@remnic/core 9.3.664 → 9.3.666

package/dist/recall-planner-llm.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/recall-planner-llm.ts"],"sourcesContent":["import { z } from \"zod\";\n\nimport type { PluginConfig, RecallPlanMode } from \"./types.js\";\nimport { planRecallMode } from \"./intent.js\";\nimport {\n  FallbackLlmClient,\n  fallbackLlmRuntimeContextFromConfig,\n  gatewayTaskChainOptions,\n  type FallbackLlmOptions,\n} from \"./fallback-llm.js\";\nimport { log } from \"./logger.js\";\n\n/**\n * LLM-based recall planning (issue #1367, Option C).\n *\n * Classifies an incoming prompt into a {@link RecallPlanMode} using an LLM\n * instead of (or alongside) the regex heuristic in {@link planRecallMode}.\n *\n * Provider-agnostic by construction: it routes through {@link FallbackLlmClient},\n * which resolves the model chain from gateway providers (OpenAI, Anthropic,\n * Ollama, Codex, …) or gateway agent personas / `taskModelChain`. Nothing here\n * is hard-coded to a single provider or to OpenAI's Responses API — the API\n * dialect is chosen per-provider by the client based on each provider's `api`\n * field. The configured `recallPlannerModel` is tried first, with the broader\n * task chain (and the gateway default) as resilient fallbacks.\n *\n * Invariants:\n *  - Never throws to the caller (gotcha #13). Any LLM failure, timeout, empty\n *    response, or unavailable backend falls back to the heuristic result and\n *    sets `fallbackUsed: true` (gotcha #34 — failures are distinct from a valid\n *    classification).\n *  - When `recallPlannerLlmEnabled` is false the LLM is never contacted.\n */\n\nexport type RecallPlannerSource = \"llm\" | \"heuristic\" | \"heuristic-fallback\";\n\nexport interface RecallPlannerLlmResult {\n  /** The mode to act on. */\n  mode: RecallPlanMode;\n  /** The heuristic mode, always computed (the fallback floor / shadow baseline). */\n  heuristicMode: RecallPlanMode;\n  /** Where `mode` came from. */\n  source: RecallPlannerSource;\n  /** Short human-readable rationale (LLM reason, or why we fell back). */\n  reason: string;\n  /** Model that actually served the classification, when an LLM was used. */\n  modelUsed?: string;\n  /** Wall-clock spent in the LLM call (0 when no call was made). */\n  latencyMs: number;\n  /** True when the LLM was enabled but we had to fall back to the heuristic. */\n  fallbackUsed: boolean;\n}\n\nconst PLANNER_SCHEMA = z.object({\n  // gotcha #2: optional fields use .optional().nullable()\n  mode: z.enum([\"no_recall\", \"minimal\", \"full\", \"graph_mode\"]),\n  reason: z.string().max(280).optional().nullable(),\n});\n\nconst SYSTEM_PROMPT = [\n  \"You are a recall-planning classifier for a long-term memory system.\",\n  \"Given the user's latest message, decide how much stored memory should be retrieved before the assistant responds.\",\n  \"Reply with a single JSON object: {\\\"mode\\\": <one of no_recall|minimal|full|graph_mode>, \\\"reason\\\": <short string>}.\",\n  \"\",\n  \"Modes:\",\n  '- \"no_recall\": low-information acknowledgements or chit-chat with nothing to look up (e.g. \"ok\", \"thanks\", \"sounds good\"). Retrieve nothing.',\n  '- \"minimal\": short, self-contained operational directives that rarely need history (e.g. \"restart the service\", \"run the tests\", \"show status\"). Retrieve a little.',\n  '- \"full\": anything memory-seeking, analytical, or a question that benefits from prior context, decisions, or facts. This is the safe default when unsure.',\n  '- \"graph_mode\": queries about timelines, sequences, history, causal chains, or root cause (\"how did we get here\", \"what led to this regression\"). Retrieve relationship/graph context.',\n  \"\",\n  \"When uncertain, prefer \\\"full\\\" over dropping recall. Never invent facts; only classify intent.\",\n].join(\"\\n\");\n\n/** Clamp a planner prompt to the configured character budget. */\nfunction clampPrompt(prompt: string, maxChars: number): string {\n  const safeMax = Number.isFinite(maxChars) && maxChars > 0 ? Math.floor(maxChars) : 4000;\n  if (prompt.length <= safeMax) return prompt;\n  return prompt.slice(0, safeMax);\n}\n\n/** Trim and cap the optional memory hints used to ground the classification. */\nfunction clampHints(hints: string[] | undefined, maxHints: number): string[] {\n  if (!Array.isArray(hints) || hints.length === 0) return [];\n  const safeMax = Number.isFinite(maxHints) && maxHints > 0 ? Math.floor(maxHints) : 0;\n  if (safeMax <= 0) return [];\n  const cleaned: string[] = [];\n  for (const hint of hints) {\n    if (typeof hint !== \"string\") continue;\n    const trimmed = hint.trim();\n    if (trimmed.length === 0) continue;\n    cleaned.push(trimmed);\n    if (cleaned.length >= safeMax) break;\n  }\n  return cleaned;\n}\n\nfunction buildMessages(\n  prompt: string,\n  hints: string[],\n  config: PluginConfig,\n): Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }> {\n  const clampedPrompt = clampPrompt(prompt, config.recallPlannerMaxPromptChars);\n  const userParts = [`User message:\\n${clampedPrompt}`];\n  if (hints.length > 0) {\n    userParts.push(\n      `\\nRecent memory topics (for grounding only, do not treat as the message):\\n- ${hints.join(\"\\n- \")}`,\n    );\n  }\n  userParts.push('\\nRespond with JSON only: {\"mode\": \"...\", \"reason\": \"...\"}.');\n  return [\n    { role: \"system\", content: SYSTEM_PROMPT },\n    { role: \"user\", content: userParts.join(\"\\n\") },\n  ];\n}\n\n/**\n * Resolve the FallbackLlmClient routing options for the recall planner.\n *\n * - The dedicated `recallPlannerModel` is tried first (as the `model`\n *   override — it is prepended to the chain by FallbackLlmClient). If it does\n *   not resolve to a configured provider it is silently skipped, so a stale\n *   default never breaks routing.\n * - In gateway mode the shared `gatewayTaskChainOptions` (taskModelChain >\n *   gatewayAgentId, gotcha #22) is layered in as the fallback chain, plus the\n *   implicit gateway default appended by the client.\n * - In plugin mode only the explicit model + gateway providers apply.\n */\n/**\n * A `recallPlannerModel` value is only usable as a FallbackLlmClient `model`\n * override when it is provider-qualified (`provider/model`). The client's\n * `parseModelString` rejects bare names, so forwarding a bare value (e.g. the\n * legacy default `\"gpt-5.5\"`) would log \"invalid model format\" on every call\n * and never resolve. Bare values are dropped so routing falls through to the\n * gateway chain / agent / default instead (issue #1367 review on PR #1428).