@aglyn/plugins-ai 1.0.0-beta.257 → 1.0.0-beta.258

package/src/lib/providers/cache-write-hour.js

@@ -0,0 +1,46 @@
+/**
+ * @license
+ * Copyright 2026 Aglyn LLC
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */ /**
+ * What an hour-long cache entry costs to write on top of what the rates
+ * price it at (AGL-3830). A rate table carries one write rate, the
+ * five-minute entry's 1.25 times the input price; the provider bills an
+ * hour-long entry at {@link AI_CACHE_WRITE_HOUR_FACTOR} times, so each token
+ * of `cacheWriteHourTokens` owes the difference. Kept out of the four-figure
+ * price on purpose: only a server reader can be handed an hour-long write,
+ * and the model selector that prices nominal usage never is. Its own module,
+ * not the catalog's, because the catalog loads on every console page and a
+ * helper beside it would load there too.
+ */ export const AI_CACHE_WRITE_HOUR_FACTOR = 2;
+export function aiCacheWriteHourSurchargeUsd(usage, rate) {
+    var _usage_cacheWriteHourTokens;
+    const tokens = Number((_usage_cacheWriteHourTokens = usage.cacheWriteHourTokens) != null ? _usage_cacheWriteHourTokens : 0);
+    if (!Number.isFinite(tokens) || tokens <= 0) return 0;
+    const usd = tokens * (rate.inputPerToken * AI_CACHE_WRITE_HOUR_FACTOR - rate.cacheWritePerToken);
+    return Math.round(usd * 1000000) / 1000000;
+}
+/**
+ * The hour-long writes of two exchanges added up, as a spread: absent where
+ * neither reported any, so a usage object that never had the field keeps
+ * the shape it had.
+ */ export function aiCacheWriteHourSum(a, b) {
+    var _a_cacheWriteHourTokens, _b_cacheWriteHourTokens;
+    if (a.cacheWriteHourTokens === undefined && b.cacheWriteHourTokens === undefined) return {};
+    return {
+        cacheWriteHourTokens: ((_a_cacheWriteHourTokens = a.cacheWriteHourTokens) != null ? _a_cacheWriteHourTokens : 0) + ((_b_cacheWriteHourTokens = b.cacheWriteHourTokens) != null ? _b_cacheWriteHourTokens : 0)
+    };
+}
+
+//# sourceMappingURL=cache-write-hour.js.map

package/src/lib/providers/cache-write-hour.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/cache-write-hour.ts"],"sourcesContent":["/**\n * @license\n * Copyright 2026 Aglyn LLC\n *\n * Licensed under the Apache License, Version 2.0 (the \"License\");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n *   http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an \"AS IS\" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n */\n\nimport type { AiTokenRates } from './catalog'\nimport type { AiUsage } from './contract'\n\n/**\n * What an hour-long cache entry costs to write on top of what the rates\n * price it at (AGL-3830). A rate table carries one write rate, the\n * five-minute entry's 1.25 times the input price; the provider bills an\n * hour-long entry at {@link AI_CACHE_WRITE_HOUR_FACTOR} times, so each token\n * of `cacheWriteHourTokens` owes the difference. Kept out of the four-figure\n * price on purpose: only a server reader can be handed an hour-long write,\n * and the model selector that prices nominal usage never is. Its own module,\n * not the catalog's, because the catalog loads on every console page and a\n * helper beside it would load there too.\n */\nexport const AI_CACHE_WRITE_HOUR_FACTOR = 2\n\nexport function aiCacheWriteHourSurchargeUsd(\n  usage: Pick<AiUsage, 'cacheWriteHourTokens'>,\n  rate: AiTokenRates,\n): number {\n  const tokens = Number(usage.cacheWriteHourTokens ?? 0)\n  if (!Number.isFinite(tokens) || tokens <= 0) return 0\n  const usd = tokens * (rate.inputPerToken * AI_CACHE_WRITE_HOUR_FACTOR - rate.cacheWritePerToken)\n  return Math.round(usd * 1_000_000) / 1_000_000\n}\n\n/**\n * The hour-long writes of two exchanges added up, as a spread: absent where\n * neither reported any, so a usage object that never had the field keeps\n * the shape it had.\n */\nexport function aiCacheWriteHourSum(\n  a: Pick<AiUsage, 'cacheWriteHourTokens'>,\n  b: Pick<AiUsage, 'cacheWriteHourTokens'>,\n): Pick<AiUsage, 'cacheWriteHourTokens'> {\n  if (a.cacheWriteHourTokens === undefined && b.cacheWriteHourTokens === undefined) return {}\n  return { cacheWriteHourTokens: (a.cacheWriteHourTokens ?? 0) + (b.cacheWriteHourTokens ?? 0) }\n}\n"],"names":["AI_CACHE_WRITE_HOUR_FACTOR","aiCacheWriteHourSurchargeUsd","usage","rate","tokens","Number","cacheWriteHourTokens","isFinite","usd","inputPerToken","cacheWritePerToken","Math","round","aiCacheWriteHourSum","a","b","undefined"],"mappings":"AAAA;;;;;;;;;;;;;;;CAeC,GAKD;;;;;;;;;;CAUC,GACD,OAAO,MAAMA,6BAA6B,EAAC;AAE3C,OAAO,SAASC,6BACdC,KAA4C,EAC5CC,IAAkB;QAEID;IAAtB,MAAME,SAASC,QAAOH,8BAAAA,MAAMI,oBAAoB,YAA1BJ,8BAA8B;IACpD,IAAI,CAACG,OAAOE,QAAQ,CAACH,WAAWA,UAAU,GAAG,OAAO;IACpD,MAAMI,MAAMJ,SAAUD,CAAAA,KAAKM,aAAa,GAAGT,6BAA6BG,KAAKO,kBAAkB,AAAD;IAC9F,OAAOC,KAAKC,KAAK,CAACJ,MAAM,WAAa;AACvC;AAEA;;;;CAIC,GACD,OAAO,SAASK,oBACdC,CAAwC,EACxCC,CAAwC;QAGRD,yBAAgCC;IADhE,IAAID,EAAER,oBAAoB,KAAKU,aAAaD,EAAET,oBAAoB,KAAKU,WAAW,OAAO,CAAC;IAC1F,OAAO;QAAEV,sBAAsB,EAACQ,0BAAAA,EAAER,oBAAoB,YAAtBQ,0BAA0B,OAAMC,0BAAAA,EAAET,oBAAoB,YAAtBS,0BAA0B;IAAG;AAC/F"}

package/src/lib/providers/contract.d.ts

@@ -215,9 +215,11 @@ export interface AiRequestBase {
     /**
      * How long a cache entry the breakpoints write should live, where the
      * provider offers a choice: the provider's default (five minutes on the
-     * Messages API) when absent. Only a development run asks for `1h`
-     * (`runtime/ai-dev-replay.ts`); it changes no answer, only what a second
-     * request within the hour pays for the prefix.
