@warlock.js/ai 5.2.3 → 5.2.4

package/esm/memory/memory.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"memory.mjs","names":[],"sources":["../../../../../../../ai/src/memory/memory.ts"],"sourcesContent":["import { resolveDefaultStore } from \"../config\";\nimport type {\n  MemoryConfig,\n  WorkingMemoryConfig,\n} from \"../contracts/memory/memory-config.type\";\nimport type {\n  MemoryItem,\n  MemoryTier,\n  RecalledMemory,\n} from \"../contracts/memory/memory-item.type\";\nimport type { MemoryContract } from \"../contracts/memory/memory.contract\";\nimport type { RecallOptions } from \"../contracts/memory/recall-options.type\";\nimport { EpisodicMemory } from \"./episodic-memory\";\nimport { ProceduralMemory } from \"./procedural-memory\";\nimport { SemanticMemory } from \"./semantic-memory\";\nimport { WorkingMemory } from \"./working-memory\";\n\nconst DEFAULT_NAME = \"memory\";\nconst DEFAULT_SEMANTIC_NAMESPACE = \"ai.memory.semantic\";\nconst DEFAULT_EPISODIC_NAMESPACE = \"ai.memory.episodic\";\nconst DEFAULT_PROCEDURAL_NAMESPACE = \"ai.memory.procedural\";\nconst DEFAULT_K = 5;\nconst DEFAULT_THRESHOLD = 0.7;\nconst DEFAULT_RECENCY_WEIGHT = 0.3;\nconst DEFAULT_HALF_LIFE_MS = 7 * 24 * 60 * 60 * 1000;\nconst DEFAULT_REINFORCEMENT_WEIGHT = 0.3;\n\n/**\n * Entries the in-process working buffer holds before it starts evicting\n * its oldest (4.15.0 — security fix for unbounded growth). Sized to hold\n * a deep multi-session scratch history while capping the tier's worst\n * case at a few MB of resident text rather than \"everything this process\n * has ever been told.\"\n */\nconst DEFAULT_WORKING_MAX_ITEMS = 1000;\n\n/**\n * Create an agent memory store (memory core M2).\n *\n * Wires up to four tiers behind the {@link MemoryContract}: **working**\n * (in-run scratch, recency), **semantic** (durable facts by cosine\n * similarity), **episodic** (durable events, similarity blended with\n * recency), and **procedural** (durable how-tos, similarity blended with\n * reinforcement). The working tier is on by default; the other three each\n * activate only when their config is supplied. The three vector tiers\n * mirror how `semanticCache` delegates similarity to the cache driver's\n * `.similar()`.\n *\n * Resolution happens once here, at construction (loud), rather than per\n * call (silent until first use): a vector-tier config with no `store` and\n * no `ai.config({ defaultStore })` throws now; enabling no tier at all\n * throws now.\n *\n * TTL-based decay / forgetting remains deferred. The working tier is\n * size-bounded (`working: { maxItems }`, default `1000`, oldest-written\n * evicted first) because it is the one tier that holds everything it is\n * told in process memory for the life of the instance; the durable tiers\n * delegate retention to their `CacheDriver`.\n *\n * **Isolation (4.15.0).** `remember({ scope })` / `recall(query, { scope })`\n * carry an opaque tenant / session key that every tier enforces as an\n * exact-equality filter before scoring — one scope's memories never\n * surface in another's recall, and identical text under two scopes stays\n * two entries. Unscoped writes form a shared pool that only an unscoped\n * recall can read; there is no \"all scopes\" query. `ai.orchestrator()`\n * derives this from the turn's `sessionId` automatically.\n *\n * @example\n * import { ai } from \"@warlock.js/ai\";\n * import { MemoryCacheDriver } from \"@warlock.js/cache\";\n *\n * const store = new MemoryCacheDriver();\n * store.setOptions({});\n *\n * const mem = ai.memory({\n *   semantic: { embedder, store },\n *   defaultTier: \"semantic\",\n * });\n *\n * await mem.remember({ text: \"User prefers concise answers.\" });\n * const hits = await mem.recall(\"how should I respond?\", { k: 3 });\n */\nexport function memory(config: MemoryConfig = {}): MemoryContract {\n  const name = config.name ?? DEFAULT_NAME;\n  const workingConfig = config.working ?? true;\n  const defaultK = config.k ?? DEFAULT_K;\n  const defaultThreshold = config.threshold ?? DEFAULT_THRESHOLD;\n\n  const working =\n    workingConfig === false\n      ? undefined\n      : new WorkingMemory(resolveWorkingMaxItems(workingConfig, name));\n\n  const semantic = config.semantic\n    ? buildSemanticTier(config.semantic, name)\n    : undefined;\n\n  const episodic = config.episodic\n    ? buildEpisodicTier(config.episodic, name)\n    : undefined;\n\n  const procedural = config.procedural\n    ? buildProceduralTier(config.procedural, name)\n    : undefined;\n\n  const tiers: Tiers = { working, semantic, episodic, procedural };\n\n  if (!working && !semantic && !episodic && !procedural) {\n    throw new Error(\n      `memory(\"${name}\"): no tier enabled — enable \\`working\\` (default) or pass a \\`semantic\\` / \\`episodic\\` / \\`procedural\\` config; a memory with no tiers can neither store nor recall`,\n    );\n  }\n\n  const defaultTier: MemoryTier = config.defaultTier ?? \"working\";\n\n  assertTierEnabled(defaultTier, tiers, name);\n\n  return {\n    name,\n    async remember(items: MemoryItem | MemoryItem[]): Promise<void> {\n      const list = Array.isArray(items) ? items : [items];\n\n      const writes: Promise<void>[] = [];\n\n      for (const item of list) {\n        const tier = item.tier ?? defaultTier;\n\n        assertTierEnabled(tier, tiers, name);\n\n        if (tier === \"working\") {\n          working!.remember(item);\n\n          continue;\n        }\n\n        if (tier === \"semantic\") {\n          writes.push(semantic!.remember(item));\n\n          continue;\n        }\n\n        if (tier === \"episodic\") {\n          writes.push(episodic!.remember(item));\n\n          continue;\n        }\n\n        writes.push(procedural!.remember(item));\n      }\n\n      await Promise.all(writes);\n    },\n    async recall(\n      query: string,\n      options: RecallOptions = {},\n    ): Promise<RecalledMemory[]> {\n      const k = options.k ?? defaultK;\n      const threshold = options.threshold ?? defaultThreshold;\n\n      if (options.tier) {\n        assertTierEnabled(options.tier, tiers, name);\n      }\n\n      const wants = (tier: MemoryTier): boolean =>\n        !options.tier || options.tier === tier;\n\n      // `options.scope` is the isolation key — each tier applies it as an\n      // exact-equality filter internally, BEFORE its own scoring and\n      // slicing, so nothing outside the scope reaches this merge.\n      const scope = options.scope;\n\n      const [workingHits, semanticHits, episodicHits, proceduralHits] =\n        await Promise.all([\n          working && wants(\"working\")\n            ? Promise.resolve(working.recall(k, scope))\n            : Promise.resolve([] as RecalledMemory[]),\n          semantic && wants(\"semantic\")\n            ? semantic.recall(query, k, threshold, scope)\n            : Promise.resolve([] as RecalledMemory[]),\n          episodic && wants(\"episodic\")\n            ? episodic.recall(query, k, threshold, scope)\n            : Promise.resolve([] as RecalledMemory[]),\n          procedural && wants(\"procedural\")\n            ? procedural.recall(query, k, threshold, scope)\n            : Promise.resolve([] as RecalledMemory[]),\n        ]);\n\n      return [\n        ...workingHits,\n        ...semanticHits,\n        ...episodicHits,\n        ...proceduralHits,\n      ]\n        .sort((first, second) => second.score - first.score)\n        .slice(0, k);\n    },\n    async clear(tier?: MemoryTier): Promise<void> {\n      const clears: Promise<void>[] = [];\n\n      if (working && (!tier || tier === \"working\")) {\n        working.clear();\n      }\n\n      if (semantic && (!tier || tier === \"semantic\")) {\n        clears.push(semantic.clear());\n      }\n\n      if (episodic && (!tier || tier === \"episodic\")) {\n        clears.push(episodic.clear());\n      }\n\n      if (procedural && (!tier || tier === \"procedural\")) {\n        clears.push(procedural.clear());\n      }\n\n      await Promise.all(clears);\n    },\n  };\n}\n\n/** The four tier instances a `memory()` composes; `undefined` when off. */\ntype Tiers = {\n  working: WorkingMemory | undefined;\n  semantic: SemanticMemory | undefined;\n  episodic: EpisodicMemory | undefined;\n  procedural: ProceduralMemory | undefined;\n};\n\n/**\n * Resolve the working tier's size bound from the `working` config\n * (`true` / a `{ maxItems }` object), validating it at construction the\n * same way every other tier's wiring fails loud-and-now rather than on\n * first use. There is deliberately no unbounded setting — the buffer is\n * process-resident for the life of the memory instance, so \"no cap\" is\n * a memory-exhaustion vector, not a configuration choice.\n */\nfunction resolveWorkingMaxItems(\n  workingConfig: true | WorkingMemoryConfig,\n  name: string,\n): number {\n  const maxItems =\n    workingConfig === true\n      ? DEFAULT_WORKING_MAX_ITEMS\n      : (workingConfig.maxItems ?? DEFAULT_WORKING_MAX_ITEMS);\n\n  if (!Number.isInteger(maxItems) || maxItems < 1) {\n    throw new Error(\n      `memory(\"${name}\"): working tier \\`maxItems\\` must be an integer >= 1 — received ${String(maxItems)}`,\n    );\n  }\n\n  return maxItems;\n}\n\n/**\n * Resolve the semantic tier's store (explicit `store` wins, else the\n * global `ai.config({ defaultStore })`) and build the tier. Throws at\n * construction when neither is available — the same loud-now contract\n * `semanticCache` follows.\n */\nfunction buildSemanticTier(\n  semanticConfig: NonNullable<MemoryConfig[\"semantic\"]>,\n  name: string,\n): SemanticMemory {\n  const store = semanticConfig.store ?? resolveDefaultStore();\n\n  if (!store) {\n    throw new Error(\n      `memory(\"${name}\"): semantic tier has no store — pass \\`semantic.store\\` or call \\`ai.config({ defaultStore })\\` at app boot before constructing the memory`,\n    );\n  }\n\n  return new SemanticMemory(\n    semanticConfig.embedder,\n    store,\n    semanticConfig.namespace ?? DEFAULT_SEMANTIC_NAMESPACE,\n  );\n}\n\n/**\n * Resolve the episodic tier's store (explicit `store` wins, else the\n * global default) and build the tier with its recency knobs. Throws at\n * construction when neither store is available — the same loud-now\n * contract the semantic tier follows.\n */\nfunction buildEpisodicTier(\n  episodicConfig: NonNullable<MemoryConfig[\"episodic\"]>,\n  name: string,\n): EpisodicMemory {\n  const store = episodicConfig.store ?? resolveDefaultStore();\n\n  if (!store) {\n    throw new Error(\n      `memory(\"${name}\"): episodic tier has no store — pass \\`episodic.store\\` or call \\`ai.config({ defaultStore })\\` at app boot before constructing the memory`,\n    );\n  }\n\n  return new EpisodicMemory(\n    episodicConfig.embedder,\n    store,\n    episodicConfig.namespace ?? DEFAULT_EPISODIC_NAMESPACE,\n    episodicConfig.recencyWeight ?? DEFAULT_RECENCY_WEIGHT,\n    episodicConfig.halfLifeMs ?? DEFAULT_HALF_LIFE_MS,\n    episodicConfig.now ?? (() => Date.now()),\n  );\n}\n\n/**\n * Resolve the procedural tier's store and build the tier with its\n * reinforcement knob. Throws at construction when no store is available.\n */\nfunction buildProceduralTier(\n  proceduralConfig: NonNullable<MemoryConfig[\"procedural\"]>,\n  name: string,\n): ProceduralMemory {\n  const store = proceduralConfig.store ?? resolveDefaultStore();\n\n  if (!store) {\n    throw new Error(\n      `memory(\"${name}\"): procedural tier has no store — pass \\`procedural.store\\` or call \\`ai.config({ defaultStore })\\` at app boot before constructing the memory`,\n    );\n  }\n\n  return new ProceduralMemory(\n    proceduralConfig.embedder,\n    store,\n    proceduralConfig.namespace ?? DEFAULT_PROCEDURAL_NAMESPACE,\n    proceduralConfig.reinforcementWeight ?? DEFAULT_REINFORCEMENT_WEIGHT,\n  );\n}\n\n/**\n * Guard that a tier referenced by config / a call is actually enabled,\n * failing fast with an actionable message instead of a downstream\n * `undefined` dereference.\n */\nfunction assertTierEnabled(tier: MemoryTier, tiers: Tiers, name: string): void {\n  if (tier === \"working\" && !tiers.working) {\n    throw new Error(\n      `memory(\"${name}\"): working tier is disabled — set \\`working: true\\` (the default) to use it`,\n    );\n  }\n\n  if (tier === \"semantic\" && !tiers.semantic) {\n    throw new Error(\n      `memory(\"${name}\"): semantic tier is not configured — pass \\`semantic\\` config to use it`,\n    );\n  }\n\n  if (tier === \"episodic\" && !tiers.episodic) {\n    throw new Error(\n      `memory(\"${name}\"): episodic tier is not configured — pass \\`episodic\\` config to use it`,\n    );\n  }\n\n  if (tier === \"procedural\" && !tiers.procedural) {\n    throw new Error(\n      `memory(\"${name}\"): procedural tier is not configured — pass \\`procedural\\` config to use it`,\n    );\n  }\n}\n"],"mappings":";;;;;;;AAiBA,MAAM,eAAe;AACrB,MAAM,6BAA6B;AACnC,MAAM,6BAA6B;AACnC,MAAM,+BAA+B;AACrC,MAAM,YAAY;AAClB,MAAM,oBAAoB;AAC1B,MAAM,yBAAyB;AAC/B,MAAM,uBAAuB,QAAc,KAAK;AAChD,MAAM,+BAA+B;;;;;;;;AASrC,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgDlC,SAAgB,OAAO,SAAuB,CAAC,GAAmB;CAChE,MAAM,OAAO,OAAO,QAAQ;CAC5B,MAAM,gBAAgB,OAAO,WAAW;CACxC,MAAM,WAAW,OAAO,KAAK;CAC7B,MAAM,mBAAmB,OAAO,aAAa;CAE7C,MAAM,UACJ,kBAAkB,QACd,SACA,IAAI,cAAc,uBAAuB,eAAe,IAAI,CAAC;CAEnE,MAAM,WAAW,OAAO,WACpB,kBAAkB,OAAO,UAAU,IAAI,IACvC;CAEJ,MAAM,WAAW,OAAO,WACpB,kBAAkB,OAAO,UAAU,IAAI,IACvC;CAEJ,MAAM,aAAa,OAAO,aACtB,oBAAoB,OAAO,YAAY,IAAI,IAC3C;CAEJ,MAAM,QAAe;EAAE;EAAS;EAAU;EAAU;CAAW;CAE/D,IAAI,CAAC,WAAW,CAAC,YAAY,CAAC,YAAY,CAAC,YACzC,MAAM,IAAI,MACR,WAAW,KAAK,sKAClB;CAGF,MAAM,cAA0B,OAAO,eAAe;CAEtD,kBAAkB,aAAa,OAAO,IAAI;CAE1C,OAAO;EACL;EACA,MAAM,SAAS,OAAiD;GAC9D,MAAM,OAAO,MAAM,QAAQ,KAAK,IAAI,QAAQ,CAAC,KAAK;GAElD,MAAM,SAA0B,CAAC;GAEjC,KAAK,MAAM,QAAQ,MAAM;IACvB,MAAM,OAAO,KAAK,QAAQ;IAE1B,kBAAkB,MAAM,OAAO,IAAI;IAEnC,IAAI,SAAS,WAAW;KACtB,QAAS,SAAS,IAAI;KAEtB;IACF;IAEA,IAAI,SAAS,YAAY;KACvB,OAAO,KAAK,SAAU,SAAS,IAAI,CAAC;KAEpC;IACF;IAEA,IAAI,SAAS,YAAY;KACvB,OAAO,KAAK,SAAU,SAAS,IAAI,CAAC;KAEpC;IACF;IAEA,OAAO,KAAK,WAAY,SAAS,IAAI,CAAC;GACxC;GAEA,MAAM,QAAQ,IAAI,MAAM;EAC1B;EACA,MAAM,OACJ,OACA,UAAyB,CAAC,GACC;GAC3B,MAAM,IAAI,QAAQ,KAAK;GACvB,MAAM,YAAY,QAAQ,aAAa;GAEvC,IAAI,QAAQ,MACV,kBAAkB,QAAQ,MAAM,OAAO,IAAI;GAG7C,MAAM,SAAS,SACb,CAAC,QAAQ,QAAQ,QAAQ,SAAS;GAKpC,MAAM,QAAQ,QAAQ;GAEtB,MAAM,CAAC,aAAa,cAAc,cAAc,kBAC9C,MAAM,QAAQ,IAAI;IAChB,WAAW,MAAM,SAAS,IACtB,QAAQ,QAAQ,QAAQ,OAAO,GAAG,KAAK,CAAC,IACxC,QAAQ,QAAQ,CAAC,CAAqB;IAC1C,YAAY,MAAM,UAAU,IACxB,SAAS,OAAO,OAAO,GAAG,WAAW,KAAK,IAC1C,QAAQ,QAAQ,CAAC,CAAqB;IAC1C,YAAY,MAAM,UAAU,IACxB,SAAS,OAAO,OAAO,GAAG,WAAW,KAAK,IAC1C,QAAQ,QAAQ,CAAC,CAAqB;IAC1C,cAAc,MAAM,YAAY,IAC5B,WAAW,OAAO,OAAO,GAAG,WAAW,KAAK,IAC5C,QAAQ,QAAQ,CAAC,CAAqB;GAC5C,CAAC;GAEH,OAAO;IACL,GAAG;IACH,GAAG;IACH,GAAG;IACH,GAAG;GACL,EACG,MAAM,OAAO,WAAW,OAAO,QAAQ,MAAM,KAAK,EAClD,MAAM,GAAG,CAAC;EACf;EACA,MAAM,MAAM,MAAkC;GAC5C,MAAM,SAA0B,CAAC;GAEjC,IAAI,YAAY,CAAC,QAAQ,SAAS,YAChC,QAAQ,MAAM;GAGhB,IAAI,aAAa,CAAC,QAAQ,SAAS,aACjC,OAAO,KAAK,SAAS,MAAM,CAAC;GAG9B,IAAI,aAAa,CAAC,QAAQ,SAAS,aACjC,OAAO,KAAK,SAAS,MAAM,CAAC;GAG9B,IAAI,eAAe,CAAC,QAAQ,SAAS,eACnC,OAAO,KAAK,WAAW,MAAM,CAAC;GAGhC,MAAM,QAAQ,IAAI,MAAM;EAC1B;CACF;AACF;;;;;;;;;AAkBA,SAAS,uBACP,eACA,MACQ;CACR,MAAM,WACJ,kBAAkB,OACd,4BACC,cAAc,YAAY;CAEjC,IAAI,CAAC,OAAO,UAAU,QAAQ,KAAK,WAAW,GAC5C,MAAM,IAAI,MACR,WAAW,KAAK,mEAAmE,OAAO,QAAQ,GACpG;CAGF,OAAO;AACT;;;;;;;AAQA,SAAS,kBACP,gBACA,MACgB;CAChB,MAAM,QAAQ,eAAe,SAAS,oBAAoB;CAE1D,IAAI,CAAC,OACH,MAAM,IAAI,MACR,WAAW,KAAK,4IAClB;CAGF,OAAO,IAAI,eACT,eAAe,UACf,OACA,eAAe,aAAa,0BAC9B;AACF;;;;;;;AAQA,SAAS,kBACP,gBACA,MACgB;CAChB,MAAM,QAAQ,eAAe,SAAS,oBAAoB;CAE1D,IAAI,CAAC,OACH,MAAM,IAAI,MACR,WAAW,KAAK,4IAClB;CAGF,OAAO,IAAI,eACT,eAAe,UACf,OACA,eAAe,aAAa,4BAC5B,eAAe,iBAAiB,wBAChC,eAAe,cAAc,sBAC7B,eAAe,cAAc,KAAK,IAAI,EACxC;AACF;;;;;AAMA,SAAS,oBACP,kBACA,MACkB;CAClB,MAAM,QAAQ,iBAAiB,SAAS,oBAAoB;CAE5D,IAAI,CAAC,OACH,MAAM,IAAI,MACR,WAAW,KAAK,gJAClB;CAGF,OAAO,IAAI,iBACT,iBAAiB,UACjB,OACA,iBAAiB,aAAa,8BAC9B,iBAAiB,uBAAuB,4BAC1C;AACF;;;;;;AAOA,SAAS,kBAAkB,MAAkB,OAAc,MAAoB;CAC7E,IAAI,SAAS,aAAa,CAAC,MAAM,SAC/B,MAAM,IAAI,MACR,WAAW,KAAK,6EAClB;CAGF,IAAI,SAAS,cAAc,CAAC,MAAM,UAChC,MAAM,IAAI,MACR,WAAW,KAAK,yEAClB;CAGF,IAAI,SAAS,cAAc,CAAC,MAAM,UAChC,MAAM,IAAI,MACR,WAAW,KAAK,yEAClB;CAGF,IAAI,SAAS,gBAAgB,CAAC,MAAM,YAClC,MAAM,IAAI,MACR,WAAW,KAAK,6EAClB;AAEJ"}