\n */\nfunction qualifiedPlannerModel(recallPlannerModel: string | undefined): string | undefined {\n  if (typeof recallPlannerModel !== \"string\") return undefined;\n  const trimmed = recallPlannerModel.trim();\n  return trimmed.includes(\"/\") ? trimmed : undefined;\n}\n\nexport function resolveRecallPlannerLlmOptions(\n  config: Pick<\n    PluginConfig,\n    \"modelSource\" | \"taskModelChain\" | \"gatewayAgentId\" | \"recallPlannerModel\" | \"recallPlannerTimeoutMs\"\n  >,\n): FallbackLlmOptions {\n  const chainOptions =\n    config.modelSource === \"gateway\" ? gatewayTaskChainOptions(config) : {};\n  return {\n    ...chainOptions,\n    model: qualifiedPlannerModel(config.recallPlannerModel),\n    temperature: 0,\n    maxTokens: 64,\n    timeoutMs:\n      typeof config.recallPlannerTimeoutMs === \"number\" && config.recallPlannerTimeoutMs > 0\n        ? config.recallPlannerTimeoutMs\n        : 1500,\n  };\n}\n\n// One-time warning per distinct routing signature so an opted-in operator with\n// no usable model learns why planning silently uses the heuristic, without\n// spamming a line on every recall.\nconst warnedNoRoutingSignatures = new Set<string>();\n\nfunction heuristicResult(\n  heuristicMode: RecallPlanMode,\n  source: RecallPlannerSource,\n  reason: string,\n  latencyMs: number,\n  fallbackUsed: boolean,\n): RecallPlannerLlmResult {\n  return { mode: heuristicMode, heuristicMode, source, reason, latencyMs, fallbackUsed };\n}\n\n/**\n * Plan the recall mode for `prompt`, optionally consulting an LLM.\n *\n * Always safe to call: returns the heuristic result when the LLM is disabled,\n * unavailable, or fails.\n *\n * @param llm  injectable client (tests pass a stub); constructed from gateway\n *             config when omitted.\n */\nexport async function planRecallModeLLM(\n  prompt: string,\n  hints: string[] | undefined,\n  config: PluginConfig,\n  llm?: FallbackLlmClient,\n  signal?: AbortSignal,\n): Promise<RecallPlannerLlmResult> {\n  const heuristicMode = planRecallMode(prompt);\n\n  if (!config.recallPlannerLlmEnabled) {\n    return heuristicResult(heuristicMode, \"heuristic\", \"llm-disabled\", 0, false);\n  }\n\n  // Participate in the recall cancellation contract: if the outer recall is\n  // already aborted (outer timeout / reset / session abort), don't start an LLM\n  // round-trip — fall back to the heuristic immediately (#1428 review).\n  if (signal?.aborted) {\n    return heuristicResult(heuristicMode, \"heuristic-fallback\", \"aborted\", 0, true);\n  }\n\n  const safePrompt = typeof prompt === \"string\" ? prompt.trim() : \"\";\n  if (safePrompt.length === 0) {\n    // Empty prompts never need an LLM round-trip.\n    return heuristicResult(heuristicMode, \"heuristic\", \"empty-prompt\", 0, false);\n  }\n\n  const client =\n    llm ??\n    new FallbackLlmClient(\n      config.gatewayConfig,\n      fallbackLlmRuntimeContextFromConfig(config),\n    );\n\n  // Forward the recall abort signal so an aborted/timed-out outer recall can\n  // cancel an in-flight planner call (FallbackLlmClient honors `signal`).\n  const options = { ...resolveRecallPlannerLlmOptions(config), signal };\n\n  // Availability check uses the same routing options so plugin-mode / empty\n  // chains short-circuit to the heuristic without a network attempt. `model`\n  // here is already provider-qualified (bare names were dropped), so a present\n  // model means the override is genuinely routable.\n  const availabilityProbe = {\n    agentId: options.agentId,\n    modelChain: options.modelChain,\n  };\n  if (!client.isAvailable(availabilityProbe) && !options.model) {\n    // Opted-in but nothing routable resolves (e.g. plugin mode with the bare\n    // default `recallPlannerModel` and no gateway chain). Warn once so it's not\n    // a silent no-op, then fall back to the heuristic.\n    const signature = `${config.modelSource}:${config.recallPlannerModel ?? \"\"}`;\n    if (!warnedNoRoutingSignatures.has(signature)) {\n      warnedNoRoutingSignatures.add(signature);\n      log.warn(\n        \"[recall-planner] recallPlannerLlmEnabled is on but no routable model resolves — \" +\n          \"set recallPlannerModel to a 'provider/model' value or configure a gateway model chain. \" +\n          \"Falling back to the heuristic planner.\",\n      );\n    }\n    return heuristicResult(heuristicMode, \"heuristic-fallback\", \"llm-no-model\", 0, true);\n  }\n\n  const clampedHints = clampHints(hints, config.recallPlannerMaxMemoryHints);\n  const messages = buildMessages(safePrompt, clampedHints, config);\n\n  const start = Date.now();\n  try {\n    const detailed = await client.parseWithSchemaDetailed(messages, PLANNER_SCHEMA, options);\n    const latencyMs = Date.now() - start;\n    if (!detailed?.result) {\n      // Distinguish failure from a valid empty (gotcha #34): a null here means\n      // no parseable classification, so fall back to the heuristic.\n      return heuristicResult(heuristicMode, \"heuristic-fallback\", \"llm-empty\", latencyMs, true);\n    }\n    const mode = detailed.result.mode;\n    const reason =\n      typeof detailed.result.reason === \"string\" && detailed.result.reason.trim().length > 0\n        ? detailed.result.reason.trim()\n        : \"llm-classified\";\n    return {\n      mode,\n      heuristicMode,\n      source: \"llm\",\n      reason,\n      modelUsed: detailed.modelUsed,\n      latencyMs,\n      fallbackUsed: false,\n    };\n  } catch (err) {\n    const latencyMs = Date.now() - start;\n    if (signal?.aborted) {\n      // Cancelled by the outer recall — expected, not an error worth warning on.