+     * Messages API) when absent. A step a site job runs asks for `1h` through
+     * its routing row (`AiRoutingRow.cacheTtl`, AGL-3830), in production too,
+     * and a development live run asks for it on every request
+     * (`runtime/ai-dev-env.ts`). It changes no answer, only what a request
+     * within the hour pays for the prefix, so a replay key does not read it.
      */
     cacheTtl?: '1h';
     /**
@@ -245,6 +247,17 @@ export interface AiUsage {
     outputTokens: number;
     cacheReadTokens: number;
     cacheWriteTokens: number;
+    /**
+     * How many of `cacheWriteTokens` went to an hour-long cache entry
+     * (AGL-3830), where the provider breaks the writes down; absent where none
+     * did or the provider does not say.
+     *
+     * Inside `cacheWriteTokens`, never beside it, like `thinkingTokens` below.
+     * An hour-long write is billed at twice the input price where a rate
+     * table's `cacheWritePerToken` is the five-minute 1.25 times, so a price
+     * reads this to add the difference (`aiCacheWriteHourSurchargeUsd`).
+     */
+    cacheWriteHourTokens?: number;
     /**
      * How many of `outputTokens` the model spent thinking, where the provider
      * reports it; absent where it does not (AGL-3143).

package/src/lib/providers/contract.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/contract.ts"],"sourcesContent":["/**\n * @license\n * Copyright 2026 Aglyn LLC\n *\n * Licensed under the Apache License, Version 2.0 (the \"License\");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n *   http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an \"AS IS\" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n */\n\nimport { UpstreamServiceError } from '@aglyn/shared-util-errors'\n\n/**\n * The provider contract (AGL-2939): the one shape every AI door speaks, and\n * the one thing an adapter has to implement. The runtime, the gate, the\n * jobs, the doctrine and the usage code import this and never a vendor\n * URL, header or model literal — Anthropic is one adapter behind it, and\n * an OpenAI-compatible endpoint is the second, which is what proves the\n * seam holds.\n *\n * ── Prompt caching, by construction ───────────────────────────────────────\n *\n * `system` is an ordered list of blocks, and caching is a PREFIX match: a\n * breakpoint covers every block before it, and one byte that differs per\n * request anywhere inside that prefix gives every distinct value its own\n * copy of it. The AGL-2352 rule is that no per-org byte may sit inside a\n * cached prefix — a white-label brand name interpolated into a cached block\n * multiplies the prefix count by the number of brands, each written off its\n * own traffic and expired before it is read. So a block marked `volatile`\n * may not sit at or before the last `cacheBreakpoint`, and the runtime\n * refuses the request before a byte leaves the process. A provider with no\n * cache honors the order and ignores the markers.\n *\n * ── Structured output ─────────────────────────────────────────────────────\n *\n * A `tools` entry is sent strict, with additional properties forbidden on\n * its schema, so a tool call's `input` validates against the schema\n * exactly. Tool choice stays automatic — forced tool use is not portable\n * across models — and the caller's system text names the tool it wants. An\n * adapter delivers `input` as an object; a stream assembles it from\n * fragments and parses the whole once, at the block's end. Nothing\n * string-matches the serialized form: models differ in how they escape it.\n * A provider bounds how much strict schema one request may carry, and an\n * adapter declares those bounds as `toolSchemaLimits`: three bounds a vendor\n * publishes and counts, and a fourth, the size of the grammar the schemas\n * compile to, that none publishes and only a live request settles.\n *\n * ── Thinking ──────────────────────────────────────────────────────────────\n *\n * `thinking: 'off'` and `'adaptive'` are sent when the provider has the\n * control; an omitted field sends NOTHING, leaving the model's own default\n * in force. The omission is deliberate rather than a gap: some models\n * reject an explicit setting, and the copy assistant serves its element\n * mode on such a model. `effort` rides the same rule.\n *\n * ── Pictures ──────────────────────────────────────────────────────────────\n *\n * A user turn may carry pictures beside its text (AGL-2916): `content` is\n * then an ordered list of parts, and a turn with nothing but text stays a\n * string, which every adapter passes through as it always has. A picture is\n * base64 bytes of one of `AI_IMAGE_MEDIA_TYPES`, never a URL, so no provider\n * fetches anything on the platform's behalf. The runtime refuses a picture\n * on an assistant turn, a picture for a model whose descriptor does not say\n * `vision`, and a picture past the size or count bounds, all before any\n * network is touched. Each adapter maps the parts onto its vendor's shape.\n */\n\n/**\n * One system block. `cacheBreakpoint` marks the end of a cacheable prefix;\n * `volatile` declares that the text carries a per-request or per-org byte\n * and therefore may never sit inside one.\n *\n * `site` declares a block that describes ONE site (AGL-3661): its business\n * profile, status and remembered preferences. As a breakpoint it is cached\n * once per site, so it must close the cached span — every platform-wide\n * block sits ahead of it, where every workspace still shares one entry —\n * and the runtime refuses a request that puts another breakpoint after it.