+{"version":3,"file":"memory.mjs","names":[],"sources":["../../../../../../../ai/src/memory/memory.ts"],"sourcesContent":["import { resolveDefaultStore } from \"../config\";\nimport type {\n  MemoryConfig,\n  WorkingMemoryConfig,\n} from \"../contracts/memory/memory-config.type\";\nimport type {\n  MemoryItem,\n  MemoryTier,\n  RecalledMemory,\n} from \"../contracts/memory/memory-item.type\";\nimport type { MemoryContract } from \"../contracts/memory/memory.contract\";\nimport type { RecallOptions } from \"../contracts/memory/recall-options.type\";\nimport { EpisodicMemory } from \"./episodic-memory\";\nimport { ProceduralMemory } from \"./procedural-memory\";\nimport { SemanticMemory } from \"./semantic-memory\";\nimport { WorkingMemory } from \"./working-memory\";\n\nconst DEFAULT_NAME = \"memory\";\nconst DEFAULT_SEMANTIC_NAMESPACE = \"ai.memory.semantic\";\nconst DEFAULT_EPISODIC_NAMESPACE = \"ai.memory.episodic\";\nconst DEFAULT_PROCEDURAL_NAMESPACE = \"ai.memory.procedural\";\nconst DEFAULT_K = 5;\nconst DEFAULT_THRESHOLD = 0.7;\nconst DEFAULT_RECENCY_WEIGHT = 0.3;\nconst DEFAULT_HALF_LIFE_MS = 7 * 24 * 60 * 60 * 1000;\nconst DEFAULT_REINFORCEMENT_WEIGHT = 0.3;\n\n/**\n * Entries the in-process working buffer holds before it starts evicting\n * its oldest (4.15.0 — security fix for unbounded growth). Sized to hold\n * a deep multi-session scratch history while capping the tier's worst\n * case at a few MB of resident text rather than \"everything this process\n * has ever been told.\"\n */\nconst DEFAULT_WORKING_MAX_ITEMS = 1000;\n\n/**\n * Create an agent memory store (memory core M2).\n *\n * Wires up to four tiers behind the {@link MemoryContract}: **working**\n * (in-run scratch, recency), **semantic** (durable facts by cosine\n * similarity), **episodic** (durable events, similarity blended with\n * recency), and **procedural** (durable how-tos, similarity blended with\n * reinforcement). The working tier is on by default; the other three each\n * activate only when their config is supplied. The three vector tiers\n * mirror how `semanticCache` delegates similarity to the cache driver's\n * `.similar()`.\n *\n * Resolution happens once here, at construction (loud), rather than per\n * call (silent until first use): a vector-tier config with no `store` and\n * no `ai.config({ defaultStore })` throws now; enabling no tier at all\n * throws now.\n *\n * TTL-based decay / forgetting remains deferred. The working tier is\n * size-bounded (`working: { maxItems }`, default `1000`, oldest-written\n * evicted first) because it is the one tier that holds everything it is\n * told in process memory for the life of the instance; the durable tiers\n * delegate retention to their `CacheDriver`.\n *\n * **Isolation (4.15.0).** `remember({ scope })` / `recall(query, { scope })`\n * carry an opaque tenant / session key that every tier enforces as an\n * exact-equality filter before scoring — one scope's memories never\n * surface in another's recall, and identical text under two scopes stays\n * two entries. Unscoped writes form a shared pool that only an unscoped\n * recall can read; there is no \"all scopes\" query. `ai.orchestrator()`\n * derives this from the turn's `sessionId` automatically.\n *\n * @example\n * import { ai } from \"@warlock.js/ai\";\n * import { MemoryCacheDriver } from \"@warlock.js/cache\";\n *\n * const store = new MemoryCacheDriver();\n * store.setOptions({});\n *\n * const mem = ai.memory({\n *   semantic: { embedder, store },\n *   defaultTier: \"semantic\",\n * });\n *\n * await mem.remember({ text: \"User prefers concise answers.\" });\n * const hits = await mem.recall(\"how should I respond?\", { k: 3 });\n */\nexport function memory(config: MemoryConfig = {}): MemoryContract {\n  const name = config.name ?? DEFAULT_NAME;\n  const workingConfig = config.working ?? true;\n  const defaultK = config.k ?? DEFAULT_K;\n  const defaultThreshold = config.threshold ?? DEFAULT_THRESHOLD;\n\n  const working =\n    workingConfig === false\n      ? undefined\n      : new WorkingMemory(resolveWorkingMaxItems(workingConfig, name));\n\n  const semantic = config.semantic\n    ? buildSemanticTier(config.semantic, name)\n    : undefined;\n\n  const episodic = config.episodic\n    ? buildEpisodicTier(config.episodic, name)\n    : undefined;\n\n  const procedural = config.procedural\n    ? buildProceduralTier(config.procedural, name)\n    : undefined;\n\n  const tiers: Tiers = { working, semantic, episodic, procedural };\n\n  if (!working && !semantic && !episodic && !procedural) {\n    throw new Error(\n      `memory(\"${name}\"): no tier enabled — enable \\`working\\` (default) or pass a \\`semantic\\` / \\`episodic\\` / \\`procedural\\` config; a memory with no tiers can neither store nor recall`,\n    );\n  }\n\n  const defaultTier: MemoryTier = config.defaultTier ?? \"working\";\n\n  assertTierEnabled(defaultTier, tiers, name);\n\n  return {\n    name,\n    async remember(items: MemoryItem | MemoryItem[]): Promise<void> {\n      const list = Array.isArray(items) ? items : [items];\n\n      const writes: Promise<void>[] = [];\n\n      for (const item of list) {\n        const tier = item.tier ?? defaultTier;\n\n        assertTierEnabled(tier, tiers, name);\n\n        if (tier === \"working\") {\n          working!.remember(item);\n\n          continue;\n        }\n\n        if (tier === \"semantic\") {\n          writes.push(semantic!.remember(item));\n\n          continue;\n        }\n\n        if (tier === \"episodic\") {\n          writes.push(episodic!.remember(item));\n\n          continue;\n        }\n\n        writes.push(procedural!.remember(item));\n      }\n\n      await Promise.all(writes);\n    },\n    async recall(\n      query: string,\n      options: RecallOptions = {},\n    ): Promise<RecalledMemory[]> {\n      const k = options.k ?? defaultK;\n      const threshold = options.threshold ?? defaultThreshold;\n\n      if (options.tier) {\n        assertTierEnabled(options.tier, tiers, name);\n      }\n\n      const wants = (tier: MemoryTier): boolean =>\n        !options.tier || options.tier === tier;\n\n      // `options.scope` is the isolation key — each tier applies it as an\n      // exact-equality filter internally, BEFORE its own scoring and\n      // slicing, so nothing outside the scope reaches this merge.\n      const scope = options.scope;\n\n      const [workingHits, semanticHits, episodicHits, proceduralHits] =\n        await Promise.all([\n          working && wants(\"working\")\n            ? Promise.resolve(working.recall(k, scope))\n            : Promise.resolve([] as RecalledMemory[]),\n          semantic && wants(\"semantic\")\n            ? semantic.recall(query, k, threshold, scope)\n            : Promise.resolve([] as RecalledMemory[]),\n          episodic && wants(\"episodic\")\n            ? episodic.recall(query, k, threshold, scope)\n            : Promise.resolve([] as RecalledMemory[]),\n          procedural && wants(\"procedural\")\n            ? procedural.recall(query, k, threshold, scope)\n            : Promise.resolve([] as RecalledMemory[]),\n        ]);\n\n      return [\n        ...workingHits,\n        ...semanticHits,\n        ...episodicHits,\n        ...proceduralHits,\n      ]\n        .sort((first, second) => second.score - first.score)\n        .slice(0, k);\n    },\n    async clear(tier?: MemoryTier): Promise<void> {\n      const clears: Promise<void>[] = [];\n\n      if (working && (!tier || tier === \"working\")) {\n        working.clear();\n      }\n\n      if (semantic && (!tier || tier === \"semantic\")) {\n        clears.push(semantic.clear());\n      }\n\n      if (episodic && (!tier || tier === \"episodic\")) {\n        clears.push(episodic.clear());\n      }\n\n      if (procedural && (!tier || tier === \"procedural\")) {\n        clears.push(procedural.clear());\n      }\n\n      await Promise.all(clears);\n    },\n  };\n}\n\n/** The four tier instances a `memory()` composes; `undefined` when off. */\ntype Tiers = {\n  working: WorkingMemory | undefined;\n  semantic: SemanticMemory | undefined;\n  episodic: EpisodicMemory | undefined;\n  procedural: ProceduralMemory | undefined;\n};\n\n/**\n * Resolve the working tier's size bound from the `working` config\n * (`true` / a `{ maxItems }` object), validating it at construction the\n * same way every other tier's wiring fails loud-and-now rather than on\n * first use. There is deliberately no unbounded setting — the buffer is\n * process-resident for the life of the memory instance, so \"no cap\" is\n * a memory-exhaustion vector, not a configuration choice.\n */\nfunction resolveWorkingMaxItems(\n  workingConfig: true | WorkingMemoryConfig,\n  name: string,\n): number {\n  const maxItems =\n    workingConfig === true\n      ? DEFAULT_WORKING_MAX_ITEMS\n      : (workingConfig.maxItems ?? DEFAULT_WORKING_MAX_ITEMS);\n\n  if (!Number.isInteger(maxItems) || maxItems < 1) {\n    throw new Error(\n      `memory(\"${name}\"): working tier \\`maxItems\\` must be an integer >= 1 — received ${String(maxItems)}`,\n    );\n  }\n\n  return maxItems;\n}\n\n/**\n * Resolve the semantic tier's store (explicit `store` wins, else the\n * global `ai.config({ defaultStore })`) and build the tier. Throws at\n * construction when neither is available — the same loud-now contract\n * `semanticCache` follows.\n */\nfunction buildSemanticTier(\n  semanticConfig: NonNullable<MemoryConfig[\"semantic\"]>,\n  name: string,\n): SemanticMemory {\n  const store = semanticConfig.store ?? resolveDefaultStore();\n\n  if (!store) {\n    throw new Error(\n      `memory(\"${name}\"): semantic tier has no store — pass \\`semantic.store\\` or call \\`ai.config({ defaultStore })\\` at app boot before constructing the memory`,\n    );\n  }\n\n  return new SemanticMemory(\n    semanticConfig.embedder,\n    store,\n    semanticConfig.namespace ?? DEFAULT_SEMANTIC_NAMESPACE,\n  );\n}\n\n/**\n * Resolve the episodic tier's store (explicit `store` wins, else the\n * global default) and build the tier with its recency knobs. Throws at\n * construction when neither store is available — the same loud-now\n * contract the semantic tier follows.\n */\nfunction buildEpisodicTier(\n  episodicConfig: NonNullable<MemoryConfig[\"episodic\"]>,\n  name: string,\n): EpisodicMemory {\n  const store = episodicConfig.store ?? resolveDefaultStore();\n\n  if (!store) {\n    throw new Error(\n      `memory(\"${name}\"): episodic tier has no store — pass \\`episodic.store\\` or call \\`ai.config({ defaultStore })\\` at app boot before constructing the memory`,\n    );\n  }\n\n  return new EpisodicMemory(\n    episodicConfig.embedder,\n    store,\n    episodicConfig.namespace ?? DEFAULT_EPISODIC_NAMESPACE,\n    episodicConfig.recencyWeight ?? DEFAULT_RECENCY_WEIGHT,\n    episodicConfig.halfLifeMs ?? DEFAULT_HALF_LIFE_MS,\n    episodicConfig.now ?? (() => Date.now()),\n  );\n}\n\n/**\n * Resolve the procedural tier's store and build the tier with its\n * reinforcement knob. Throws at construction when no store is available.\n */\nfunction buildProceduralTier(\n  proceduralConfig: NonNullable<MemoryConfig[\"procedural\"]>,\n  name: string,\n): ProceduralMemory {\n  const store = proceduralConfig.store ?? resolveDefaultStore();\n\n  if (!store) {\n    throw new Error(\n      `memory(\"${name}\"): procedural tier has no store — pass \\`procedural.store\\` or call \\`ai.config({ defaultStore })\\` at app boot before constructing the memory`,\n    );\n  }\n\n  return new ProceduralMemory(\n    proceduralConfig.embedder,\n    store,\n    proceduralConfig.namespace ?? DEFAULT_PROCEDURAL_NAMESPACE,\n    proceduralConfig.reinforcementWeight ?? DEFAULT_REINFORCEMENT_WEIGHT,\n  );\n}\n\n/**\n * Guard that a tier referenced by config / a call is actually enabled,\n * failing fast with an actionable message instead of a downstream\n * `undefined` dereference.\n */\nfunction assertTierEnabled(tier: MemoryTier, tiers: Tiers, name: string): void {\n  if (tier === \"working\" && !tiers.working) {\n    throw new Error(\n      `memory(\"${name}\"): working tier is disabled — set \\`working: true\\` (the default) to use it`,\n    );\n  }\n\n  if (tier === \"semantic\" && !tiers.semantic) {\n    throw new Error(\n      `memory(\"${name}\"): semantic tier is not configured — pass \\`semantic\\` config to use it`,\n    );\n  }\n\n  if (tier === \"episodic\" && !tiers.episodic) {\n    throw new Error(\n      `memory(\"${name}\"): episodic tier is not configured — pass \\`episodic\\` config to use it`,\n    );\n  }\n\n  if (tier === \"procedural\" && !tiers.procedural) {\n    throw new Error(\n      `memory(\"${name}\"): procedural tier is not configured — pass \\`procedural\\` config to use it`,\n    );\n  }\n}\n"],"mappings":";;;;;;;AAiBA,MAAM,eAAe;AACrB,MAAM,6BAA6B;AACnC,MAAM,6BAA6B;AACnC,MAAM,+BAA+B;AACrC,MAAM,YAAY;AAClB,MAAM,oBAAoB;AAC1B,MAAM,yBAAyB;AAC/B,MAAM,uBAAuB,QAAc,KAAK;AAChD,MAAM,+BAA+B;;;;;;;;AASrC,MAAM,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgDlC,SAAgB,OAAO,SAAuB,CAAC,GAAmB;CAChE,MAAM,OAAO,OAAO,QAAQ;CAC5B,MAAM,gBAAgB,OAAO,WAAW;CACxC,MAAM,WAAW,OAAO,KAAK;CAC7B,MAAM,mBAAmB,OAAO,aAAa;CAE7C,MAAM,UACJ,kBAAkB,QACd,SACA,IAAI,cAAc,uBAAuB,eAAe,IAAI,CAAC;CAEnE,MAAM,WAAW,OAAO,WACpB,kBAAkB,OAAO,UAAU,IAAI,IACvC;CAEJ,MAAM,WAAW,OAAO,WACpB,kBAAkB,OAAO,UAAU,IAAI,IACvC;CAEJ,MAAM,aAAa,OAAO,aACtB,oBAAoB,OAAO,YAAY,IAAI,IAC3C;CAEJ,MAAM,QAAe;EAAE;EAAS;EAAU;EAAU;CAAW;CAE/D,IAAI,CAAC,WAAW,CAAC,YAAY,CAAC,YAAY,CAAC,YACzC,MAAM,IAAI,MACR,WAAW,KAAK,sKAClB;CAGF,MAAM,cAA0B,OAAO,eAAe;CAEtD,kBAAkB,aAAa,OAAO,IAAI;CAE1C,OAAO;EACL;EACA,MAAM,SAAS,OAAiD;GAC9D,MAAM,OAAO,MAAM,QAAQ,KAAK,IAAI,QAAQ,CAAC,KAAK;GAElD,MAAM,SAA0B,CAAC;GAEjC,KAAK,MAAM,QAAQ,MAAM;IACvB,MAAM,OAAO,KAAK,QAAQ;IAE1B,kBAAkB,MAAM,OAAO,IAAI;IAEnC,IAAI,SAAS,WAAW;KACtB,QAAS,SAAS,IAAI;KAEtB;IACF;IAEA,IAAI,SAAS,YAAY;KACvB,OAAO,KAAK,SAAU,SAAS,IAAI,CAAC;KAEpC;IACF;IAEA,IAAI,SAAS,YAAY;KACvB,OAAO,KAAK,SAAU,SAAS,IAAI,CAAC;KAEpC;IACF;IAEA,OAAO,KAAK,WAAY,SAAS,IAAI,CAAC;GACxC;GAEA,MAAM,QAAQ,IAAI,MAAM;EAC1B;EACA,MAAM,OACJ,OACA,UAAyB,CAAC,GACC;GAC3B,MAAM,IAAI,QAAQ,KAAK;GACvB,MAAM,YAAY,QAAQ,aAAa;GAEvC,IAAI,QAAQ,MACV,kBAAkB,QAAQ,MAAM,OAAO,IAAI;GAG7C,MAAM,SAAS,SACb,CAAC,QAAQ,QAAQ,QAAQ,SAAS;GAKpC,MAAM,QAAQ,QAAQ;GAEtB,MAAM,CAAC,aAAa,cAAc,cAAc,kBAC9C,MAAM,QAAQ,IAAI;IAChB,WAAW,MAAM,SAAS,IACtB,QAAQ,QAAQ,QAAQ,OAAO,GAAG,KAAK,CAAC,IACxC,QAAQ,QAAQ,CAAC,CAAqB;IAC1C,YAAY,MAAM,UAAU,IACxB,SAAS,OAAO,OAAO,GAAG,WAAW,KAAK,IAC1C,QAAQ,QAAQ,CAAC,CAAqB;IAC1C,YAAY,MAAM,UAAU,IACxB,SAAS,OAAO,OAAO,GAAG,WAAW,KAAK,IAC1C,QAAQ,QAAQ,CAAC,CAAqB;IAC1C,cAAc,MAAM,YAAY,IAC5B,WAAW,OAAO,OAAO,GAAG,WAAW,KAAK,IAC5C,QAAQ,QAAQ,CAAC,CAAqB;GAC5C,CAAC;GAEH,OAAO;IACL,GAAG;IACH,GAAG;IACH,GAAG;IACH,GAAG;GACL,CAAC,CACE,MAAM,OAAO,WAAW,OAAO,QAAQ,MAAM,KAAK,CAAC,CACnD,MAAM,GAAG,CAAC;EACf;EACA,MAAM,MAAM,MAAkC;GAC5C,MAAM,SAA0B,CAAC;GAEjC,IAAI,YAAY,CAAC,QAAQ,SAAS,YAChC,QAAQ,MAAM;GAGhB,IAAI,aAAa,CAAC,QAAQ,SAAS,aACjC,OAAO,KAAK,SAAS,MAAM,CAAC;GAG9B,IAAI,aAAa,CAAC,QAAQ,SAAS,aACjC,OAAO,KAAK,SAAS,MAAM,CAAC;GAG9B,IAAI,eAAe,CAAC,QAAQ,SAAS,eACnC,OAAO,KAAK,WAAW,MAAM,CAAC;GAGhC,MAAM,QAAQ,IAAI,MAAM;EAC1B;CACF;AACF;;;;;;;;;AAkBA,SAAS,uBACP,eACA,MACQ;CACR,MAAM,WACJ,kBAAkB,OACd,4BACC,cAAc,YAAY;CAEjC,IAAI,CAAC,OAAO,UAAU,QAAQ,KAAK,WAAW,GAC5C,MAAM,IAAI,MACR,WAAW,KAAK,mEAAmE,OAAO,QAAQ,GACpG;CAGF,OAAO;AACT;;;;;;;AAQA,SAAS,kBACP,gBACA,MACgB;CAChB,MAAM,QAAQ,eAAe,SAAS,oBAAoB;CAE1D,IAAI,CAAC,OACH,MAAM,IAAI,MACR,WAAW,KAAK,4IAClB;CAGF,OAAO,IAAI,eACT,eAAe,UACf,OACA,eAAe,aAAa,0BAC9B;AACF;;;;;;;AAQA,SAAS,kBACP,gBACA,MACgB;CAChB,MAAM,QAAQ,eAAe,SAAS,oBAAoB;CAE1D,IAAI,CAAC,OACH,MAAM,IAAI,MACR,WAAW,KAAK,4IAClB;CAGF,OAAO,IAAI,eACT,eAAe,UACf,OACA,eAAe,aAAa,4BAC5B,eAAe,iBAAiB,wBAChC,eAAe,cAAc,sBAC7B,eAAe,cAAc,KAAK,IAAI,EACxC;AACF;;;;;AAMA,SAAS,oBACP,kBACA,MACkB;CAClB,MAAM,QAAQ,iBAAiB,SAAS,oBAAoB;CAE5D,IAAI,CAAC,OACH,MAAM,IAAI,MACR,WAAW,KAAK,gJAClB;CAGF,OAAO,IAAI,iBACT,iBAAiB,UACjB,OACA,iBAAiB,aAAa,8BAC9B,iBAAiB,uBAAuB,4BAC1C;AACF;;;;;;AAOA,SAAS,kBAAkB,MAAkB,OAAc,MAAoB;CAC7E,IAAI,SAAS,aAAa,CAAC,MAAM,SAC/B,MAAM,IAAI,MACR,WAAW,KAAK,6EAClB;CAGF,IAAI,SAAS,cAAc,CAAC,MAAM,UAChC,MAAM,IAAI,MACR,WAAW,KAAK,yEAClB;CAGF,IAAI,SAAS,cAAc,CAAC,MAAM,UAChC,MAAM,IAAI,MACR,WAAW,KAAK,yEAClB;CAGF,IAAI,SAAS,gBAAgB,CAAC,MAAM,YAClC,MAAM,IAAI,MACR,WAAW,KAAK,6EAClB;AAEJ"}