\n      return heuristicResult(heuristicMode, \"heuristic-fallback\", \"aborted\", latencyMs, true);\n    }\n    const message = err instanceof Error ? err.message : String(err);\n    log.warn(`[recall-planner] LLM failed, falling back to heuristic: ${message}`);\n    return heuristicResult(\n      heuristicMode,\n      \"heuristic-fallback\",\n      `llm-error:${message}`,\n      latencyMs,\n      true,\n    );\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;AAAA,SAAS,SAAS;AAqDlB,IAAM,iBAAiB,EAAE,OAAO;AAAA;AAAA,EAE9B,MAAM,EAAE,KAAK,CAAC,aAAa,WAAW,QAAQ,YAAY,CAAC;AAAA,EAC3D,QAAQ,EAAE,OAAO,EAAE,IAAI,GAAG,EAAE,SAAS,EAAE,SAAS;AAClD,CAAC;AAED,IAAM,gBAAgB;AAAA,EACpB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,EAAE,KAAK,IAAI;AAGX,SAAS,YAAY,QAAgB,UAA0B;AAC7D,QAAM,UAAU,OAAO,SAAS,QAAQ,KAAK,WAAW,IAAI,KAAK,MAAM,QAAQ,IAAI;AACnF,MAAI,OAAO,UAAU,QAAS,QAAO;AACrC,SAAO,OAAO,MAAM,GAAG,OAAO;AAChC;AAGA,SAAS,WAAW,OAA6B,UAA4B;AAC3E,MAAI,CAAC,MAAM,QAAQ,KAAK,KAAK,MAAM,WAAW,EAAG,QAAO,CAAC;AACzD,QAAM,UAAU,OAAO,SAAS,QAAQ,KAAK,WAAW,IAAI,KAAK,MAAM,QAAQ,IAAI;AACnF,MAAI,WAAW,EAAG,QAAO,CAAC;AAC1B,QAAM,UAAoB,CAAC;AAC3B,aAAW,QAAQ,OAAO;AACxB,QAAI,OAAO,SAAS,SAAU;AAC9B,UAAM,UAAU,KAAK,KAAK;AAC1B,QAAI,QAAQ,WAAW,EAAG;AAC1B,YAAQ,KAAK,OAAO;AACpB,QAAI,QAAQ,UAAU,QAAS;AAAA,EACjC;AACA,SAAO;AACT;AAEA,SAAS,cACP,QACA,OACA,QACmE;AACnE,QAAM,gBAAgB,YAAY,QAAQ,OAAO,2BAA2B;AAC5E,QAAM,YAAY,CAAC;AAAA,EAAkB,aAAa,EAAE;AACpD,MAAI,MAAM,SAAS,GAAG;AACpB,cAAU;AAAA,MACR;AAAA;AAAA,IAAgF,MAAM,KAAK,MAAM,CAAC;AAAA,IACpG;AAAA,EACF;AACA,YAAU,KAAK,6DAA6D;AAC5E,SAAO;AAAA,IACL,EAAE,MAAM,UAAU,SAAS,cAAc;AAAA,IACzC,EAAE,MAAM,QAAQ,SAAS,UAAU,KAAK,IAAI,EAAE;AAAA,EAChD;AACF;AAsBA,SAAS,sBAAsB,oBAA4D;AACzF,MAAI,OAAO,uBAAuB,SAAU,QAAO;AACnD,QAAM,UAAU,mBAAmB,KAAK;AACxC,SAAO,QAAQ,SAAS,GAAG,IAAI,UAAU;AAC3C;AAEO,SAAS,+BACd,QAIoB;AACpB,QAAM,eACJ,OAAO,gBAAgB,YAAY,wBAAwB,MAAM,IAAI,CAAC;AACxE,SAAO;AAAA,IACL,GAAG;AAAA,IACH,OAAO,sBAAsB,OAAO,kBAAkB;AAAA,IACtD,aAAa;AAAA,IACb,WAAW;AAAA,IACX,WACE,OAAO,OAAO,2BAA2B,YAAY,OAAO,yBAAyB,IACjF,OAAO,yBACP;AAAA,EACR;AACF;AAKA,IAAM,4BAA4B,oBAAI,IAAY;AAElD,SAAS,gBACP,eACA,QACA,QACA,WACA,cACwB;AACxB,SAAO,EAAE,MAAM,eAAe,eAAe,QAAQ,QAAQ,WAAW,aAAa;AACvF;AAWA,eAAsB,kBACpB,QACA,OACA,QACA,KACA,QACiC;AACjC,QAAM,gBAAgB,eAAe,MAAM;AAE3C,MAAI,CAAC,OAAO,yBAAyB;AACnC,WAAO,gBAAgB,eAAe,aAAa,gBAAgB,GAAG,KAAK;AAAA,EAC7E;AAKA,MAAI,QAAQ,SAAS;AACnB,WAAO,gBAAgB,eAAe,sBAAsB,WAAW,GAAG,IAAI;AAAA,EAChF;AAEA,QAAM,aAAa,OAAO,WAAW,WAAW,OAAO,KAAK,IAAI;AAChE,MAAI,WAAW,WAAW,GAAG;AAE3B,WAAO,gBAAgB,eAAe,aAAa,gBAAgB,GAAG,KAAK;AAAA,EAC7E;AAEA,QAAM,SACJ,OACA,IAAI;AAAA,IACF,OAAO;AAAA,IACP,oCAAoC,MAAM;AAAA,EAC5C;AAIF,QAAM,UAAU,EAAE,GAAG,+BAA+B,MAAM,GAAG,OAAO;AAMpE,QAAM,oBAAoB;AAAA,IACxB,SAAS,QAAQ;AAAA,IACjB,YAAY,QAAQ;AAAA,EACtB;AACA,MAAI,CAAC,OAAO,YAAY,iBAAiB,KAAK,CAAC,QAAQ,OAAO;AAI5D,UAAM,YAAY,GAAG,OAAO,WAAW,IAAI,OAAO,sBAAsB,EAAE;AAC1E,QAAI,CAAC,0BAA0B,IAAI,SAAS,GAAG;AAC7C,gCAA0B,IAAI,SAAS;AACvC,UAAI;AAAA,QACF;AAAA,MAGF;AAAA,IACF;AACA,WAAO,gBAAgB,eAAe,sBAAsB,gBAAgB,GAAG,IAAI;AAAA,EACrF;AAEA,QAAM,eAAe,WAAW,OAAO,OAAO,2BAA2B;AACzE,QAAM,WAAW,cAAc,YAAY,cAAc,MAAM;AAE/D,QAAM,QAAQ,KAAK,IAAI;AACvB,MAAI;AACF,UAAM,WAAW,MAAM,OAAO,wBAAwB,UAAU,gBAAgB,OAAO;AACvF,UAAM,YAAY,KAAK,IAAI,IAAI;AAC/B,QAAI,CAAC,UAAU,QAAQ;AAGrB,aAAO,gBAAgB,eAAe,sBAAsB,aAAa,WAAW,IAAI;AAAA,IAC1F;AACA,UAAM,OAAO,SAAS,OAAO;AAC7B,UAAM,SACJ,OAAO,SAAS,OAAO,WAAW,YAAY,SAAS,OAAO,OAAO,KAAK,EAAE,SAAS,IACjF,SAAS,OAAO,OAAO,KAAK,IAC5B;AACN,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA,QAAQ;AAAA,MACR;AAAA,MACA,WAAW,SAAS;AAAA,MACpB;AAAA,MACA,cAAc;AAAA,IAChB;AAAA,EACF,SAAS,KAAK;AACZ,UAAM,YAAY,KAAK,IAAI,IAAI;AAC/B,QAAI,QAAQ,SAAS;AAEnB,aAAO,gBAAgB,eAAe,sBAAsB,WAAW,WAAW,IAAI;AAAA,IACxF;AACA,UAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,QAAI,KAAK,2DAA2D,OAAO,EAAE;AAC7E,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA,aAAa,OAAO;AAAA,MACpB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;","names":[]}
+{"version":3,"sources":["../src/recall-planner-llm.ts"],"sourcesContent":["import { z } from \"zod\";\n\nimport type { PluginConfig, RecallPlanMode } from \"./types.js\";\nimport type { CapabilitySet } from \"./capabilities.js\";\nimport { planRecallMode } from \"./intent.js\";\nimport {\n  FallbackLlmClient,\n  fallbackLlmRuntimeContextFromConfig,\n  gatewayTaskChainOptions,\n  type FallbackLlmOptions,\n} from \"./fallback-llm.js\";\nimport { log } from \"./logger.js\";\n\n/**\n * LLM-based recall planning (issue #1367, Option C).\n *\n * Classifies an incoming prompt into a {@link RecallPlanMode} using an LLM\n * instead of (or alongside) the regex heuristic in {@link planRecallMode}.\n *\n * Provider-agnostic by construction: it routes through {@link FallbackLlmClient},\n * which resolves the model chain from gateway providers (OpenAI, Anthropic,\n * Ollama, Codex, …) or gateway agent personas / `taskModelChain`. Nothing here\n * is hard-coded to a single provider or to OpenAI's Responses API — the API\n * dialect is chosen per-provider by the client based on each provider's `api`\n * field. The configured `recallPlannerModel` is tried first, with the broader\n * task chain (and the gateway default) as resilient fallbacks.\n *\n * Invariants:\n *  - Never throws to the caller (gotcha #13). Any LLM failure, timeout, empty\n *    response, or unavailable backend falls back to the heuristic result and\n *    sets `fallbackUsed: true` (gotcha #34 — failures are distinct from a valid\n *    classification).\n *  - When `recallPlannerLlmEnabled` is false the LLM is never contacted.\n */\n\nexport type RecallPlannerSource = \"llm\" | \"heuristic\" | \"heuristic-fallback\";\n\nexport interface RecallPlannerLlmResult {\n  /** The mode to act on. */\n  mode: RecallPlanMode;\n  /** The heuristic mode, always computed (the fallback floor / shadow baseline). */\n  heuristicMode: RecallPlanMode;\n  /** Where `mode` came from. */\n  source: RecallPlannerSource;\n  /** Short human-readable rationale (LLM reason, or why we fell back). */\n  reason: string;\n  /** Model that actually served the classification, when an LLM was used. */\n  modelUsed?: string;\n  /** Wall-clock spent in the LLM call (0 when no call was made). */\n  latencyMs: number;\n  /** True when the LLM was enabled but we had to fall back to the heuristic. */\n  fallbackUsed: boolean;\n}\n\nconst PLANNER_SCHEMA = z.object({\n  // gotcha #2: optional fields use .optional().nullable()\n  mode: z.enum([\"no_recall\", \"minimal\", \"full\", \"graph_mode\"]),\n  reason: z.string().max(280).optional().nullable(),\n});\n\nconst SYSTEM_PROMPT = [\n  \"You are a recall-planning classifier for a long-term memory system.\",\n  \"Given the user's latest message, decide how much stored memory should be retrieved before the assistant responds.\",\n  \"Reply with a single JSON object: {\\\"mode\\\": <one of no_recall|minimal|full|graph_mode>, \\\"reason\\\": <short string>}.\",\n  \"\",\n  \"Modes:\",\n  '- \"no_recall\": low-information acknowledgements or chit-chat with nothing to look up (e.g. \"ok\", \"thanks\", \"sounds good\"). Retrieve nothing.',\n  '- \"minimal\": short, self-contained operational directives that rarely need history (e.g. \"restart the service\", \"run the tests\", \"show status\"). Retrieve a little.',\n  '- \"full\": anything memory-seeking, analytical, or a question that benefits from prior context, decisions, or facts. This is the safe default when unsure.',\n  '- \"graph_mode\": queries about timelines, sequences, history, causal chains, or root cause (\"how did we get here\", \"what led to this regression\"). Retrieve relationship/graph context.',\n  \"\",\n  \"When uncertain, prefer \\\"full\\\" over dropping recall. Never invent facts; only classify intent.