\n */\nexport interface AiSystemBlock {\n  text: string\n  cacheBreakpoint?: true\n  volatile?: true\n  site?: true\n}\n\n/** A structured-output tool. `strict` is required, not optional. */\nexport interface AiTool {\n  name: string\n  description: string\n  /** A JSON schema of `type: 'object'`; `additionalProperties` is forced off. */\n  inputSchema: Record<string, unknown>\n  strict: true\n}\n\n/**\n * How much of a request's strict tools a provider can compile (AGL-3096). A\n * provider that constrains decoding to a tool's schema compiles every strict\n * schema of a request together, before the model runs, and refuses a request\n * whose schemas are past its bounds. That refusal is permanent: the schema is\n * the same on every retry. Each bound is a TOTAL over every strict tool one\n * request sends, as `aiToolSchemaCounts` counts them. A bound left out is one\n * the provider does not state.\n *\n * A provider may also refuse a request whose schemas compile to too large a\n * grammar, which none of the published counts can see: a request counting 1\n * strict tool, 0 optional parameters and 1 union was refused for exactly that\n * (AGL-3096). `compiledSchemaBytes` stands in for it, on live evidence.\n */\nexport interface AiToolSchemaLimits {\n  /** The most strict tools one request may send. */\n  strictTools?: number\n  /** The most properties, across every strict schema, left out of their object's `required`. */\n  optionalParameters?: number\n  /** The most schemas, across every strict schema, that are a union: an `anyOf` or `oneOf`, or a list of types. */\n  unionParameters?: number\n  /**\n   * The most bytes of schema a request's grammar may be compiled from, as\n   * `aiToolSchemaCompiledBytes` measures them.\n   *\n   * Unlike the bounds above, this one is EMPIRICAL. A provider refuses a\n   * request whose compiled grammar is too large without publishing the size,\n   * so the number here is not read off a vendor's page: it is the largest a\n   * live request has been watched to carry, and every number above it is\n   * unknown rather than safe. Raising it on an offline measurement alone\n   * puts the guard back where it could not see the break.\n   */\n  compiledSchemaBytes?: number\n}\n\n/** A request's strict tools, counted on the measures `AiToolSchemaLimits` bounds. */\nexport type AiToolSchemaCounts = Required<AiToolSchemaLimits>\n\n/**\n * One schema as a grammar compiler reads it: every local `$ref` written out\n * at each place it is used, and every description, title and `$defs` dropped,\n * since none of them constrains a single token. A schema that refers to\n * itself is written out once around and then left as the reference it is.\n */\nfunction compiledForm(node: unknown, root: Record<string, unknown>, resolving: ReadonlySet<string>): unknown {\n  if (Array.isArray(node)) return node.map((entry) => compiledForm(entry, root, resolving))\n  if (!isSchemaObject(node)) return node\n  const ref = node['$ref']\n  if (typeof ref === 'string') {\n    if (resolving.has(ref)) return { $ref: ref }\n    return compiledForm(localSchemaAt(root, ref), root, new Set([...resolving, ref]))\n  }\n  const written: Record<string, unknown> = {}\n  for (const [key, value] of Object.entries(node)) {\n    if (key === 'description' || key === 'title' || key === '$defs') continue\n    written[key] = compiledForm(value, root, resolving)\n  }\n  return written\n}\n\n/**\n * The bytes of schema a request's grammar is compiled from, over all its\n * strict tools. It is a stand-in for the size of the grammar itself, which\n * nothing outside the provider can measure, and it errs toward more: a `$ref`\n * counts at each place it is used, because a compiler that writes it out\n * there pays for it there.\n */\nexport function aiToolSchemaCompiledBytes(tools: readonly AiTool[]): number {\n  return tools.reduce(\n    (bytes, tool) => bytes + JSON.stringify(compiledForm(tool.inputSchema, tool.inputSchema, new Set())).length,\n    0,\n  )\n}\n\nfunction isSchemaObject(value: unknown): value is Record<string, unknown> {\n  return typeof value === 'object' && value !== null && !Array.isArray(value)\n}\n\n/** The schema a local `$ref` (`#/…`) points at, or `undefined`. */\nfunction localSchemaAt(root: Record<string, unknown>, ref: string): unknown {\n  if (!ref.startsWith('#/')) return undefined\n  return ref\n    .slice(2)\n    .split('/')\n    .map((key) => key.replace(/~1/g, '/').replace(/~0/g, '~'))\n    .reduce<unknown>((node, key) => (isSchemaObject(node) ? node[key] : undefined), root)\n}\n\nfunction countSchema(\n  node: unknown,\n  root: Record<string, unknown>,\n  counts: AiToolSchemaCounts,\n  resolving: ReadonlySet<string>,\n): void {\n  if (!isSchemaObject(node)) return\n  if (Array.isArray(node['anyOf']) || Array.isArray(node['oneOf']) || Array.isArray(node['type'])) {\n    counts.unionParameters += 1\n  }\n  const ref = node['$ref']\n  if (typeof ref === 'string' && !resolving.has(ref)) {\n    countSchema(localSchemaAt(root, ref), root, counts, new Set([...resolving, ref]))\n  }\n  const properties = node['properties']\n  if (isSchemaObject(properties)) {\n    const required = Array.isArray(node['required']) ? node['required'] : []\n    for (const [key, property] of Object.entries(properties)) {\n      if (!required.includes(key)) counts.optionalParameters += 1\n      countSchema(property, root, counts, resolving)\n    }\n  }\n  const items = node['items']\n  for (const item of Array.isArray(items) ? items : [items]) countSchema(item, root, counts, resolving)\n  for (const key of ['anyOf', 'allOf', 'oneOf']) {\n    const branches = node[key]\n    if (Array.isArray(branches)) for (const branch of branches) countSchema(branch, root, counts, resolving)\n  }\n}\n\n/**\n * The strict tools of one request, counted on each measure a provider bounds.