package/esm/memory/procedural-memory.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"procedural-memory.mjs","names":[],"sources":["../../../../../../../ai/src/memory/procedural-memory.ts"],"sourcesContent":["import type { CacheDriver, CacheSimilarHit } from \"@warlock.js/cache\";\nimport type { EmbedderContract } from \"../contracts/embedder.contract\";\nimport type {\n  MemoryItem,\n  RecalledMemory,\n} from \"../contracts/memory/memory-item.type\";\nimport { deriveMemoryId } from \"./derive-id\";\n\n/**\n * Shape persisted per procedure. `uses` is the reinforcement counter —\n * how many times the procedure has been remembered/re-affirmed — and\n * feeds the reinforcement half of the blended recall score. The vector\n * lives in the driver's index, so it is not duplicated here.\n */\ntype StoredProcedure = {\n  id: string;\n  text: string;\n  uses: number;\n  /** Isolation key the procedure was written under; absent = the shared pool. */\n  scope?: string;\n  metadata?: Record<string, unknown>;\n};\n\n/**\n * Extra candidates pulled from `similar()` before re-ranking by the\n * reinforcement-blended score and slicing to `k` — reinforcement can\n * promote a well-worn procedure past a slightly-closer one-off, which the\n * raw top-`k` by similarity would miss.\n */\nconst RECALL_OVERSCAN = 5;\n\n/**\n * Procedural recall tier (memory core M2).\n *\n * Holds durable *how-to* knowledge — learned procedures, policies, and\n * playbooks — and retrieves the ones relevant to a query, **blended with\n * reinforcement** so procedures that have proven themselves (remembered /\n * re-affirmed more often) outrank one-offs at equal similarity. That\n * reinforcement weighting is the difference from the semantic tier (which\n * treats every fact equally): procedural memory gets *stronger with use*.\n *\n * Reinforcement is explicit and side-effect-free on read: re-remembering\n * a procedure (same id, or same text → same derived id) increments its\n * `uses`, so a caller strengthens a procedure by remembering it again\n * after a successful application. Recall never mutates.\n *\n * Like the other vector tiers it delegates similarity to the\n * `@warlock.js/cache` driver's `similar()`. The blended `score` stays in\n * `[0, 1]` so procedural hits merge and sort alongside the other tiers.\n *\n * Internal to the `memory()` factory — never exported on the package\n * surface.\n */\nexport class ProceduralMemory {\n  public constructor(\n    private readonly embedder: EmbedderContract,\n    private readonly store: CacheDriver<any, any>,\n    private readonly namespace: string,\n    private readonly reinforcementWeight: number,\n  ) {}\n\n  /**\n   * Embed the procedure text and index it, incrementing its `uses` when\n   * it already exists (reinforcement) or seeding it at `1` when new.\n   * Metadata on a reinforcing write wins; an omitted metadata keeps the\n   * prior value rather than wiping it.\n   */\n  public async remember(item: MemoryItem): Promise<void> {\n    const id = item.id ?? deriveMemoryId(item.text);\n    const { vector } = await this.embedder.embed(item.text);\n\n    const key = this.keyFor(id, item.scope);\n    const existing = await this.store.get<StoredProcedure>(key);\n    const uses = (existing?.uses ?? 0) + 1;\n\n    const value: StoredProcedure = {\n      id,\n      text: item.text,\n      uses,\n      scope: item.scope,\n      metadata: item.metadata ?? existing?.metadata,\n    };\n\n    await this.store.set(key, value, { vector });\n  }\n\n  /**\n   * Embed `query`, pull the nearest procedures clearing the similarity\n   * `threshold`, then re-rank each by a reinforcement-blended score and\n   * return the top `k`. The similarity floor still gates relevance;\n   * reinforcement only reorders procedures that already cleared it.\n   *\n   * Procedures written under a different `scope` (another tenant /\n   * session) are dropped here, before scoring and slicing, so they can\n   * neither leak nor consume a slot. An unscoped recall reads only\n   * unscoped procedures.\n   */\n  public async recall(\n    query: string,\n    k: number,\n    threshold: number,\n    scope?: string,\n  ): Promise<RecalledMemory[]> {\n    const { vector } = await this.embedder.embed(query);\n\n    const hits = await this.store.similar<StoredProcedure>(vector, {\n      topK: Math.max(k * RECALL_OVERSCAN, k),\n      threshold,\n    });\n\n    const prefix = `${this.namespace}.`;\n\n    return hits\n      .filter(\n        (hit: CacheSimilarHit<StoredProcedure>) =>\n          hit.key.startsWith(prefix) && hit.value?.scope === scope,\n      )\n      .map((hit: CacheSimilarHit<StoredProcedure>) => ({\n        id: hit.value.id,\n        text: hit.value.text,\n        tier: \"procedural\" as const,\n        score: this.blend(hit.score, hit.value.uses),\n        metadata: hit.value.metadata,\n      }))\n      .sort((first, second) => second.score - first.score)\n      .slice(0, k);\n  }\n\n  /** Drop every procedure written under this instance's namespace. */\n  public async clear(): Promise<void> {\n    await this.store.removeNamespace(this.namespace);\n  }\n\n  /**\n   * Combine raw similarity with a saturating reinforcement proxy:\n   * `(1 - w)·similarity + w·(uses / (uses + 1))`. A first-time procedure\n   * contributes `0.5`; each reinforcement nudges it toward `1` with\n   * diminishing returns. With `reinforcementWeight` 0 the score is pure\n   * similarity.\n   */\n  private blend(similarity: number, uses: number): number {\n    const reinforcement = uses / (uses + 1);\n\n    return (\n      (1 - this.reinforcementWeight) * similarity +\n      this.reinforcementWeight * reinforcement\n    );\n  }\n\n  /**\n   * Namespaced key for an entry — dot separator, matching `similar()`\n   * keys, plus a hashed scope segment so reinforcement counters never\n   * cross a scope boundary (one tenant re-affirming a procedure must not\n   * strengthen — or overwrite — another tenant's identical text).\n   * Unscoped keys keep their pre-4.15.0 shape.\n   */\n  private keyFor(id: string, scope?: string): string {\n    return scope === undefined\n      ? `${this.namespace}.${id}`\n      : `${this.namespace}.${deriveMemoryId(scope)}.${id}`;\n  }\n}\n"],"mappings":";;;;;;;;;AA6BA,MAAM,kBAAkB;;;;;;;;;;;;;;;;;;;;;;;AAwBxB,IAAa,mBAAb,MAA8B;CAC5B,AAAO,YACL,AAAiB,UACjB,AAAiB,OACjB,AAAiB,WACjB,AAAiB,qBACjB;EAJiB;EACA;EACA;EACA;CAChB;;;;;;;CAQH,MAAa,SAAS,MAAiC;EACrD,MAAM,KAAK,KAAK,MAAM,eAAe,KAAK,IAAI;EAC9C,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,MAAM,KAAK,IAAI;EAEtD,MAAM,MAAM,KAAK,OAAO,IAAI,KAAK,KAAK;EACtC,MAAM,WAAW,MAAM,KAAK,MAAM,IAAqB,GAAG;EAC1D,MAAM,QAAQ,UAAU,QAAQ,KAAK;EAErC,MAAM,QAAyB;GAC7B;GACA,MAAM,KAAK;GACX;GACA,OAAO,KAAK;GACZ,UAAU,KAAK,YAAY,UAAU;EACvC;EAEA,MAAM,KAAK,MAAM,IAAI,KAAK,OAAO,EAAE,OAAO,CAAC;CAC7C;;;;;;;;;;;;CAaA,MAAa,OACX,OACA,GACA,WACA,OAC2B;EAC3B,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,MAAM,KAAK;EAElD,MAAM,OAAO,MAAM,KAAK,MAAM,QAAyB,QAAQ;GAC7D,MAAM,KAAK,IAAI,IAAI,iBAAiB,CAAC;GACrC;EACF,CAAC;EAED,MAAM,SAAS,GAAG,KAAK,UAAU;EAEjC,OAAO,KACJ,QACE,QACC,IAAI,IAAI,WAAW,MAAM,KAAK,IAAI,OAAO,UAAU,KACvD,EACC,KAAK,SAA2C;GAC/C,IAAI,IAAI,MAAM;GACd,MAAM,IAAI,MAAM;GAChB,MAAM;GACN,OAAO,KAAK,MAAM,IAAI,OAAO,IAAI,MAAM,IAAI;GAC3C,UAAU,IAAI,MAAM;EACtB,EAAE,EACD,MAAM,OAAO,WAAW,OAAO,QAAQ,MAAM,KAAK,EAClD,MAAM,GAAG,CAAC;CACf;;CAGA,MAAa,QAAuB;EAClC,MAAM,KAAK,MAAM,gBAAgB,KAAK,SAAS;CACjD;;;;;;;;CASA,AAAQ,MAAM,YAAoB,MAAsB;EACtD,MAAM,gBAAgB,QAAQ,OAAO;EAErC,QACG,IAAI,KAAK,uBAAuB,aACjC,KAAK,sBAAsB;CAE/B;;;;;;;;CASA,AAAQ,OAAO,IAAY,OAAwB;EACjD,OAAO,UAAU,SACb,GAAG,KAAK,UAAU,GAAG,OACrB,GAAG,KAAK,UAAU,GAAG,eAAe,KAAK,EAAE,GAAG;CACpD;AACF"}
+{"version":3,"file":"procedural-memory.mjs","names":[],"sources":["../../../../../../../ai/src/memory/procedural-memory.ts"],"sourcesContent":["import type { CacheDriver, CacheSimilarHit } from \"@warlock.js/cache\";\nimport type { EmbedderContract } from \"../contracts/embedder.contract\";\nimport type {\n  MemoryItem,\n  RecalledMemory,\n} from \"../contracts/memory/memory-item.type\";\nimport { deriveMemoryId } from \"./derive-id\";\n\n/**\n * Shape persisted per procedure. `uses` is the reinforcement counter —\n * how many times the procedure has been remembered/re-affirmed — and\n * feeds the reinforcement half of the blended recall score. The vector\n * lives in the driver's index, so it is not duplicated here.\n */\ntype StoredProcedure = {\n  id: string;\n  text: string;\n  uses: number;\n  /** Isolation key the procedure was written under; absent = the shared pool. */\n  scope?: string;\n  metadata?: Record<string, unknown>;\n};\n\n/**\n * Extra candidates pulled from `similar()` before re-ranking by the\n * reinforcement-blended score and slicing to `k` — reinforcement can\n * promote a well-worn procedure past a slightly-closer one-off, which the\n * raw top-`k` by similarity would miss.\n */\nconst RECALL_OVERSCAN = 5;\n\n/**\n * Procedural recall tier (memory core M2).\n *\n * Holds durable *how-to* knowledge — learned procedures, policies, and\n * playbooks — and retrieves the ones relevant to a query, **blended with\n * reinforcement** so procedures that have proven themselves (remembered /\n * re-affirmed more often) outrank one-offs at equal similarity. That\n * reinforcement weighting is the difference from the semantic tier (which\n * treats every fact equally): procedural memory gets *stronger with use*.\n *\n * Reinforcement is explicit and side-effect-free on read: re-remembering\n * a procedure (same id, or same text → same derived id) increments its\n * `uses`, so a caller strengthens a procedure by remembering it again\n * after a successful application. Recall never mutates.\n *\n * Like the other vector tiers it delegates similarity to the\n * `@warlock.js/cache` driver's `similar()`. The blended `score` stays in\n * `[0, 1]` so procedural hits merge and sort alongside the other tiers.\n *\n * Internal to the `memory()` factory — never exported on the package\n * surface.\n */\nexport class ProceduralMemory {\n  public constructor(\n    private readonly embedder: EmbedderContract,\n    private readonly store: CacheDriver<any, any>,\n    private readonly namespace: string,\n    private readonly reinforcementWeight: number,\n  ) {}\n\n  /**\n   * Embed the procedure text and index it, incrementing its `uses` when\n   * it already exists (reinforcement) or seeding it at `1` when new.\n   * Metadata on a reinforcing write wins; an omitted metadata keeps the\n   * prior value rather than wiping it.\n   */\n  public async remember(item: MemoryItem): Promise<void> {\n    const id = item.id ?? deriveMemoryId(item.text);\n    const { vector } = await this.embedder.embed(item.text);\n\n    const key = this.keyFor(id, item.scope);\n    const existing = await this.store.get<StoredProcedure>(key);\n    const uses = (existing?.uses ?? 0) + 1;\n\n    const value: StoredProcedure = {\n      id,\n      text: item.text,\n      uses,\n      scope: item.scope,\n      metadata: item.metadata ?? existing?.metadata,\n    };\n\n    await this.store.set(key, value, { vector });\n  }\n\n  /**\n   * Embed `query`, pull the nearest procedures clearing the similarity\n   * `threshold`, then re-rank each by a reinforcement-blended score and\n   * return the top `k`. The similarity floor still gates relevance;\n   * reinforcement only reorders procedures that already cleared it.\n   *\n   * Procedures written under a different `scope` (another tenant /\n   * session) are dropped here, before scoring and slicing, so they can\n   * neither leak nor consume a slot. An unscoped recall reads only\n   * unscoped procedures.\n   */\n  public async recall(\n    query: string,\n    k: number,\n    threshold: number,\n    scope?: string,\n  ): Promise<RecalledMemory[]> {\n    const { vector } = await this.embedder.embed(query);\n\n    const hits = await this.store.similar<StoredProcedure>(vector, {\n      topK: Math.max(k * RECALL_OVERSCAN, k),\n      threshold,\n    });\n\n    const prefix = `${this.namespace}.`;\n\n    return hits\n      .filter(\n        (hit: CacheSimilarHit<StoredProcedure>) =>\n          hit.key.startsWith(prefix) && hit.value?.scope === scope,\n      )\n      .map((hit: CacheSimilarHit<StoredProcedure>) => ({\n        id: hit.value.id,\n        text: hit.value.text,\n        tier: \"procedural\" as const,\n        score: this.blend(hit.score, hit.value.uses),\n        metadata: hit.value.metadata,\n      }))\n      .sort((first, second) => second.score - first.score)\n      .slice(0, k);\n  }\n\n  /** Drop every procedure written under this instance's namespace. */\n  public async clear(): Promise<void> {\n    await this.store.removeNamespace(this.namespace);\n  }\n\n  /**\n   * Combine raw similarity with a saturating reinforcement proxy:\n   * `(1 - w)·similarity + w·(uses / (uses + 1))`. A first-time procedure\n   * contributes `0.5`; each reinforcement nudges it toward `1` with\n   * diminishing returns. With `reinforcementWeight` 0 the score is pure\n   * similarity.\n   */\n  private blend(similarity: number, uses: number): number {\n    const reinforcement = uses / (uses + 1);\n\n    return (\n      (1 - this.reinforcementWeight) * similarity +\n      this.reinforcementWeight * reinforcement\n    );\n  }\n\n  /**\n   * Namespaced key for an entry — dot separator, matching `similar()`\n   * keys, plus a hashed scope segment so reinforcement counters never\n   * cross a scope boundary (one tenant re-affirming a procedure must not\n   * strengthen — or overwrite — another tenant's identical text).\n   * Unscoped keys keep their pre-4.15.0 shape.\n   */\n  private keyFor(id: string, scope?: string): string {\n    return scope === undefined\n      ? `${this.namespace}.${id}`\n      : `${this.namespace}.${deriveMemoryId(scope)}.${id}`;\n  }\n}\n"],"mappings":";;;;;;;;;AA6BA,MAAM,kBAAkB;;;;;;;;;;;;;;;;;;;;;;;AAwBxB,IAAa,mBAAb,MAA8B;CAC5B,AAAO,YACL,AAAiB,UACjB,AAAiB,OACjB,AAAiB,WACjB,AAAiB,qBACjB;EAJiB;EACA;EACA;EACA;CAChB;;;;;;;CAQH,MAAa,SAAS,MAAiC;EACrD,MAAM,KAAK,KAAK,MAAM,eAAe,KAAK,IAAI;EAC9C,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,MAAM,KAAK,IAAI;EAEtD,MAAM,MAAM,KAAK,OAAO,IAAI,KAAK,KAAK;EACtC,MAAM,WAAW,MAAM,KAAK,MAAM,IAAqB,GAAG;EAC1D,MAAM,QAAQ,UAAU,QAAQ,KAAK;EAErC,MAAM,QAAyB;GAC7B;GACA,MAAM,KAAK;GACX;GACA,OAAO,KAAK;GACZ,UAAU,KAAK,YAAY,UAAU;EACvC;EAEA,MAAM,KAAK,MAAM,IAAI,KAAK,OAAO,EAAE,OAAO,CAAC;CAC7C;;;;;;;;;;;;CAaA,MAAa,OACX,OACA,GACA,WACA,OAC2B;EAC3B,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,MAAM,KAAK;EAElD,MAAM,OAAO,MAAM,KAAK,MAAM,QAAyB,QAAQ;GAC7D,MAAM,KAAK,IAAI,IAAI,iBAAiB,CAAC;GACrC;EACF,CAAC;EAED,MAAM,SAAS,GAAG,KAAK,UAAU;EAEjC,OAAO,KACJ,QACE,QACC,IAAI,IAAI,WAAW,MAAM,KAAK,IAAI,OAAO,UAAU,KACvD,CAAC,CACA,KAAK,SAA2C;GAC/C,IAAI,IAAI,MAAM;GACd,MAAM,IAAI,MAAM;GAChB,MAAM;GACN,OAAO,KAAK,MAAM,IAAI,OAAO,IAAI,MAAM,IAAI;GAC3C,UAAU,IAAI,MAAM;EACtB,EAAE,CAAC,CACF,MAAM,OAAO,WAAW,OAAO,QAAQ,MAAM,KAAK,CAAC,CACnD,MAAM,GAAG,CAAC;CACf;;CAGA,MAAa,QAAuB;EAClC,MAAM,KAAK,MAAM,gBAAgB,KAAK,SAAS;CACjD;;;;;;;;CASA,AAAQ,MAAM,YAAoB,MAAsB;EACtD,MAAM,gBAAgB,QAAQ,OAAO;EAErC,QACG,IAAI,KAAK,uBAAuB,aACjC,KAAK,sBAAsB;CAE/B;;;;;;;;CASA,AAAQ,OAAO,IAAY,OAAwB;EACjD,OAAO,UAAU,SACb,GAAG,KAAK,UAAU,GAAG,OACrB,GAAG,KAAK,UAAU,GAAG,eAAe,KAAK,EAAE,GAAG;CACpD;AACF"}

