBAAc,QAAQ;AACtB,YAAI,MAAM,sCAAsC,OAAO,eAAe;AAAA,MACxE,OAAO;AACL,YAAI;AAAA,UACF,gCAAgC,OAAO;AAAA,QACzC;AAAA,MACF;AAAA,IACF;AAEA,QAAI,CAAC,aAAa;AAChB,oBAAc,KAAK,eAAe,QAAQ,UAAU;AAAA,IACtD;AAGA,UAAM,eAAyB,CAAC;AAEhC,QAAI,aAAa,SAAS;AACxB,mBAAa,KAAK,YAAY,OAAO;AAAA,IACvC;AAEA,QAAI,MAAM,QAAQ,aAAa,SAAS,GAAG;AACzC,iBAAW,MAAM,YAAY,WAAW;AACtC,YAAI,OAAO,OAAO,YAAY,CAAC,aAAa,SAAS,EAAE,GAAG;AACxD,uBAAa,KAAK,EAAE;AAAA,QACtB;AAAA,MACF;AAAA,IACF;AAGA,eAAW,eAAe,cAAc;AACtC,YAAM,WAAW,KAAK,iBAAiB,aAAa,SAAS;AAC7D,UAAI,UAAU;AACZ,cAAM,KAAK,QAAQ;AAAA,MACrB;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA,EAKQ,iBACN,aACA,WACiB;AAEjB,UAAM,QAAQ,YAAY,MAAM,GAAG;AACnC,QAAI,MAAM,SAAS,GAAG;AACpB,UAAI,KAAK,uCAAuC,WAAW,EAAE;AAC7D,aAAO;AAAA,IACT;AAEA,UAAM,aAAa,MAAM,CAAC;AAC1B,UAAM,UAAU,MAAM,MAAM,CAAC,EAAE,KAAK,GAAG;AAEvC,UAAM,iBAAiB,UAAU,UAAU;AAC3C,QAAI,CAAC,gBAAgB;AACnB,UAAI,KAAK,qCAAqC,UAAU,EAAE;AAC1D,aAAO;AAAA,IACT;AAEA,WAAO,EAAE,YAAY,SAAS,gBAAgB,YAAY;AAAA,EAC5D;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,cACZ,YACA,gBAC6B;AAC7B,WAAO,sBAAsB,YAAY,eAAe,QAAQ,KAAK,aAAa;AAAA,EACpF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAc,SACZ,OACA,UACA,SAC2E;AAI3E,UAAM,SAAS,MAAM,eAAe;AACpC,UAAM,kBAAkB,WAAW,uBAC7B,OAAO,WAAW,YAAY,WAAW;AAC/C,UAAM,iBAAiB,MAAM,KAAK,cAAc,MAAM,YAAY,MAAM,cAAc;AAEtF,QAAI,mBAAmB,CAAC,gBAAgB;AACtC,YAAM,IAAI,MAAM,yBAAyB,MAAM,UAAU,yCAAyC;AAAA,IACpG;AAEA,UAAM,wBAAwB,iBAC1B,EAAE,GAAG,MAAM,gBAAgB,QAAQ,eAAe,IAClD,MAAM;AAEV,YAAQ,MAAM,eAAe,KAAK;AAAA,MAChC,KAAK;AACH,eAAO,MAAM,KAAK,cAAc,uBAAuB,MAAM,SAAS,UAAU,OAAO;AAAA,MACzF,KAAK;AAAA,MACL;AACE,eAAO,MAAM,KAAK;AAAA,UAChB;AAAA,UACA,MAAM;AAAA,UACN;AAAA,UACA;AAAA,UACA,kCAAkC,MAAM,eAAe,OAAO;AAAA,QAChE;AAAA,IACJ;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAc,WACZ,QACA,SACA,UACA,SACA,cAC2E;AAC3E,UAAM,OAAO,OAAO,QAAQ,QAAQ,OAAO,EAAE;AAC7C,UAAM,MAAM,KAAK,SAAS,KAAK,IAC3B,GAAG,IAAI,sBACP,GAAG,IAAI;AAEX,UAAM,UAAkC;AAAA,MACtC,gBAAgB;AAAA,MAChB,GAAG,OAAO;AAAA,IACZ;AAGA,QAAI,OAAO,UAAU,OAAO,OAAO,WAAW,UAAU;AACtD,UAAI,OAAO,eAAe,OAAO;AAC/B,gBAAQ,eAAe,IAAI,UAAU,OAAO,MAAM;AAAA,MACpD;AAAA,IACF;AAEA,UAAM,OAAO;AAAA,MACX,OAAO;AAAA,MACP;AAAA,MACA,aAAa,QAAQ,eAAe;AAAA,MACpC,GAAG,8BAA8B,SAAS,QAAQ,aAAa,MAAM;AAAA,QACnE;AAAA,MACF,CAAC;AAAA,IACH;AAEA,UAAM,WAAW,MAAM,MAAM,KAAK;AAAA,MAChC,QAAQ;AAAA,MACR;AAAA,MACA,MAAM,KAAK,UAAU,IAAI;AAAA,IAC3B,CAAC;AAED,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,QAAQ,MAAM,SAAS,KAAK;AAClC,YAAM,IAAI,MAAM,qBAAqB,SAAS,MAAM,IAAI,KAAK,EAAE;AAAA,IACjE;AAEA,UAAM,OAAQ,MAAM,SAAS,KAAK;AAalC,UAAM,UAAU,KAAK,UAAU,CAAC,GAAG,SAAS;AAC5C,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI,MAAM,gCAAgC;AAAA,IAClD;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO,KAAK,QACR;AAAA,QACE,aAAa,KAAK,MAAM;AAAA,QACxB,cAAc,KAAK,MAAM;AAAA,QACzB,aAAa,KAAK,MAAM;AAAA,MAC1B,IACA;AAAA,IACN;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAc,cACZ,QACA,SACA,UACA,SAC2E;AAC3E,UAAM,MAAM,GAAG,OAAO,QAAQ,QAAQ,OAAO,EAAE,CAAC;AAEhD,UAAM,UAAkC;AAAA,MACtC,gBAAgB;AAAA,MAChB,qBAAqB;AAAA,MACrB,GAAG,OAAO;AAAA,IACZ;AAGA,QAAI,OAAO,UAAU,OAAO,OAAO,WAAW,UAAU;AACtD,cAAQ,WAAW,IAAI,OAAO;AAAA,IAChC;AAGA,UAAM,gBAAgB,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACjE,UAAM,oBAAoB,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,QAAQ;AAGpE,UAAM,oBAAoB,kBAAkB,IAAI,CAAC,OAAO;AAAA,MACtD,MAAM,EAAE;AAAA,MACR,SAAS,EAAE;AAAA,IACb,EAAE;AAEF,UAAM,OAAgC;AAAA,MACpC,OAAO;AAAA,MACP,UAAU;AAAA,MACV,YAAY,QAAQ,aAAa;AAAA,MACjC,aAAa,QAAQ,eAAe;AAAA,IACtC;AAEA,QAAI,eAAe;AACjB,WAAK,SAAS;AAAA,IAChB;AAEA,UAAM,WAAW,MAAM,MAAM,KAAK;AAAA,MAChC,QAAQ;AAAA,MACR;AAAA,MACA,MAAM,KAAK,UAAU,IAAI;AAAA,IAC3B,CAAC;AAED,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,QAAQ,MAAM,SAAS,KAAK;AAClC,YAAM,IAAI,MAAM,wBAAwB,SAAS,MAAM,IAAI,KAAK,EAAE;AAAA,IACpE;AAEA,UAAM,OAAQ,MAAM,SAAS,KAAK;AAWlC,UAAM,UAAU,KAAK,UAAU,CAAC,GAAG;AACnC,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI,MAAM,mCAAmC;AAAA,IACrD;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO,KAAK,QACR;AAAA,QACE,aAAa,KAAK,MAAM;AAAA,QACxB,cAAc,KAAK,MAAM;AAAA,QACzB,cAAc,KAAK,MAAM,gBAAgB,MAAM,KAAK,MAAM,iBAAiB;AAAA,MAC7E,IACA;AAAA,IACN;AAAA,EACF;AACF;","names":[]}

package/dist/{chunk-EM2BU45W.js → chunk-PIQWOQEX.js}

@@ -830,7 +830,7 @@ var CompoundingEngine = class {
     let promotionCandidates = this.config.compoundingSemanticEnabled ? this.derivePromotionCandidates(outcomeSummary, mistakes.registry, rubrics) : [];
     if (this.config.cmcConsolidationEnabled) {
       try {
-        const { deriveCausalPromotionCandidates } = await import("./causal-consolidation-5RFIDR74.js");
+        const { deriveCausalPromotionCandidates } = await import("./causal-consolidation-62M3TOOX.js");
         const causalCandidates = await deriveCausalPromotionCandidates({
           memoryDir: this.config.memoryDir,
           causalTrajectoryStoreDir: this.config.causalTrajectoryStoreDir,
@@ -851,7 +851,7 @@ var CompoundingEngine = class {
     }
     if (this.config.calibrationEnabled) {
       try {
-        const { runCalibrationConsolidation } = await import("./calibration-UY5SQU5K.js");
+        const { runCalibrationConsolidation } = await import("./calibration-WH4XEHIA.js");
         const calRules = await runCalibrationConsolidation({
           memoryDir: this.config.memoryDir,
           gatewayConfig: this.config.gatewayConfig,
@@ -1839,4 +1839,4 @@ export {
   defaultTierMigrationCycleBudget,
   CompoundingEngine
 };
-//# sourceMappingURL=chunk-EM2BU45W.js.map
+//# sourceMappingURL=chunk-PIQWOQEX.js.map

package/dist/{chunk-3ILF2OUT.js → chunk-S6X3YWYF.js}

@@ -3,7 +3,7 @@ import {
   CompoundingEngine,
   SharedContextManager,
   defaultTierMigrationCycleBudget
-} from "./chunk-EM2BU45W.js";
+} from "./chunk-PIQWOQEX.js";
 import {
   searchCausalTrajectories
 } from "./chunk-IMMYYNXG.js";
@@ -36,7 +36,7 @@ import {
   buildChatCompletionTokenLimit,
   extractJsonCandidates,
   shouldAssumeOpenAiChatCompletions
-} from "./chunk-FP3MUEPD.js";
+} from "./chunk-4722S3A6.js";
 import {
   BoxBuilder,
   assertIsoRecordedAt,