\n * Every tool the contract sends is strict. Each schema is walked whole: every\n * property, every `items`, every branch of an `anyOf`, `allOf` or `oneOf`, and\n * a local `$ref` at each place it is used. A union nested in an array's items\n * or in another union's branch counts on its own, as it does for a provider\n * that compiles it; a `$ref` counted where it is used errs toward more.\n */\nexport function aiToolSchemaCounts(tools: readonly AiTool[]): AiToolSchemaCounts {\n  const counts: AiToolSchemaCounts = {\n    strictTools: tools.length,\n    optionalParameters: 0,\n    unionParameters: 0,\n    compiledSchemaBytes: aiToolSchemaCompiledBytes(tools),\n  }\n  for (const tool of tools) countSchema(tool.inputSchema, tool.inputSchema, counts, new Set())\n  return counts\n}\n\n/**\n * Each bound, as one, as many, and where its number comes from — a published\n * bound reads as the provider's own, a measured one as what has been proved.\n */\nconst TOOL_SCHEMA_MEASURES: ReadonlyArray<[keyof AiToolSchemaLimits, string, string, string]> = [\n  ['strictTools', 'strict tool', 'strict tools', 'the provider compiles'],\n  ['optionalParameters', 'optional parameter', 'optional parameters', 'the provider compiles'],\n  ['unionParameters', 'union-typed parameter', 'union-typed parameters', 'the provider compiles'],\n  ['compiledSchemaBytes', 'byte of compiled schema', 'bytes of compiled schema', 'a live request has proved'],\n]\n\n/** Each bound a request's tools are past, as a sentence; empty when every one holds. */\nexport function aiToolSchemaBreaches(\n  tools: readonly AiTool[],\n  limits: AiToolSchemaLimits | undefined,\n): string[] {\n  if (!limits) return []\n  const counts = aiToolSchemaCounts(tools)\n  return TOOL_SCHEMA_MEASURES.flatMap(([measure, one, many, source]) => {\n    const limit = limits[measure]\n    const count = counts[measure]\n    return limit !== undefined && count > limit\n      ? [`${count} ${count === 1 ? one : many} in one request, over the ${limit} ${source}`]\n      : []\n  })\n}\n\nexport type AiEffort = 'low' | 'medium' | 'high'\nexport type AiThinking = 'off' | 'adaptive'\n\n/**\n * The picture formats a request may carry (AGL-2916): the four every model\n * the catalog marks as reading images accepts. A door converts anything else\n * — an AVIF, an SVG — before it attaches the picture, or attaches none.\n */\nexport const AI_IMAGE_MEDIA_TYPES = ['image/jpeg', 'image/png', 'image/webp', 'image/gif'] as const\nexport type AiImageMediaType = (typeof AI_IMAGE_MEDIA_TYPES)[number]\n\n/**\n * The largest picture one part may carry, in decoded bytes: the ceiling the\n * providers behind the contract accept for one image. A door that attaches a\n * picture resizes it far below this; the runtime refuses anything above it\n * before a byte leaves the process, rather than letting the provider do it\n * after the request was paid for in time.\n */\nexport const AI_IMAGE_MAX_BYTES = 5 * 1024 * 1024\n\n/** The most pictures one request may carry. */\nexport const AI_REQUEST_MAX_IMAGES = 8\n\n/** Text in a message made of parts. */\nexport interface AiTextPart {\n  type: 'text'\n  text: string\n}\n\n/**\n * A picture in a message (AGL-2916): its bytes, base64-encoded, and their\n * format. Only a USER turn carries one, and only to a model whose descriptor\n * says it reads images; the runtime refuses any other request before a byte\n * leaves the process, so a door cannot send a picture by accident to a model\n * that would reject it, or to a turn the model wrote.\n */\nexport interface AiImagePart {\n  type: 'image'\n  mediaType: AiImageMediaType\n  /** The picture's bytes, base64-encoded, with no `data:` prefix. */\n  data: string\n}\n\nexport type AiMessagePart = AiTextPart | AiImagePart\n\nexport interface AiMessage {\n  role: 'user' | 'assistant'\n  /**\n   * The turn's text, or its parts in order — text and pictures — for a door\n   * that shows the model an image. A door with nothing but text sends a\n   * string, which every adapter passes through as it always has.\n   */\n  content: string | readonly AiMessagePart[]\n}\n\n/** A message's text: its parts' text joined, with every picture left out. */\nexport function aiMessageText(message: Pick<AiMessage, 'content'>): string {\n  if (typeof message.content === 'string') return message.content\n  return message.content\n    .filter((part): part is AiTextPart => part.type === 'text')\n    .map((part) => part.text)\n    .join('\\n\\n')\n}\n\n/** Every picture a list of messages carries, in order. */\nexport function aiMessageImages(messages: readonly Pick<AiMessage, 'content'>[]): AiImagePart[] {\n  return messages.flatMap((message) =>\n    typeof message.content === 'string'\n      ? []\n      : message.content.filter((part): part is AiImagePart => part.type === 'image'),\n  )\n}\n\n/** The decoded size of a base64 string, without decoding it. */\nexport function aiBase64Bytes(data: string): number {\n  const padding = data.endsWith('==') ? 2 : data.endsWith('=') ? 1 : 0\n  return Math.max(0, Math.floor((data.length * 3) / 4) - padding)\n}\n\n/** What every door hands the runtime; `provider` and `model` resolve there. */\nexport interface AiRequestBase {\n  model: string\n  system: AiSystemBlock[]\n  messages: AiMessage[]\n  tools?: AiTool[]\n  maxTokens: number\n  effort?: AiEffort\n  thinking?: AiThinking\n  signal?: AbortSignal\n  /**\n   * How long a cache entry the breakpoints write should live, where the\n   * provider offers a choice: the provider's default (five minutes on the\n   * Messages API) when absent. Only a development run asks for `1h`\n   * (`runtime/ai-dev-replay.ts`); it changes no answer, only what a second\n   * request within the hour pays for the prefix.