package/esm/memory/semantic-memory.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"semantic-memory.mjs","names":[],"sources":["../../../../../../../ai/src/memory/semantic-memory.ts"],"sourcesContent":["import type { CacheDriver, CacheSimilarHit } from \"@warlock.js/cache\";\nimport type { EmbedderContract } from \"../contracts/embedder.contract\";\nimport type {\n  MemoryItem,\n  RecalledMemory,\n} from \"../contracts/memory/memory-item.type\";\nimport { deriveMemoryId } from \"./derive-id\";\n\n/**\n * Extra candidates pulled from `similar()` on a SCOPED recall before the\n * scope filter runs — the driver's ranking spans every scope in the\n * index, so a bare top-`k` can come back entirely foreign. Mirrors the\n * episodic / procedural tiers' overscan constant.\n */\nconst RECALL_OVERSCAN = 5;\n\n/**\n * Shape persisted per semantic memory in the cache driver. The vector\n * itself is stored by the driver's own index (passed via\n * `set({ vector })`), so it is not duplicated in the value.\n */\ntype StoredMemory = {\n  id: string;\n  text: string;\n  /** Isolation key the entry was written under; absent = the shared pool. */\n  scope?: string;\n  metadata?: Record<string, unknown>;\n};\n\n/**\n * Semantic recall tier (memory core M1).\n *\n * Owns: embedding remembered text, writing it to a `@warlock.js/cache`\n * driver with `set({ vector })`, and retrieving by cosine similarity via\n * the driver's `similar()`. Does NOT own: the similarity algorithm or\n * the ANN index — those belong to the cache driver. This mirrors the\n * delegation model of `middleware/builtins/semantic-cache.ts`: memory is\n * embedding-agnostic and store-agnostic, gluing an {@link EmbedderContract}\n * to a {@link CacheDriver}.\n *\n * The driver may be shared across memory instances, so every key carries\n * the configured `namespace` and recall filters hits to that prefix —\n * foreign entries indexed by another instance never leak into a query.\n *\n * Internal to the `memory()` factory — never exported on the package\n * surface.\n */\nexport class SemanticMemory {\n  public constructor(\n    private readonly embedder: EmbedderContract,\n    private readonly store: CacheDriver<any, any>,\n    private readonly namespace: string,\n  ) {}\n\n  /**\n   * Embed the item's text and index it under a namespaced, id-derived\n   * key. Re-remembering the same id overwrites the prior vector +\n   * value (the driver upserts by key).\n   */\n  public async remember(item: MemoryItem): Promise<void> {\n    const id = item.id ?? deriveMemoryId(item.text);\n    const { vector } = await this.embedder.embed(item.text);\n\n    const value: StoredMemory = {\n      id,\n      text: item.text,\n      scope: item.scope,\n      metadata: item.metadata,\n    };\n\n    await this.store.set(this.keyFor(id, item.scope), value, { vector });\n  }\n\n  /**\n   * Embed `query`, ask the driver for the `k` nearest entries clearing\n   * `threshold`, and return those within this instance's namespace AND\n   * this call's `scope` as scored {@link RecalledMemory}. Hits indexed\n   * under a different namespace (a shared driver) or a different scope\n   * (another tenant / session) are filtered out here, before the caller\n   * ever sees them — an unscoped recall reads only unscoped entries.\n   */\n  public async recall(\n    query: string,\n    k: number,\n    threshold: number,\n    scope?: string,\n  ): Promise<RecalledMemory[]> {\n    const { vector } = await this.embedder.embed(query);\n\n    // A scoped recall overscans: the driver ranks across every scope in\n    // the index, so a plain top-`k` could be filled entirely by foreign\n    // scopes and starve this one. Pull extra candidates, filter, then cap.\n    const hits = await this.store.similar<StoredMemory>(vector, {\n      topK: scope === undefined ? k : Math.max(k * RECALL_OVERSCAN, k),\n      threshold,\n    });\n\n    const prefix = `${this.namespace}.`;\n\n    return hits\n      .filter(\n        (hit: CacheSimilarHit<StoredMemory>) =>\n          hit.key.startsWith(prefix) && hit.value?.scope === scope,\n      )\n      .map((hit: CacheSimilarHit<StoredMemory>) => ({\n        id: hit.value.id,\n        text: hit.value.text,\n        tier: \"semantic\" as const,\n        score: hit.score,\n        metadata: hit.value.metadata,\n      }))\n      .slice(0, Math.max(0, k));\n  }\n\n  /** Drop every semantic entry written under this instance's namespace. */\n  public async clear(): Promise<void> {\n    await this.store.removeNamespace(this.namespace);\n  }\n\n  /**\n   * Namespaced key for an entry. The cache's `parseKey` normalizes `:`\n   * to `.`, so a dot separator keeps the prefix used here aligned with\n   * the `hit.key` the driver returns from `similar()`.\n   *\n   * A scoped entry gets an extra hashed segment so two scopes writing\n   * identical text (same derived id) don't overwrite each other; the\n   * unscoped key shape is unchanged, so entries written before 4.15.0\n   * still resolve. The hash is a write-separation device only — recall\n   * authorization is the exact `value.scope` equality check, so even a\n   * hash collision cannot widen what a scope can read.\n   */\n  private keyFor(id: string, scope?: string): string {\n    return scope === undefined\n      ? `${this.namespace}.${id}`\n      : `${this.namespace}.${deriveMemoryId(scope)}.${id}`;\n  }\n}\n"],"mappings":";;;;;;;;;AAcA,MAAM,kBAAkB;;;;;;;;;;;;;;;;;;;AAiCxB,IAAa,iBAAb,MAA4B;CAC1B,AAAO,YACL,AAAiB,UACjB,AAAiB,OACjB,AAAiB,WACjB;EAHiB;EACA;EACA;CAChB;;;;;;CAOH,MAAa,SAAS,MAAiC;EACrD,MAAM,KAAK,KAAK,MAAM,eAAe,KAAK,IAAI;EAC9C,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,MAAM,KAAK,IAAI;EAEtD,MAAM,QAAsB;GAC1B;GACA,MAAM,KAAK;GACX,OAAO,KAAK;GACZ,UAAU,KAAK;EACjB;EAEA,MAAM,KAAK,MAAM,IAAI,KAAK,OAAO,IAAI,KAAK,KAAK,GAAG,OAAO,EAAE,OAAO,CAAC;CACrE;;;;;;;;;CAUA,MAAa,OACX,OACA,GACA,WACA,OAC2B;EAC3B,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,MAAM,KAAK;EAKlD,MAAM,OAAO,MAAM,KAAK,MAAM,QAAsB,QAAQ;GAC1D,MAAM,UAAU,SAAY,IAAI,KAAK,IAAI,IAAI,iBAAiB,CAAC;GAC/D;EACF,CAAC;EAED,MAAM,SAAS,GAAG,KAAK,UAAU;EAEjC,OAAO,KACJ,QACE,QACC,IAAI,IAAI,WAAW,MAAM,KAAK,IAAI,OAAO,UAAU,KACvD,EACC,KAAK,SAAwC;GAC5C,IAAI,IAAI,MAAM;GACd,MAAM,IAAI,MAAM;GAChB,MAAM;GACN,OAAO,IAAI;GACX,UAAU,IAAI,MAAM;EACtB,EAAE,EACD,MAAM,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;CAC5B;;CAGA,MAAa,QAAuB;EAClC,MAAM,KAAK,MAAM,gBAAgB,KAAK,SAAS;CACjD;;;;;;;;;;;;;CAcA,AAAQ,OAAO,IAAY,OAAwB;EACjD,OAAO,UAAU,SACb,GAAG,KAAK,UAAU,GAAG,OACrB,GAAG,KAAK,UAAU,GAAG,eAAe,KAAK,EAAE,GAAG;CACpD;AACF"}
+{"version":3,"file":"semantic-memory.mjs","names":[],"sources":["../../../../../../../ai/src/memory/semantic-memory.ts"],"sourcesContent":["import type { CacheDriver, CacheSimilarHit } from \"@warlock.js/cache\";\nimport type { EmbedderContract } from \"../contracts/embedder.contract\";\nimport type {\n  MemoryItem,\n  RecalledMemory,\n} from \"../contracts/memory/memory-item.type\";\nimport { deriveMemoryId } from \"./derive-id\";\n\n/**\n * Extra candidates pulled from `similar()` on a SCOPED recall before the\n * scope filter runs — the driver's ranking spans every scope in the\n * index, so a bare top-`k` can come back entirely foreign. Mirrors the\n * episodic / procedural tiers' overscan constant.\n */\nconst RECALL_OVERSCAN = 5;\n\n/**\n * Shape persisted per semantic memory in the cache driver. The vector\n * itself is stored by the driver's own index (passed via\n * `set({ vector })`), so it is not duplicated in the value.\n */\ntype StoredMemory = {\n  id: string;\n  text: string;\n  /** Isolation key the entry was written under; absent = the shared pool. */\n  scope?: string;\n  metadata?: Record<string, unknown>;\n};\n\n/**\n * Semantic recall tier (memory core M1).\n *\n * Owns: embedding remembered text, writing it to a `@warlock.js/cache`\n * driver with `set({ vector })`, and retrieving by cosine similarity via\n * the driver's `similar()`. Does NOT own: the similarity algorithm or\n * the ANN index — those belong to the cache driver. This mirrors the\n * delegation model of `middleware/builtins/semantic-cache.ts`: memory is\n * embedding-agnostic and store-agnostic, gluing an {@link EmbedderContract}\n * to a {@link CacheDriver}.\n *\n * The driver may be shared across memory instances, so every key carries\n * the configured `namespace` and recall filters hits to that prefix —\n * foreign entries indexed by another instance never leak into a query.\n *\n * Internal to the `memory()` factory — never exported on the package\n * surface.\n */\nexport class SemanticMemory {\n  public constructor(\n    private readonly embedder: EmbedderContract,\n    private readonly store: CacheDriver<any, any>,\n    private readonly namespace: string,\n  ) {}\n\n  /**\n   * Embed the item's text and index it under a namespaced, id-derived\n   * key. Re-remembering the same id overwrites the prior vector +\n   * value (the driver upserts by key).\n   */\n  public async remember(item: MemoryItem): Promise<void> {\n    const id = item.id ?? deriveMemoryId(item.text);\n    const { vector } = await this.embedder.embed(item.text);\n\n    const value: StoredMemory = {\n      id,\n      text: item.text,\n      scope: item.scope,\n      metadata: item.metadata,\n    };\n\n    await this.store.set(this.keyFor(id, item.scope), value, { vector });\n  }\n\n  /**\n   * Embed `query`, ask the driver for the `k` nearest entries clearing\n   * `threshold`, and return those within this instance's namespace AND\n   * this call's `scope` as scored {@link RecalledMemory}. Hits indexed\n   * under a different namespace (a shared driver) or a different scope\n   * (another tenant / session) are filtered out here, before the caller\n   * ever sees them — an unscoped recall reads only unscoped entries.\n   */\n  public async recall(\n    query: string,\n    k: number,\n    threshold: number,\n    scope?: string,\n  ): Promise<RecalledMemory[]> {\n    const { vector } = await this.embedder.embed(query);\n\n    // A scoped recall overscans: the driver ranks across every scope in\n    // the index, so a plain top-`k` could be filled entirely by foreign\n    // scopes and starve this one. Pull extra candidates, filter, then cap.\n    const hits = await this.store.similar<StoredMemory>(vector, {\n      topK: scope === undefined ? k : Math.max(k * RECALL_OVERSCAN, k),\n      threshold,\n    });\n\n    const prefix = `${this.namespace}.`;\n\n    return hits\n      .filter(\n        (hit: CacheSimilarHit<StoredMemory>) =>\n          hit.key.startsWith(prefix) && hit.value?.scope === scope,\n      )\n      .map((hit: CacheSimilarHit<StoredMemory>) => ({\n        id: hit.value.id,\n        text: hit.value.text,\n        tier: \"semantic\" as const,\n        score: hit.score,\n        metadata: hit.value.metadata,\n      }))\n      .slice(0, Math.max(0, k));\n  }\n\n  /** Drop every semantic entry written under this instance's namespace. */\n  public async clear(): Promise<void> {\n    await this.store.removeNamespace(this.namespace);\n  }\n\n  /**\n   * Namespaced key for an entry. The cache's `parseKey` normalizes `:`\n   * to `.`, so a dot separator keeps the prefix used here aligned with\n   * the `hit.key` the driver returns from `similar()`.\n   *\n   * A scoped entry gets an extra hashed segment so two scopes writing\n   * identical text (same derived id) don't overwrite each other; the\n   * unscoped key shape is unchanged, so entries written before 4.15.0\n   * still resolve. The hash is a write-separation device only — recall\n   * authorization is the exact `value.scope` equality check, so even a\n   * hash collision cannot widen what a scope can read.\n   */\n  private keyFor(id: string, scope?: string): string {\n    return scope === undefined\n      ? `${this.namespace}.${id}`\n      : `${this.namespace}.${deriveMemoryId(scope)}.${id}`;\n  }\n}\n"],"mappings":";;;;;;;;;AAcA,MAAM,kBAAkB;;;;;;;;;;;;;;;;;;;AAiCxB,IAAa,iBAAb,MAA4B;CAC1B,AAAO,YACL,AAAiB,UACjB,AAAiB,OACjB,AAAiB,WACjB;EAHiB;EACA;EACA;CAChB;;;;;;CAOH,MAAa,SAAS,MAAiC;EACrD,MAAM,KAAK,KAAK,MAAM,eAAe,KAAK,IAAI;EAC9C,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,MAAM,KAAK,IAAI;EAEtD,MAAM,QAAsB;GAC1B;GACA,MAAM,KAAK;GACX,OAAO,KAAK;GACZ,UAAU,KAAK;EACjB;EAEA,MAAM,KAAK,MAAM,IAAI,KAAK,OAAO,IAAI,KAAK,KAAK,GAAG,OAAO,EAAE,OAAO,CAAC;CACrE;;;;;;;;;CAUA,MAAa,OACX,OACA,GACA,WACA,OAC2B;EAC3B,MAAM,EAAE,WAAW,MAAM,KAAK,SAAS,MAAM,KAAK;EAKlD,MAAM,OAAO,MAAM,KAAK,MAAM,QAAsB,QAAQ;GAC1D,MAAM,UAAU,SAAY,IAAI,KAAK,IAAI,IAAI,iBAAiB,CAAC;GAC/D;EACF,CAAC;EAED,MAAM,SAAS,GAAG,KAAK,UAAU;EAEjC,OAAO,KACJ,QACE,QACC,IAAI,IAAI,WAAW,MAAM,KAAK,IAAI,OAAO,UAAU,KACvD,CAAC,CACA,KAAK,SAAwC;GAC5C,IAAI,IAAI,MAAM;GACd,MAAM,IAAI,MAAM;GAChB,MAAM;GACN,OAAO,IAAI;GACX,UAAU,IAAI,MAAM;EACtB,EAAE,CAAC,CACF,MAAM,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;CAC5B;;CAGA,MAAa,QAAuB;EAClC,MAAM,KAAK,MAAM,gBAAgB,KAAK,SAAS;CACjD;;;;;;;;;;;;;CAcA,AAAQ,OAAO,IAAY,OAAwB;EACjD,OAAO,UAAU,SACb,GAAG,KAAK,UAAU,GAAG,OACrB,GAAG,KAAK,UAAU,GAAG,eAAe,KAAK,EAAE,GAAG;CACpD;AACF"}

package/esm/memory/working-memory.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"working-memory.mjs","names":[],"sources":["../../../../../../../ai/src/memory/working-memory.ts"],"sourcesContent":["import type {\n  MemoryItem,\n  RecalledMemory,\n} from \"../contracts/memory/memory-item.type\";\nimport { deriveMemoryId } from \"./derive-id\";\n\n/**\n * In-run working memory — the volatile scratch tier (memory core M1).\n *\n * Owns: an insertion-ordered buffer of remembered items keyed by id,\n * with overwrite-in-place on a repeated id. Does NOT own: durability,\n * cross-process sharing, embeddings, or similarity — working memory is\n * a plain in-process buffer the orchestrator threads across the turns of\n * a single run.\n *\n * Recall here is not semantic: with no vector index, \"relevant\" reduces\n * to \"recent.\" `recall()` returns the most-recently-remembered items\n * first, each scored on a `[0, 1]` recency proxy so a caller can merge\n * working hits with semantic hits and sort on one `score` field.\n *\n * **Bounded (4.15.0).** The buffer holds at most `maxItems` entries\n * across every scope; the oldest-written entry is evicted on overflow\n * (FIFO). The tier lives in process memory for the lifetime of the\n * `memory()` instance — which the orchestrator resolves once and reuses\n * for every session — so an unbounded buffer was a memory-exhaustion\n * vector for any long-lived, internet-reachable deployment.\n *\n * Internal to the `memory()` factory — never exported on the package\n * surface.\n */\nexport class WorkingMemory {\n  /**\n   * Hard ceiling on buffered entries, across all scopes. Enforced on\n   * every `remember()`; see {@link evictOverflow} for the policy.\n   */\n  private readonly maxItems: number;\n\n  public constructor(maxItems: number) {\n    this.maxItems = maxItems;\n  }\n\n  /**\n   * Scoped key → entry. A `Map` preserves insertion order, so iteration\n   * yields oldest-first; recall reverses it for most-recent-first.\n   *\n   * The map key folds in the item's `scope` so two scopes remembering\n   * identical text (same derived id) stay two independent entries\n   * instead of clobbering one another; the entry keeps its logical `id`\n   * and its `scope` so recall can filter and still report the id the\n   * caller knows.\n   */\n  private readonly entries = new Map<\n    string,\n    {\n      id: string;\n      text: string;\n      scope?: string;\n      metadata?: Record<string, unknown>;\n    }\n  >();\n\n  /**\n   * Append an item to the buffer (or overwrite the entry sharing its\n   * id *within the same scope*). Re-inserting an existing key keeps its\n   * original position; delete + set would move it to the end and lie\n   * about recency, so the value is updated in place.\n   *\n   * Overflowing `maxItems` evicts from the front — see\n   * {@link evictOverflow}.\n   */\n  public remember(item: MemoryItem): void {\n    const id = item.id ?? deriveMemoryId(item.text);\n\n    this.entries.set(scopedKey(item.scope, id), {\n      id,\n      text: item.text,\n      scope: item.scope,\n      metadata: item.metadata,\n    });\n\n    this.evictOverflow();\n  }\n\n  /**\n   * Enforce the size bound by dropping oldest-written entries first\n   * (FIFO over the `Map`'s insertion order).\n   *\n   * **Why FIFO, not LRU.** Recall here is a pure recency proxy — it\n   * reverses insertion order and slices the newest `k` — and never\n   * reorders anything, so the front of the buffer is by construction the\n   * region recall reaches last. FIFO therefore evicts exactly the\n   * entries a bounded recall would never have returned. True LRU would\n   * need read-time reordering, which would also rewrite the `score`\n   * every recall reports (a re-read entry would masquerade as freshly\n   * remembered), trading a real correctness property for no gain.\n   *\n   * **Known limitation (documented, not a regression).** The bound is\n   * global, not per-scope: a session writing heavily can push another\n   * session's older entries out of the buffer. That is a recall-quality\n   * degradation on a volatile scratch tier, never a disclosure — the\n   * scope filter in {@link recall} still applies — and a per-scope quota\n   * would not help anyway, since an attacker holding many sessions\n   * evicts through the global bound regardless. Durable recall belongs\n   * in the semantic / episodic tiers.\n   */\n  private evictOverflow(): void {\n    while (this.entries.size > this.maxItems) {\n      const oldest = this.entries.keys().next();\n\n      if (oldest.done) {\n        return;\n      }\n\n      this.entries.delete(oldest.value);\n    }\n  }\n\n  /**\n   * Return up to `k` most-recently-remembered items *within `scope`*,\n   * newest first. The scope match is exact equality (an unscoped recall\n   * sees only unscoped entries) and is applied BEFORE the slice, so a\n   * foreign scope's entries can never consume a slot or leak out.\n   *\n   * The `score` is a linear recency proxy: the newest item scores `1`,\n   * the oldest of the returned slice trends toward `0`. Working memory\n   * ignores any similarity threshold — it has no vector to compare.\n   */\n  public recall(k: number, scope?: string): RecalledMemory[] {\n    const ordered = [...this.entries.values()]\n      .reverse()\n      .filter((entry) => entry.scope === scope);\n\n    const slice = ordered.slice(0, Math.max(0, k));\n\n    return slice.map((entry, index) => ({\n      id: entry.id,\n      text: entry.text,\n      tier: \"working\" as const,\n      score: slice.length <= 1 ? 1 : 1 - index / slice.length,\n      metadata: entry.metadata,\n    }));\n  }\n\n  /** Drop every working-tier entry, across every scope. */\n  public clear(): void {\n    this.entries.clear();\n  }\n}\n\n/**\n * Map key for a buffer entry: the isolation `scope` (empty for the\n * unscoped pool) length-prefixed and joined to the logical id. The\n * length prefix makes the encoding injective — no crafted scope/id pair\n * can collide with a different scope's entry the way a plain `:` join\n * would allow.\n */\nfunction scopedKey(scope: string | undefined, id: string): string {\n  return `${scope?.length ?? 0}:${scope ?? \"\"}:${id}`;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AA8BA,IAAa,gBAAb,MAA2B;CAOzB,AAAO,YAAY,UAAkB;iCAcV,IAAI,IAQ7B;EArBA,KAAK,WAAW;CAClB;;;;;;;;;;CA+BA,AAAO,SAAS,MAAwB;EACtC,MAAM,KAAK,KAAK,MAAM,eAAe,KAAK,IAAI;EAE9C,KAAK,QAAQ,IAAI,UAAU,KAAK,OAAO,EAAE,GAAG;GAC1C;GACA,MAAM,KAAK;GACX,OAAO,KAAK;GACZ,UAAU,KAAK;EACjB,CAAC;EAED,KAAK,cAAc;CACrB;;;;;;;;;;;;;;;;;;;;;;;CAwBA,AAAQ,gBAAsB;EAC5B,OAAO,KAAK,QAAQ,OAAO,KAAK,UAAU;GACxC,MAAM,SAAS,KAAK,QAAQ,KAAK,EAAE,KAAK;GAExC,IAAI,OAAO,MACT;GAGF,KAAK,QAAQ,OAAO,OAAO,KAAK;EAClC;CACF;;;;;;;;;;;CAYA,AAAO,OAAO,GAAW,OAAkC;EAKzD,MAAM,QAJU,CAAC,GAAG,KAAK,QAAQ,OAAO,CAAC,EACtC,QAAQ,EACR,QAAQ,UAAU,MAAM,UAAU,KAEjB,EAAE,MAAM,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;EAE7C,OAAO,MAAM,KAAK,OAAO,WAAW;GAClC,IAAI,MAAM;GACV,MAAM,MAAM;GACZ,MAAM;GACN,OAAO,MAAM,UAAU,IAAI,IAAI,IAAI,QAAQ,MAAM;GACjD,UAAU,MAAM;EAClB,EAAE;CACJ;;CAGA,AAAO,QAAc;EACnB,KAAK,QAAQ,MAAM;CACrB;AACF;;;;;;;;AASA,SAAS,UAAU,OAA2B,IAAoB;CAChE,OAAO,GAAG,OAAO,UAAU,EAAE,GAAG,SAAS,GAAG,GAAG;AACjD"}
+{"version":3,"file":"working-memory.mjs","names":[],"sources":["../../../../../../../ai/src/memory/working-memory.ts"],"sourcesContent":["import type {\n  MemoryItem,\n  RecalledMemory,\n} from \"../contracts/memory/memory-item.type\";\nimport { deriveMemoryId } from \"./derive-id\";\n\n/**\n * In-run working memory — the volatile scratch tier (memory core M1).\n *\n * Owns: an insertion-ordered buffer of remembered items keyed by id,\n * with overwrite-in-place on a repeated id. Does NOT own: durability,\n * cross-process sharing, embeddings, or similarity — working memory is\n * a plain in-process buffer the orchestrator threads across the turns of\n * a single run.\n *\n * Recall here is not semantic: with no vector index, \"relevant\" reduces\n * to \"recent.\" `recall()` returns the most-recently-remembered items\n * first, each scored on a `[0, 1]` recency proxy so a caller can merge\n * working hits with semantic hits and sort on one `score` field.\n *\n * **Bounded (4.15.0).** The buffer holds at most `maxItems` entries\n * across every scope; the oldest-written entry is evicted on overflow\n * (FIFO). The tier lives in process memory for the lifetime of the\n * `memory()` instance — which the orchestrator resolves once and reuses\n * for every session — so an unbounded buffer was a memory-exhaustion\n * vector for any long-lived, internet-reachable deployment.\n *\n * Internal to the `memory()` factory — never exported on the package\n * surface.\n */\nexport class WorkingMemory {\n  /**\n   * Hard ceiling on buffered entries, across all scopes. Enforced on\n   * every `remember()`; see {@link evictOverflow} for the policy.\n   */\n  private readonly maxItems: number;\n\n  public constructor(maxItems: number) {\n    this.maxItems = maxItems;\n  }\n\n  /**\n   * Scoped key → entry. A `Map` preserves insertion order, so iteration\n   * yields oldest-first; recall reverses it for most-recent-first.\n   *\n   * The map key folds in the item's `scope` so two scopes remembering\n   * identical text (same derived id) stay two independent entries\n   * instead of clobbering one another; the entry keeps its logical `id`\n   * and its `scope` so recall can filter and still report the id the\n   * caller knows.\n   */\n  private readonly entries = new Map<\n    string,\n    {\n      id: string;\n      text: string;\n      scope?: string;\n      metadata?: Record<string, unknown>;\n    }\n  >();\n\n  /**\n   * Append an item to the buffer (or overwrite the entry sharing its\n   * id *within the same scope*). Re-inserting an existing key keeps its\n   * original position; delete + set would move it to the end and lie\n   * about recency, so the value is updated in place.\n   *\n   * Overflowing `maxItems` evicts from the front — see\n   * {@link evictOverflow}.\n   */\n  public remember(item: MemoryItem): void {\n    const id = item.id ?? deriveMemoryId(item.text);\n\n    this.entries.set(scopedKey(item.scope, id), {\n      id,\n      text: item.text,\n      scope: item.scope,\n      metadata: item.metadata,\n    });\n\n    this.evictOverflow();\n  }\n\n  /**\n   * Enforce the size bound by dropping oldest-written entries first\n   * (FIFO over the `Map`'s insertion order).\n   *\n   * **Why FIFO, not LRU.** Recall here is a pure recency proxy — it\n   * reverses insertion order and slices the newest `k` — and never\n   * reorders anything, so the front of the buffer is by construction the\n   * region recall reaches last. FIFO therefore evicts exactly the\n   * entries a bounded recall would never have returned. True LRU would\n   * need read-time reordering, which would also rewrite the `score`\n   * every recall reports (a re-read entry would masquerade as freshly\n   * remembered), trading a real correctness property for no gain.\n   *\n   * **Known limitation (documented, not a regression).** The bound is\n   * global, not per-scope: a session writing heavily can push another\n   * session's older entries out of the buffer. That is a recall-quality\n   * degradation on a volatile scratch tier, never a disclosure — the\n   * scope filter in {@link recall} still applies — and a per-scope quota\n   * would not help anyway, since an attacker holding many sessions\n   * evicts through the global bound regardless. Durable recall belongs\n   * in the semantic / episodic tiers.\n   */\n  private evictOverflow(): void {\n    while (this.entries.size > this.maxItems) {\n      const oldest = this.entries.keys().next();\n\n      if (oldest.done) {\n        return;\n      }\n\n      this.entries.delete(oldest.value);\n    }\n  }\n\n  /**\n   * Return up to `k` most-recently-remembered items *within `scope`*,\n   * newest first. The scope match is exact equality (an unscoped recall\n   * sees only unscoped entries) and is applied BEFORE the slice, so a\n   * foreign scope's entries can never consume a slot or leak out.\n   *\n   * The `score` is a linear recency proxy: the newest item scores `1`,\n   * the oldest of the returned slice trends toward `0`. Working memory\n   * ignores any similarity threshold — it has no vector to compare.\n   */\n  public recall(k: number, scope?: string): RecalledMemory[] {\n    const ordered = [...this.entries.values()]\n      .reverse()\n      .filter((entry) => entry.scope === scope);\n\n    const slice = ordered.slice(0, Math.max(0, k));\n\n    return slice.map((entry, index) => ({\n      id: entry.id,\n      text: entry.text,\n      tier: \"working\" as const,\n      score: slice.length <= 1 ? 1 : 1 - index / slice.length,\n      metadata: entry.metadata,\n    }));\n  }\n\n  /** Drop every working-tier entry, across every scope. */\n  public clear(): void {\n    this.entries.clear();\n  }\n}\n\n/**\n * Map key for a buffer entry: the isolation `scope` (empty for the\n * unscoped pool) length-prefixed and joined to the logical id. The\n * length prefix makes the encoding injective — no crafted scope/id pair\n * can collide with a different scope's entry the way a plain `:` join\n * would allow.\n */\nfunction scopedKey(scope: string | undefined, id: string): string {\n  return `${scope?.length ?? 0}:${scope ?? \"\"}:${id}`;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AA8BA,IAAa,gBAAb,MAA2B;CAOzB,AAAO,YAAY,UAAkB;iCAcV,IAAI,IAQ7B;EArBA,KAAK,WAAW;CAClB;;;;;;;;;;CA+BA,AAAO,SAAS,MAAwB;EACtC,MAAM,KAAK,KAAK,MAAM,eAAe,KAAK,IAAI;EAE9C,KAAK,QAAQ,IAAI,UAAU,KAAK,OAAO,EAAE,GAAG;GAC1C;GACA,MAAM,KAAK;GACX,OAAO,KAAK;GACZ,UAAU,KAAK;EACjB,CAAC;EAED,KAAK,cAAc;CACrB;;;;;;;;;;;;;;;;;;;;;;;CAwBA,AAAQ,gBAAsB;EAC5B,OAAO,KAAK,QAAQ,OAAO,KAAK,UAAU;GACxC,MAAM,SAAS,KAAK,QAAQ,KAAK,CAAC,CAAC,KAAK;GAExC,IAAI,OAAO,MACT;GAGF,KAAK,QAAQ,OAAO,OAAO,KAAK;EAClC;CACF;;;;;;;;;;;CAYA,AAAO,OAAO,GAAW,OAAkC;EAKzD,MAAM,QAJU,CAAC,GAAG,KAAK,QAAQ,OAAO,CAAC,CAAC,CACvC,QAAQ,CAAC,CACT,QAAQ,UAAU,MAAM,UAAU,KAEjB,CAAC,CAAC,MAAM,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;EAE7C,OAAO,MAAM,KAAK,OAAO,WAAW;GAClC,IAAI,MAAM;GACV,MAAM,MAAM;GACZ,MAAM;GACN,OAAO,MAAM,UAAU,IAAI,IAAI,IAAI,QAAQ,MAAM;GACjD,UAAU,MAAM;EAClB,EAAE;CACJ;;CAGA,AAAO,QAAc;EACnB,KAAK,QAAQ,MAAM;CACrB;AACF;;;;;;;;AASA,SAAS,UAAU,OAA2B,IAAoB;CAChE,OAAO,GAAG,OAAO,UAAU,EAAE,GAAG,SAAS,GAAG,GAAG;AACjD"}