\",\n].join(\"\\n\");\n\n/** Clamp a planner prompt to the configured character budget. */\nfunction clampPrompt(prompt: string, maxChars: number): string {\n  const safeMax = Number.isFinite(maxChars) && maxChars > 0 ? Math.floor(maxChars) : 4000;\n  if (prompt.length <= safeMax) return prompt;\n  return prompt.slice(0, safeMax);\n}\n\n/** Trim and cap the optional memory hints used to ground the classification. */\nfunction clampHints(hints: string[] | undefined, maxHints: number): string[] {\n  if (!Array.isArray(hints) || hints.length === 0) return [];\n  const safeMax = Number.isFinite(maxHints) && maxHints > 0 ? Math.floor(maxHints) : 0;\n  if (safeMax <= 0) return [];\n  const cleaned: string[] = [];\n  for (const hint of hints) {\n    if (typeof hint !== \"string\") continue;\n    const trimmed = hint.trim();\n    if (trimmed.length === 0) continue;\n    cleaned.push(trimmed);\n    if (cleaned.length >= safeMax) break;\n  }\n  return cleaned;\n}\n\nfunction buildMessages(\n  prompt: string,\n  hints: string[],\n  config: PluginConfig,\n): Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }> {\n  const clampedPrompt = clampPrompt(prompt, config.recallPlannerMaxPromptChars);\n  const userParts = [`User message:\\n${clampedPrompt}`];\n  if (hints.length > 0) {\n    userParts.push(\n      `\\nRecent memory topics (for grounding only, do not treat as the message):\\n- ${hints.join(\"\\n- \")}`,\n    );\n  }\n  userParts.push('\\nRespond with JSON only: {\"mode\": \"...\", \"reason\": \"...\"}.');\n  return [\n    { role: \"system\", content: SYSTEM_PROMPT },\n    { role: \"user\", content: userParts.join(\"\\n\") },\n  ];\n}\n\n/**\n * Resolve the FallbackLlmClient routing options for the recall planner.\n *\n * - The dedicated `recallPlannerModel` is tried first (as the `model`\n *   override — it is prepended to the chain by FallbackLlmClient). If it does\n *   not resolve to a configured provider it is silently skipped, so a stale\n *   default never breaks routing.\n * - In gateway mode the shared `gatewayTaskChainOptions` (taskModelChain >\n *   gatewayAgentId, gotcha #22) is layered in as the fallback chain, plus the\n *   implicit gateway default appended by the client.\n * - In plugin mode only the explicit model + gateway providers apply.\n */\n/**\n * A `recallPlannerModel` value is only usable as a FallbackLlmClient `model`\n * override when it is provider-qualified (`provider/model`). The client's\n * `parseModelString` rejects bare names, so forwarding a bare value (e.g. the\n * legacy default `\"gpt-5.5\"`) would log \"invalid model format\" on every call\n * and never resolve. Bare values are dropped so routing falls through to the\n * gateway chain / agent / default instead (issue #1367 review on PR #1428).\n */\nfunction qualifiedPlannerModel(recallPlannerModel: string | undefined): string | undefined {\n  if (typeof recallPlannerModel !== \"string\") return undefined;\n  const trimmed = recallPlannerModel.trim();\n  return trimmed.includes(\"/\") ? trimmed : undefined;\n}\n\nexport function resolveRecallPlannerLlmOptions(\n  config: Pick<\n    PluginConfig,\n    \"modelSource\" | \"taskModelChain\" | \"gatewayAgentId\" | \"recallPlannerModel\" | \"recallPlannerTimeoutMs\"\n  >,\n): FallbackLlmOptions {\n  const chainOptions =\n    config.modelSource === \"gateway\" ? gatewayTaskChainOptions(config) : {};\n  return {\n    ...chainOptions,\n    model: qualifiedPlannerModel(config.recallPlannerModel),\n    temperature: 0,\n    maxTokens: 64,\n    timeoutMs:\n      typeof config.recallPlannerTimeoutMs === \"number\" && config.recallPlannerTimeoutMs > 0\n        ? config.recallPlannerTimeoutMs\n        : 1500,\n  };\n}\n\n// One-time warning per distinct routing signature so an opted-in operator with\n// no usable model learns why planning silently uses the heuristic, without\n// spamming a line on every recall.\nconst warnedNoRoutingSignatures = new Set<string>();\n\nfunction heuristicResult(\n  heuristicMode: RecallPlanMode,\n  source: RecallPlannerSource,\n  reason: string,\n  latencyMs: number,\n  fallbackUsed: boolean,\n): RecallPlannerLlmResult {\n  return { mode: heuristicMode, heuristicMode, source, reason, latencyMs, fallbackUsed };\n}\n\n/**\n * Plan the recall mode for `prompt`, optionally consulting an LLM.\n *\n * Always safe to call: returns the heuristic result when the LLM is disabled,\n * unavailable, or fails.\n *\n * @param llm  injectable client (tests pass a stub); constructed from gateway\n *             config when omitted.\n */\nexport async function planRecallModeLLM(\n  prompt: string,\n  hints: string[] | undefined,\n  config: PluginConfig,\n  llm?: FallbackLlmClient,\n  signal?: AbortSignal,\n  caps?: CapabilitySet,\n): Promise<RecallPlannerLlmResult> {\n  const heuristicMode = planRecallMode(prompt);\n\n  // `caps` is OPTIONAL and additive (issue #1523). Prefer the resolved\n  // capability when supplied; fall back to the config flag so existing callers\n  // that pass only `config` keep identical gating.\n  const plannerLlmEnabled = caps?.recallPlannerLlm ?? config.recallPlannerLlmEnabled;\n  if (!plannerLlmEnabled) {\n    return heuristicResult(heuristicMode, \"heuristic\", \"llm-disabled\", 0, false);\n  }\n\n  // Participate in the recall cancellation contract: if the outer recall is\n  // already aborted (outer timeout / reset / session abort), don't start an LLM\n  // round-trip — fall back to the heuristic immediately (#1428 review).\n  if (signal?.aborted) {\n    return heuristicResult(heuristicMode, \"heuristic-fallback\", \"aborted\", 0, true);\n  }\n\n  const safePrompt = typeof prompt === \"string\" ? prompt.trim() : \"\";\n  if (safePrompt.length === 0) {\n    // Empty prompts never need an LLM round-trip.\n    return heuristicResult(heuristicMode, \"heuristic\", \"empty-prompt\", 0, false);\n  }\n\n  const client =\n    llm ??\n    new FallbackLlmClient(\n      config.gatewayConfig,\n      fallbackLlmRuntimeContextFromConfig(config),\n    );\n\n  // Forward the recall abort signal so an aborted/timed-out outer recall can\n  // cancel an in-flight planner call (FallbackLlmClient honors `signal`).