\n   */\n  cacheTtl?: '1h'\n  /**\n   * Cache the conversation through its last message as well as the system\n   * blocks (AGL-3809), for a request the door expects to continue: a\n   * generation that looks the site up or is asked again re-sends its whole\n   * first turn, which without this is billed at the full input rate on every\n   * continuation. One more breakpoint, on the last message, counted against\n   * the same limit as the system's; a provider without a cache ignores it.\n   * It changes no answer, so a replay key does not read it.\n   */\n  cacheConversation?: boolean\n}\n\n/** The request an adapter receives: the runtime has already validated it. */\nexport interface AiProviderRequest extends AiRequestBase {\n  /** The provider's own key, resolved by the runtime from the environment or the org's setting. */\n  apiKey: string\n}\n\n/**\n * Token usage in the meter's shape. Cache reads and writes are separate\n * because they are priced separately; a provider with no cache reports 0.\n */\nexport interface AiUsage {\n  inputTokens: number\n  outputTokens: number\n  cacheReadTokens: number\n  cacheWriteTokens: number\n  /**\n   * How many of `outputTokens` the model spent thinking, where the provider\n   * reports it; absent where it does not (AGL-3143).\n   *\n   * Inside `outputTokens`, never beside it: a rate table prices the four\n   * figures above and would charge these twice if it read this one.\n   *\n   * It is here because it is the figure that says where a generation's\n   * output went. A tool call that comes back far smaller than the tokens it\n   * cost has three shapes, and only this separates the first from the other\n   * two: the model thought and then answered briefly, which is a model\n   * working as asked; a runaway string the partial parse discarded; or\n   * decoding that wrote nothing usable. Kept on every answer, because a\n   * generation burns its budget thinking without ever reaching its ceiling,\n   * and `rawOutput` is only there when it does.\n   */\n  thinkingTokens?: number\n}\n\nexport interface AiToolUse {\n  name: string\n  input: Record<string, unknown>\n}\n\n/** A completed, non-streaming answer. */\nexport interface AiCompletion {\n  kind: 'completion'\n  /** Every text block, joined. */\n  text: string\n  toolUse: AiToolUse[]\n  usage: AiUsage\n  /**\n   * The exchange at the serving model's BILLED rates — what a customer's\n   * credits are drawn from, and never a margin's cost side (AGL-3015). What\n   * the same exchange cost us is `estimateAiProviderCostUsd`, recorded by\n   * the meter; a result travels no further than the credit path, so it\n   * carries the one figure that path needs.\n   */\n  estCostUsd: number\n  stopReason: string | null\n  /**\n   * WHERE THE OUTPUT WENT (AGL-3143). The provider's own output for this\n   * call, as it answered it, kept ONLY when the call stopped at its ceiling\n   * and cut to `AI_RAW_OUTPUT_MAX_CHARS`.\n   *\n   * A tool call truncated mid-answer reaches `toolUse` as whatever the\n   * provider could still parse, and the parsed input alone cannot say where\n   * the tokens went: a runaway string the partial parse discarded and a\n   * decoding stall that wrote nothing usable both leave a small input\n   * against a spent ceiling. This is what tells them apart. These are the\n   * model's own words, so they belong to the trace alone: never to a reader,\n   * and never to a log, which carries the figures that separate the two\n   * shapes — `outputTokens`, `parsedChars`, `rawChars`.\n   */\n  rawOutput?: string\n}\n\n/**\n * Whether a model call stopped because it reached the output ceiling it was\n * asked for (AGL-3042). `max_tokens` is the contract's word, and the\n * OpenAI-compatible adapter maps its endpoint's `length` onto it; `length` is\n * read as well, for an adapter that passes its provider's own word through.\n */\nexport function aiStoppedAtCeiling(stopReason: string | null): boolean {\n  return stopReason === 'max_tokens' || stopReason === 'length'\n}\n\n/** The most characters of raw output a call cut off at its ceiling keeps. */\nexport const AI_RAW_OUTPUT_MAX_CHARS = 20_000\n\n/**\n * A cut-off call's output as the trace keeps it: what the provider answered,\n * serialized where it is not already text, and cut to the bound with an\n * ellipsis so a reader can tell a cut trace from a short answer.\n */\nexport function aiRawOutputOf(output: unknown): string {\n  const text = typeof output === 'string' ? output : (JSON.stringify(output) ?? '')\n  return text.length > AI_RAW_OUTPUT_MAX_CHARS ? `${text.slice(0, AI_RAW_OUTPUT_MAX_CHARS)}…` : text\n}\n\n/**\n * WHERE THE OUTPUT WENT, as a log may carry it (AGL-3143): the figures that\n * separate the shapes a call cut off at its ceiling comes in, and nothing\n * else.\n *\n * Thinking accounts for the spend first, because the thinking budget is\n * drawn from the same ceiling as the answer: a call can spend nearly all of\n * one thinking and still be cut off with a correct, tiny tool input, which\n * is a ceiling to raise rather than a decoder to fix. What thinking does not\n * account for is told apart by how much was written — long, and a runaway\n * string the partial parse discarded; short, and decoding wrote nothing\n * usable.\n *\n * It takes the raw output rather than a count on purpose. The bytes are the\n * model's own words and therefore the site's content, so they belong to the\n * trace alone and never to a log; passing them through one function that can\n * only ever return their LENGTH is what holds that rule at every door,\n * instead of asking each one to remember it.