package/esm/middleware/builtins/budget.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"budget.mjs","names":[],"sources":["../../../../../../../../ai/src/middleware/builtins/budget.ts"],"sourcesContent":["import type {\n  AgentMiddleware,\n  MiddlewareExecuteContext,\n} from \"../../contracts/middleware\";\nimport { BudgetExceededError, type BudgetUnit } from \"../../errors\";\nimport { namespacedState } from \"../utils\";\nimport type {\n  BudgetContract,\n  BudgetContractDimension,\n  BudgetContractViolation,\n} from \"./budget-contract.type\";\n\nexport type {\n  BudgetContract,\n  BudgetContractDimension,\n  BudgetContractFallback,\n  BudgetContractViolation,\n  BudgetContractViolationMode,\n} from \"./budget-contract.type\";\n\n/**\n * Per-model pricing used to compute USD cost from token counts.\n * Caller-supplied — no bundled table. Keys are model names (the\n * `ModelContract.name` value); values are input / output token\n * prices expressed as **USD per 1K tokens** to match every major\n * provider's published pricing sheet.\n */\nexport type BudgetPricing = Record<\n  string,\n  {\n    /** USD per 1,000 input tokens. */\n    inputPer1K: number;\n    /** USD per 1,000 output tokens. */\n    outputPer1K: number;\n  }\n>;\n\n/**\n * Configuration for `budget()`. At least one of `maxTokens` or\n * `maxCostUSD` must be supplied — a budget with no cap is a no-op.\n */\nexport type BudgetOptions = {\n  /**\n   * Hard cap on cumulative total tokens (input + output, summed\n   * across every trip of the run). Inclusive — exceeding triggers\n   * the configured `onExceeded`.\n   */\n  maxTokens?: number;\n  /**\n   * Hard cap on cumulative USD cost. Requires `pricing` for the\n   * agent's configured model — without a pricing entry the USD check\n   * silently skips (tokens-only enforcement still applies).\n   */\n  maxCostUSD?: number;\n  /**\n   * Per-model pricing table used to compute USD cost. Only consulted\n   * when `maxCostUSD` is set. Model names must match the running\n   * agent's `ModelContract.name` exactly.\n   */\n  pricing?: BudgetPricing;\n  /**\n   * Behavior when a cap is breached. `\"abort\"` throws\n   * `BudgetExceededError` — surfaces on `result.error`, stops the\n   * run at the next trip boundary. `\"warn\"` logs a warning and\n   * lets the run continue (useful for observability-first rollouts\n   * before flipping the switch to abort). Default `\"abort\"`.\n   */\n  onExceeded?: \"abort\" | \"warn\";\n  /**\n   * Override the middleware name. Useful when two budgets coexist\n   * (e.g. a per-request cap plus a session-wide cap via different\n   * instances). Default `\"budget\"`.\n   */\n  name?: string;\n  /**\n   * Declarative SLO / cost contract enforced on top of (and\n   * independently of) the legacy `maxTokens` / `maxCostUSD` caps.\n   * Adds a wall-clock `maxLatencyMs` dimension and a per-contract\n   * `onViolation` reaction (`\"abort\"` hard-stops, `\"fallback\"` records\n   * a signal + fires `fallback` and lets the run continue). Omit to\n   * keep the classic budget behavior unchanged.\n   *\n   * Read a recorded fallback signal back with\n   * {@link readBudgetFallbackSignal}.\n   */\n  contract?: BudgetContract;\n};\n\ntype BudgetCounters = {\n  tokens: number;\n  costUSD: number;\n  warned: boolean;\n  /**\n   * Wall-clock epoch ms captured at `execute.before`. Used to derive\n   * cumulative run latency for the contract's `maxLatencyMs` clause.\n   */\n  startedAt: number;\n  /**\n   * Set once a `\"fallback\"` contract clause has fired, so the signal +\n   * callback are emitted at most once per run even if later trips keep\n   * breaching.\n   */\n  fallbackFired: boolean;\n};\n\n/**\n * Recorded contract fallback signal, stashed under the `<name>.fallback`\n * state key when a `\"fallback\"` clause trips. A fallback orchestrator\n * reads it via {@link readBudgetFallbackSignal} to decide how to degrade.\n */\nexport type BudgetFallbackSignal = BudgetContractViolation;\n\n/**\n * The `BudgetUnit` to stamp on the thrown error per contract dimension.\n * Latency has no native unit — it borrows `\"requests\"` as the least-wrong\n * operational measure, while the authoritative detail rides on the\n * error's `context.dimension`.\n */\nconst DIMENSION_UNIT: Record<BudgetContractDimension, BudgetUnit> = {\n  tokens: \"tokens\",\n  cost: \"usd\",\n  latency: \"requests\",\n};\n\nfunction breach(\n  limit: number,\n  actual: number,\n  unit: BudgetUnit,\n  name: string,\n): never {\n  throw new BudgetExceededError(\n    `budget \"${name}\" exceeded — ${actual} ${unit} (cap: ${limit})`,\n    { limit, actual, unit },\n  );\n}\n\nfunction breachContract(\n  name: string,\n  dimension: BudgetContractDimension,\n  limit: number,\n  actual: number,\n): never {\n  throw new BudgetExceededError(\n    `budget \"${name}\" contract exceeded — ${dimension} ${actual} (cap: ${limit})`,\n    {\n      limit,\n      actual,\n      unit: DIMENSION_UNIT[dimension],\n      context: { dimension, limit, actual, source: \"contract\" },\n    },\n  );\n}\n\n/**\n * Read the contract fallback signal recorded by a `budget()` middleware\n * running under `contract.onViolation: \"fallback\"`. Returns `undefined`\n * when no clause was breached.\n *\n * **Role.** The middleware cannot itself switch models on a soft breach,\n * so it records a typed {@link BudgetFallbackSignal} in the shared state\n * bag and lets the run continue. A fallback orchestrator (or the\n * `execute.after` hook of an outer middleware) reads it back here and\n * decides how to degrade the next run — cheaper model, cached answer,\n * truncated context.\n *\n * @param state - The middleware state bag (`ctx.state`).\n * @param name - The budget middleware's name. Default `\"budget\"`,\n *   matching `BudgetOptions.name`'s default.\n *\n * @example\n * const guard = budget({ contract: { maxCostUSD: 0.05, onViolation: \"fallback\" } });\n *\n * // In an outer middleware's execute.after, after the run:\n * const signal = readBudgetFallbackSignal(ctx.state);\n * if (signal?.dimension === \"cost\") {\n *   await rerunOnCheaperModel();\n * }\n */\nexport function readBudgetFallbackSignal(\n  state: MiddlewareExecuteContext[\"state\"],\n  name = \"budget\",\n): BudgetFallbackSignal | undefined {\n  return namespacedState<BudgetFallbackSignal>(\n    { state },\n    `${name}.fallback`,\n  ).get();\n}\n\n/**\n * Enforced token and / or USD budget for an agent run.\n *\n * **Role.** Guards against runaway tool loops, misconfigured\n * prompts, and unexpected provider price swings by capping\n * cumulative usage across every LLM trip of a single execution.\n * Aborts the run with a typed `BudgetExceededError` the moment a cap\n * is breached, rather than letting the damage grow trip by trip.\n *\n * **Scope.** Per-execution. A fresh counter is created at\n * `execute.before` and lives in the middleware state bag until the\n * run ends. Two concurrent `agent.execute()` calls on the same\n * agent therefore enforce the cap independently.\n *\n * **Token accounting.** After each successful trip, the middleware\n * adds `response.usage.total` to its running total and checks\n * against `maxTokens`. Synthetic trips (cache hits) contribute\n * `usage.total` as returned by the cache — cache middleware is\n * expected to surface zero usage on a hit, which naturally excludes\n * those trips from the budget.\n *\n * **USD accounting.** When `maxCostUSD` + `pricing[modelName]` are\n * both present, the middleware converts per-trip input / output\n * tokens to USD and accumulates. Missing pricing silently degrades\n * to tokens-only — explicit rather than guessing.\n *\n * **Warn mode.** `onExceeded: \"warn\"` logs a single warning the first\n * time a cap is breached and lets the run continue. Useful for\n * measuring real-world traffic against a proposed cap before flipping\n * to `\"abort\"` in production.\n *\n * **Contract / SLO mode.** Pass `contract` to enforce a declarative\n * service-level objective — `maxCostUSD`, `maxLatencyMs`, `maxTokens` —\n * on top of the legacy caps, with a single `onViolation` reaction:\n * `\"abort\"` hard-stops with `BudgetExceededError`; `\"fallback\"` records\n * a typed signal (read it via {@link readBudgetFallbackSignal}), fires\n * the optional `fallback` callback, and lets the run continue so an\n * outer layer can degrade gracefully. The contract's clauses are\n * evaluated independently of — and after — the top-level caps; the\n * top-level caps stay fully functional with or without a contract.\n *\n * @example\n * const budgetMiddleware = budget({ maxTokens: 50_000 });\n *\n * const myAgent = agent({\n *   model,\n *   middleware: [budgetMiddleware],\n * });\n *\n * @example\n * // With USD cap and custom pricing\n * const guard = budget({\n *   maxCostUSD: 0.5,\n *   pricing: {\n *     \"gpt-4o\": { inputPer1K: 0.005, outputPer1K: 0.015 },\n *   },\n * });\n *\n * @example\n * // SLO contract — soft-fallback on any breach\n * const sloGuard = budget({\n *   pricing: { \"gpt-4o\": { inputPer1K: 0.005, outputPer1K: 0.015 } },\n *   contract: {\n *     maxCostUSD: 0.05,\n *     maxLatencyMs: 8_000,\n *     maxTokens: 40_000,\n *     onViolation: \"fallback\",\n *     fallback: (violation) => routeToCheaperModel(violation.dimension),\n *   },\n * });\n */\nexport function budget(options: BudgetOptions): AgentMiddleware {\n  const name = options.name ?? \"budget\";\n  const onExceeded = options.onExceeded ?? \"abort\";\n  const hasTokenCap = typeof options.maxTokens === \"number\";\n  const hasCostCap = typeof options.maxCostUSD === \"number\";\n\n  const contract = options.contract;\n  const contractMode = contract?.onViolation ?? \"abort\";\n  const hasContractTokenCap = typeof contract?.maxTokens === \"number\";\n  const hasContractCostCap = typeof contract?.maxCostUSD === \"number\";\n  const hasContractLatencyCap = typeof contract?.maxLatencyMs === \"number\";\n  const contractNeedsCost = hasCostCap || hasContractCostCap;\n  // Warn once per model when a cost cap is configured but the running model\n  // has no pricing entry — without this the USD cap silently never enforces\n  // (costUSD stays 0), a fail-open the JSDoc on `maxCostUSD` documents.\n  const warnedUnpricedModels = new Set<string>();\n\n  return {\n    name,\n    execute: {\n      before(context) {\n        const counters = namespacedState<BudgetCounters>(context, name);\n        counters.set({\n          tokens: 0,\n          costUSD: 0,\n          warned: false,\n          startedAt: Date.now(),\n          fallbackFired: false,\n        });\n      },\n    },\n    trip: {\n      async after(context, response) {\n        const counters = namespacedState<BudgetCounters>(context, name).get();\n\n        if (!counters) {\n          return;\n        }\n\n        counters.tokens += response.usage.total;\n\n        if (contractNeedsCost) {\n          const pricing = options.pricing?.[context.model.name];\n\n          if (pricing) {\n            const tripCost =\n              (response.usage.input / 1000) * pricing.inputPer1K +\n              (response.usage.output / 1000) * pricing.outputPer1K;\n            counters.costUSD += tripCost;\n          } else if (!warnedUnpricedModels.has(context.model.name)) {\n            // A cost cap is set but no pricing matched the running model, so\n            // costUSD can never grow and the USD cap silently never fires.\n            // Surface the fail-open once per model instead of swallowing it.\n            warnedUnpricedModels.add(context.model.name);\n            console.warn(\n              `ai.middleware.budget(\"${name}\"): a USD cost cap is set but no pricing entry ` +\n                `matches the running model \"${context.model.name}\" — the cap cannot be enforced ` +\n                `for it. Add a pricing entry for \"${context.model.name}\" to options.pricing.`,\n            );\n          }\n        }\n\n        if (hasTokenCap && counters.tokens > options.maxTokens!) {\n          if (onExceeded === \"abort\") {\n            breach(options.maxTokens!, counters.tokens, \"tokens\", name);\n          }\n\n          if (!counters.warned) {\n            counters.warned = true;\n          }\n        }\n\n        if (hasCostCap && counters.costUSD > options.maxCostUSD!) {\n          if (onExceeded === \"abort\") {\n            breach(options.maxCostUSD!, counters.costUSD, \"usd\", name);\n          }\n\n          if (!counters.warned) {\n            counters.warned = true;\n          }\n        }\n\n        if (!contract) {\n          return;\n        }\n\n        if (hasContractTokenCap && counters.tokens > contract.maxTokens!) {\n          await enforceContract(\n            context,\n            counters,\n            name,\n            contractMode,\n            contract,\n            \"tokens\",\n            contract.maxTokens!,\n            counters.tokens,\n          );\n        }\n\n        if (hasContractCostCap && counters.costUSD > contract.maxCostUSD!) {\n          await enforceContract(\n            context,\n            counters,\n            name,\n            contractMode,\n            contract,\n            \"cost\",\n            contract.maxCostUSD!,\n            counters.costUSD,\n          );\n        }\n\n        if (hasContractLatencyCap) {\n          const elapsedMs = Date.now() - counters.startedAt;\n\n          if (elapsedMs > contract.maxLatencyMs!) {\n            await enforceContract(\n              context,\n              counters,\n              name,\n              contractMode,\n              contract,\n              \"latency\",\n              contract.maxLatencyMs!,\n              elapsedMs,\n            );\n          }\n        }\n      },\n    },\n  };\n}\n\n/**\n * Apply the contract's reaction to a single breached clause. `\"abort\"`\n * throws `BudgetExceededError` (stops the run); `\"fallback\"` records the\n * signal once, fires the callback, and returns so the run continues.\n *\n * The callback is invoked at most once per run (guarded by\n * `counters.fallbackFired`) and its rejections are swallowed — a buggy\n * fallback hook must never crash the agent.\n */\nasync function enforceContract(\n  context: MiddlewareExecuteContext,\n  counters: BudgetCounters,\n  name: string,\n  mode: NonNullable<BudgetContract[\"onViolation\"]>,\n  contract: BudgetContract,\n  dimension: BudgetContractDimension,\n  limit: number,\n  actual: number,\n): Promise<void> {\n  if (mode === \"abort\") {\n    breachContract(name, dimension, limit, actual);\n  }\n\n  if (counters.fallbackFired) {\n    return;\n  }\n\n  counters.fallbackFired = true;\n\n  const violation: BudgetContractViolation = {\n    dimension,\n    limit,\n    actual,\n    mode,\n  };\n\n  namespacedState<BudgetContractViolation>(context, `${name}.fallback`).set(\n    violation,\n  );\n\n  if (!contract.fallback) {\n    return;\n  }\n\n  try {\n    await contract.fallback(violation, context);\n  } catch {\n    // A fallback callback is a notification hook — its failure must\n    // never crash the run. Swallow deliberately.\n  }\n}\n"],"mappings":";;;;;;;;;;;AAsHA,MAAM,iBAA8D;CAClE,QAAQ;CACR,MAAM;CACN,SAAS;AACX;AAEA,SAAS,OACP,OACA,QACA,MACA,MACO;CACP,MAAM,IAAI,oBACR,WAAW,KAAK,eAAe,OAAO,GAAG,KAAK,SAAS,MAAM,IAC7D;EAAE;EAAO;EAAQ;CAAK,CACxB;AACF;AAEA,SAAS,eACP,MACA,WACA,OACA,QACO;CACP,MAAM,IAAI,oBACR,WAAW,KAAK,wBAAwB,UAAU,GAAG,OAAO,SAAS,MAAM,IAC3E;EACE;EACA;EACA,MAAM,eAAe;EACrB,SAAS;GAAE;GAAW;GAAO;GAAQ,QAAQ;EAAW;CAC1D,CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,yBACd,OACA,OAAO,UAC2B;CAClC,OAAO,gBACL,EAAE,MAAM,GACR,GAAG,KAAK,UACV,EAAE,IAAI;AACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyEA,SAAgB,OAAO,SAAyC;CAC9D,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,cAAc,OAAO,QAAQ,cAAc;CACjD,MAAM,aAAa,OAAO,QAAQ,eAAe;CAEjD,MAAM,WAAW,QAAQ;CACzB,MAAM,eAAe,UAAU,eAAe;CAC9C,MAAM,sBAAsB,OAAO,UAAU,cAAc;CAC3D,MAAM,qBAAqB,OAAO,UAAU,eAAe;CAC3D,MAAM,wBAAwB,OAAO,UAAU,iBAAiB;CAChE,MAAM,oBAAoB,cAAc;CAIxC,MAAM,uCAAuB,IAAI,IAAY;CAE7C,OAAO;EACL;EACA,SAAS,EACP,OAAO,SAAS;GAEd,AADiB,gBAAgC,SAAS,IACnD,EAAE,IAAI;IACX,QAAQ;IACR,SAAS;IACT,QAAQ;IACR,WAAW,KAAK,IAAI;IACpB,eAAe;GACjB,CAAC;EACH,EACF;EACA,MAAM,EACJ,MAAM,MAAM,SAAS,UAAU;GAC7B,MAAM,WAAW,gBAAgC,SAAS,IAAI,EAAE,IAAI;GAEpE,IAAI,CAAC,UACH;GAGF,SAAS,UAAU,SAAS,MAAM;GAElC,IAAI,mBAAmB;IACrB,MAAM,UAAU,QAAQ,UAAU,QAAQ,MAAM;IAEhD,IAAI,SAAS;KACX,MAAM,WACH,SAAS,MAAM,QAAQ,MAAQ,QAAQ,aACvC,SAAS,MAAM,SAAS,MAAQ,QAAQ;KAC3C,SAAS,WAAW;IACtB,OAAO,IAAI,CAAC,qBAAqB,IAAI,QAAQ,MAAM,IAAI,GAAG;KAIxD,qBAAqB,IAAI,QAAQ,MAAM,IAAI;KAC3C,QAAQ,KACN,yBAAyB,KAAK,4EACE,QAAQ,MAAM,KAAK,kEACb,QAAQ,MAAM,KAAK,sBAC3D;IACF;GACF;GAEA,IAAI,eAAe,SAAS,SAAS,QAAQ,WAAY;IACvD,IAAI,eAAe,SACjB,OAAO,QAAQ,WAAY,SAAS,QAAQ,UAAU,IAAI;IAG5D,IAAI,CAAC,SAAS,QACZ,SAAS,SAAS;GAEtB;GAEA,IAAI,cAAc,SAAS,UAAU,QAAQ,YAAa;IACxD,IAAI,eAAe,SACjB,OAAO,QAAQ,YAAa,SAAS,SAAS,OAAO,IAAI;IAG3D,IAAI,CAAC,SAAS,QACZ,SAAS,SAAS;GAEtB;GAEA,IAAI,CAAC,UACH;GAGF,IAAI,uBAAuB,SAAS,SAAS,SAAS,WACpD,MAAM,gBACJ,SACA,UACA,MACA,cACA,UACA,UACA,SAAS,WACT,SAAS,MACX;GAGF,IAAI,sBAAsB,SAAS,UAAU,SAAS,YACpD,MAAM,gBACJ,SACA,UACA,MACA,cACA,UACA,QACA,SAAS,YACT,SAAS,OACX;GAGF,IAAI,uBAAuB;IACzB,MAAM,YAAY,KAAK,IAAI,IAAI,SAAS;IAExC,IAAI,YAAY,SAAS,cACvB,MAAM,gBACJ,SACA,UACA,MACA,cACA,UACA,WACA,SAAS,cACT,SACF;GAEJ;EACF,EACF;CACF;AACF;;;;;;;;;;AAWA,eAAe,gBACb,SACA,UACA,MACA,MACA,UACA,WACA,OACA,QACe;CACf,IAAI,SAAS,SACX,eAAe,MAAM,WAAW,OAAO,MAAM;CAG/C,IAAI,SAAS,eACX;CAGF,SAAS,gBAAgB;CAEzB,MAAM,YAAqC;EACzC;EACA;EACA;EACA;CACF;CAEA,gBAAyC,SAAS,GAAG,KAAK,UAAU,EAAE,IACpE,SACF;CAEA,IAAI,CAAC,SAAS,UACZ;CAGF,IAAI;EACF,MAAM,SAAS,SAAS,WAAW,OAAO;CAC5C,QAAQ,CAGR;AACF"}
+{"version":3,"file":"budget.mjs","names":[],"sources":["../../../../../../../../ai/src/middleware/builtins/budget.ts"],"sourcesContent":["import type {\n  AgentMiddleware,\n  MiddlewareExecuteContext,\n} from \"../../contracts/middleware\";\nimport { BudgetExceededError, type BudgetUnit } from \"../../errors\";\nimport { namespacedState } from \"../utils\";\nimport type {\n  BudgetContract,\n  BudgetContractDimension,\n  BudgetContractViolation,\n} from \"./budget-contract.type\";\n\nexport type {\n  BudgetContract,\n  BudgetContractDimension,\n  BudgetContractFallback,\n  BudgetContractViolation,\n  BudgetContractViolationMode,\n} from \"./budget-contract.type\";\n\n/**\n * Per-model pricing used to compute USD cost from token counts.\n * Caller-supplied — no bundled table. Keys are model names (the\n * `ModelContract.name` value); values are input / output token\n * prices expressed as **USD per 1K tokens** to match every major\n * provider's published pricing sheet.\n */\nexport type BudgetPricing = Record<\n  string,\n  {\n    /** USD per 1,000 input tokens. */\n    inputPer1K: number;\n    /** USD per 1,000 output tokens. */\n    outputPer1K: number;\n  }\n>;\n\n/**\n * Configuration for `budget()`. At least one of `maxTokens` or\n * `maxCostUSD` must be supplied — a budget with no cap is a no-op.\n */\nexport type BudgetOptions = {\n  /**\n   * Hard cap on cumulative total tokens (input + output, summed\n   * across every trip of the run). Inclusive — exceeding triggers\n   * the configured `onExceeded`.\n   */\n  maxTokens?: number;\n  /**\n   * Hard cap on cumulative USD cost. Requires `pricing` for the\n   * agent's configured model — without a pricing entry the USD check\n   * silently skips (tokens-only enforcement still applies).\n   */\n  maxCostUSD?: number;\n  /**\n   * Per-model pricing table used to compute USD cost. Only consulted\n   * when `maxCostUSD` is set. Model names must match the running\n   * agent's `ModelContract.name` exactly.\n   */\n  pricing?: BudgetPricing;\n  /**\n   * Behavior when a cap is breached. `\"abort\"` throws\n   * `BudgetExceededError` — surfaces on `result.error`, stops the\n   * run at the next trip boundary. `\"warn\"` logs a warning and\n   * lets the run continue (useful for observability-first rollouts\n   * before flipping the switch to abort). Default `\"abort\"`.\n   */\n  onExceeded?: \"abort\" | \"warn\";\n  /**\n   * Override the middleware name. Useful when two budgets coexist\n   * (e.g. a per-request cap plus a session-wide cap via different\n   * instances). Default `\"budget\"`.\n   */\n  name?: string;\n  /**\n   * Declarative SLO / cost contract enforced on top of (and\n   * independently of) the legacy `maxTokens` / `maxCostUSD` caps.\n   * Adds a wall-clock `maxLatencyMs` dimension and a per-contract\n   * `onViolation` reaction (`\"abort\"` hard-stops, `\"fallback\"` records\n   * a signal + fires `fallback` and lets the run continue). Omit to\n   * keep the classic budget behavior unchanged.\n   *\n   * Read a recorded fallback signal back with\n   * {@link readBudgetFallbackSignal}.\n   */\n  contract?: BudgetContract;\n};\n\ntype BudgetCounters = {\n  tokens: number;\n  costUSD: number;\n  warned: boolean;\n  /**\n   * Wall-clock epoch ms captured at `execute.before`. Used to derive\n   * cumulative run latency for the contract's `maxLatencyMs` clause.\n   */\n  startedAt: number;\n  /**\n   * Set once a `\"fallback\"` contract clause has fired, so the signal +\n   * callback are emitted at most once per run even if later trips keep\n   * breaching.\n   */\n  fallbackFired: boolean;\n};\n\n/**\n * Recorded contract fallback signal, stashed under the `<name>.fallback`\n * state key when a `\"fallback\"` clause trips. A fallback orchestrator\n * reads it via {@link readBudgetFallbackSignal} to decide how to degrade.\n */\nexport type BudgetFallbackSignal = BudgetContractViolation;\n\n/**\n * The `BudgetUnit` to stamp on the thrown error per contract dimension.\n * Latency has no native unit — it borrows `\"requests\"` as the least-wrong\n * operational measure, while the authoritative detail rides on the\n * error's `context.dimension`.\n */\nconst DIMENSION_UNIT: Record<BudgetContractDimension, BudgetUnit> = {\n  tokens: \"tokens\",\n  cost: \"usd\",\n  latency: \"requests\",\n};\n\nfunction breach(\n  limit: number,\n  actual: number,\n  unit: BudgetUnit,\n  name: string,\n): never {\n  throw new BudgetExceededError(\n    `budget \"${name}\" exceeded — ${actual} ${unit} (cap: ${limit})`,\n    { limit, actual, unit },\n  );\n}\n\nfunction breachContract(\n  name: string,\n  dimension: BudgetContractDimension,\n  limit: number,\n  actual: number,\n): never {\n  throw new BudgetExceededError(\n    `budget \"${name}\" contract exceeded — ${dimension} ${actual} (cap: ${limit})`,\n    {\n      limit,\n      actual,\n      unit: DIMENSION_UNIT[dimension],\n      context: { dimension, limit, actual, source: \"contract\" },\n    },\n  );\n}\n\n/**\n * Read the contract fallback signal recorded by a `budget()` middleware\n * running under `contract.onViolation: \"fallback\"`. Returns `undefined`\n * when no clause was breached.\n *\n * **Role.** The middleware cannot itself switch models on a soft breach,\n * so it records a typed {@link BudgetFallbackSignal} in the shared state\n * bag and lets the run continue. A fallback orchestrator (or the\n * `execute.after` hook of an outer middleware) reads it back here and\n * decides how to degrade the next run — cheaper model, cached answer,\n * truncated context.\n *\n * @param state - The middleware state bag (`ctx.state`).\n * @param name - The budget middleware's name. Default `\"budget\"`,\n *   matching `BudgetOptions.name`'s default.\n *\n * @example\n * const guard = budget({ contract: { maxCostUSD: 0.05, onViolation: \"fallback\" } });\n *\n * // In an outer middleware's execute.after, after the run:\n * const signal = readBudgetFallbackSignal(ctx.state);\n * if (signal?.dimension === \"cost\") {\n *   await rerunOnCheaperModel();\n * }\n */\nexport function readBudgetFallbackSignal(\n  state: MiddlewareExecuteContext[\"state\"],\n  name = \"budget\",\n): BudgetFallbackSignal | undefined {\n  return namespacedState<BudgetFallbackSignal>(\n    { state },\n    `${name}.fallback`,\n  ).get();\n}\n\n/**\n * Enforced token and / or USD budget for an agent run.\n *\n * **Role.** Guards against runaway tool loops, misconfigured\n * prompts, and unexpected provider price swings by capping\n * cumulative usage across every LLM trip of a single execution.\n * Aborts the run with a typed `BudgetExceededError` the moment a cap\n * is breached, rather than letting the damage grow trip by trip.\n *\n * **Scope.** Per-execution. A fresh counter is created at\n * `execute.before` and lives in the middleware state bag until the\n * run ends. Two concurrent `agent.execute()` calls on the same\n * agent therefore enforce the cap independently.\n *\n * **Token accounting.** After each successful trip, the middleware\n * adds `response.usage.total` to its running total and checks\n * against `maxTokens`. Synthetic trips (cache hits) contribute\n * `usage.total` as returned by the cache — cache middleware is\n * expected to surface zero usage on a hit, which naturally excludes\n * those trips from the budget.\n *\n * **USD accounting.** When `maxCostUSD` + `pricing[modelName]` are\n * both present, the middleware converts per-trip input / output\n * tokens to USD and accumulates. Missing pricing silently degrades\n * to tokens-only — explicit rather than guessing.\n *\n * **Warn mode.** `onExceeded: \"warn\"` logs a single warning the first\n * time a cap is breached and lets the run continue. Useful for\n * measuring real-world traffic against a proposed cap before flipping\n * to `\"abort\"` in production.\n *\n * **Contract / SLO mode.** Pass `contract` to enforce a declarative\n * service-level objective — `maxCostUSD`, `maxLatencyMs`, `maxTokens` —\n * on top of the legacy caps, with a single `onViolation` reaction:\n * `\"abort\"` hard-stops with `BudgetExceededError`; `\"fallback\"` records\n * a typed signal (read it via {@link readBudgetFallbackSignal}), fires\n * the optional `fallback` callback, and lets the run continue so an\n * outer layer can degrade gracefully. The contract's clauses are\n * evaluated independently of — and after — the top-level caps; the\n * top-level caps stay fully functional with or without a contract.\n *\n * @example\n * const budgetMiddleware = budget({ maxTokens: 50_000 });\n *\n * const myAgent = agent({\n *   model,\n *   middleware: [budgetMiddleware],\n * });\n *\n * @example\n * // With USD cap and custom pricing\n * const guard = budget({\n *   maxCostUSD: 0.5,\n *   pricing: {\n *     \"gpt-4o\": { inputPer1K: 0.005, outputPer1K: 0.015 },\n *   },\n * });\n *\n * @example\n * // SLO contract — soft-fallback on any breach\n * const sloGuard = budget({\n *   pricing: { \"gpt-4o\": { inputPer1K: 0.005, outputPer1K: 0.015 } },\n *   contract: {\n *     maxCostUSD: 0.05,\n *     maxLatencyMs: 8_000,\n *     maxTokens: 40_000,\n *     onViolation: \"fallback\",\n *     fallback: (violation) => routeToCheaperModel(violation.dimension),\n *   },\n * });\n */\nexport function budget(options: BudgetOptions): AgentMiddleware {\n  const name = options.name ?? \"budget\";\n  const onExceeded = options.onExceeded ?? \"abort\";\n  const hasTokenCap = typeof options.maxTokens === \"number\";\n  const hasCostCap = typeof options.maxCostUSD === \"number\";\n\n  const contract = options.contract;\n  const contractMode = contract?.onViolation ?? \"abort\";\n  const hasContractTokenCap = typeof contract?.maxTokens === \"number\";\n  const hasContractCostCap = typeof contract?.maxCostUSD === \"number\";\n  const hasContractLatencyCap = typeof contract?.maxLatencyMs === \"number\";\n  const contractNeedsCost = hasCostCap || hasContractCostCap;\n  // Warn once per model when a cost cap is configured but the running model\n  // has no pricing entry — without this the USD cap silently never enforces\n  // (costUSD stays 0), a fail-open the JSDoc on `maxCostUSD` documents.\n  const warnedUnpricedModels = new Set<string>();\n\n  return {\n    name,\n    execute: {\n      before(context) {\n        const counters = namespacedState<BudgetCounters>(context, name);\n        counters.set({\n          tokens: 0,\n          costUSD: 0,\n          warned: false,\n          startedAt: Date.now(),\n          fallbackFired: false,\n        });\n      },\n    },\n    trip: {\n      async after(context, response) {\n        const counters = namespacedState<BudgetCounters>(context, name).get();\n\n        if (!counters) {\n          return;\n        }\n\n        counters.tokens += response.usage.total;\n\n        if (contractNeedsCost) {\n          const pricing = options.pricing?.[context.model.name];\n\n          if (pricing) {\n            const tripCost =\n              (response.usage.input / 1000) * pricing.inputPer1K +\n              (response.usage.output / 1000) * pricing.outputPer1K;\n            counters.costUSD += tripCost;\n          } else if (!warnedUnpricedModels.has(context.model.name)) {\n            // A cost cap is set but no pricing matched the running model, so\n            // costUSD can never grow and the USD cap silently never fires.\n            // Surface the fail-open once per model instead of swallowing it.\n            warnedUnpricedModels.add(context.model.name);\n            console.warn(\n              `ai.middleware.budget(\"${name}\"): a USD cost cap is set but no pricing entry ` +\n                `matches the running model \"${context.model.name}\" — the cap cannot be enforced ` +\n                `for it. Add a pricing entry for \"${context.model.name}\" to options.pricing.`,\n            );\n          }\n        }\n\n        if (hasTokenCap && counters.tokens > options.maxTokens!) {\n          if (onExceeded === \"abort\") {\n            breach(options.maxTokens!, counters.tokens, \"tokens\", name);\n          }\n\n          if (!counters.warned) {\n            counters.warned = true;\n          }\n        }\n\n        if (hasCostCap && counters.costUSD > options.maxCostUSD!) {\n          if (onExceeded === \"abort\") {\n            breach(options.maxCostUSD!, counters.costUSD, \"usd\", name);\n          }\n\n          if (!counters.warned) {\n            counters.warned = true;\n          }\n        }\n\n        if (!contract) {\n          return;\n        }\n\n        if (hasContractTokenCap && counters.tokens > contract.maxTokens!) {\n          await enforceContract(\n            context,\n            counters,\n            name,\n            contractMode,\n            contract,\n            \"tokens\",\n            contract.maxTokens!,\n            counters.tokens,\n          );\n        }\n\n        if (hasContractCostCap && counters.costUSD > contract.maxCostUSD!) {\n          await enforceContract(\n            context,\n            counters,\n            name,\n            contractMode,\n            contract,\n            \"cost\",\n            contract.maxCostUSD!,\n            counters.costUSD,\n          );\n        }\n\n        if (hasContractLatencyCap) {\n          const elapsedMs = Date.now() - counters.startedAt;\n\n          if (elapsedMs > contract.maxLatencyMs!) {\n            await enforceContract(\n              context,\n              counters,\n              name,\n              contractMode,\n              contract,\n              \"latency\",\n              contract.maxLatencyMs!,\n              elapsedMs,\n            );\n          }\n        }\n      },\n    },\n  };\n}\n\n/**\n * Apply the contract's reaction to a single breached clause. `\"abort\"`\n * throws `BudgetExceededError` (stops the run); `\"fallback\"` records the\n * signal once, fires the callback, and returns so the run continues.\n *\n * The callback is invoked at most once per run (guarded by\n * `counters.fallbackFired`) and its rejections are swallowed — a buggy\n * fallback hook must never crash the agent.\n */\nasync function enforceContract(\n  context: MiddlewareExecuteContext,\n  counters: BudgetCounters,\n  name: string,\n  mode: NonNullable<BudgetContract[\"onViolation\"]>,\n  contract: BudgetContract,\n  dimension: BudgetContractDimension,\n  limit: number,\n  actual: number,\n): Promise<void> {\n  if (mode === \"abort\") {\n    breachContract(name, dimension, limit, actual);\n  }\n\n  if (counters.fallbackFired) {\n    return;\n  }\n\n  counters.fallbackFired = true;\n\n  const violation: BudgetContractViolation = {\n    dimension,\n    limit,\n    actual,\n    mode,\n  };\n\n  namespacedState<BudgetContractViolation>(context, `${name}.fallback`).set(\n    violation,\n  );\n\n  if (!contract.fallback) {\n    return;\n  }\n\n  try {\n    await contract.fallback(violation, context);\n  } catch {\n    // A fallback callback is a notification hook — its failure must\n    // never crash the run. Swallow deliberately.\n  }\n}\n"],"mappings":";;;;;;;;;;;AAsHA,MAAM,iBAA8D;CAClE,QAAQ;CACR,MAAM;CACN,SAAS;AACX;AAEA,SAAS,OACP,OACA,QACA,MACA,MACO;CACP,MAAM,IAAI,oBACR,WAAW,KAAK,eAAe,OAAO,GAAG,KAAK,SAAS,MAAM,IAC7D;EAAE;EAAO;EAAQ;CAAK,CACxB;AACF;AAEA,SAAS,eACP,MACA,WACA,OACA,QACO;CACP,MAAM,IAAI,oBACR,WAAW,KAAK,wBAAwB,UAAU,GAAG,OAAO,SAAS,MAAM,IAC3E;EACE;EACA;EACA,MAAM,eAAe;EACrB,SAAS;GAAE;GAAW;GAAO;GAAQ,QAAQ;EAAW;CAC1D,CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,yBACd,OACA,OAAO,UAC2B;CAClC,OAAO,gBACL,EAAE,MAAM,GACR,GAAG,KAAK,UACV,CAAC,CAAC,IAAI;AACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyEA,SAAgB,OAAO,SAAyC;CAC9D,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,aAAa,QAAQ,cAAc;CACzC,MAAM,cAAc,OAAO,QAAQ,cAAc;CACjD,MAAM,aAAa,OAAO,QAAQ,eAAe;CAEjD,MAAM,WAAW,QAAQ;CACzB,MAAM,eAAe,UAAU,eAAe;CAC9C,MAAM,sBAAsB,OAAO,UAAU,cAAc;CAC3D,MAAM,qBAAqB,OAAO,UAAU,eAAe;CAC3D,MAAM,wBAAwB,OAAO,UAAU,iBAAiB;CAChE,MAAM,oBAAoB,cAAc;CAIxC,MAAM,uCAAuB,IAAI,IAAY;CAE7C,OAAO;EACL;EACA,SAAS,EACP,OAAO,SAAS;GAEd,AADiB,gBAAgC,SAAS,IACnD,CAAC,CAAC,IAAI;IACX,QAAQ;IACR,SAAS;IACT,QAAQ;IACR,WAAW,KAAK,IAAI;IACpB,eAAe;GACjB,CAAC;EACH,EACF;EACA,MAAM,EACJ,MAAM,MAAM,SAAS,UAAU;GAC7B,MAAM,WAAW,gBAAgC,SAAS,IAAI,CAAC,CAAC,IAAI;GAEpE,IAAI,CAAC,UACH;GAGF,SAAS,UAAU,SAAS,MAAM;GAElC,IAAI,mBAAmB;IACrB,MAAM,UAAU,QAAQ,UAAU,QAAQ,MAAM;IAEhD,IAAI,SAAS;KACX,MAAM,WACH,SAAS,MAAM,QAAQ,MAAQ,QAAQ,aACvC,SAAS,MAAM,SAAS,MAAQ,QAAQ;KAC3C,SAAS,WAAW;IACtB,OAAO,IAAI,CAAC,qBAAqB,IAAI,QAAQ,MAAM,IAAI,GAAG;KAIxD,qBAAqB,IAAI,QAAQ,MAAM,IAAI;KAC3C,QAAQ,KACN,yBAAyB,KAAK,4EACE,QAAQ,MAAM,KAAK,kEACb,QAAQ,MAAM,KAAK,sBAC3D;IACF;GACF;GAEA,IAAI,eAAe,SAAS,SAAS,QAAQ,WAAY;IACvD,IAAI,eAAe,SACjB,OAAO,QAAQ,WAAY,SAAS,QAAQ,UAAU,IAAI;IAG5D,IAAI,CAAC,SAAS,QACZ,SAAS,SAAS;GAEtB;GAEA,IAAI,cAAc,SAAS,UAAU,QAAQ,YAAa;IACxD,IAAI,eAAe,SACjB,OAAO,QAAQ,YAAa,SAAS,SAAS,OAAO,IAAI;IAG3D,IAAI,CAAC,SAAS,QACZ,SAAS,SAAS;GAEtB;GAEA,IAAI,CAAC,UACH;GAGF,IAAI,uBAAuB,SAAS,SAAS,SAAS,WACpD,MAAM,gBACJ,SACA,UACA,MACA,cACA,UACA,UACA,SAAS,WACT,SAAS,MACX;GAGF,IAAI,sBAAsB,SAAS,UAAU,SAAS,YACpD,MAAM,gBACJ,SACA,UACA,MACA,cACA,UACA,QACA,SAAS,YACT,SAAS,OACX;GAGF,IAAI,uBAAuB;IACzB,MAAM,YAAY,KAAK,IAAI,IAAI,SAAS;IAExC,IAAI,YAAY,SAAS,cACvB,MAAM,gBACJ,SACA,UACA,MACA,cACA,UACA,WACA,SAAS,cACT,SACF;GAEJ;EACF,EACF;CACF;AACF;;;;;;;;;;AAWA,eAAe,gBACb,SACA,UACA,MACA,MACA,UACA,WACA,OACA,QACe;CACf,IAAI,SAAS,SACX,eAAe,MAAM,WAAW,OAAO,MAAM;CAG/C,IAAI,SAAS,eACX;CAGF,SAAS,gBAAgB;CAEzB,MAAM,YAAqC;EACzC;EACA;EACA;EACA;CACF;CAEA,gBAAyC,SAAS,GAAG,KAAK,UAAU,CAAC,CAAC,IACpE,SACF;CAEA,IAAI,CAAC,SAAS,UACZ;CAGF,IAAI;EACF,MAAM,SAAS,SAAS,WAAW,OAAO;CAC5C,QAAQ,CAGR;AACF"}