\n  const options = { ...resolveRecallPlannerLlmOptions(config), signal };\n\n  // Availability check uses the same routing options so plugin-mode / empty\n  // chains short-circuit to the heuristic without a network attempt. `model`\n  // here is already provider-qualified (bare names were dropped), so a present\n  // model means the override is genuinely routable.\n  const availabilityProbe = {\n    agentId: options.agentId,\n    modelChain: options.modelChain,\n  };\n  if (!client.isAvailable(availabilityProbe) && !options.model) {\n    // Opted-in but nothing routable resolves (e.g. plugin mode with the bare\n    // default `recallPlannerModel` and no gateway chain). Warn once so it's not\n    // a silent no-op, then fall back to the heuristic.\n    const signature = `${config.modelSource}:${config.recallPlannerModel ?? \"\"}`;\n    if (!warnedNoRoutingSignatures.has(signature)) {\n      warnedNoRoutingSignatures.add(signature);\n      log.warn(\n        \"[recall-planner] recallPlannerLlmEnabled is on but no routable model resolves — \" +\n          \"set recallPlannerModel to a 'provider/model' value or configure a gateway model chain. \" +\n          \"Falling back to the heuristic planner.\",\n      );\n    }\n    return heuristicResult(heuristicMode, \"heuristic-fallback\", \"llm-no-model\", 0, true);\n  }\n\n  const clampedHints = clampHints(hints, config.recallPlannerMaxMemoryHints);\n  const messages = buildMessages(safePrompt, clampedHints, config);\n\n  const start = Date.now();\n  try {\n    const detailed = await client.parseWithSchemaDetailed(messages, PLANNER_SCHEMA, options);\n    const latencyMs = Date.now() - start;\n    if (!detailed?.result) {\n      // Distinguish failure from a valid empty (gotcha #34): a null here means\n      // no parseable classification, so fall back to the heuristic.\n      return heuristicResult(heuristicMode, \"heuristic-fallback\", \"llm-empty\", latencyMs, true);\n    }\n    const mode = detailed.result.mode;\n    const reason =\n      typeof detailed.result.reason === \"string\" && detailed.result.reason.trim().length > 0\n        ? detailed.result.reason.trim()\n        : \"llm-classified\";\n    return {\n      mode,\n      heuristicMode,\n      source: \"llm\",\n      reason,\n      modelUsed: detailed.modelUsed,\n      latencyMs,\n      fallbackUsed: false,\n    };\n  } catch (err) {\n    const latencyMs = Date.now() - start;\n    if (signal?.aborted) {\n      // Cancelled by the outer recall — expected, not an error worth warning on.\n      return heuristicResult(heuristicMode, \"heuristic-fallback\", \"aborted\", latencyMs, true);\n    }\n    const message = err instanceof Error ? err.message : String(err);\n    log.warn(`[recall-planner] LLM failed, falling back to heuristic: ${message}`);\n    return heuristicResult(\n      heuristicMode,\n      \"heuristic-fallback\",\n      `llm-error:${message}`,\n      latencyMs,\n      true,\n    );\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;AAAA,SAAS,SAAS;AAsDlB,IAAM,iBAAiB,EAAE,OAAO;AAAA;AAAA,EAE9B,MAAM,EAAE,KAAK,CAAC,aAAa,WAAW,QAAQ,YAAY,CAAC;AAAA,EAC3D,QAAQ,EAAE,OAAO,EAAE,IAAI,GAAG,EAAE,SAAS,EAAE,SAAS;AAClD,CAAC;AAED,IAAM,gBAAgB;AAAA,EACpB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,EAAE,KAAK,IAAI;AAGX,SAAS,YAAY,QAAgB,UAA0B;AAC7D,QAAM,UAAU,OAAO,SAAS,QAAQ,KAAK,WAAW,IAAI,KAAK,MAAM,QAAQ,IAAI;AACnF,MAAI,OAAO,UAAU,QAAS,QAAO;AACrC,SAAO,OAAO,MAAM,GAAG,OAAO;AAChC;AAGA,SAAS,WAAW,OAA6B,UAA4B;AAC3E,MAAI,CAAC,MAAM,QAAQ,KAAK,KAAK,MAAM,WAAW,EAAG,QAAO,CAAC;AACzD,QAAM,UAAU,OAAO,SAAS,QAAQ,KAAK,WAAW,IAAI,KAAK,MAAM,QAAQ,IAAI;AACnF,MAAI,WAAW,EAAG,QAAO,CAAC;AAC1B,QAAM,UAAoB,CAAC;AAC3B,aAAW,QAAQ,OAAO;AACxB,QAAI,OAAO,SAAS,SAAU;AAC9B,UAAM,UAAU,KAAK,KAAK;AAC1B,QAAI,QAAQ,WAAW,EAAG;AAC1B,YAAQ,KAAK,OAAO;AACpB,QAAI,QAAQ,UAAU,QAAS;AAAA,EACjC;AACA,SAAO;AACT;AAEA,SAAS,cACP,QACA,OACA,QACmE;AACnE,QAAM,gBAAgB,YAAY,QAAQ,OAAO,2BAA2B;AAC5E,QAAM,YAAY,CAAC;AAAA,EAAkB,aAAa,EAAE;AACpD,MAAI,MAAM,SAAS,GAAG;AACpB,cAAU;AAAA,MACR;AAAA;AAAA,IAAgF,MAAM,KAAK,MAAM,CAAC;AAAA,IACpG;AAAA,EACF;AACA,YAAU,KAAK,6DAA6D;AAC5E,SAAO;AAAA,IACL,EAAE,MAAM,UAAU,SAAS,cAAc;AAAA,IACzC,EAAE,MAAM,QAAQ,SAAS,UAAU,KAAK,IAAI,EAAE;AAAA,EAChD;AACF;AAsBA,SAAS,sBAAsB,oBAA4D;AACzF,MAAI,OAAO,uBAAuB,SAAU,QAAO;AACnD,QAAM,UAAU,mBAAmB,KAAK;AACxC,SAAO,QAAQ,SAAS,GAAG,IAAI,UAAU;AAC3C;AAEO,SAAS,+BACd,QAIoB;AACpB,QAAM,eACJ,OAAO,gBAAgB,YAAY,wBAAwB,MAAM,IAAI,CAAC;AACxE,SAAO;AAAA,IACL,GAAG;AAAA,IACH,OAAO,sBAAsB,OAAO,kBAAkB;AAAA,IACtD,aAAa;AAAA,IACb,WAAW;AAAA,IACX,WACE,OAAO,OAAO,2BAA2B,YAAY,OAAO,yBAAyB,IACjF,OAAO,yBACP;AAAA,EACR;AACF;AAKA,IAAM,4BAA4B,oBAAI,IAAY;AAElD,SAAS,gBACP,eACA,QACA,QACA,WACA,cACwB;AACxB,SAAO,EAAE,MAAM,eAAe,eAAe,QAAQ,QAAQ,WAAW,aAAa;AACvF;AAWA,eAAsB,kBACpB,QACA,OACA,QACA,KACA,QACA,MACiC;AACjC,QAAM,gBAAgB,eAAe,MAAM;AAK3C,QAAM,oBAAoB,MAAM,oBAAoB,OAAO;AAC3D,MAAI,CAAC,mBAAmB;AACtB,WAAO,gBAAgB,eAAe,aAAa,gBAAgB,GAAG,KAAK;AAAA,EAC7E;AAKA,MAAI,QAAQ,SAAS;AACnB,WAAO,gBAAgB,eAAe,sBAAsB,WAAW,GAAG,IAAI;AAAA,EAChF;AAEA,QAAM,aAAa,OAAO,WAAW,WAAW,OAAO,KAAK,IAAI;AAChE,MAAI,WAAW,WAAW,GAAG;AAE3B,WAAO,gBAAgB,eAAe,aAAa,gBAAgB,GAAG,KAAK;AAAA,EAC7E;AAEA,QAAM,SACJ,OACA,IAAI;AAAA,IACF,OAAO;AAAA,IACP,oCAAoC,MAAM;AAAA,EAC5C;AAIF,QAAM,UAAU,EAAE,GAAG,+BAA+B,MAAM,GAAG,OAAO;AAMpE,QAAM,oBAAoB;AAAA,IACxB,SAAS,QAAQ;AAAA,IACjB,YAAY,QAAQ;AAAA,EACtB;AACA,MAAI,CAAC,OAAO,YAAY,iBAAiB,KAAK,CAAC,QAAQ,OAAO;AAI5D,UAAM,YAAY,GAAG,OAAO,WAAW,IAAI,OAAO,sBAAsB,EAAE;AAC1E,QAAI,CAAC,0BAA0B,IAAI,SAAS,GAAG;AAC7C,gCAA0B,IAAI,SAAS;AACvC,UAAI;AAAA,QACF;AAAA,MAGF;AAAA,IACF;AACA,WAAO,gBAAgB,eAAe,sBAAsB,gBAAgB,GAAG,IAAI;AAAA,EACrF;AAEA,QAAM,eAAe,WAAW,OAAO,OAAO,2BAA2B;AACzE,QAAM,WAAW,cAAc,YAAY,cAAc,MAAM;AAE/D,QAAM,QAAQ,KAAK,IAAI;AACvB,MAAI;AACF,UAAM,WAAW,MAAM,OAAO,wBAAwB,UAAU,gBAAgB,OAAO;AACvF,UAAM,YAAY,KAAK,IAAI,IAAI;AAC/B,QAAI,CAAC,UAAU,QAAQ;AAGrB,aAAO,gBAAgB,eAAe,sBAAsB,aAAa,WAAW,IAAI;AAAA,IAC1F;AACA,UAAM,OAAO,SAAS,OAAO;AAC7B,UAAM,SACJ,OAAO,SAAS,OAAO,WAAW,YAAY,SAAS,OAAO,OAAO,KAAK,EAAE,SAAS,IACjF,SAAS,OAAO,OAAO,KAAK,IAC5B;AACN,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA,QAAQ;AAAA,MACR;AAAA,MACA,WAAW,SAAS;AAAA,MACpB;AAAA,MACA,cAAc;AAAA,IAChB;AAAA,EACF,SAAS,KAAK;AACZ,UAAM,YAAY,KAAK,IAAI,IAAI;AAC/B,QAAI,QAAQ,SAAS;AAEnB,aAAO,gBAAgB,eAAe,sBAAsB,WAAW,WAAW,IAAI;AAAA,IACxF;AACA,UAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,QAAI,KAAK,2DAA2D,OAAO,EAAE;AAC7E,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA,aAAa,OAAO;AAAA,MACpB;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;","names":[]}