@@ -3917,7 +3917,7 @@ var LocalLlmClient = class _LocalLlmClient {
     const operation = options.operation ?? "unspecified";
     const startedAtMs = Date.now();
     if (queueMeta) {
-      log.info(
+      log.debug(
         `local LLM queue start: priority=${queueMeta.priority} waitMs=${startedAtMs - queueMeta.enqueuedAtMs} op=${operation}`
       );
     }
@@ -4116,7 +4116,7 @@ var LocalLlmClient = class _LocalLlmClient {
       if (queueMeta) {
         const finishedAtMs = Date.now();
         const waitMs = startedAtMs - queueMeta.enqueuedAtMs;
-        log.info(
+        log.debug(
           `local LLM queue finish: priority=${queueMeta.priority} waitMs=${waitMs} runMs=${finishedAtMs - startedAtMs} totalMs=${finishedAtMs - queueMeta.enqueuedAtMs} op=${operation}`
         );
       }
@@ -4129,7 +4129,7 @@ var LocalLlmClient = class _LocalLlmClient {
   async getLoadedModelInfo() {
     const baseUrl = this.config.localLlmUrl.replace("localhost", "127.0.0.1").replace(/\/+$/, "");
     const modelsUrl = baseUrl.endsWith("/v1") ? `${baseUrl}/models` : `${baseUrl}/v1/models`;
-    log.info(`Fetching model info from ${modelsUrl}`);
+    log.debug(`Fetching model info from ${modelsUrl}`);
     try {
       const result = await this.fetchWithTimeout(modelsUrl, 3e3);
       if (!result.ok) {
@@ -4758,8 +4758,8 @@ var ExtractionEngine = class {
     this.fallbackLlm = new FallbackLlmClient(gatewayConfig);
     this.modelRegistry = modelRegistry ?? new ModelRegistry(config.memoryDir);
     if (config.modelSource === "gateway") {
-      log.info(
-        `extraction engine: using gateway model source` + (config.gatewayAgentId ? ` (agent: ${config.gatewayAgentId})` : " (defaults)")
+      log.debug(
+        `extraction engine: gateway model source active; extraction uses the gateway chain as its primary path` + (config.gatewayAgentId ? ` (agent: ${config.gatewayAgentId})` : " (defaults)")
       );
     }
   }
@@ -5330,7 +5330,13 @@ var ExtractionEngine = class {
     }
     const fallbackTraceId = crypto.randomUUID();
     const fallbackStartTime = Date.now();
-    log.info("extraction: falling back to gateway default AI");
+    if (this.useGatewayModelSource) {
+      log.debug(
+        `extraction: using gateway model chain as primary path` + (this.config.gatewayAgentId ? ` (agent: ${this.config.gatewayAgentId})` : " (defaults)")
+      );
+    } else {
+      log.info("extraction: falling back to gateway default AI");
+    }
     try {
       const messages = [
         { role: "system", content: this.buildExtractionInstructions(existingEntities) },
@@ -5541,7 +5547,7 @@ ${truncatedConversation}`;
     const tokenParams = buildChatCompletionTokenLimit(this.config.model, this.config.extractionMaxOutputTokens, {
       assumeOpenAI: this.directClientUsesOpenAiTokenSemantics()
     });
-    log.info(`extractWithDirectClient: calling model=${this.config.model} tokenParams=${JSON.stringify(tokenParams)}`);