\n */\nexport function aiCutOffFigures(call: {\n  usage: AiUsage\n  /** The tool input that still parsed, or null where none did. */\n  parsed: unknown\n  rawOutput: string | null | undefined\n}): {\n  outputTokens: number\n  thinkingTokens: number | null\n  parsedChars: number\n  rawChars: number | null\n} {\n  const parsed =\n    call.parsed === null || call.parsed === undefined ? '' : (JSON.stringify(call.parsed) ?? '')\n  return {\n    outputTokens: call.usage.outputTokens,\n    thinkingTokens: call.usage.thinkingTokens ?? null,\n    parsedChars: parsed.length,\n    rawChars: call.rawOutput?.length ?? null,\n  }\n}\n\n/**\n * The model declined. Tokens were spent, so the usage is real and the\n * caller meters it; the text is whatever partial answer preceded the stop,\n * usually nothing. Every adapter maps its provider's own decline signal —\n * a refusal stop, a content filter — onto this one kind.\n */\nexport interface AiRefusal {\n  kind: 'refusal'\n  text: string\n  usage: AiUsage\n  estCostUsd: number\n  stopReason: 'refusal'\n}\n\nexport type AiResult = AiCompletion | AiRefusal\n\nexport type AiStreamEvent =\n  | { type: 'delta'; text: string }\n  | { type: 'tool'; name: string; input: Record<string, unknown> }\n  /**\n   * The provider's in-stream error. The stream is still open and ends with\n   * `done`, so the caller decides what the reader is told; the provider's\n   * own words went to the log, not here.\n   */\n  | { type: 'error'; retryable: boolean }\n  | {\n      type: 'done'\n      usage: AiUsage\n      estCostUsd: number\n      stopReason: string | null\n      /**\n       * WHERE THE OUTPUT WENT, for a stream (AGL-3143): the bytes the\n       * stream's tool calls were written in, kept ONLY when the stream\n       * stopped at its ceiling and cut to `AI_RAW_OUTPUT_MAX_CHARS`. The\n       * streaming counterpart of `AiCompletion.rawOutput`.\n       *\n       * A `tool` event carries what `aiToolInputOf` could still parse, which\n       * for a call cut mid-answer is a prefix of the object or nothing at\n       * all; the fragments those events were assembled from are the only\n       * record of what the model actually emitted, and without them a\n       * runaway string and a decoding stall look identical from the\n       * outside. These are the model's own words, so they belong to the\n       * trace alone: never to a reader, and never to a log, which carries\n       * the figures that separate the two shapes.\n       */\n      rawOutput?: string\n    }\n\n/** What a provider says about one of its models; the catalog carries both rates. */\nexport interface AiModelDescriptor {\n  id: string\n  /** The provider id the model is served by. */\n  provider: string\n  label: string\n  capabilities: {\n    streaming: boolean\n    tools: boolean\n    /** Whether the provider accepts an explicit `thinking` setting for this model. */\n    thinking: boolean\n    /** Whether the provider caches a system prefix for this model. */\n    promptCache: boolean\n    /**\n     * Whether the model reads a picture in a user turn (AGL-2916). Absent\n     * reads as no, so a provider registered before the field existed is never\n     * sent an image it did not say it can read.\n     */\n    vision?: boolean\n  }\n}\n\n/** The one sentence a customer may read when a provider fails. */\nexport const AI_UPSTREAM_FAILURE_COPY = 'The AI request failed — try again.'\n\n/**\n * A non-2xx answer from the provider. `message` is the fixed customer-safe\n * sentence; the provider's payload is in the log under `requestId`.\n */\nexport class AiUpstreamError extends UpstreamServiceError {\n  override readonly name = 'AiUpstreamError'\n  /**\n   * A failure about the PLATFORM's account with the provider rather than the\n   * request (AGL-3377): the key was refused, or the account is out of\n   * credit. Every AI feature on the install fails until the operator fixes\n   * it, so the runtime tells them. Each adapter classifies its own vendor's\n   * answers; absent is an ordinary failure.\n   */\n  readonly accountProblem: AiAccountProblem | null\n\n  constructor(\n    status: number | null,\n    retryable: boolean,\n    requestId: string | null,\n    accountProblem: AiAccountProblem | null = null,\n  ) {\n    super(AI_UPSTREAM_FAILURE_COPY, { status, retryable, requestId })\n    this.accountProblem = accountProblem\n  }\n}\n\n/** See {@link AiUpstreamError.accountProblem}. */\nexport type AiAccountProblem = 'credentials' | 'credit'\n\n/**\n * A request the runtime refuses to send. Thrown synchronously before any\n * network is touched, and it names the block at fault: this is a\n * programming error in the caller's prompt assembly, not a condition to\n * handle at runtime.\n */\nexport class AiRequestShapeError extends Error {\n  constructor(message: string) {\n    super(message)\n    this.name = 'AiRequestShapeError'\n  }\n}\n\n/**\n * A provider: one vendor's wire protocol behind the contract. `id` is the\n * routing key (`anthropic`, `openai-compatible`); `models` is what the\n * provider can serve; `complete` and `stream` are the two shapes every door\n * uses. An adapter maps its vendor's errors onto `AiUpstreamError` and its\n * declines onto `AiRefusal`, and reports usage in the meter's shape.\n */\nexport interface AiProvider {\n  readonly id: string\n  readonly label: string\n  /** The environment variable that carries this provider's key, for operator messages. */\n  readonly apiKeyEnv: string\n  /**\n   * The key from the server environment, or `undefined` when the provider\n   * is not configured to answer. Each adapter reads its own variable by\n   * name in its own module, so every key read is one the key-exposure\n   * guard can see (AGL-2240) and no door names a vendor's variable.