package/esm/middleware/builtins/semantic-cache.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"semantic-cache.mjs","names":[],"sources":["../../../../../../../../ai/src/middleware/builtins/semantic-cache.ts"],"sourcesContent":["import type { CacheDriver } from \"@warlock.js/cache\";\nimport { resolveDefaultStore } from \"../../config\";\nimport type { Message } from \"../../contracts/conversation-message.type\";\nimport type { EmbedderContract } from \"../../contracts/embedder.contract\";\nimport type { AgentMiddleware } from \"../../contracts/middleware\";\nimport type { MiddlewareTripContext } from \"../../contracts/middleware/middleware-context.type\";\nimport type { ModelResponse } from \"../../contracts/model.contract\";\nimport { extractUserText } from \"../utils\";\n\n/**\n * Isolation boundary for cache reads and writes.\n *\n * - `\"session\"` (default) — key every entry off the run's\n *   `AgentExecuteOptions.sessionId`, so one session never receives a\n *   response cached for another. Calls made WITHOUT a `sessionId` share\n *   one unscoped pool (the pre-4.15.0 behavior); an unscoped read never\n *   sees a session-scoped entry and vice versa.\n * - `\"shared\"` — one pool for every caller, regardless of session. The\n *   explicit opt-in for genuinely public Q&A (docs bots, FAQ) where the\n *   cross-user hit rate is the point and no response can carry one\n *   caller's private context.\n * - a resolver — derive the key yourself, e.g. per tenant\n *   (`ctx => ctx.options?.toolCtx?.tenantId`). Returning `undefined`\n *   falls back to the unscoped pool, so return a constant sentinel (or\n *   throw) if you need the call to fail closed instead.\n */\nexport type SemanticCacheScope =\n  | \"session\"\n  | \"shared\"\n  | ((context: MiddlewareTripContext) => string | undefined);\n\n/**\n * Configuration for `semanticCache()`.\n */\nexport type SemanticCacheOptions = {\n  /** Embedder used to produce the query vector from the prompt text. */\n  embedder: EmbedderContract;\n  /**\n   * Vector-capable cache driver from `@warlock.js/cache`. Production\n   * deployments pick a driver with a real ANN index (`pg` with\n   * pgvector, `redis` with RediSearch). Dev / test environments use\n   * `new MemoryCacheDriver()` — zero config, correct, but O(N) per\n   * query. Drivers without similarity support throw\n   * `CacheUnsupportedError` from `set({ vector })` / `similar()`.\n   *\n   * Falls back to `ai.config({ defaultStore })` when omitted. When\n   * neither is set, the factory throws at construction time —\n   * semantic cache cannot operate without a store.\n   */\n  store?: CacheDriver<any, any>;\n  /**\n   * Minimum cosine similarity for a vector hit. Between 0 and 1 —\n   * 0.95 is a solid default for question-answering caches.\n   */\n  threshold: number;\n  /**\n   * Optional TTL in milliseconds. Entries whose `storedAt` is older\n   * than this are treated as misses on read and overwritten on the\n   * next write. Default: no expiry — entries live until the store\n   * evicts them (per its own TTL/eviction policy).\n   */\n  ttlMs?: number;\n  /**\n   * Namespace prefix applied to every key the cache writes. Lets\n   * multiple agents share one driver without collision. Default\n   * `\"ai.cache\"`.\n   */\n  namespace?: string;\n  /**\n   * Per-caller isolation boundary. Default `\"session\"` — a cached\n   * response is served back only to the session that produced it.\n   *\n   * A `semanticCache` is normally built once at app boot and shared by\n   * every end user, and a hit is returned as the model's answer with no\n   * LLM call in between; without a scope that pools every caller's Q&A\n   * pairs into one namespace, which is both a disclosure path (user B's\n   * near-enough prompt gets served user A's answer, personal context\n   * included) and a poisoning path (an attacker seeds an entry near a\n   * predictable future query). Set `\"shared\"` to opt back into pooling\n   * where that is actually desirable. See {@link SemanticCacheScope}.\n   */\n  scope?: SemanticCacheScope;\n  /**\n   * Middleware name — also the state-bag key prefix inside a single\n   * execution. Default `\"semantic-cache\"`.\n   */\n  name?: string;\n};\n\ntype CachedEntry = {\n  response: ModelResponse;\n  storedAt: number;\n  /**\n   * Isolation key the entry was written under; absent = the unscoped\n   * pool (also the shape of every entry written before 4.15.0).\n   */\n  scope?: string;\n};\n\ntype PendingWrite = {\n  promptKey: string;\n  vector: number[];\n  scope?: string;\n};\n\nconst DEFAULT_NAMESPACE = \"ai.cache\";\n\n/**\n * Extra candidates pulled from `similar()` on a SCOPED lookup before the\n * scope filter runs. The driver ranks across every scope in the index,\n * so a bare `topK: 1` can come back as a foreign entry and mask this\n * scope's own legitimate hit. Mirrors the memory tiers' overscan.\n */\nconst SIMILAR_OVERSCAN = 5;\n\n/**\n * Build a stable fingerprint for a prompt covering the full message\n * list (system + history + user turn). Ensures two prompts sharing\n * the user text but differing in prior context do not collide on\n * the exact-match fast path.\n *\n * FNV-1a variant — cheap, collision-resistant enough for a cache,\n * dependency-free. NOT a cryptographic hash: collisions would\n * surface as wrong cache hits, not a security issue in the current\n * trust model.\n */\nfunction hashPrompt(messages: ReadonlyArray<Message>): string {\n  return fnv1a(\n    messages\n      .map((message) => {\n        const role = message.role;\n        const content = Array.isArray(message.content)\n          ? message.content\n              .filter((part) => part.type === \"text\")\n              .map((part) => (part as { text: string }).text)\n              .join(\"|\")\n          : message.content;\n\n        return `${role}:${content}`;\n      })\n      .join(\"||\"),\n  );\n}\n\n/** FNV-1a over a string — see {@link hashPrompt} for the caveats. */\nfunction fnv1a(serialized: string): string {\n  let hash = 0x811c9dc5;\n\n  for (let index = 0; index < serialized.length; index++) {\n    hash ^= serialized.charCodeAt(index);\n    hash = Math.imul(hash, 0x01000193);\n  }\n\n  return (hash >>> 0).toString(16);\n}\n\n/**\n * Resolve the isolation key this trip reads and writes under.\n *\n * Derived from the run's own `sessionId` (or the developer's resolver) —\n * never from the prompt, the model's output, or anything the LLM can\n * write to. `\"shared\"` and an unidentified run both resolve to\n * `undefined`, i.e. the unscoped pool, which a scoped lookup can never\n * read.\n */\nfunction resolveScope(\n  scope: SemanticCacheScope,\n  context: MiddlewareTripContext,\n): string | undefined {\n  if (scope === \"shared\") {\n    return undefined;\n  }\n\n  const key =\n    typeof scope === \"function\"\n      ? scope(context)\n      : sessionScope(context.options?.sessionId);\n\n  return key ? key : undefined;\n}\n\n/**\n * The default `\"session\"` key: the session id under a reserved prefix so\n * a custom resolver returning a bare tenant id can't collide with a\n * session pool. Mirrors the orchestrator's `sessionMemoryScope`.\n */\nfunction sessionScope(sessionId: string | undefined): string | undefined {\n  return sessionId ? `session:${sessionId}` : undefined;\n}\n\nfunction isFresh(entry: CachedEntry, ttlMs: number | undefined): boolean {\n  if (ttlMs === undefined) {\n    return true;\n  }\n\n  return Date.now() - entry.storedAt <= ttlMs;\n}\n\n/**\n * Semantic-similarity response cache for an agent run.\n *\n * **Role.** Skips LLM round-trips when the current prompt is\n * semantically close to one the agent has already answered. For\n * FAQ / support-style traffic this often eliminates 60–80% of\n * model calls — the production win is massive for cost and\n * latency.\n *\n * **Delegation to `@warlock.js/cache`.** This middleware does NOT\n * implement similarity search itself. It delegates to the supplied\n * `CacheDriver`. Production deployments pick a driver with an ANN\n * index (`pg` + pgvector, `redis` + RediSearch). Dev / test\n * environments pass `new MemoryCacheDriver()` — zero config, correct,\n * but O(N) per query. Drivers without similarity support throw\n * `CacheUnsupportedError` from `set({ vector })` / `similar()`.\n *\n * **Two-tier lookup.**\n * 1. *Exact-match key* — a cheap FNV hash over the entire message\n *    list. `store.get(hash)` returns the entry without an embedding\n *    round trip when the prompt hasn't changed at all.\n * 2. *Vector-match* — on exact-match miss, embed the prompt and\n *    call `store.similar(vector, { topK: 1, threshold })`. The\n *    driver uses its native similarity index; anything clearing\n *    `threshold` is returned as a hit.\n *\n * **Write-on-miss.** When both tiers miss, `trip.before` stashes\n * the prompt hash + vector in `ctx.state`; `trip.after` reads back\n * the pending entry and calls\n * `store.set(hash, entry, { vector })`. If an outer middleware\n * (guardrail) throws in `trip.after` before the cache's `trip.after`\n * runs, the pending entry is never written — bad responses stay out\n * of the cache **as long as the canonical install order is followed**\n * (cache outermost).\n *\n * **Synthetic-response on hit.** Returns a `ModelResponse` with\n * `usage: { input: 0, output: 0, total: 0 }` so budget /\n * observability correctly exclude the saved trip.\n *\n * **Per-session scoping (4.15.0).** One `semanticCache` instance\n * normally serves every end user, and a hit is returned as the answer\n * with no model call in between — so entries are keyed by the run's\n * `sessionId` (`scope`, default `\"session\"`) and a lookup only ever\n * sees entries written under the same key. Runs made without a\n * `sessionId` share one unscoped pool; pass `sessionId` on\n * `agent.execute()` (composites thread their own through automatically)\n * to get the isolation, or set `scope: \"shared\"` to pool deliberately.\n * Note the cost/benefit shift: scoping trades cross-user hit rate for\n * isolation, so public-FAQ deployments where no response can carry a\n * caller's private context should opt into `\"shared\"` explicitly.\n *\n * @example\n * import { semanticCache } from \"@warlock.js/ai\";\n * import { MemoryCacheDriver } from \"@warlock.js/cache\";\n *\n * const store = new MemoryCacheDriver();\n * store.setOptions({});\n *\n * const cache = semanticCache({\n *   embedder: openai.embedder({ name: \"text-embedding-3-small\" }),\n *   store,\n *   threshold: 0.95,\n *   ttlMs: 60 * 60 * 1000,\n * });\n *\n * const myAgent = agent({ model, middleware: [cache] });\n */\nexport function semanticCache(options: SemanticCacheOptions): AgentMiddleware {\n  const name = options.name ?? \"semantic-cache\";\n  const namespace = options.namespace ?? DEFAULT_NAMESPACE;\n  const scopeMode: SemanticCacheScope = options.scope ?? \"session\";\n  const pendingKey = `${name}.pending`;\n\n  // Resolve the effective store at factory time, not per-call. Every\n  // subsequent hook closes over `store` so the resolution happens once.\n  // Throws now (loud, at construction) instead of later during the\n  // first trip (silent until the agent actually runs).\n  const store = options.store ?? resolveDefaultStore();\n\n  if (!store) {\n    throw new Error(\n      `semanticCache: no store supplied — pass \\`store\\` in options or call \\`ai.config({ defaultStore })\\` at app boot before constructing the middleware`,\n    );\n  }\n\n  // Cache's parseKey replaces \":\" with \".\" so the namespace boundary\n  // matches what `similar()` actually returns in `hit.key`. Using a\n  // dot here keeps prefix checks aligned with stored keys.\n  //\n  // A scoped entry gets an extra hashed segment, so two sessions asking\n  // the identical question stay two entries instead of overwriting each\n  // other; the scope is hashed because a `sessionId` is caller-supplied\n  // and may contain the key delimiter. The unscoped key shape is\n  // unchanged, so pre-4.15.0 entries still resolve. The hash is a\n  // write-separation device only — a read is authorized by the exact\n  // `entry.scope` equality check below, so even a hash collision cannot\n  // widen what a session can read.\n  const keyFor = (hash: string, scope: string | undefined): string =>\n    scope === undefined\n      ? `${namespace}.${hash}`\n      : `${namespace}.${fnv1a(scope)}.${hash}`;\n\n  return {\n    name,\n    log: true,\n    trip: {\n      async before(context) {\n        // Only cache the first trip's response. Subsequent trips\n        // happen because the previous trip requested tool calls — the\n        // message list now carries tool results the original prompt\n        // never saw, so a semantic match on the unchanged user text\n        // would serve back the prior `tool_calls` response and loop\n        // the agent forever. The first turn is also the only one\n        // where a \"same question → same final answer\" caching story\n        // is sound.\n        if (context.tripIndex !== 0) {\n          return;\n        }\n\n        const promptText = extractUserText(context.messages);\n\n        if (!promptText) {\n          return;\n        }\n\n        const scope = resolveScope(scopeMode, context);\n        const promptKey = hashPrompt(context.messages);\n\n        const exact = await store.get<CachedEntry>(keyFor(promptKey, scope));\n\n        // The key already carries the scope; re-checking the stored\n        // `scope` is the actual authorization step, so a key collision\n        // or a hand-written entry can't serve across the boundary.\n        if (exact && exact.scope === scope && isFresh(exact, options.ttlMs)) {\n          return toSyntheticResponse(exact.response);\n        }\n\n        const query = await options.embedder.embed(promptText);\n\n        const hits = await store.similar<CachedEntry>(query.vector, {\n          topK: scope === undefined ? 1 : SIMILAR_OVERSCAN,\n          threshold: options.threshold,\n        });\n\n        // Only entries written inside this cache's namespace AND this\n        // caller's scope are eligible. A shared driver would otherwise\n        // leak a foreign namespace's entries; a shared namespace would\n        // leak another session's answer to this one.\n        const hit = hits.find(\n          (candidate) =>\n            candidate.key.startsWith(`${namespace}.`) &&\n            candidate.value?.scope === scope &&\n            isFresh(candidate.value, options.ttlMs),\n        );\n\n        if (hit) {\n          return toSyntheticResponse(hit.value.response);\n        }\n\n        const pending: PendingWrite = {\n          promptKey,\n          vector: query.vector,\n          scope,\n        };\n        context.state.set(pendingKey, pending);\n\n        return;\n      },\n      async after(context, response) {\n        const pending = context.state.get(pendingKey) as PendingWrite | undefined;\n\n        if (!pending) {\n          return;\n        }\n\n        // Mid-stream tool-call responses must not be cached — the\n        // useful answer comes from the trip *after* the tool returns.\n        // Crucially, leave the pending entry in place so a later trip\n        // (the one that actually finishes with `stop`) can read it\n        // and write the final response under the *original* trip-0\n        // prompt key. Deleting here would orphan the pending and the\n        // post-tool answer would never make it into the store.\n        if (response.finishReason === \"tool_calls\") {\n          return;\n        }\n\n        context.state.delete(pendingKey);\n\n        const entry: CachedEntry = {\n          response,\n          storedAt: Date.now(),\n          scope: pending.scope,\n        };\n\n        await store.set(keyFor(pending.promptKey, pending.scope), entry, {\n          vector: pending.vector,\n        });\n\n        return;\n      },\n    },\n  };\n}\n\nfunction toSyntheticResponse(response: ModelResponse): ModelResponse {\n  return {\n    content: response.content,\n    finishReason: response.finishReason,\n    usage: { input: 0, output: 0, total: 0 },\n    toolCalls: response.toolCalls,\n  };\n}\n"],"mappings":";;;;AAyGA,MAAM,oBAAoB;;;;;;;AAQ1B,MAAM,mBAAmB;;;;;;;;;;;;AAazB,SAAS,WAAW,UAA0C;CAC5D,OAAO,MACL,SACG,KAAK,YAAY;EAShB,OAAO,GARM,QAAQ,KAQN,GAPC,MAAM,QAAQ,QAAQ,OAAO,IACzC,QAAQ,QACL,QAAQ,SAAS,KAAK,SAAS,MAAM,EACrC,KAAK,SAAU,KAA0B,IAAI,EAC7C,KAAK,GAAG,IACX,QAAQ;CAGd,CAAC,EACA,KAAK,IAAI,CACd;AACF;;AAGA,SAAS,MAAM,YAA4B;CACzC,IAAI,OAAO;CAEX,KAAK,IAAI,QAAQ,GAAG,QAAQ,WAAW,QAAQ,SAAS;EACtD,QAAQ,WAAW,WAAW,KAAK;EACnC,OAAO,KAAK,KAAK,MAAM,QAAU;CACnC;CAEA,QAAQ,SAAS,GAAG,SAAS,EAAE;AACjC;;;;;;;;;;AAWA,SAAS,aACP,OACA,SACoB;CACpB,IAAI,UAAU,UACZ;CAGF,MAAM,MACJ,OAAO,UAAU,aACb,MAAM,OAAO,IACb,aAAa,QAAQ,SAAS,SAAS;CAE7C,OAAO,MAAM,MAAM;AACrB;;;;;;AAOA,SAAS,aAAa,WAAmD;CACvE,OAAO,YAAY,WAAW,cAAc;AAC9C;AAEA,SAAS,QAAQ,OAAoB,OAAoC;CACvE,IAAI,UAAU,QACZ,OAAO;CAGT,OAAO,KAAK,IAAI,IAAI,MAAM,YAAY;AACxC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqEA,SAAgB,cAAc,SAAgD;CAC5E,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,YAAY,QAAQ,aAAa;CACvC,MAAM,YAAgC,QAAQ,SAAS;CACvD,MAAM,aAAa,GAAG,KAAK;CAM3B,MAAM,QAAQ,QAAQ,SAAS,oBAAoB;CAEnD,IAAI,CAAC,OACH,MAAM,IAAI,MACR,qJACF;CAeF,MAAM,UAAU,MAAc,UAC5B,UAAU,SACN,GAAG,UAAU,GAAG,SAChB,GAAG,UAAU,GAAG,MAAM,KAAK,EAAE,GAAG;CAEtC,OAAO;EACL;EACA,KAAK;EACL,MAAM;GACJ,MAAM,OAAO,SAAS;IASpB,IAAI,QAAQ,cAAc,GACxB;IAGF,MAAM,aAAa,gBAAgB,QAAQ,QAAQ;IAEnD,IAAI,CAAC,YACH;IAGF,MAAM,QAAQ,aAAa,WAAW,OAAO;IAC7C,MAAM,YAAY,WAAW,QAAQ,QAAQ;IAE7C,MAAM,QAAQ,MAAM,MAAM,IAAiB,OAAO,WAAW,KAAK,CAAC;IAKnE,IAAI,SAAS,MAAM,UAAU,SAAS,QAAQ,OAAO,QAAQ,KAAK,GAChE,OAAO,oBAAoB,MAAM,QAAQ;IAG3C,MAAM,QAAQ,MAAM,QAAQ,SAAS,MAAM,UAAU;IAWrD,MAAM,OAAM,MATO,MAAM,QAAqB,MAAM,QAAQ;KAC1D,MAAM,UAAU,SAAY,IAAI;KAChC,WAAW,QAAQ;IACrB,CAAC,GAMgB,MACd,cACC,UAAU,IAAI,WAAW,GAAG,UAAU,EAAE,KACxC,UAAU,OAAO,UAAU,SAC3B,QAAQ,UAAU,OAAO,QAAQ,KAAK,CAC1C;IAEA,IAAI,KACF,OAAO,oBAAoB,IAAI,MAAM,QAAQ;IAG/C,MAAM,UAAwB;KAC5B;KACA,QAAQ,MAAM;KACd;IACF;IACA,QAAQ,MAAM,IAAI,YAAY,OAAO;GAGvC;GACA,MAAM,MAAM,SAAS,UAAU;IAC7B,MAAM,UAAU,QAAQ,MAAM,IAAI,UAAU;IAE5C,IAAI,CAAC,SACH;IAUF,IAAI,SAAS,iBAAiB,cAC5B;IAGF,QAAQ,MAAM,OAAO,UAAU;IAE/B,MAAM,QAAqB;KACzB;KACA,UAAU,KAAK,IAAI;KACnB,OAAO,QAAQ;IACjB;IAEA,MAAM,MAAM,IAAI,OAAO,QAAQ,WAAW,QAAQ,KAAK,GAAG,OAAO,EAC/D,QAAQ,QAAQ,OAClB,CAAC;GAGH;EACF;CACF;AACF;AAEA,SAAS,oBAAoB,UAAwC;CACnE,OAAO;EACL,SAAS,SAAS;EAClB,cAAc,SAAS;EACvB,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACvC,WAAW,SAAS;CACtB;AACF"}
+{"version":3,"file":"semantic-cache.mjs","names":[],"sources":["../../../../../../../../ai/src/middleware/builtins/semantic-cache.ts"],"sourcesContent":["import type { CacheDriver } from \"@warlock.js/cache\";\nimport { resolveDefaultStore } from \"../../config\";\nimport type { Message } from \"../../contracts/conversation-message.type\";\nimport type { EmbedderContract } from \"../../contracts/embedder.contract\";\nimport type { AgentMiddleware } from \"../../contracts/middleware\";\nimport type { MiddlewareTripContext } from \"../../contracts/middleware/middleware-context.type\";\nimport type { ModelResponse } from \"../../contracts/model.contract\";\nimport { extractUserText } from \"../utils\";\n\n/**\n * Isolation boundary for cache reads and writes.\n *\n * - `\"session\"` (default) — key every entry off the run's\n *   `AgentExecuteOptions.sessionId`, so one session never receives a\n *   response cached for another. Calls made WITHOUT a `sessionId` share\n *   one unscoped pool (the pre-4.15.0 behavior); an unscoped read never\n *   sees a session-scoped entry and vice versa.\n * - `\"shared\"` — one pool for every caller, regardless of session. The\n *   explicit opt-in for genuinely public Q&A (docs bots, FAQ) where the\n *   cross-user hit rate is the point and no response can carry one\n *   caller's private context.\n * - a resolver — derive the key yourself, e.g. per tenant\n *   (`ctx => ctx.options?.toolCtx?.tenantId`). Returning `undefined`\n *   falls back to the unscoped pool, so return a constant sentinel (or\n *   throw) if you need the call to fail closed instead.\n */\nexport type SemanticCacheScope =\n  | \"session\"\n  | \"shared\"\n  | ((context: MiddlewareTripContext) => string | undefined);\n\n/**\n * Configuration for `semanticCache()`.\n */\nexport type SemanticCacheOptions = {\n  /** Embedder used to produce the query vector from the prompt text. */\n  embedder: EmbedderContract;\n  /**\n   * Vector-capable cache driver from `@warlock.js/cache`. Production\n   * deployments pick a driver with a real ANN index (`pg` with\n   * pgvector, `redis` with RediSearch). Dev / test environments use\n   * `new MemoryCacheDriver()` — zero config, correct, but O(N) per\n   * query. Drivers without similarity support throw\n   * `CacheUnsupportedError` from `set({ vector })` / `similar()`.\n   *\n   * Falls back to `ai.config({ defaultStore })` when omitted. When\n   * neither is set, the factory throws at construction time —\n   * semantic cache cannot operate without a store.\n   */\n  store?: CacheDriver<any, any>;\n  /**\n   * Minimum cosine similarity for a vector hit. Between 0 and 1 —\n   * 0.95 is a solid default for question-answering caches.\n   */\n  threshold: number;\n  /**\n   * Optional TTL in milliseconds. Entries whose `storedAt` is older\n   * than this are treated as misses on read and overwritten on the\n   * next write. Default: no expiry — entries live until the store\n   * evicts them (per its own TTL/eviction policy).\n   */\n  ttlMs?: number;\n  /**\n   * Namespace prefix applied to every key the cache writes. Lets\n   * multiple agents share one driver without collision. Default\n   * `\"ai.cache\"`.\n   */\n  namespace?: string;\n  /**\n   * Per-caller isolation boundary. Default `\"session\"` — a cached\n   * response is served back only to the session that produced it.\n   *\n   * A `semanticCache` is normally built once at app boot and shared by\n   * every end user, and a hit is returned as the model's answer with no\n   * LLM call in between; without a scope that pools every caller's Q&A\n   * pairs into one namespace, which is both a disclosure path (user B's\n   * near-enough prompt gets served user A's answer, personal context\n   * included) and a poisoning path (an attacker seeds an entry near a\n   * predictable future query). Set `\"shared\"` to opt back into pooling\n   * where that is actually desirable. See {@link SemanticCacheScope}.\n   */\n  scope?: SemanticCacheScope;\n  /**\n   * Middleware name — also the state-bag key prefix inside a single\n   * execution. Default `\"semantic-cache\"`.\n   */\n  name?: string;\n};\n\ntype CachedEntry = {\n  response: ModelResponse;\n  storedAt: number;\n  /**\n   * Isolation key the entry was written under; absent = the unscoped\n   * pool (also the shape of every entry written before 4.15.0).\n   */\n  scope?: string;\n};\n\ntype PendingWrite = {\n  promptKey: string;\n  vector: number[];\n  scope?: string;\n};\n\nconst DEFAULT_NAMESPACE = \"ai.cache\";\n\n/**\n * Extra candidates pulled from `similar()` on a SCOPED lookup before the\n * scope filter runs. The driver ranks across every scope in the index,\n * so a bare `topK: 1` can come back as a foreign entry and mask this\n * scope's own legitimate hit. Mirrors the memory tiers' overscan.\n */\nconst SIMILAR_OVERSCAN = 5;\n\n/**\n * Build a stable fingerprint for a prompt covering the full message\n * list (system + history + user turn). Ensures two prompts sharing\n * the user text but differing in prior context do not collide on\n * the exact-match fast path.\n *\n * FNV-1a variant — cheap, collision-resistant enough for a cache,\n * dependency-free. NOT a cryptographic hash: collisions would\n * surface as wrong cache hits, not a security issue in the current\n * trust model.\n */\nfunction hashPrompt(messages: ReadonlyArray<Message>): string {\n  return fnv1a(\n    messages\n      .map((message) => {\n        const role = message.role;\n        const content = Array.isArray(message.content)\n          ? message.content\n              .filter((part) => part.type === \"text\")\n              .map((part) => (part as { text: string }).text)\n              .join(\"|\")\n          : message.content;\n\n        return `${role}:${content}`;\n      })\n      .join(\"||\"),\n  );\n}\n\n/** FNV-1a over a string — see {@link hashPrompt} for the caveats. */\nfunction fnv1a(serialized: string): string {\n  let hash = 0x811c9dc5;\n\n  for (let index = 0; index < serialized.length; index++) {\n    hash ^= serialized.charCodeAt(index);\n    hash = Math.imul(hash, 0x01000193);\n  }\n\n  return (hash >>> 0).toString(16);\n}\n\n/**\n * Resolve the isolation key this trip reads and writes under.\n *\n * Derived from the run's own `sessionId` (or the developer's resolver) —\n * never from the prompt, the model's output, or anything the LLM can\n * write to. `\"shared\"` and an unidentified run both resolve to\n * `undefined`, i.e. the unscoped pool, which a scoped lookup can never\n * read.\n */\nfunction resolveScope(\n  scope: SemanticCacheScope,\n  context: MiddlewareTripContext,\n): string | undefined {\n  if (scope === \"shared\") {\n    return undefined;\n  }\n\n  const key =\n    typeof scope === \"function\"\n      ? scope(context)\n      : sessionScope(context.options?.sessionId);\n\n  return key ? key : undefined;\n}\n\n/**\n * The default `\"session\"` key: the session id under a reserved prefix so\n * a custom resolver returning a bare tenant id can't collide with a\n * session pool. Mirrors the orchestrator's `sessionMemoryScope`.\n */\nfunction sessionScope(sessionId: string | undefined): string | undefined {\n  return sessionId ? `session:${sessionId}` : undefined;\n}\n\nfunction isFresh(entry: CachedEntry, ttlMs: number | undefined): boolean {\n  if (ttlMs === undefined) {\n    return true;\n  }\n\n  return Date.now() - entry.storedAt <= ttlMs;\n}\n\n/**\n * Semantic-similarity response cache for an agent run.\n *\n * **Role.** Skips LLM round-trips when the current prompt is\n * semantically close to one the agent has already answered. For\n * FAQ / support-style traffic this often eliminates 60–80% of\n * model calls — the production win is massive for cost and\n * latency.\n *\n * **Delegation to `@warlock.js/cache`.** This middleware does NOT\n * implement similarity search itself. It delegates to the supplied\n * `CacheDriver`. Production deployments pick a driver with an ANN\n * index (`pg` + pgvector, `redis` + RediSearch). Dev / test\n * environments pass `new MemoryCacheDriver()` — zero config, correct,\n * but O(N) per query. Drivers without similarity support throw\n * `CacheUnsupportedError` from `set({ vector })` / `similar()`.\n *\n * **Two-tier lookup.**\n * 1. *Exact-match key* — a cheap FNV hash over the entire message\n *    list. `store.get(hash)` returns the entry without an embedding\n *    round trip when the prompt hasn't changed at all.\n * 2. *Vector-match* — on exact-match miss, embed the prompt and\n *    call `store.similar(vector, { topK: 1, threshold })`. The\n *    driver uses its native similarity index; anything clearing\n *    `threshold` is returned as a hit.\n *\n * **Write-on-miss.** When both tiers miss, `trip.before` stashes\n * the prompt hash + vector in `ctx.state`; `trip.after` reads back\n * the pending entry and calls\n * `store.set(hash, entry, { vector })`. If an outer middleware\n * (guardrail) throws in `trip.after` before the cache's `trip.after`\n * runs, the pending entry is never written — bad responses stay out\n * of the cache **as long as the canonical install order is followed**\n * (cache outermost).\n *\n * **Synthetic-response on hit.** Returns a `ModelResponse` with\n * `usage: { input: 0, output: 0, total: 0 }` so budget /\n * observability correctly exclude the saved trip.\n *\n * **Per-session scoping (4.15.0).** One `semanticCache` instance\n * normally serves every end user, and a hit is returned as the answer\n * with no model call in between — so entries are keyed by the run's\n * `sessionId` (`scope`, default `\"session\"`) and a lookup only ever\n * sees entries written under the same key. Runs made without a\n * `sessionId` share one unscoped pool; pass `sessionId` on\n * `agent.execute()` (composites thread their own through automatically)\n * to get the isolation, or set `scope: \"shared\"` to pool deliberately.\n * Note the cost/benefit shift: scoping trades cross-user hit rate for\n * isolation, so public-FAQ deployments where no response can carry a\n * caller's private context should opt into `\"shared\"` explicitly.\n *\n * @example\n * import { semanticCache } from \"@warlock.js/ai\";\n * import { MemoryCacheDriver } from \"@warlock.js/cache\";\n *\n * const store = new MemoryCacheDriver();\n * store.setOptions({});\n *\n * const cache = semanticCache({\n *   embedder: openai.embedder({ name: \"text-embedding-3-small\" }),\n *   store,\n *   threshold: 0.95,\n *   ttlMs: 60 * 60 * 1000,\n * });\n *\n * const myAgent = agent({ model, middleware: [cache] });\n */\nexport function semanticCache(options: SemanticCacheOptions): AgentMiddleware {\n  const name = options.name ?? \"semantic-cache\";\n  const namespace = options.namespace ?? DEFAULT_NAMESPACE;\n  const scopeMode: SemanticCacheScope = options.scope ?? \"session\";\n  const pendingKey = `${name}.pending`;\n\n  // Resolve the effective store at factory time, not per-call. Every\n  // subsequent hook closes over `store` so the resolution happens once.\n  // Throws now (loud, at construction) instead of later during the\n  // first trip (silent until the agent actually runs).\n  const store = options.store ?? resolveDefaultStore();\n\n  if (!store) {\n    throw new Error(\n      `semanticCache: no store supplied — pass \\`store\\` in options or call \\`ai.config({ defaultStore })\\` at app boot before constructing the middleware`,\n    );\n  }\n\n  // Cache's parseKey replaces \":\" with \".\" so the namespace boundary\n  // matches what `similar()` actually returns in `hit.key`. Using a\n  // dot here keeps prefix checks aligned with stored keys.\n  //\n  // A scoped entry gets an extra hashed segment, so two sessions asking\n  // the identical question stay two entries instead of overwriting each\n  // other; the scope is hashed because a `sessionId` is caller-supplied\n  // and may contain the key delimiter. The unscoped key shape is\n  // unchanged, so pre-4.15.0 entries still resolve. The hash is a\n  // write-separation device only — a read is authorized by the exact\n  // `entry.scope` equality check below, so even a hash collision cannot\n  // widen what a session can read.\n  const keyFor = (hash: string, scope: string | undefined): string =>\n    scope === undefined\n      ? `${namespace}.${hash}`\n      : `${namespace}.${fnv1a(scope)}.${hash}`;\n\n  return {\n    name,\n    log: true,\n    trip: {\n      async before(context) {\n        // Only cache the first trip's response. Subsequent trips\n        // happen because the previous trip requested tool calls — the\n        // message list now carries tool results the original prompt\n        // never saw, so a semantic match on the unchanged user text\n        // would serve back the prior `tool_calls` response and loop\n        // the agent forever. The first turn is also the only one\n        // where a \"same question → same final answer\" caching story\n        // is sound.\n        if (context.tripIndex !== 0) {\n          return;\n        }\n\n        const promptText = extractUserText(context.messages);\n\n        if (!promptText) {\n          return;\n        }\n\n        const scope = resolveScope(scopeMode, context);\n        const promptKey = hashPrompt(context.messages);\n\n        const exact = await store.get<CachedEntry>(keyFor(promptKey, scope));\n\n        // The key already carries the scope; re-checking the stored\n        // `scope` is the actual authorization step, so a key collision\n        // or a hand-written entry can't serve across the boundary.\n        if (exact && exact.scope === scope && isFresh(exact, options.ttlMs)) {\n          return toSyntheticResponse(exact.response);\n        }\n\n        const query = await options.embedder.embed(promptText);\n\n        const hits = await store.similar<CachedEntry>(query.vector, {\n          topK: scope === undefined ? 1 : SIMILAR_OVERSCAN,\n          threshold: options.threshold,\n        });\n\n        // Only entries written inside this cache's namespace AND this\n        // caller's scope are eligible. A shared driver would otherwise\n        // leak a foreign namespace's entries; a shared namespace would\n        // leak another session's answer to this one.\n        const hit = hits.find(\n          (candidate) =>\n            candidate.key.startsWith(`${namespace}.`) &&\n            candidate.value?.scope === scope &&\n            isFresh(candidate.value, options.ttlMs),\n        );\n\n        if (hit) {\n          return toSyntheticResponse(hit.value.response);\n        }\n\n        const pending: PendingWrite = {\n          promptKey,\n          vector: query.vector,\n          scope,\n        };\n        context.state.set(pendingKey, pending);\n\n        return;\n      },\n      async after(context, response) {\n        const pending = context.state.get(pendingKey) as PendingWrite | undefined;\n\n        if (!pending) {\n          return;\n        }\n\n        // Mid-stream tool-call responses must not be cached — the\n        // useful answer comes from the trip *after* the tool returns.\n        // Crucially, leave the pending entry in place so a later trip\n        // (the one that actually finishes with `stop`) can read it\n        // and write the final response under the *original* trip-0\n        // prompt key. Deleting here would orphan the pending and the\n        // post-tool answer would never make it into the store.\n        if (response.finishReason === \"tool_calls\") {\n          return;\n        }\n\n        context.state.delete(pendingKey);\n\n        const entry: CachedEntry = {\n          response,\n          storedAt: Date.now(),\n          scope: pending.scope,\n        };\n\n        await store.set(keyFor(pending.promptKey, pending.scope), entry, {\n          vector: pending.vector,\n        });\n\n        return;\n      },\n    },\n  };\n}\n\nfunction toSyntheticResponse(response: ModelResponse): ModelResponse {\n  return {\n    content: response.content,\n    finishReason: response.finishReason,\n    usage: { input: 0, output: 0, total: 0 },\n    toolCalls: response.toolCalls,\n  };\n}\n"],"mappings":";;;;AAyGA,MAAM,oBAAoB;;;;;;;AAQ1B,MAAM,mBAAmB;;;;;;;;;;;;AAazB,SAAS,WAAW,UAA0C;CAC5D,OAAO,MACL,SACG,KAAK,YAAY;EAShB,OAAO,GARM,QAAQ,KAQN,GAPC,MAAM,QAAQ,QAAQ,OAAO,IACzC,QAAQ,QACL,QAAQ,SAAS,KAAK,SAAS,MAAM,CAAC,CACtC,KAAK,SAAU,KAA0B,IAAI,CAAC,CAC9C,KAAK,GAAG,IACX,QAAQ;CAGd,CAAC,CAAC,CACD,KAAK,IAAI,CACd;AACF;;AAGA,SAAS,MAAM,YAA4B;CACzC,IAAI,OAAO;CAEX,KAAK,IAAI,QAAQ,GAAG,QAAQ,WAAW,QAAQ,SAAS;EACtD,QAAQ,WAAW,WAAW,KAAK;EACnC,OAAO,KAAK,KAAK,MAAM,QAAU;CACnC;CAEA,QAAQ,SAAS,EAAC,CAAE,SAAS,EAAE;AACjC;;;;;;;;;;AAWA,SAAS,aACP,OACA,SACoB;CACpB,IAAI,UAAU,UACZ;CAGF,MAAM,MACJ,OAAO,UAAU,aACb,MAAM,OAAO,IACb,aAAa,QAAQ,SAAS,SAAS;CAE7C,OAAO,MAAM,MAAM;AACrB;;;;;;AAOA,SAAS,aAAa,WAAmD;CACvE,OAAO,YAAY,WAAW,cAAc;AAC9C;AAEA,SAAS,QAAQ,OAAoB,OAAoC;CACvE,IAAI,UAAU,QACZ,OAAO;CAGT,OAAO,KAAK,IAAI,IAAI,MAAM,YAAY;AACxC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqEA,SAAgB,cAAc,SAAgD;CAC5E,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,YAAY,QAAQ,aAAa;CACvC,MAAM,YAAgC,QAAQ,SAAS;CACvD,MAAM,aAAa,GAAG,KAAK;CAM3B,MAAM,QAAQ,QAAQ,SAAS,oBAAoB;CAEnD,IAAI,CAAC,OACH,MAAM,IAAI,MACR,qJACF;CAeF,MAAM,UAAU,MAAc,UAC5B,UAAU,SACN,GAAG,UAAU,GAAG,SAChB,GAAG,UAAU,GAAG,MAAM,KAAK,EAAE,GAAG;CAEtC,OAAO;EACL;EACA,KAAK;EACL,MAAM;GACJ,MAAM,OAAO,SAAS;IASpB,IAAI,QAAQ,cAAc,GACxB;IAGF,MAAM,aAAa,gBAAgB,QAAQ,QAAQ;IAEnD,IAAI,CAAC,YACH;IAGF,MAAM,QAAQ,aAAa,WAAW,OAAO;IAC7C,MAAM,YAAY,WAAW,QAAQ,QAAQ;IAE7C,MAAM,QAAQ,MAAM,MAAM,IAAiB,OAAO,WAAW,KAAK,CAAC;IAKnE,IAAI,SAAS,MAAM,UAAU,SAAS,QAAQ,OAAO,QAAQ,KAAK,GAChE,OAAO,oBAAoB,MAAM,QAAQ;IAG3C,MAAM,QAAQ,MAAM,QAAQ,SAAS,MAAM,UAAU;IAWrD,MAAM,OAAM,MATO,MAAM,QAAqB,MAAM,QAAQ;KAC1D,MAAM,UAAU,SAAY,IAAI;KAChC,WAAW,QAAQ;IACrB,CAAC,EAMe,CAAC,MACd,cACC,UAAU,IAAI,WAAW,GAAG,UAAU,EAAE,KACxC,UAAU,OAAO,UAAU,SAC3B,QAAQ,UAAU,OAAO,QAAQ,KAAK,CAC1C;IAEA,IAAI,KACF,OAAO,oBAAoB,IAAI,MAAM,QAAQ;IAG/C,MAAM,UAAwB;KAC5B;KACA,QAAQ,MAAM;KACd;IACF;IACA,QAAQ,MAAM,IAAI,YAAY,OAAO;GAGvC;GACA,MAAM,MAAM,SAAS,UAAU;IAC7B,MAAM,UAAU,QAAQ,MAAM,IAAI,UAAU;IAE5C,IAAI,CAAC,SACH;IAUF,IAAI,SAAS,iBAAiB,cAC5B;IAGF,QAAQ,MAAM,OAAO,UAAU;IAE/B,MAAM,QAAqB;KACzB;KACA,UAAU,KAAK,IAAI;KACnB,OAAO,QAAQ;IACjB;IAEA,MAAM,MAAM,IAAI,OAAO,QAAQ,WAAW,QAAQ,KAAK,GAAG,OAAO,EAC/D,QAAQ,QAAQ,OAClB,CAAC;GAGH;EACF;CACF;AACF;AAEA,SAAS,oBAAoB,UAAwC;CACnE,OAAO;EACL,SAAS,SAAS;EAClB,cAAc,SAAS;EACvB,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACvC,WAAW,SAAS;CACtB;AACF"}