package/dist/search/factory.js

@@ -2,24 +2,24 @@ import {
   createConversationIndexRuntime,
   createConversationSearchBackend,
   createSearchBackend
-} from "../chunk-ROHLEUTH.js";
+} from "../chunk-23EBQ27U.js";
+import "../chunk-OUWAQVDJ.js";
 import "../chunk-DOCTITOP.js";
 import "../chunk-CYEPCZN5.js";
 import "../chunk-Q5ZU3RNY.js";
 import "../chunk-JOASJWQR.js";
-import "../chunk-RN7MUWON.js";
-import "../chunk-OUWAQVDJ.js";
 import "../chunk-AER6MT24.js";
+import "../chunk-RN7MUWON.js";
 import "../chunk-CINZGPSJ.js";
-import "../chunk-5GPPACXK.js";
-import "../chunk-7OGJQP7T.js";
+import "../chunk-CRO4LCQ6.js";
+import "../chunk-7DTASS5T.js";
 import "../chunk-E6ZDCOHM.js";
 import "../chunk-OIF36KGD.js";
-import "../chunk-7DTASS5T.js";
 import "../chunk-6RHNCKHG.js";
 import "../chunk-YNQ6DFSV.js";
 import "../chunk-O75CRYGF.js";
 import "../chunk-3ONXXHQO.js";
+import "../chunk-5GPPACXK.js";
 import "../chunk-AWJ2FHCF.js";
 import "../chunk-42JKGUFJ.js";
 import "../chunk-JUC24CTX.js";

package/dist/search/index.js

@@ -2,7 +2,10 @@ import {
   createConversationIndexRuntime,
   createConversationSearchBackend,
   createSearchBackend
-} from "../chunk-ROHLEUTH.js";
+} from "../chunk-23EBQ27U.js";
+import {
+  LanceDbBackend
+} from "../chunk-OUWAQVDJ.js";
 import {
   MeilisearchBackend
 } from "../chunk-DOCTITOP.js";
@@ -15,25 +18,22 @@ import {
 import {
   RemoteSearchBackend
 } from "../chunk-JOASJWQR.js";
-import {
-  EmbedHelper
-} from "../chunk-RN7MUWON.js";
-import {
-  LanceDbBackend
-} from "../chunk-OUWAQVDJ.js";
 import {
   scanMemoryDir
 } from "../chunk-AER6MT24.js";
+import {
+  EmbedHelper
+} from "../chunk-RN7MUWON.js";
 import "../chunk-CINZGPSJ.js";
-import "../chunk-5GPPACXK.js";
-import "../chunk-7OGJQP7T.js";
+import "../chunk-CRO4LCQ6.js";
+import "../chunk-7DTASS5T.js";
 import "../chunk-E6ZDCOHM.js";
 import "../chunk-OIF36KGD.js";
-import "../chunk-7DTASS5T.js";
 import "../chunk-6RHNCKHG.js";
 import "../chunk-YNQ6DFSV.js";
 import "../chunk-O75CRYGF.js";
 import "../chunk-3ONXXHQO.js";
+import "../chunk-5GPPACXK.js";
 import "../chunk-AWJ2FHCF.js";
 import "../chunk-42JKGUFJ.js";
 import "../chunk-JUC24CTX.js";

package/dist/search/lancedb-backend.js

@@ -3,8 +3,8 @@ import {
 } from "../chunk-OUWAQVDJ.js";
 import "../chunk-AER6MT24.js";
 import "../chunk-CINZGPSJ.js";
-import "../chunk-5GPPACXK.js";
 import "../chunk-YNQ6DFSV.js";
+import "../chunk-5GPPACXK.js";
 import "../chunk-2ODBA7MQ.js";
 import "../chunk-PVGDJXVK.js";
 import "../chunk-VS2IYZRU.js";

package/dist/search/meilisearch-backend.js

@@ -3,8 +3,8 @@ import {
 } from "../chunk-DOCTITOP.js";
 import "../chunk-AER6MT24.js";
 import "../chunk-CINZGPSJ.js";
-import "../chunk-5GPPACXK.js";
 import "../chunk-YNQ6DFSV.js";
+import "../chunk-5GPPACXK.js";
 import "../chunk-2ODBA7MQ.js";
 import "../chunk-PVGDJXVK.js";
 import "../chunk-VS2IYZRU.js";

package/dist/search/orama-backend.js

@@ -4,8 +4,8 @@ import {
 } from "../chunk-Q5ZU3RNY.js";
 import "../chunk-AER6MT24.js";
 import "../chunk-CINZGPSJ.js";
-import "../chunk-5GPPACXK.js";
 import "../chunk-YNQ6DFSV.js";
+import "../chunk-5GPPACXK.js";
 import "../chunk-2ODBA7MQ.js";
 import "../chunk-PVGDJXVK.js";
 import "../chunk-VS2IYZRU.js";

package/dist/semantic-consolidation.js

@@ -7,16 +7,17 @@ import {
   materializeAfterSemanticConsolidation,
   parseConsolidationResponse,
   parseOperatorAwareConsolidationResponse
-} from "./chunk-XW3W4PV4.js";
-import "./chunk-C7AF236A.js";
+} from "./chunk-JTPXSXHC.js";
+import "./chunk-AARDBQTA.js";
 import "./chunk-LN4YGHTM.js";
 import {
   resolveExtensionsRoot
 } from "./chunk-JLNBQWZ2.js";
 import "./chunk-3UXOZBHV.js";
-import "./chunk-VH6EIKVS.js";
+import "./chunk-WKMCC4NQ.js";
 import "./chunk-M7XQSUBB.js";
 import "./chunk-5UZXUTVO.js";
+import "./chunk-5GPPACXK.js";
 import "./chunk-J6A3CX5N.js";
 import "./chunk-AZBV4RRY.js";
 import "./chunk-AWJ2FHCF.js";
@@ -28,13 +29,13 @@ import "./chunk-6KYMPV2O.js";
 import "./chunk-DM2T26WE.js";
 import "./chunk-LDXUBPMO.js";
 import "./chunk-FVQJYWH7.js";
+import "./chunk-VF4XKTX3.js";
 import {
   CONSOLIDATION_OPERATORS,
   isConsolidationOperator,
   isSemanticConsolidationLlmOperator,
   isValidDerivedFromEntry
 } from "./chunk-G7D6GZ5J.js";
-import "./chunk-VF4XKTX3.js";
 import "./chunk-U3PN77QT.js";
 import "./chunk-4DJQYKMN.js";
 import "./chunk-EYIEWJNI.js";