+    log.debug(`extractWithDirectClient: calling model=${this.config.model} tokenParams=${JSON.stringify(tokenParams)}`);
     const response = await this.client.chat.completions.create({
       model: this.config.model,
       messages: [
@@ -21737,7 +21743,7 @@ var Orchestrator = class _Orchestrator {
     })() : this.localLlm;
     this.fastGatewayLlm = config.modelSource === "gateway" ? new FallbackLlmClient(config.gatewayConfig) : null;
     if (config.modelSource === "gateway") {
-      log.info(
+      log.debug(
         `orchestrator: gateway model source active` + (config.gatewayAgentId ? ` (primary: ${config.gatewayAgentId})` : "") + (config.fastGatewayAgentId ? ` (fast: ${config.fastGatewayAgentId})` : "")
       );
     }
@@ -22309,7 +22315,7 @@ var Orchestrator = class _Orchestrator {
       );
       return result;
     }
-    const { FallbackLlmClient: FallbackLlmClient2 } = await import("./fallback-llm-3DXDWI5H.js");
+    const { FallbackLlmClient: FallbackLlmClient2 } = await import("./fallback-llm-PKOQ6PB5.js");
     const useGateway = this.config.modelSource === "gateway";
     const modelSetting = this.config.semanticConsolidationModel;
     if (modelSetting === "fast" && this.fastLlm && !useGateway) {
@@ -23039,7 +23045,7 @@ ${r.snippet.trim()}
    * Warns the user if there's a mismatch.
    */
   async validateLocalLlmModel() {
-    log.info("Local LLM: Validating model configuration...");
+    log.debug("Local LLM: validating model configuration");
     try {
       const modelInfo = await this.localLlm.getLoadedModelInfo();
       if (!modelInfo) {
@@ -23053,7 +23059,7 @@ ${r.snippet.trim()}
       }
       const configuredMaxContext = this.config.localLlmMaxContext;
       if (modelInfo.contextWindow) {
-        log.info(
+        log.debug(
           `Local LLM: ${modelInfo.id} loaded with ${modelInfo.contextWindow.toLocaleString()} token context window`
         );
         if (configuredMaxContext && configuredMaxContext > modelInfo.contextWindow) {
@@ -23063,7 +23069,7 @@ ${r.snippet.trim()}
           this.config.localLlmMaxContext = modelInfo.contextWindow;
         }
       } else {
-        log.info(
+        log.debug(
           `Local LLM: ${modelInfo.id} loaded (context window not reported by server)`
         );
         if (!configuredMaxContext) {
@@ -24546,7 +24552,7 @@ ${trimmedBody}`;
         return null;
       }
       try {
-        const { getCalibrationRulesForRecall, buildCalibrationRecallSection } = await import("./calibration-UY5SQU5K.js");
+        const { getCalibrationRulesForRecall, buildCalibrationRecallSection } = await import("./calibration-WH4XEHIA.js");
         const rules = await getCalibrationRulesForRecall(this.config.memoryDir);
         if (rules.length === 0) {
           recordRecallSectionMetric({
@@ -31534,4 +31540,4 @@ export {
   EngramAccessInputError,
   EngramAccessService
 };
-//# sourceMappingURL=chunk-3ILF2OUT.js.map
+//# sourceMappingURL=chunk-S6X3YWYF.js.map