package/esm/middleware/helpers/compose.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"compose.d.mts","names":[],"sources":["../../../../../../../../ai/src/middleware/helpers/compose.ts"],"mappings":";;;;;AAsCA;;;;;;;;;;;;;;;;;AAEkB;;;;;;;;;;;;;;;;iBAFF,iBAAA,CAAA,GACX,OAAA,EAAS,aAAA,CAAc,eAAA,GAAkB,aAAA,CAAc,eAAA,KACzD,eAAA"}
+{"version":3,"file":"compose.d.mts","names":[],"sources":["../../../../../../../../ai/src/middleware/helpers/compose.ts"],"mappings":";;;;;AAsCA;;;;;;;;;;;;;;;;;AAEkB;;;;;;;;;;;;;;;;iBAFF,iBAAA,IACX,OAAA,EAAS,aAAA,CAAc,eAAA,GAAkB,aAAA,CAAc,eAAA,KACzD,eAAA"}

package/esm/middleware/helpers/for-tool.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"for-tool.mjs","names":[],"sources":["../../../../../../../../ai/src/middleware/helpers/for-tool.ts"],"sourcesContent":["import type { AgentMiddleware } from \"../../contracts/middleware\";\n\n/**\n * Scope a middleware's `tool`-level hooks to only fire for a\n * specific tool name (or a set of names). `execute` and `trip`\n * hooks pass through unchanged.\n *\n * **Role.** Tool-specific concerns — \"rate-limit `search_web`\",\n * \"cache results for `fetch_faq`\" — are common. Rather than adding\n * a `middleware` field to `ai.tool()` (see decisions §27), the\n * framework keeps one contract (`AgentMiddleware`) and offers this\n * helper for the locality problem. The middleware body stays agnostic\n * of the tool name; `forTool` handles the filtering.\n *\n * **What gets filtered.** Only `tool.before` / `tool.after` /\n * `tool.onError`. Each hook is wrapped so that `ctx.tool.name`\n * must be in the allowed set or the wrapped hook is a no-op.\n * `execute` and `trip` hooks are NOT touched — they run normally.\n *\n * **Why not filter execute/trip too?** Because a middleware that\n * reaches across levels (a tool-specific budget that initializes a\n * counter in `execute.before` and checks it in `tool.before`) still\n * needs `execute.before` to fire unconditionally. Scoping all hooks\n * would break cross-level middleware; scoping only `tool` hooks\n * matches the mental model of \"this middleware cares about these\n * tools.\"\n *\n * **Single-name vs multi-name.** A string matches one tool; a string\n * array matches any of the listed tools. No wildcards, no regex —\n * keep it boring.\n *\n * @example\n * // Single tool\n * const scoped = ai.middleware.forTool(\n *   \"search_web\",\n *   toolRateLimit({ maxCalls: 3 }),\n * );\n *\n * @example\n * // Multiple tools sharing a rule\n * const scoped = ai.middleware.forTool(\n *   [\"paid_api\", \"expensive_db\"],\n *   toolRateLimit({ maxCalls: 5 }),\n * );\n *\n * ai.agent({\n *   model,\n *   tools: [webTool, paidApiTool, expensiveDbTool],\n *   middleware: [scoped],\n * });\n */\nexport function forTool(\n  toolNames: string | ReadonlyArray<string>,\n  middleware: AgentMiddleware,\n): AgentMiddleware {\n  const allowed = new Set(\n    typeof toolNames === \"string\" ? [toolNames] : toolNames,\n  );\n  const scope =\n    allowed.size === 1 ? Array.from(allowed)[0] : Array.from(allowed).join(\"+\");\n\n  if (!middleware.tool) {\n    return middleware;\n  }\n\n  const innerBefore = middleware.tool.before;\n  const innerAfter = middleware.tool.after;\n  const innerOnError = middleware.tool.onError;\n\n  return {\n    ...middleware,\n    name: `${middleware.name}[for:${scope}]`,\n    tool: {\n      before: innerBefore\n        ? async ctx => {\n            if (!allowed.has(ctx.tool.name)) {\n              return;\n            }\n\n            return innerBefore(ctx);\n          }\n        : undefined,\n      after: innerAfter\n        ? async (ctx, result) => {\n            if (!allowed.has(ctx.tool.name)) {\n              return;\n            }\n\n            return innerAfter(ctx, result);\n          }\n        : undefined,\n      onError: innerOnError\n        ? async (ctx, error) => {\n            if (!allowed.has(ctx.tool.name)) {\n              return;\n            }\n\n            return innerOnError(ctx, error);\n          }\n        : undefined,\n    },\n  };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmDA,SAAgB,QACd,WACA,YACiB;CACjB,MAAM,UAAU,IAAI,IAClB,OAAO,cAAc,WAAW,CAAC,SAAS,IAAI,SAChD;CACA,MAAM,QACJ,QAAQ,SAAS,IAAI,MAAM,KAAK,OAAO,EAAE,KAAK,MAAM,KAAK,OAAO,EAAE,KAAK,GAAG;CAE5E,IAAI,CAAC,WAAW,MACd,OAAO;CAGT,MAAM,cAAc,WAAW,KAAK;CACpC,MAAM,aAAa,WAAW,KAAK;CACnC,MAAM,eAAe,WAAW,KAAK;CAErC,OAAO;EACL,GAAG;EACH,MAAM,GAAG,WAAW,KAAK,OAAO,MAAM;EACtC,MAAM;GACJ,QAAQ,cACJ,OAAM,QAAO;IACX,IAAI,CAAC,QAAQ,IAAI,IAAI,KAAK,IAAI,GAC5B;IAGF,OAAO,YAAY,GAAG;GACxB,IACA;GACJ,OAAO,aACH,OAAO,KAAK,WAAW;IACrB,IAAI,CAAC,QAAQ,IAAI,IAAI,KAAK,IAAI,GAC5B;IAGF,OAAO,WAAW,KAAK,MAAM;GAC/B,IACA;GACJ,SAAS,eACL,OAAO,KAAK,UAAU;IACpB,IAAI,CAAC,QAAQ,IAAI,IAAI,KAAK,IAAI,GAC5B;IAGF,OAAO,aAAa,KAAK,KAAK;GAChC,IACA;EACN;CACF;AACF"}
+{"version":3,"file":"for-tool.mjs","names":[],"sources":["../../../../../../../../ai/src/middleware/helpers/for-tool.ts"],"sourcesContent":["import type { AgentMiddleware } from \"../../contracts/middleware\";\n\n/**\n * Scope a middleware's `tool`-level hooks to only fire for a\n * specific tool name (or a set of names). `execute` and `trip`\n * hooks pass through unchanged.\n *\n * **Role.** Tool-specific concerns — \"rate-limit `search_web`\",\n * \"cache results for `fetch_faq`\" — are common. Rather than adding\n * a `middleware` field to `ai.tool()` (see decisions §27), the\n * framework keeps one contract (`AgentMiddleware`) and offers this\n * helper for the locality problem. The middleware body stays agnostic\n * of the tool name; `forTool` handles the filtering.\n *\n * **What gets filtered.** Only `tool.before` / `tool.after` /\n * `tool.onError`. Each hook is wrapped so that `ctx.tool.name`\n * must be in the allowed set or the wrapped hook is a no-op.\n * `execute` and `trip` hooks are NOT touched — they run normally.\n *\n * **Why not filter execute/trip too?** Because a middleware that\n * reaches across levels (a tool-specific budget that initializes a\n * counter in `execute.before` and checks it in `tool.before`) still\n * needs `execute.before` to fire unconditionally. Scoping all hooks\n * would break cross-level middleware; scoping only `tool` hooks\n * matches the mental model of \"this middleware cares about these\n * tools.\"\n *\n * **Single-name vs multi-name.** A string matches one tool; a string\n * array matches any of the listed tools. No wildcards, no regex —\n * keep it boring.\n *\n * @example\n * // Single tool\n * const scoped = ai.middleware.forTool(\n *   \"search_web\",\n *   toolRateLimit({ maxCalls: 3 }),\n * );\n *\n * @example\n * // Multiple tools sharing a rule\n * const scoped = ai.middleware.forTool(\n *   [\"paid_api\", \"expensive_db\"],\n *   toolRateLimit({ maxCalls: 5 }),\n * );\n *\n * ai.agent({\n *   model,\n *   tools: [webTool, paidApiTool, expensiveDbTool],\n *   middleware: [scoped],\n * });\n */\nexport function forTool(\n  toolNames: string | ReadonlyArray<string>,\n  middleware: AgentMiddleware,\n): AgentMiddleware {\n  const allowed = new Set(\n    typeof toolNames === \"string\" ? [toolNames] : toolNames,\n  );\n  const scope =\n    allowed.size === 1 ? Array.from(allowed)[0] : Array.from(allowed).join(\"+\");\n\n  if (!middleware.tool) {\n    return middleware;\n  }\n\n  const innerBefore = middleware.tool.before;\n  const innerAfter = middleware.tool.after;\n  const innerOnError = middleware.tool.onError;\n\n  return {\n    ...middleware,\n    name: `${middleware.name}[for:${scope}]`,\n    tool: {\n      before: innerBefore\n        ? async ctx => {\n            if (!allowed.has(ctx.tool.name)) {\n              return;\n            }\n\n            return innerBefore(ctx);\n          }\n        : undefined,\n      after: innerAfter\n        ? async (ctx, result) => {\n            if (!allowed.has(ctx.tool.name)) {\n              return;\n            }\n\n            return innerAfter(ctx, result);\n          }\n        : undefined,\n      onError: innerOnError\n        ? async (ctx, error) => {\n            if (!allowed.has(ctx.tool.name)) {\n              return;\n            }\n\n            return innerOnError(ctx, error);\n          }\n        : undefined,\n    },\n  };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmDA,SAAgB,QACd,WACA,YACiB;CACjB,MAAM,UAAU,IAAI,IAClB,OAAO,cAAc,WAAW,CAAC,SAAS,IAAI,SAChD;CACA,MAAM,QACJ,QAAQ,SAAS,IAAI,MAAM,KAAK,OAAO,CAAC,CAAC,KAAK,MAAM,KAAK,OAAO,CAAC,CAAC,KAAK,GAAG;CAE5E,IAAI,CAAC,WAAW,MACd,OAAO;CAGT,MAAM,cAAc,WAAW,KAAK;CACpC,MAAM,aAAa,WAAW,KAAK;CACnC,MAAM,eAAe,WAAW,KAAK;CAErC,OAAO;EACL,GAAG;EACH,MAAM,GAAG,WAAW,KAAK,OAAO,MAAM;EACtC,MAAM;GACJ,QAAQ,cACJ,OAAM,QAAO;IACX,IAAI,CAAC,QAAQ,IAAI,IAAI,KAAK,IAAI,GAC5B;IAGF,OAAO,YAAY,GAAG;GACxB,IACA;GACJ,OAAO,aACH,OAAO,KAAK,WAAW;IACrB,IAAI,CAAC,QAAQ,IAAI,IAAI,KAAK,IAAI,GAC5B;IAGF,OAAO,WAAW,KAAK,MAAM;GAC/B,IACA;GACJ,SAAS,eACL,OAAO,KAAK,UAAU;IACpB,IAAI,CAAC,QAAQ,IAAI,IAAI,KAAK,IAAI,GAC5B;IAGF,OAAO,aAAa,KAAK,KAAK;GAChC,IACA;EACN;CACF;AACF"}

package/esm/middleware/pipeline.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"pipeline.d.mts","names":[],"sources":["../../../../../../../ai/src/middleware/pipeline.ts"],"mappings":";;;;;;;;;AAmBA;;;;KAAY,eAAA;AAOZ;;;;;AAAA,KAAY,wBAAA;EACV,OAAA,EAAS,wBAAA;EACT,IAAA,EAAM,qBAAA;EACN,IAAA,EAAM,qBAAA;EACN,UAAA,EAAY,2BAAA;AAAA;;;;;;;;AAA2B;AA+CzC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAMkB;iBANI,WAAA,eAA0B,eAAA,UAAA,CAC9C,WAAA,EAAa,aAAA,CAAc,eAAA,GAC3B,KAAA,EAAO,KAAA,EACP,OAAA,EAAS,wBAAA,CAAyB,KAAA,GAClC,KAAA,QAAa,OAAA,CAAQ,OAAA,GACrB,MAAA,GAAS,MAAA,GACR,OAAA,CAAQ,OAAA"}
+{"version":3,"file":"pipeline.d.mts","names":[],"sources":["../../../../../../../ai/src/middleware/pipeline.ts"],"mappings":";;;;;;;;;AAmBA;;;;KAAY,eAAA;AAOZ;;;;;AAAA,KAAY,wBAAA;EACV,OAAA,EAAS,wBAAA;EACT,IAAA,EAAM,qBAAA;EACN,IAAA,EAAM,qBAAA;EACN,UAAA,EAAY,2BAAA;AAAA;;;;;;;;AAA2B;AA+CzC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAMkB;iBANI,WAAA,eAA0B,eAAA,WAC9C,WAAA,EAAa,aAAA,CAAc,eAAA,GAC3B,KAAA,EAAO,KAAA,EACP,OAAA,EAAS,wBAAA,CAAyB,KAAA,GAClC,KAAA,QAAa,OAAA,CAAQ,OAAA,GACrB,MAAA,GAAS,MAAA,GACR,OAAA,CAAQ,OAAA"}

package/esm/middleware/utils/extract-user-text.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"extract-user-text.mjs","names":[],"sources":["../../../../../../../../ai/src/middleware/utils/extract-user-text.ts"],"sourcesContent":["import type { Message } from \"../../contracts/conversation-message.type\";\n\n/**\n * Pull the text a content-inspection middleware should care about\n * from the outbound message list.\n *\n * **Role.** Built-ins that inspect \"what the user just said\" — the\n * guardrail on `trip.before`, the semantic cache on `trip.before`,\n * future consumers like PII redactors — all need the same string:\n * the most recent `user`-role message's text content. This helper\n * is the single authority on how that string is resolved.\n *\n * **Behavior.**\n * - Walks `messages` from the end backwards so the LAST user turn\n *   wins (correct when the agent has history + a fresh prompt).\n * - Returns a plain string directly when `content` is a string.\n * - Joins `text` parts with `\"\\n\"` when `content` is a multipart\n *   `ContentPart[]`. Non-text parts (images, audio, pdf) are skipped —\n *   callers concerned with multimodal content inspect `request`\n *   / attachments separately.\n * - Returns `\"\"` when there is no user message at all (e.g. a trip\n *   composed entirely of tool results).\n *\n * **Coverage limit (D3).** Because only `text` parts are extracted, any\n * guardrail / PII detector built on this helper inspects **text only** —\n * image / PDF / audio attachment content is NOT scanned. A guardrail is\n * therefore not a multimodal safety control: for non-text inputs add an\n * attachment-level policy (e.g. an OCR / moderation pass before the call)\n * rather than relying on input detectors.\n *\n * @example\n * const prompt = extractUserText(context.messages);\n * if (!prompt) return;\n * const verdict = await inputCheck(prompt);\n */\nexport function extractUserText(messages: ReadonlyArray<Message>): string {\n  for (let index = messages.length - 1; index >= 0; index--) {\n    const message = messages[index];\n\n    if (message.role !== \"user\") {\n      continue;\n    }\n\n    if (typeof message.content === \"string\") {\n      return message.content;\n    }\n\n    if (Array.isArray(message.content)) {\n      return message.content\n        .filter((part) => part.type === \"text\")\n        .map((part) => (part as { text: string }).text)\n        .join(\"\\n\");\n    }\n  }\n\n  return \"\";\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmCA,SAAgB,gBAAgB,UAA0C;CACxE,KAAK,IAAI,QAAQ,SAAS,SAAS,GAAG,SAAS,GAAG,SAAS;EACzD,MAAM,UAAU,SAAS;EAEzB,IAAI,QAAQ,SAAS,QACnB;EAGF,IAAI,OAAO,QAAQ,YAAY,UAC7B,OAAO,QAAQ;EAGjB,IAAI,MAAM,QAAQ,QAAQ,OAAO,GAC/B,OAAO,QAAQ,QACZ,QAAQ,SAAS,KAAK,SAAS,MAAM,EACrC,KAAK,SAAU,KAA0B,IAAI,EAC7C,KAAK,IAAI;CAEhB;CAEA,OAAO;AACT"}
+{"version":3,"file":"extract-user-text.mjs","names":[],"sources":["../../../../../../../../ai/src/middleware/utils/extract-user-text.ts"],"sourcesContent":["import type { Message } from \"../../contracts/conversation-message.type\";\n\n/**\n * Pull the text a content-inspection middleware should care about\n * from the outbound message list.\n *\n * **Role.** Built-ins that inspect \"what the user just said\" — the\n * guardrail on `trip.before`, the semantic cache on `trip.before`,\n * future consumers like PII redactors — all need the same string:\n * the most recent `user`-role message's text content. This helper\n * is the single authority on how that string is resolved.\n *\n * **Behavior.**\n * - Walks `messages` from the end backwards so the LAST user turn\n *   wins (correct when the agent has history + a fresh prompt).\n * - Returns a plain string directly when `content` is a string.\n * - Joins `text` parts with `\"\\n\"` when `content` is a multipart\n *   `ContentPart[]`. Non-text parts (images, audio, pdf) are skipped —\n *   callers concerned with multimodal content inspect `request`\n *   / attachments separately.\n * - Returns `\"\"` when there is no user message at all (e.g. a trip\n *   composed entirely of tool results).\n *\n * **Coverage limit (D3).** Because only `text` parts are extracted, any\n * guardrail / PII detector built on this helper inspects **text only** —\n * image / PDF / audio attachment content is NOT scanned. A guardrail is\n * therefore not a multimodal safety control: for non-text inputs add an\n * attachment-level policy (e.g. an OCR / moderation pass before the call)\n * rather than relying on input detectors.\n *\n * @example\n * const prompt = extractUserText(context.messages);\n * if (!prompt) return;\n * const verdict = await inputCheck(prompt);\n */\nexport function extractUserText(messages: ReadonlyArray<Message>): string {\n  for (let index = messages.length - 1; index >= 0; index--) {\n    const message = messages[index];\n\n    if (message.role !== \"user\") {\n      continue;\n    }\n\n    if (typeof message.content === \"string\") {\n      return message.content;\n    }\n\n    if (Array.isArray(message.content)) {\n      return message.content\n        .filter((part) => part.type === \"text\")\n        .map((part) => (part as { text: string }).text)\n        .join(\"\\n\");\n    }\n  }\n\n  return \"\";\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmCA,SAAgB,gBAAgB,UAA0C;CACxE,KAAK,IAAI,QAAQ,SAAS,SAAS,GAAG,SAAS,GAAG,SAAS;EACzD,MAAM,UAAU,SAAS;EAEzB,IAAI,QAAQ,SAAS,QACnB;EAGF,IAAI,OAAO,QAAQ,YAAY,UAC7B,OAAO,QAAQ;EAGjB,IAAI,MAAM,QAAQ,QAAQ,OAAO,GAC/B,OAAO,QAAQ,QACZ,QAAQ,SAAS,KAAK,SAAS,MAAM,CAAC,CACtC,KAAK,SAAU,KAA0B,IAAI,CAAC,CAC9C,KAAK,IAAI;CAEhB;CAEA,OAAO;AACT"}

package/esm/middleware/utils/namespaced-state.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"namespaced-state.d.mts","names":[],"sources":["../../../../../../../../ai/src/middleware/utils/namespaced-state.ts"],"mappings":";;;;;AAyCA;;;;;;;;;;;;;;;;AAG4B;AAwB5B;;;;;;;;;;;;;AAIK;;;;;;iBA/BW,eAAA,GAAA,CACd,GAAA;EAAA,SAAgB,KAAA,EAAO,eAAA;AAAA,GACvB,SAAA,WACC,uBAAA,CAAwB,CAAA;;;;;;;;KAwBf,uBAAA;EACV,GAAA,IAAO,CAAA;EACP,GAAA,CAAI,KAAA,EAAO,CAAC;EACZ,MAAA;EACA,GAAA;AAAA"}
+{"version":3,"file":"namespaced-state.d.mts","names":[],"sources":["../../../../../../../../ai/src/middleware/utils/namespaced-state.ts"],"mappings":";;;;;AAyCA;;;;;;;;;;;;;;;;AAG4B;AAwB5B;;;;;;;;;;;;;AAIK;;;;;;iBA/BW,eAAA,IACd,GAAA;EAAA,SAAgB,KAAA,EAAO,eAAA;AAAA,GACvB,SAAA,WACC,uBAAA,CAAwB,CAAA;;;;;;;;KAwBf,uBAAA;EACV,GAAA,IAAO,CAAA;EACP,GAAA,CAAI,KAAA,EAAO,CAAC;EACZ,MAAA;EACA,GAAA;AAAA"}

package/esm/mock/mock-agent.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"mock-agent.d.mts","names":[],"sources":["../../../../../../../ai/src/mock/mock-agent.ts"],"mappings":";;;;;;;AA6BA;;;;;;;;;;;;;;;;;;;;iBAAgB,SAAA,mBAAA,CACd,OAAA;EACE,IAAA;EACA,SAAA,GAAY,iBAAA;EACZ,KAAA,GAAQ,YAAA;;;;;;EAMR,SAAA;AAAA,IAED,aAAA,CAAc,OAAA"}
+{"version":3,"file":"mock-agent.d.mts","names":[],"sources":["../../../../../../../ai/src/mock/mock-agent.ts"],"mappings":";;;;;;;AA6BA;;;;;;;;;;;;;;;;;;;;iBAAgB,SAAA,oBACd,OAAA;EACE,IAAA;EACA,SAAA,GAAY,iBAAA;EACZ,KAAA,GAAQ,YAAA;;;;;;EAMR,SAAA;AAAA,IAED,aAAA,CAAc,OAAA"}