package/dist/semantic-rule-promotion.js

@@ -1,10 +1,11 @@
 import {
   promoteSemanticRuleFromMemory,
   setSemanticRulePromotionTestHooks
-} from "./chunk-IHG6CC7T.js";
-import "./chunk-VH6EIKVS.js";
+} from "./chunk-BQJUPECT.js";
+import "./chunk-WKMCC4NQ.js";
 import "./chunk-M7XQSUBB.js";
 import "./chunk-5UZXUTVO.js";
+import "./chunk-5GPPACXK.js";
 import "./chunk-J6A3CX5N.js";
 import "./chunk-AZBV4RRY.js";
 import "./chunk-AWJ2FHCF.js";
@@ -16,8 +17,8 @@ import "./chunk-6KYMPV2O.js";
 import "./chunk-DM2T26WE.js";
 import "./chunk-LDXUBPMO.js";
 import "./chunk-FVQJYWH7.js";
-import "./chunk-G7D6GZ5J.js";
 import "./chunk-VF4XKTX3.js";
+import "./chunk-G7D6GZ5J.js";
 import "./chunk-4DJQYKMN.js";
 import "./chunk-2ODBA7MQ.js";
 import "./chunk-J2HSAU72.js";

package/dist/semantic-rule-verifier.js

@@ -1,10 +1,11 @@
 import {
   compareVerifiedSemanticRuleResults,
   searchVerifiedSemanticRules
-} from "./chunk-DSLUOQDY.js";
-import "./chunk-VH6EIKVS.js";
+} from "./chunk-XMWF6AU3.js";
+import "./chunk-WKMCC4NQ.js";
 import "./chunk-M7XQSUBB.js";
 import "./chunk-5UZXUTVO.js";
+import "./chunk-5GPPACXK.js";
 import "./chunk-J6A3CX5N.js";
 import "./chunk-AZBV4RRY.js";
 import "./chunk-AWJ2FHCF.js";
@@ -16,8 +17,8 @@ import "./chunk-6KYMPV2O.js";
 import "./chunk-DM2T26WE.js";
 import "./chunk-LDXUBPMO.js";
 import "./chunk-FVQJYWH7.js";
-import "./chunk-G7D6GZ5J.js";
 import "./chunk-VF4XKTX3.js";
+import "./chunk-G7D6GZ5J.js";
 import "./chunk-4DJQYKMN.js";
 import "./chunk-ZBJMUXZH.js";
 import "./chunk-2ODBA7MQ.js";

package/dist/storage.js

@@ -9,9 +9,10 @@ import {
   parseEntityFile,
   serializeEntityFile,
   stripAttributesSuffix
-} from "./chunk-VH6EIKVS.js";
+} from "./chunk-WKMCC4NQ.js";
 import "./chunk-M7XQSUBB.js";
 import "./chunk-5UZXUTVO.js";
+import "./chunk-5GPPACXK.js";
 import "./chunk-J6A3CX5N.js";
 import "./chunk-AZBV4RRY.js";
 import "./chunk-AWJ2FHCF.js";
@@ -23,8 +24,8 @@ import "./chunk-6KYMPV2O.js";
 import "./chunk-DM2T26WE.js";
 import "./chunk-LDXUBPMO.js";
 import "./chunk-FVQJYWH7.js";
-import "./chunk-G7D6GZ5J.js";
 import "./chunk-VF4XKTX3.js";
+import "./chunk-G7D6GZ5J.js";
 import "./chunk-4DJQYKMN.js";
 import "./chunk-2ODBA7MQ.js";
 import "./chunk-J2HSAU72.js";

package/dist/transfer/backup.js

@@ -4,11 +4,11 @@ import {
 import "../chunk-765K3SAT.js";
 import "../chunk-X7Y7WX73.js";
 import "../chunk-J4EB7DNW.js";
+import "../chunk-UI3NYK34.js";
+import "../chunk-GCGJW34D.js";
 import "../chunk-BJMBJZ2Y.js";
 import "../chunk-UKJAGEXH.js";
 import "../chunk-FP2373TW.js";
-import "../chunk-UI3NYK34.js";
-import "../chunk-GCGJW34D.js";
 import "../chunk-A6XUJE5D.js";
 import "../chunk-PZ5AY32C.js";
 export {

package/dist/transfer/capsule-export.js

@@ -5,11 +5,11 @@ import {
 import "../chunk-WEHSQBFR.js";
 import "../chunk-X7Y7WX73.js";
 import "../chunk-J4EB7DNW.js";
+import "../chunk-UI3NYK34.js";
+import "../chunk-GCGJW34D.js";
 import "../chunk-BJMBJZ2Y.js";
 import "../chunk-UKJAGEXH.js";
 import "../chunk-FP2373TW.js";
-import "../chunk-UI3NYK34.js";
-import "../chunk-GCGJW34D.js";
 import "../chunk-A6XUJE5D.js";
 import "../chunk-PZ5AY32C.js";
 export {

package/dist/transfer/capsule-import.js

@@ -4,10 +4,10 @@ import {
 import "../chunk-VF4XKTX3.js";
 import "../chunk-WEHSQBFR.js";
 import "../chunk-X7Y7WX73.js";
+import "../chunk-GCGJW34D.js";
 import "../chunk-BJMBJZ2Y.js";
 import "../chunk-UKJAGEXH.js";
 import "../chunk-FP2373TW.js";
-import "../chunk-GCGJW34D.js";
 import "../chunk-A6XUJE5D.js";
 import "../chunk-VS2IYZRU.js";
 import "../chunk-PZ5AY32C.js";

package/dist/verified-recall.js

@@ -1,11 +1,12 @@
 import {
   compareVerifiedEpisodeResults,
   searchVerifiedEpisodes
-} from "./chunk-YW52BQSU.js";
+} from "./chunk-2TCHDANJ.js";
 import "./chunk-HQ6NIBL6.js";
-import "./chunk-VH6EIKVS.js";
+import "./chunk-WKMCC4NQ.js";
 import "./chunk-M7XQSUBB.js";
 import "./chunk-5UZXUTVO.js";
+import "./chunk-5GPPACXK.js";
 import "./chunk-J6A3CX5N.js";
 import "./chunk-AZBV4RRY.js";
 import "./chunk-AWJ2FHCF.js";
@@ -17,8 +18,8 @@ import "./chunk-6KYMPV2O.js";
 import "./chunk-DM2T26WE.js";
 import "./chunk-LDXUBPMO.js";
 import "./chunk-FVQJYWH7.js";
-import "./chunk-G7D6GZ5J.js";
 import "./chunk-VF4XKTX3.js";
+import "./chunk-G7D6GZ5J.js";
 import "./chunk-4DJQYKMN.js";
 import "./chunk-ZBJMUXZH.js";
 import "./chunk-2ODBA7MQ.js";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@remnic/core",
-  "version": "9.3.664",
+  "version": "9.3.666",
   "description": "Framework-agnostic Remnic memory engine — orchestrator, storage, extraction, search, trust zones",
   "type": "module",
   "main": "dist/index.js",

package/src/capabilities.test.ts

@@ -0,0 +1,97 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import { parseConfig } from "./config.js";
+import { resolveCapabilities, type CapabilitySet } from "./capabilities.js";
+
+/**
+ * Characterization tests for the recall-operation CapabilitySet (issue #1523).
+ *
+ * These guard against composition drift: every capability field must project
+ * from its `<field>Enabled` config flag, so a future edit to
+ * `resolveCapabilities` that accidentally maps a field to the wrong flag (the
+ * rule-39 gate-divergence class, moved up one layer) fails loudly here.
+ */
+
+/**
+ * Map of CapabilitySet field → the PluginConfig flag it projects from.
+ * Kept explicit (rather than derived by string concat) so the two graph flags
+ * with non-`<field>Enabled` names are covered too.
+ */
+const FIELD_TO_FLAG: Record<keyof CapabilitySet, string> = {
+  rerankCache: "rerankCacheEnabled",
+  recallDirectAnswer: "recallDirectAnswerEnabled",
+  recallMemoryWorthFilter: "recallMemoryWorthFilterEnabled",
+  recallMmr: "recallMmrEnabled",
+  recallReasoningTraceBoost: "recallReasoningTraceBoostEnabled",
+  recallPlannerLlm: "recallPlannerLlmEnabled",
+  recallPlanner: "recallPlannerEnabled",
+  recallConfidenceGate: "recallConfidenceGateEnabled",
+  graphRecall: "graphRecallEnabled",
+  graphAssistInFullMode: "graphAssistInFullModeEnabled",
+  graphExpandedIntent: "graphExpandedIntentEnabled",
+};
+
+const FIELDS = Object.keys(FIELD_TO_FLAG) as Array<keyof CapabilitySet>;
+
+test("resolveCapabilities projects every field from its <field>Enabled flag (true variant)", () => {
+  // Build a config where every migrated flag is explicitly true.
+  const overrides: Record<string, boolean> = {};
+  for (const flag of Object.values(FIELD_TO_FLAG)) overrides[flag] = true;
+  const config = parseConfig(overrides);
+  const caps = resolveCapabilities(config);
+
+  for (const field of FIELDS) {
+    const flag = FIELD_TO_FLAG[field];
+    assert.equal(
+      caps[field],
+      (config as unknown as Record<string, boolean>)[flag],
+      `caps.${field} must equal config.${flag} (true variant)`,
+    );
+    assert.equal(caps[field], true, `caps.${field} should be true here`);
+  }
+});
+
+test("resolveCapabilities projects every field from its <field>Enabled flag (false variant)", () => {
+  const overrides: Record<string, boolean> = {};
+  for (const flag of Object.values(FIELD_TO_FLAG)) overrides[flag] = false;
+  const config = parseConfig(overrides);
+  const caps = resolveCapabilities(config);
+
+  for (const field of FIELDS) {
+    const flag = FIELD_TO_FLAG[field];
+    // The two optional graph flags carry default-when-undefined semantics, but
+    // when explicitly set to a concrete boolean the projection must match it.
+    assert.equal(
+      caps[field],
+      (config as unknown as Record<string, boolean>)[flag],
+      `caps.${field} must equal config.${flag} (false variant)`,
+    );
+    assert.equal(caps[field], false, `caps.${field} should be false here`);
+  }
+});
+
+test("resolveCapabilities preserves optional-flag defaults when the flag is undefined", () => {
+  // parseConfig with no overrides exercises the documented defaults. The two
+  // optional graph flags encode asymmetric defaults on purpose:
+  //   graphAssistInFullModeEnabled → default-ON  (`!== false`)
+  //   graphExpandedIntentEnabled    → default-OFF (`=== true`)
+  const config = parseConfig({});
+  const caps = resolveCapabilities(config);
+
+  assert.equal(
+    caps.graphAssistInFullMode,
+    config.graphAssistInFullModeEnabled !== false,
+    "graphAssistInFullMode must be default-on unless explicitly false",
+  );
+  assert.equal(
+    caps.graphExpandedIntent,
+    config.graphExpandedIntentEnabled === true,
+    "graphExpandedIntent must be default-off unless explicitly true",
+  );
+});
+
+test("resolveCapabilities returns a frozen object", () => {
+  const caps = resolveCapabilities(parseConfig({}));
+  assert.equal(Object.isFrozen(caps), true, "CapabilitySet must be frozen");
+});

package/src/capabilities.ts

@@ -0,0 +1,86 @@
+/**
+ * CapabilitySet — recall-operation feature gates resolved once, then threaded.
+ *
+ * Issue #1523 (Phase 1 of epic #1520). Root cause this addresses: 161+
+ * scattered `config.<flag>Enabled` reads mean each gate is re-derived at every
+ * call site, and reviews keep finding parallel code paths where one branch
+ * checks a gate the other forgot (CLAUDE.md rule 39 — the "gate divergence"
+ * defect class). The fix is to resolve a frozen capability projection ONCE at
+ * the top of the recall operation and pass it down explicitly.
+ *
+ * Scope of THIS module (first migration PR): only the recall-operation-scoped
+ * flags below. Flags that are also read in graph construction, writes, CLI, or
+ * the summarizer are deliberately deferred to a follow-up so we never leave a
+ * single flag half-migrated (some sites on `caps.`, some on `config.`).
+ *
+ * Field naming: each field is the config flag name with the trailing `Enabled`
+ * removed (`recallMmrEnabled` → `recallMmr`).
+ *
+ * This is plumbing, not a feature — there is deliberately NO `enabled` gate for
+ * the CapabilitySet itself (rule 30 governs behavior changes; resolving and
+ * threading a capability projection must stay behavior-preserving).
+ */
+
+import type { PluginConfig } from "./types.js";
+
+/**
+ * Frozen projection of recall-operation feature gates.
+ *
+ * Every field is `readonly boolean`. The composition that maps a config flag to
+ * a capability (including default-when-undefined semantics for optional flags)
+ * lives ONLY in {@link resolveCapabilities} — call sites must read the
+ * capability, never re-derive it from raw config.
+ */
+export interface CapabilitySet {
+  /** `rerankCacheEnabled` — cache reranker scores across recall passes. */
+  readonly rerankCache: boolean;
+  /** `recallDirectAnswerEnabled` — observation-mode direct-answer tier. */
+  readonly recallDirectAnswer: boolean;
+  /** `recallMemoryWorthFilterEnabled` — Memory-Worth score reweighting. */
+  readonly recallMemoryWorthFilter: boolean;
+  /** `recallMmrEnabled` — maximal-marginal-relevance diversification. */
+  readonly recallMmr: boolean;
+  /** `recallReasoningTraceBoostEnabled` — boost reasoning-trace memories. */
+  readonly recallReasoningTraceBoost: boolean;
+  /** `recallPlannerLlmEnabled` — LLM-backed recall-mode planner. */
+  readonly recallPlannerLlm: boolean;
+  /** `recallPlannerEnabled` — recall-mode planner (heuristic + optional LLM). */
+  readonly recallPlanner: boolean;
+  /** `recallConfidenceGateEnabled` — Synapse-style confidence gate. */
+  readonly recallConfidenceGate: boolean;
+  /** `graphRecallEnabled` — graph-mode recall tier (gates planner graph mode). */
+  readonly graphRecall: boolean;
+  /** `graphAssistInFullModeEnabled` — graph-assist overlay in full mode. */
+  readonly graphAssistInFullMode: boolean;
+  /** `graphExpandedIntentEnabled` — promote broad-intent asks to graph mode. */
+  readonly graphExpandedIntent: boolean;
+}
+
+/**
+ * Resolve the recall-operation {@link CapabilitySet} from parsed config.
+ *
+ * Call this ONCE per recall operation (at the `recall()` / `recallInternal`
+ * entry) and thread the result down. Composition lives here and only here.
+ *
+ * Session toggles are intentionally not a parameter yet: `session-toggles.ts`
+ * is agent-scoped (per session/agent enable-disable of the whole plugin), not
+ * flag-scoped — none of the flags projected here have a per-session override,
+ * so there is nothing for a toggle argument to compose at this layer.
+ */
+export function resolveCapabilities(config: PluginConfig): CapabilitySet {
+  return Object.freeze({
+    rerankCache: config.rerankCacheEnabled,
+    recallDirectAnswer: config.recallDirectAnswerEnabled,
+    recallMemoryWorthFilter: config.recallMemoryWorthFilterEnabled,
+    recallMmr: config.recallMmrEnabled,
+    recallReasoningTraceBoost: config.recallReasoningTraceBoostEnabled,
+    recallPlannerLlm: config.recallPlannerLlmEnabled,
+    recallPlanner: config.recallPlannerEnabled,
+    recallConfidenceGate: config.recallConfidenceGateEnabled,
+    graphRecall: config.graphRecallEnabled,
+    // Optional flags: preserve the exact default-when-undefined semantics the
+    // migrated call sites used (`!== false` = default-on, `=== true` = default-off).
+    graphAssistInFullMode: config.graphAssistInFullModeEnabled !== false,
+    graphExpandedIntent: config.graphExpandedIntentEnabled === true,
+  });
+}