package/esm/mock/mock-agent.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"mock-agent.mjs","names":[],"sources":["../../../../../../../ai/src/mock/mock-agent.ts"],"sourcesContent":["import { agent } from \"../agent/agent\";\nimport type { AgentConfig } from \"../agent/agent-config.type\";\nimport type { AgentContract } from \"../contracts/agent/agent.contract\";\nimport type { ToolContract } from \"../tool/tool\";\nimport type { MockModelResponse } from \"./mock-config.type\";\nimport { MockSDK } from \"./mock-sdk\";\n\n/**\n * Test helper that wires `MockSDK` → mock model → `agent()` in one call.\n *\n * Replaces the 4-line ritual specs typically write:\n * ```ts\n * const mock = MockSDK({ responses: [...] });\n * const myAgent = agent({ name: \"x\", model: mock.model({ name: \"m\" }) });\n * ```\n *\n * Defaults to a single empty `\"stop\"` response — enough to exercise an\n * agent that doesn't need a scripted reply (composition wiring,\n * anonymous-name tests, etc.). Pass `responses` to script outputs.\n *\n * `name` is optional — when omitted the resulting agent runs through\n * the normal anonymous-name fingerprint\n * (`anon_<provider>_<model>[_<tool1>+<tool2>...]`), which is the\n * default for `agent({ model })`.\n *\n * @example\n * const a = mockAgent({ name: \"writer\", responses: [{ content: \"hi\", finishReason: \"stop\" }] });\n * const result = await a.execute(\"anything\");\n */\nexport function mockAgent<TOutput = unknown>(\n  options: {\n    name?: string;\n    responses?: MockModelResponse[];\n    tools?: ToolContract<unknown, unknown>[];\n    /**\n     * Override the model name reported by the mock model. Defaults to\n     * `\"mock-model\"` (the MockSDK default). Useful when a test needs\n     * deterministic provider/model fingerprinting.\n     */\n    modelName?: string;\n  } = {},\n): AgentContract<TOutput> {\n  const responses = options.responses ?? [{ content: \"\", finishReason: \"stop\" as const }];\n  const sdk = MockSDK({ responses });\n  const model = sdk.model({ name: options.modelName ?? \"mock-model\" });\n\n  const config: AgentConfig<TOutput> = { model };\n\n  if (options.name !== undefined) {\n    config.name = options.name;\n  }\n\n  if (options.tools !== undefined) {\n    config.tools = options.tools;\n  }\n\n  return agent<TOutput>(config);\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;AA6BA,SAAgB,UACd,UAUI,CAAC,GACmB;CAKxB,MAAM,SAA+B,EAAE,OAH3B,QAAQ,EAAE,WADJ,QAAQ,aAAa,CAAC;EAAE,SAAS;EAAI,cAAc;CAAgB,CAAC,EACtD,CAChB,EAAE,MAAM,EAAE,MAAM,QAAQ,aAAa,aAAa,CAEvB,EAAE;CAE7C,IAAI,QAAQ,SAAS,QACnB,OAAO,OAAO,QAAQ;CAGxB,IAAI,QAAQ,UAAU,QACpB,OAAO,QAAQ,QAAQ;CAGzB,OAAO,MAAe,MAAM;AAC9B"}
+{"version":3,"file":"mock-agent.mjs","names":[],"sources":["../../../../../../../ai/src/mock/mock-agent.ts"],"sourcesContent":["import { agent } from \"../agent/agent\";\nimport type { AgentConfig } from \"../agent/agent-config.type\";\nimport type { AgentContract } from \"../contracts/agent/agent.contract\";\nimport type { ToolContract } from \"../tool/tool\";\nimport type { MockModelResponse } from \"./mock-config.type\";\nimport { MockSDK } from \"./mock-sdk\";\n\n/**\n * Test helper that wires `MockSDK` → mock model → `agent()` in one call.\n *\n * Replaces the 4-line ritual specs typically write:\n * ```ts\n * const mock = MockSDK({ responses: [...] });\n * const myAgent = agent({ name: \"x\", model: mock.model({ name: \"m\" }) });\n * ```\n *\n * Defaults to a single empty `\"stop\"` response — enough to exercise an\n * agent that doesn't need a scripted reply (composition wiring,\n * anonymous-name tests, etc.). Pass `responses` to script outputs.\n *\n * `name` is optional — when omitted the resulting agent runs through\n * the normal anonymous-name fingerprint\n * (`anon_<provider>_<model>[_<tool1>+<tool2>...]`), which is the\n * default for `agent({ model })`.\n *\n * @example\n * const a = mockAgent({ name: \"writer\", responses: [{ content: \"hi\", finishReason: \"stop\" }] });\n * const result = await a.execute(\"anything\");\n */\nexport function mockAgent<TOutput = unknown>(\n  options: {\n    name?: string;\n    responses?: MockModelResponse[];\n    tools?: ToolContract<unknown, unknown>[];\n    /**\n     * Override the model name reported by the mock model. Defaults to\n     * `\"mock-model\"` (the MockSDK default). Useful when a test needs\n     * deterministic provider/model fingerprinting.\n     */\n    modelName?: string;\n  } = {},\n): AgentContract<TOutput> {\n  const responses = options.responses ?? [{ content: \"\", finishReason: \"stop\" as const }];\n  const sdk = MockSDK({ responses });\n  const model = sdk.model({ name: options.modelName ?? \"mock-model\" });\n\n  const config: AgentConfig<TOutput> = { model };\n\n  if (options.name !== undefined) {\n    config.name = options.name;\n  }\n\n  if (options.tools !== undefined) {\n    config.tools = options.tools;\n  }\n\n  return agent<TOutput>(config);\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;AA6BA,SAAgB,UACd,UAUI,CAAC,GACmB;CAKxB,MAAM,SAA+B,EAAE,OAH3B,QAAQ,EAAE,WADJ,QAAQ,aAAa,CAAC;EAAE,SAAS;EAAI,cAAc;CAAgB,CAAC,EACtD,CAChB,CAAC,CAAC,MAAM,EAAE,MAAM,QAAQ,aAAa,aAAa,CAEvB,EAAE;CAE7C,IAAI,QAAQ,SAAS,QACnB,OAAO,OAAO,QAAQ;CAGxB,IAAI,QAAQ,UAAU,QACpB,OAAO,QAAQ,QAAQ;CAGzB,OAAO,MAAe,MAAM;AAC9B"}

package/esm/mock/mock-model.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"mock-model.d.mts","names":[],"sources":["../../../../../../../ai/src/mock/mock-model.ts"],"mappings":";;;;;KAUK,YAAA;EACH,QAAA,EAAU,OAAA;EACV,OAAA,GAAU,gBAAgB;AAAA;;;;;;;;AAAA;AA+B5B;;;;;;;;;;;;;;;;;;;;cAAa,SAAA,YAAqB,aAAA;EAAA,SAQd,IAAA;EAAA,iBACC,SAAA;EAAA,SARH,QAAA;EAAA,SACA,YAAA,GAAe,iBAAA;EAAA,QAEvB,aAAA;EAAA,QACA,KAAA;cAGU,IAAA,UACC,SAAA,EAAW,iBAAA,IAC5B,YAAA,GAAe,iBAAA;EADa;;;;;EAAA,IAWnB,WAAA,CAAA,GAAe,YAAA;EAQf;;;;EAAA,IAAA,SAAA,CAAA;EA8CW;;;;;;EAAA,QApCd,YAAA;EA6DN;;;;;EAAA,QAhDM,aAAA;EA8FI;AAAA;;;;;EAvEC,QAAA,CAAS,QAAA,EAAU,OAAA,IAAW,OAAA,GAAU,gBAAA,GAAmB,OAAA,CAAQ,aAAA;;;;;;;;;EAwBlE,MAAA,CACZ,QAAA,EAAU,OAAA,IACV,OAAA,GAAU,gBAAA,GACT,aAAA,CAAc,gBAAA;;;;;;EA4CV,KAAA,CAAA;AAAA"}
+{"version":3,"file":"mock-model.d.mts","names":[],"sources":["../../../../../../../ai/src/mock/mock-model.ts"],"mappings":";;;;;KAUK,YAAA;EACH,QAAA,EAAU,OAAA;EACV,OAAA,GAAU,gBAAgB;AAAA;;;;;;;;AAAA;AA+B5B;;;;;;;;;;;;;;;;;;;;cAAa,SAAA,YAAqB,aAAA;EAAA,SAQd,IAAA;EAAA,iBACC,SAAA;EAAA,SARH,QAAA;EAAA,SACA,YAAA,GAAe,iBAAA;EAAA,QAEvB,aAAA;EAAA,QACA,KAAA;cAGU,IAAA,UACC,SAAA,EAAW,iBAAA,IAC5B,YAAA,GAAe,iBAAA;EADa;;;;;EAAA,IAWnB,WAAA,IAAe,YAAA;EAQf;;;;EAAA,IAAA,SAAA;EA8CW;;;;;;EAAA,QApCd,YAAA;EA6DN;;;;;EAAA,QAhDM,aAAA;EA8FI;AAAA;;;;;EAvEC,QAAA,CAAS,QAAA,EAAU,OAAA,IAAW,OAAA,GAAU,gBAAA,GAAmB,OAAA,CAAQ,aAAA;;;;;;;;;EAwBlE,MAAA,CACZ,QAAA,EAAU,OAAA,IACV,OAAA,GAAU,gBAAA,GACT,aAAA,CAAc,gBAAA;;;;;;EA4CV,KAAA;AAAA"}

package/esm/mock/mock-model.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"mock-model.mjs","names":[],"sources":["../../../../../../../ai/src/mock/mock-model.ts"],"sourcesContent":["import type { Message } from \"../contracts/conversation-message.type\";\nimport type {\n  ModelCallOptions,\n  ModelCapabilities,\n  ModelContract,\n  ModelResponse,\n  ModelStreamChunk,\n} from \"../contracts/model.contract\";\nimport type { MockModelResponse } from \"./mock-config.type\";\n\ntype RecordedCall = {\n  messages: Message[];\n  options?: ModelCallOptions;\n};\n\n/**\n * Deterministic in-memory `ModelContract` implementation for tests.\n *\n * **Role.** Stands in for a real provider model so agent/workflow/supervisor\n * tests can assert behavior without hitting the network, spending tokens, or\n * depending on non-deterministic LLM output.\n *\n * **Responsibility.**\n * - Owns: a scripted queue of `MockModelResponse` entries, a call-history\n *   log for assertions, and the index pointer that advances through the\n *   queue on each `complete()` / `stream()` call.\n * - Does NOT own: any real inference, tokenization, or network I/O — when\n *   the queue is exhausted, the final entry is reused so tests never crash\n *   on accidental over-consumption.\n *\n * Every AI-related test in this repo uses `MockSDK` / `MockModel` — real\n * provider APIs are never hit from the test suite (see §6 of code-style.md).\n *\n * @example\n * const model = new MockModel(\"mock-gpt\", [\n *   { content: \"Hello!\", finishReason: \"stop\" },\n *   { content: \"Second turn.\", finishReason: \"stop\" },\n * ]);\n *\n * const first = await model.complete([{ role: \"user\", content: \"hi\" }]);\n * expect(first.content).toBe(\"Hello!\");\n * expect(model.callCount).toBe(1);\n */\nexport class MockModel implements ModelContract {\n  public readonly provider = \"mock\";\n  public readonly capabilities?: ModelCapabilities;\n\n  private responseIndex = 0;\n  private calls: RecordedCall[] = [];\n\n  public constructor(\n    public readonly name: string,\n    private readonly responses: MockModelResponse[],\n    capabilities?: ModelCapabilities,\n  ) {\n    this.capabilities = capabilities;\n  }\n\n  /**\n   * Full history of calls made to this model. Each entry is the exact\n   * `{ messages, options }` pair that was passed — useful for asserting\n   * that an agent built the right prompt or forwarded the right tool list.\n   */\n  public get callHistory(): RecordedCall[] {\n    return this.calls;\n  }\n\n  /**\n   * Number of times `complete()` or `stream()` has been invoked. Convenient\n   * shorthand for `callHistory.length` in assertions.\n   */\n  public get callCount(): number {\n    return this.calls.length;\n  }\n\n  /**\n   * Advance the scripted response queue by one and return the entry at the\n   * current pointer. If the queue is exhausted, the final scripted entry is\n   * returned repeatedly so over-consumption in tests produces predictable\n   * output instead of `undefined`.\n   */\n  private nextResponse(): MockModelResponse {\n    const response = this.responses[Math.min(this.responseIndex, this.responses.length - 1)];\n\n    this.responseIndex++;\n\n    return response ?? { content: \"Mock response\", finishReason: \"stop\" };\n  }\n\n  /**\n   * Convert a scripted `MockModelResponse` into a full `ModelResponse` with\n   * synthesized usage numbers when the script didn't supply them. Input\n   * usage is a fixed estimate; output usage is derived from content length.\n   */\n  private buildResponse(mock: MockModelResponse): ModelResponse {\n    const estimatedInput = 10;\n    const estimatedOutput = Math.ceil(mock.content.length / 4);\n\n    return {\n      content: mock.content,\n      finishReason: mock.finishReason ?? \"stop\",\n      usage: {\n        input: mock.usage?.input ?? estimatedInput,\n        output: mock.usage?.output ?? estimatedOutput,\n        total: (mock.usage?.input ?? estimatedInput) + (mock.usage?.output ?? estimatedOutput),\n        ...(mock.usage?.cachedTokens !== undefined ? { cachedTokens: mock.usage.cachedTokens } : {}),\n      },\n      toolCalls: mock.toolCalls,\n    };\n  }\n\n  /**\n   * Record the call, optionally delay (to simulate latency), and either\n   * throw the scripted error or return the scripted response. Mirrors the\n   * real provider's `complete()` contract so agents cannot tell the\n   * difference at runtime.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.calls.push({ messages, options });\n\n    const mock = this.nextResponse();\n\n    if (mock.delay) {\n      await new Promise((resolve) => setTimeout(resolve, mock.delay));\n    }\n\n    if (mock.error) {\n      throw mock.error;\n    }\n\n    return this.buildResponse(mock);\n  }\n\n  /**\n   * Record the call, optionally delay, then emit the scripted response as a\n   * sequence of stream chunks: the scripted `deltas` when the entry\n   * supplies them, otherwise content split word-by-word, as `delta`\n   * chunks, each scripted tool call as a `tool-call` chunk, and finally a\n   * `done` chunk with finish reason + usage. Throws eagerly if the scripted\n   * entry carries an `error`.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.calls.push({ messages, options });\n\n    const mock = this.nextResponse();\n\n    if (mock.delay) {\n      await new Promise((resolve) => setTimeout(resolve, mock.delay));\n    }\n\n    if (mock.error) {\n      throw mock.error;\n    }\n\n    const chunks = mock.deltas ?? mock.content.split(\" \").map((word) => word + \" \");\n\n    for (const chunk of chunks) {\n      yield { type: \"delta\", content: chunk };\n    }\n\n    if (mock.toolCalls) {\n      for (const toolCall of mock.toolCalls) {\n        yield {\n          type: \"tool-call\",\n          id: toolCall.id,\n          name: toolCall.name,\n          input: toolCall.input,\n        };\n      }\n    }\n\n    const response = this.buildResponse(mock);\n\n    yield {\n      type: \"done\",\n      finishReason: response.finishReason,\n      usage: response.usage,\n    };\n  }\n\n  /**\n   * Reset call history and response pointer back to their initial state.\n   * Intended for test-suite `beforeEach` hooks so a single `MockModel`\n   * instance can be reused across cases without cross-test leakage.\n   */\n  public reset(): void {\n    this.calls = [];\n    this.responseIndex = 0;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2CA,IAAa,YAAb,MAAgD;CAO9C,AAAO,YACL,AAAgB,MAChB,AAAiB,WACjB,cACA;EAHgB;EACC;kBARQ;uBAGH;eACQ,CAAC;EAO/B,KAAK,eAAe;CACtB;;;;;;CAOA,IAAW,cAA8B;EACvC,OAAO,KAAK;CACd;;;;;CAMA,IAAW,YAAoB;EAC7B,OAAO,KAAK,MAAM;CACpB;;;;;;;CAQA,AAAQ,eAAkC;EACxC,MAAM,WAAW,KAAK,UAAU,KAAK,IAAI,KAAK,eAAe,KAAK,UAAU,SAAS,CAAC;EAEtF,KAAK;EAEL,OAAO,YAAY;GAAE,SAAS;GAAiB,cAAc;EAAO;CACtE;;;;;;CAOA,AAAQ,cAAc,MAAwC;EAC5D,MAAM,iBAAiB;EACvB,MAAM,kBAAkB,KAAK,KAAK,KAAK,QAAQ,SAAS,CAAC;EAEzD,OAAO;GACL,SAAS,KAAK;GACd,cAAc,KAAK,gBAAgB;GACnC,OAAO;IACL,OAAO,KAAK,OAAO,SAAS;IAC5B,QAAQ,KAAK,OAAO,UAAU;IAC9B,QAAQ,KAAK,OAAO,SAAS,mBAAmB,KAAK,OAAO,UAAU;IACtE,GAAI,KAAK,OAAO,iBAAiB,SAAY,EAAE,cAAc,KAAK,MAAM,aAAa,IAAI,CAAC;GAC5F;GACA,WAAW,KAAK;EAClB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,MAAM,KAAK;GAAE;GAAU;EAAQ,CAAC;EAErC,MAAM,OAAO,KAAK,aAAa;EAE/B,IAAI,KAAK,OACP,MAAM,IAAI,SAAS,YAAY,WAAW,SAAS,KAAK,KAAK,CAAC;EAGhE,IAAI,KAAK,OACP,MAAM,KAAK;EAGb,OAAO,KAAK,cAAc,IAAI;CAChC;;;;;;;;;CAUA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,MAAM,KAAK;GAAE;GAAU;EAAQ,CAAC;EAErC,MAAM,OAAO,KAAK,aAAa;EAE/B,IAAI,KAAK,OACP,MAAM,IAAI,SAAS,YAAY,WAAW,SAAS,KAAK,KAAK,CAAC;EAGhE,IAAI,KAAK,OACP,MAAM,KAAK;EAGb,MAAM,SAAS,KAAK,UAAU,KAAK,QAAQ,MAAM,GAAG,EAAE,KAAK,SAAS,OAAO,GAAG;EAE9E,KAAK,MAAM,SAAS,QAClB,MAAM;GAAE,MAAM;GAAS,SAAS;EAAM;EAGxC,IAAI,KAAK,WACP,KAAK,MAAM,YAAY,KAAK,WAC1B,MAAM;GACJ,MAAM;GACN,IAAI,SAAS;GACb,MAAM,SAAS;GACf,OAAO,SAAS;EAClB;EAIJ,MAAM,WAAW,KAAK,cAAc,IAAI;EAExC,MAAM;GACJ,MAAM;GACN,cAAc,SAAS;GACvB,OAAO,SAAS;EAClB;CACF;;;;;;CAOA,AAAO,QAAc;EACnB,KAAK,QAAQ,CAAC;EACd,KAAK,gBAAgB;CACvB;AACF"}
+{"version":3,"file":"mock-model.mjs","names":[],"sources":["../../../../../../../ai/src/mock/mock-model.ts"],"sourcesContent":["import type { Message } from \"../contracts/conversation-message.type\";\nimport type {\n  ModelCallOptions,\n  ModelCapabilities,\n  ModelContract,\n  ModelResponse,\n  ModelStreamChunk,\n} from \"../contracts/model.contract\";\nimport type { MockModelResponse } from \"./mock-config.type\";\n\ntype RecordedCall = {\n  messages: Message[];\n  options?: ModelCallOptions;\n};\n\n/**\n * Deterministic in-memory `ModelContract` implementation for tests.\n *\n * **Role.** Stands in for a real provider model so agent/workflow/supervisor\n * tests can assert behavior without hitting the network, spending tokens, or\n * depending on non-deterministic LLM output.\n *\n * **Responsibility.**\n * - Owns: a scripted queue of `MockModelResponse` entries, a call-history\n *   log for assertions, and the index pointer that advances through the\n *   queue on each `complete()` / `stream()` call.\n * - Does NOT own: any real inference, tokenization, or network I/O — when\n *   the queue is exhausted, the final entry is reused so tests never crash\n *   on accidental over-consumption.\n *\n * Every AI-related test in this repo uses `MockSDK` / `MockModel` — real\n * provider APIs are never hit from the test suite (see §6 of code-style.md).\n *\n * @example\n * const model = new MockModel(\"mock-gpt\", [\n *   { content: \"Hello!\", finishReason: \"stop\" },\n *   { content: \"Second turn.\", finishReason: \"stop\" },\n * ]);\n *\n * const first = await model.complete([{ role: \"user\", content: \"hi\" }]);\n * expect(first.content).toBe(\"Hello!\");\n * expect(model.callCount).toBe(1);\n */\nexport class MockModel implements ModelContract {\n  public readonly provider = \"mock\";\n  public readonly capabilities?: ModelCapabilities;\n\n  private responseIndex = 0;\n  private calls: RecordedCall[] = [];\n\n  public constructor(\n    public readonly name: string,\n    private readonly responses: MockModelResponse[],\n    capabilities?: ModelCapabilities,\n  ) {\n    this.capabilities = capabilities;\n  }\n\n  /**\n   * Full history of calls made to this model. Each entry is the exact\n   * `{ messages, options }` pair that was passed — useful for asserting\n   * that an agent built the right prompt or forwarded the right tool list.\n   */\n  public get callHistory(): RecordedCall[] {\n    return this.calls;\n  }\n\n  /**\n   * Number of times `complete()` or `stream()` has been invoked. Convenient\n   * shorthand for `callHistory.length` in assertions.\n   */\n  public get callCount(): number {\n    return this.calls.length;\n  }\n\n  /**\n   * Advance the scripted response queue by one and return the entry at the\n   * current pointer. If the queue is exhausted, the final scripted entry is\n   * returned repeatedly so over-consumption in tests produces predictable\n   * output instead of `undefined`.\n   */\n  private nextResponse(): MockModelResponse {\n    const response = this.responses[Math.min(this.responseIndex, this.responses.length - 1)];\n\n    this.responseIndex++;\n\n    return response ?? { content: \"Mock response\", finishReason: \"stop\" };\n  }\n\n  /**\n   * Convert a scripted `MockModelResponse` into a full `ModelResponse` with\n   * synthesized usage numbers when the script didn't supply them. Input\n   * usage is a fixed estimate; output usage is derived from content length.\n   */\n  private buildResponse(mock: MockModelResponse): ModelResponse {\n    const estimatedInput = 10;\n    const estimatedOutput = Math.ceil(mock.content.length / 4);\n\n    return {\n      content: mock.content,\n      finishReason: mock.finishReason ?? \"stop\",\n      usage: {\n        input: mock.usage?.input ?? estimatedInput,\n        output: mock.usage?.output ?? estimatedOutput,\n        total: (mock.usage?.input ?? estimatedInput) + (mock.usage?.output ?? estimatedOutput),\n        ...(mock.usage?.cachedTokens !== undefined ? { cachedTokens: mock.usage.cachedTokens } : {}),\n      },\n      toolCalls: mock.toolCalls,\n    };\n  }\n\n  /**\n   * Record the call, optionally delay (to simulate latency), and either\n   * throw the scripted error or return the scripted response. Mirrors the\n   * real provider's `complete()` contract so agents cannot tell the\n   * difference at runtime.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.calls.push({ messages, options });\n\n    const mock = this.nextResponse();\n\n    if (mock.delay) {\n      await new Promise((resolve) => setTimeout(resolve, mock.delay));\n    }\n\n    if (mock.error) {\n      throw mock.error;\n    }\n\n    return this.buildResponse(mock);\n  }\n\n  /**\n   * Record the call, optionally delay, then emit the scripted response as a\n   * sequence of stream chunks: the scripted `deltas` when the entry\n   * supplies them, otherwise content split word-by-word, as `delta`\n   * chunks, each scripted tool call as a `tool-call` chunk, and finally a\n   * `done` chunk with finish reason + usage. Throws eagerly if the scripted\n   * entry carries an `error`.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.calls.push({ messages, options });\n\n    const mock = this.nextResponse();\n\n    if (mock.delay) {\n      await new Promise((resolve) => setTimeout(resolve, mock.delay));\n    }\n\n    if (mock.error) {\n      throw mock.error;\n    }\n\n    const chunks = mock.deltas ?? mock.content.split(\" \").map((word) => word + \" \");\n\n    for (const chunk of chunks) {\n      yield { type: \"delta\", content: chunk };\n    }\n\n    if (mock.toolCalls) {\n      for (const toolCall of mock.toolCalls) {\n        yield {\n          type: \"tool-call\",\n          id: toolCall.id,\n          name: toolCall.name,\n          input: toolCall.input,\n        };\n      }\n    }\n\n    const response = this.buildResponse(mock);\n\n    yield {\n      type: \"done\",\n      finishReason: response.finishReason,\n      usage: response.usage,\n    };\n  }\n\n  /**\n   * Reset call history and response pointer back to their initial state.\n   * Intended for test-suite `beforeEach` hooks so a single `MockModel`\n   * instance can be reused across cases without cross-test leakage.\n   */\n  public reset(): void {\n    this.calls = [];\n    this.responseIndex = 0;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2CA,IAAa,YAAb,MAAgD;CAO9C,AAAO,YACL,AAAgB,MAChB,AAAiB,WACjB,cACA;EAHgB;EACC;kBARQ;uBAGH;eACQ,CAAC;EAO/B,KAAK,eAAe;CACtB;;;;;;CAOA,IAAW,cAA8B;EACvC,OAAO,KAAK;CACd;;;;;CAMA,IAAW,YAAoB;EAC7B,OAAO,KAAK,MAAM;CACpB;;;;;;;CAQA,AAAQ,eAAkC;EACxC,MAAM,WAAW,KAAK,UAAU,KAAK,IAAI,KAAK,eAAe,KAAK,UAAU,SAAS,CAAC;EAEtF,KAAK;EAEL,OAAO,YAAY;GAAE,SAAS;GAAiB,cAAc;EAAO;CACtE;;;;;;CAOA,AAAQ,cAAc,MAAwC;EAC5D,MAAM,iBAAiB;EACvB,MAAM,kBAAkB,KAAK,KAAK,KAAK,QAAQ,SAAS,CAAC;EAEzD,OAAO;GACL,SAAS,KAAK;GACd,cAAc,KAAK,gBAAgB;GACnC,OAAO;IACL,OAAO,KAAK,OAAO,SAAS;IAC5B,QAAQ,KAAK,OAAO,UAAU;IAC9B,QAAQ,KAAK,OAAO,SAAS,mBAAmB,KAAK,OAAO,UAAU;IACtE,GAAI,KAAK,OAAO,iBAAiB,SAAY,EAAE,cAAc,KAAK,MAAM,aAAa,IAAI,CAAC;GAC5F;GACA,WAAW,KAAK;EAClB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,MAAM,KAAK;GAAE;GAAU;EAAQ,CAAC;EAErC,MAAM,OAAO,KAAK,aAAa;EAE/B,IAAI,KAAK,OACP,MAAM,IAAI,SAAS,YAAY,WAAW,SAAS,KAAK,KAAK,CAAC;EAGhE,IAAI,KAAK,OACP,MAAM,KAAK;EAGb,OAAO,KAAK,cAAc,IAAI;CAChC;;;;;;;;;CAUA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,MAAM,KAAK;GAAE;GAAU;EAAQ,CAAC;EAErC,MAAM,OAAO,KAAK,aAAa;EAE/B,IAAI,KAAK,OACP,MAAM,IAAI,SAAS,YAAY,WAAW,SAAS,KAAK,KAAK,CAAC;EAGhE,IAAI,KAAK,OACP,MAAM,KAAK;EAGb,MAAM,SAAS,KAAK,UAAU,KAAK,QAAQ,MAAM,GAAG,CAAC,CAAC,KAAK,SAAS,OAAO,GAAG;EAE9E,KAAK,MAAM,SAAS,QAClB,MAAM;GAAE,MAAM;GAAS,SAAS;EAAM;EAGxC,IAAI,KAAK,WACP,KAAK,MAAM,YAAY,KAAK,WAC1B,MAAM;GACJ,MAAM;GACN,IAAI,SAAS;GACb,MAAM,SAAS;GACf,OAAO,SAAS;EAClB;EAIJ,MAAM,WAAW,KAAK,cAAc,IAAI;EAExC,MAAM;GACJ,MAAM;GACN,cAAc,SAAS;GACvB,OAAO,SAAS;EAClB;CACF;;;;;;CAOA,AAAO,QAAc;EACnB,KAAK,QAAQ,CAAC;EACd,KAAK,gBAAgB;CACvB;AACF"}

package/esm/mock/mock-router.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"mock-router.d.mts","names":[],"sources":["../../../../../../../ai/src/mock/mock-router.ts"],"mappings":";;;;;;AAWA;;;;;KAAY,kBAAA,UAA4B,MAAA,qBACpC,IAAA,KACE,OAAA,EAAS,YAAA,CAAa,MAAA,MAAY,IAAA;;;;;;;;;;;;;KAc5B,oBAAA;AAAZ;;;AAAA,KAKY,iBAAA;EALoB,wEAO9B,WAAA,GAAc,oBAAoB;AAAA;;;AAAA;AA+BpC;;;;;;;;;;;;;;;;;;;;;;;;AAG0C;iBAH1B,UAAA,UAAoB,MAAA,kBAAA,CAClC,SAAA,EAAW,kBAAA,CAAmB,MAAA,KAC9B,OAAA,GAAS,iBAAA,IACP,OAAA,EAAS,YAAA,CAAa,MAAA,MAAY,IAAA"}
+{"version":3,"file":"mock-router.d.mts","names":[],"sources":["../../../../../../../ai/src/mock/mock-router.ts"],"mappings":";;;;;;AAWA;;;;;KAAY,kBAAA,UAA4B,MAAA,qBACpC,IAAA,KACE,OAAA,EAAS,YAAA,CAAa,MAAA,MAAY,IAAA;;;;;;;;;;;;;KAc5B,oBAAA;AAAZ;;;AAAA,KAKY,iBAAA;EALoB,wEAO9B,WAAA,GAAc,oBAAoB;AAAA;;;AAAA;AA+BpC;;;;;;;;;;;;;;;;;;;;;;;;AAG0C;iBAH1B,UAAA,UAAoB,MAAA,mBAClC,SAAA,EAAW,kBAAA,CAAmB,MAAA,KAC9B,OAAA,GAAS,iBAAA,IACP,OAAA,EAAS,YAAA,CAAa,MAAA,MAAY,IAAA"}