\n   */\n  readApiKey(): string | undefined\n  /** The host the provider is reached at, for the subprocessor inventory. */\n  readonly endpointHost: string\n  /**\n   * How much of one request's strict tools this provider compiles, where it\n   * states a bound. The doors never read it: a spec holds every tool set a\n   * door sends to every registered provider's limits, so a schema a provider\n   * would refuse is red before it ships rather than a 400 on every request.\n   */\n  readonly toolSchemaLimits?: AiToolSchemaLimits\n  models(): readonly AiModelDescriptor[]\n  complete(request: AiProviderRequest): Promise<AiResult>\n  /**\n   * Resolves once the provider has ACCEPTED the request — after the\n   * headers, before the first token — with an iterator of events. A non-2xx\n   * answer rejects with `AiUpstreamError` before the iterator exists, so a\n   * door can still send a plain error status and hand back its reservation.\n   */\n  stream(request: AiProviderRequest): Promise<AsyncIterable<AiStreamEvent>>\n  /**\n   * Many requests answered together, asynchronously, at the provider's batch\n   * discount, where it has one: each entry is that request's result, or the\n   * error it failed with, in the order asked. Only a development run's live\n   * eval calls it (`runtime/ai-live-batch.ts`); no door does.\n   */\n  completeBatch?(requests: readonly AiProviderRequest[], options?: AiBatchOptions): Promise<Array<AiResult | Error>>\n}\n\n/** How a batch is waited on. */\nexport interface AiBatchOptions {\n  /** Between two looks at a batch still processing; 10 s when absent. */\n  pollMs?: number\n  /**\n   * The most a batch is waited for; an hour when absent. Past it the batch\n   * is canceled, which keeps every answer it already has, and each request\n   * it did not answer comes back as an `AiBatchUnfinishedError` (AGL-3781).\n   */\n  timeoutMs?: number\n  /** How long a canceled batch is given to end, so its answers can be read; 10 minutes when absent. */\n  cancelGraceMs?: number\n}\n\n/**\n * A request a batch did not answer (AGL-3781): canceled when its wait ran\n * out, expired, errored, or missing from the results. Not the request's own\n * failure — the caller answers it with a request of its own rather than\n * losing it, and a failure that request meets is the real one.\n */\nexport class AiBatchUnfinishedError extends Error {\n  override readonly name = 'AiBatchUnfinishedError'\n  constructor(\n    readonly batchId: string,\n    readonly reason: string,\n  ) {\n    super(`batch ${batchId} did not answer this request (${reason})`)\n  }\n}\n\n/** A count off the wire: finite, non-negative, whole; anything else is 0. */\nexport function aiTokenCount(value: unknown): number {\n  const parsed = Number(value ?? 0)\n  return Number.isFinite(parsed) && parsed > 0 ? Math.floor(parsed) : 0\n}\n\n/**\n * A tool call's `input` as an object. A provider sends an object or, for\n * the one shape a stream assembles by hand, a JSON string.\n *\n * A string that does not parse to an object is an EMPTY input, not a throw.\n * The case that produces one is a stream cut off mid-call by the output\n * ceiling, and a throw there would end the stream before its `done` event —\n * the event that carries the usage a door must meter for tokens already\n * spent. An empty input is one every door's own validation refuses.\n */\nexport function aiToolInputOf(value: unknown): Record<string, unknown> {\n  if (typeof value === 'string') {\n    if (!value.trim()) return {}\n    try {\n      const parsed: unknown = JSON.parse(value)\n      return parsed && typeof parsed === 'object' && !Array.isArray(parsed)\n        ? (parsed as Record<string, unknown>)\n        : {}\n    } catch {\n      return {}\n    }\n  }\n  return value && typeof value === 'object' ? (value as Record<string, unknown>) : {}\n}\n"],"names":["UpstreamServiceError","compiledForm","node","root","resolving","Array","isArray","map","entry","isSchemaObject","ref","has","$ref","localSchemaAt","Set","written","key","value","Object","entries","aiToolSchemaCompiledBytes","tools","reduce","bytes","tool","JSON","stringify","inputSchema","length","startsWith","undefined","slice","split","replace","countSchema","counts","unionParameters","properties","required","property","includes","optionalParameters","items","item","branches","branch","aiToolSchemaCounts","strictTools","compiledSchemaBytes","TOOL_SCHEMA_MEASURES","aiToolSchemaBreaches","limits","flatMap","measure","one","many","source","limit","count","AI_IMAGE_MEDIA_TYPES","AI_IMAGE_MAX_BYTES","AI_REQUEST_MAX_IMAGES","aiMessageText","message","content","filter","part","type","text","join","aiMessageImages","messages","aiBase64Bytes","data","padding","endsWith","Math","max","floor","aiStoppedAtCeiling","stopReason","AI_RAW_OUTPUT_MAX_CHARS","aiRawOutputOf","output","aiCutOffFigures","call","parsed","outputTokens","usage","thinkingTokens","parsedChars","rawChars","rawOutput","AI_UPSTREAM_FAILURE_COPY","AiUpstreamError","status","retryable","requestId","accountProblem","name","AiRequestShapeError","Error","AiBatchUnfinishedError","batchId","reason","aiTokenCount","Number","isFinite","aiToolInputOf","trim","parse"],"mappings":"AAAA;;;;;;;;;;;;;;;CAeC,GAED,SAASA,oBAAoB,QAAQ,4BAA2B;AA0HhE;;;;;CAKC,GACD,SAASC,aAAaC,IAAa,EAAEC,IAA6B,EAAEC,SAA8B;IAChG,IAAIC,MAAMC,OAAO,CAACJ,OAAO,OAAOA,KAAKK,GAAG,CAAC,CAACC,QAAUP,aAAaO,OAAOL,MAAMC;IAC9E,IAAI,CAACK,eAAeP,OAAO,OAAOA;IAClC,MAAMQ,MAAMR,IAAI,CAAC,OAAO;IACxB,IAAI,OAAOQ,QAAQ,UAAU;QAC3B,IAAIN,UAAUO,GAAG,CAACD,MAAM,OAAO;YAAEE,MAAMF;QAAI;QAC3C,OAAOT,aAAaY,cAAcV,MAAMO,MAAMP,MAAM,IAAIW,IAAI;eAAIV;YAAWM;SAAI;IACjF;IACA,MAAMK,UAAmC,CAAC;IAC1C,KAAK,MAAM,CAACC,KAAKC,MAAM,IAAIC,OAAOC,OAAO,CAACjB,MAAO;QAC/C,IAAIc,QAAQ,iBAAiBA,QAAQ,WAAWA,QAAQ,SAAS;QACjED,OAAO,CAACC,IAAI,GAAGf,aAAagB,OAAOd,MAAMC;IAC3C;IACA,OAAOW;AACT;AAEA;;;;;;CAMC,GACD,OAAO,SAASK,0BAA0BC,KAAwB;IAChE,OAAOA,MAAMC,MAAM,CACjB,CAACC,OAAOC,OAASD,QAAQE,KAAKC,SAAS,CAACzB,aAAauB,KAAKG,WAAW,EAAEH,KAAKG,WAAW,EAAE,IAAIb,QAAQc,MAAM,EAC3G;AAEJ;AAEA,SAASnB,eAAeQ,KAAc;IACpC,OAAO,OAAOA,UAAU,YAAYA,UAAU,QAAQ,CAACZ,MAAMC,OAAO,CAACW;AACvE;AAEA,iEAAiE,GACjE,SAASJ,cAAcV,IAA6B,EAAEO,GAAW;IAC/D,IAAI,CAACA,IAAImB,UAAU,CAAC,OAAO,OAAOC;IAClC,OAAOpB,IACJqB,KAAK,CAAC,GACNC,KAAK,CAAC,KACNzB,GAAG,CAAC,CAACS,MAAQA,IAAIiB,OAAO,CAAC,OAAO,KAAKA,OAAO,CAAC,OAAO,MACpDX,MAAM,CAAU,CAACpB,MAAMc,MAASP,eAAeP,QAAQA,IAAI,CAACc,IAAI,GAAGc,WAAY3B;AACpF;AAEA,SAAS+B,YACPhC,IAAa,EACbC,IAA6B,EAC7BgC,MAA0B,EAC1B/B,SAA8B;IAE9B,IAAI,CAACK,eAAeP,OAAO;IAC3B,IAAIG,MAAMC,OAAO,CAACJ,IAAI,CAAC,QAAQ,KAAKG,MAAMC,OAAO,CAACJ,IAAI,CAAC,QAAQ,KAAKG,MAAMC,OAAO,CAACJ,IAAI,CAAC,OAAO,GAAG;QAC/FiC,OAAOC,eAAe,IAAI;IAC5B;IACA,MAAM1B,MAAMR,IAAI,CAAC,OAAO;IACxB,IAAI,OAAOQ,QAAQ,YAAY,CAACN,UAAUO,GAAG,CAACD,MAAM;QAClDwB,YAAYrB,cAAcV,MAAMO,MAAMP,MAAMgC,QAAQ,IAAIrB,IAAI;eAAIV;YAAWM;SAAI;IACjF;IACA,MAAM2B,aAAanC,IAAI,CAAC,aAAa;IACrC,IAAIO,eAAe4B,aAAa;QAC9B,MAAMC,WAAWjC,MAAMC,OAAO,CAACJ,IAAI,CAAC,WAAW,IAAIA,IAAI,CAAC,WAAW,GAAG,EAAE;QACxE,KAAK,MAAM,CAACc,KAAKuB,SAAS,IAAIrB,OAAOC,OAAO,CAACkB,YAAa;YACxD,IAAI,CAACC,SAASE,QAAQ,CAACxB,MAAMmB,OAAOM,kBAAkB,IAAI;YAC1DP,YAAYK,UAAUpC,MAAMgC,QAAQ/B;QACtC;IACF;IACA,MAAMsC,QAAQxC,IAAI,CAAC,QAAQ;IAC3B,KAAK,MAAMyC,QAAQtC,MAAMC,OAAO,CAACoC,SAASA,QAAQ;QAACA;KAAM,CAAER,YAAYS,MAAMxC,MAAMgC,QAAQ/B;IAC3F,KAAK,MAAMY,OAAO;QAAC;QAAS;QAAS;KAAQ,CAAE;QAC7C,MAAM4B,WAAW1C,IAAI,CAACc,IAAI;QAC1B,IAAIX,MAAMC,OAAO,CAACsC,WAAW,KAAK,MAAMC,UAAUD,SAAUV,YAAYW,QAAQ1C,MAAMgC,QAAQ/B;IAChG;AACF;AAEA;;;;;;;CAOC,GACD,OAAO,SAAS0C,mBAAmBzB,KAAwB;IACzD,MAAMc,SAA6B;QACjCY,aAAa1B,MAAMO,MAAM;QACzBa,oBAAoB;QACpBL,iBAAiB;QACjBY,qBAAqB5B,0BAA0BC;IACjD;IACA,KAAK,MAAMG,QAAQH,MAAOa,YAAYV,KAAKG,WAAW,EAAEH,KAAKG,WAAW,EAAEQ,QAAQ,IAAIrB;IACtF,OAAOqB;AACT;AAEA;;;CAGC,GACD,MAAMc,uBAA0F;IAC9F;QAAC;QAAe;QAAe;QAAgB;KAAwB;IACvE;QAAC;QAAsB;QAAsB;QAAuB;KAAwB;IAC5F;QAAC;QAAmB;QAAyB;QAA0B;KAAwB;IAC/F;QAAC;QAAuB;QAA2B;QAA4B;KAA4B;CAC5G;AAED,sFAAsF,GACtF,OAAO,SAASC,qBACd7B,KAAwB,EACxB8B,MAAsC;IAEtC,IAAI,CAACA,QAAQ,OAAO,EAAE;IACtB,MAAMhB,SAASW,mBAAmBzB;IAClC,OAAO4B,qBAAqBG,OAAO,CAAC,CAAC,CAACC,SAASC,KAAKC,MAAMC,OAAO;QAC/D,MAAMC,QAAQN,MAAM,CAACE,QAAQ;QAC7B,MAAMK,QAAQvB,MAAM,CAACkB,QAAQ;QAC7B,OAAOI,UAAU3B,aAAa4B,QAAQD,QAClC;YAAC,GAAGC,MAAM,CAAC,EAAEA,UAAU,IAAIJ,MAAMC,KAAK,0BAA0B,EAAEE,MAAM,CAAC,EAAED,QAAQ;SAAC,GACpF,EAAE;IACR;AACF;AAKA;;;;CAIC,GACD,OAAO,MAAMG,uBAAuB;IAAC;IAAc;IAAa;IAAc;CAAY,CAAS;AAGnG;;;;;;CAMC,GACD,OAAO,MAAMC,qBAAqB,IAAI,OAAO,KAAI;AAEjD,6CAA6C,GAC7C,OAAO,MAAMC,wBAAwB,EAAC;AAkCtC,2EAA2E,GAC3E,OAAO,SAASC,cAAcC,OAAmC;IAC/D,IAAI,OAAOA,QAAQC,OAAO,KAAK,UAAU,OAAOD,QAAQC,OAAO;IAC/D,OAAOD,QAAQC,OAAO,CACnBC,MAAM,CAAC,CAACC,OAA6BA,KAAKC,IAAI,KAAK,QACnD5D,GAAG,CAAC,CAAC2D,OAASA,KAAKE,IAAI,EACvBC,IAAI,CAAC;AACV;AAEA,wDAAwD,GACxD,OAAO,SAASC,gBAAgBC,QAA+C;IAC7E,OAAOA,SAASnB,OAAO,CAAC,CAACW,UACvB,OAAOA,QAAQC,OAAO,KAAK,WACvB,EAAE,GACFD,QAAQC,OAAO,CAACC,MAAM,CAAC,CAACC,OAA8BA,KAAKC,IAAI,KAAK;AAE5E;AAEA,8DAA8D,GAC9D,OAAO,SAASK,cAAcC,IAAY;IACxC,MAAMC,UAAUD,KAAKE,QAAQ,CAAC,QAAQ,IAAIF,KAAKE,QAAQ,CAAC,OAAO,IAAI;IACnE,OAAOC,KAAKC,GAAG,CAAC,GAAGD,KAAKE,KAAK,CAAC,AAACL,KAAK7C,MAAM,GAAG,IAAK,KAAK8C;AACzD;AAwGA;;;;;CAKC,GACD,OAAO,SAASK,mBAAmBC,UAAyB;IAC1D,OAAOA,eAAe,gBAAgBA,eAAe;AACvD;AAEA,2EAA2E,GAC3E,OAAO,MAAMC,0BAA0B,MAAM;AAE7C;;;;CAIC,GACD,OAAO,SAASC,cAAcC,MAAe;QACS1D;IAApD,MAAM2C,OAAO,OAAOe,WAAW,WAAWA,UAAU1D,kBAAAA,KAAKC,SAAS,CAACyD,mBAAf1D,kBAA0B;IAC9E,OAAO2C,KAAKxC,MAAM,GAAGqD,0BAA0B,GAAGb,KAAKrC,KAAK,CAAC,GAAGkD,yBAAyB,CAAC,CAAC,GAAGb;AAChG;AAEA;;;;;;;;;;;;;;;;;;CAkBC,GACD,OAAO,SAASgB,gBAAgBC,IAK/B;QAO6D5D,iBAG1C4D;QAENA;IANZ,MAAMC,SACJD,KAAKC,MAAM,KAAK,QAAQD,KAAKC,MAAM,KAAKxD,YAAY,MAAML,kBAAAA,KAAKC,SAAS,CAAC2D,KAAKC,MAAM,aAA1B7D,kBAA+B;IAC3F,OAAO;QACL8D,cAAcF,KAAKG,KAAK,CAACD,YAAY;QACrCE,cAAc,GAAEJ,6BAAAA,KAAKG,KAAK,CAACC,cAAc,YAAzBJ,6BAA6B;QAC7CK,aAAaJ,OAAO1D,MAAM;QAC1B+D,QAAQ,WAAEN,kBAAAA,KAAKO,SAAS,qBAAdP,gBAAgBzD,MAAM,mBAAI;IACtC;AACF;AAwEA,gEAAgE,GAChE,OAAO,MAAMiE,2BAA2B,qCAAoC;AAE5E;;;CAGC,GACD,OAAO,MAAMC,wBAAwB9F;IAWnC,YACE+F,MAAqB,EACrBC,SAAkB,EAClBC,SAAwB,EACxBC,iBAA0C,IAAI,CAC9C;QACA,KAAK,CAACL,0BAA0B;YAAEE;YAAQC;YAAWC;QAAU,SAhB/CE,OAAO;QAiBvB,IAAI,CAACD,cAAc,GAAGA;IACxB;AACF;AAKA;;;;;CAKC,GACD,OAAO,MAAME,4BAA4BC;IACvC,YAAYtC,OAAe,CAAE;QAC3B,KAAK,CAACA;QACN,IAAI,CAACoC,IAAI,GAAG;IACd;AACF;AA8DA;;;;;CAKC,GACD,OAAO,MAAMG,+BAA+BD;IAE1C,YACE,AAASE,OAAe,EACxB,AAASC,MAAc,CACvB;QACA,KAAK,CAAC,CAAC,MAAM,EAAED,QAAQ,8BAA8B,EAAEC,OAAO,CAAC,CAAC,QAHvDD,UAAAA,cACAC,SAAAA,aAHOL,OAAO;IAMzB;AACF;AAEA,2EAA2E,GAC3E,OAAO,SAASM,aAAaxF,KAAc;IACzC,MAAMqE,SAASoB,OAAOzF,gBAAAA,QAAS;IAC/B,OAAOyF,OAAOC,QAAQ,CAACrB,WAAWA,SAAS,IAAIV,KAAKE,KAAK,CAACQ,UAAU;AACtE;AAEA;;;;;;;;;CASC,GACD,OAAO,SAASsB,cAAc3F,KAAc;IAC1C,IAAI,OAAOA,UAAU,UAAU;QAC7B,IAAI,CAACA,MAAM4F,IAAI,IAAI,OAAO,CAAC;QAC3B,IAAI;YACF,MAAMvB,SAAkB7D,KAAKqF,KAAK,CAAC7F;YACnC,OAAOqE,UAAU,OAAOA,WAAW,YAAY,CAACjF,MAAMC,OAAO,CAACgF,UACzDA,SACD,CAAC;QACP,EAAE,eAAM;YACN,OAAO,CAAC;QACV;IACF;IACA,OAAOrE,SAAS,OAAOA,UAAU,WAAYA,QAAoC,CAAC;AACpF"}
+{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/contract.ts"],"sourcesContent":["/**\n * @license\n * Copyright 2026 Aglyn LLC\n *\n * Licensed under the Apache License, Version 2.0 (the \"License\");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n *   http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an \"AS IS\" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n */\n\nimport { UpstreamServiceError } from '@aglyn/shared-util-errors'\n\n/**\n * The provider contract (AGL-2939): the one shape every AI door speaks, and\n * the one thing an adapter has to implement. The runtime, the gate, the\n * jobs, the doctrine and the usage code import this and never a vendor\n * URL, header or model literal — Anthropic is one adapter behind it, and\n * an OpenAI-compatible endpoint is the second, which is what proves the\n * seam holds.\n *\n * ── Prompt caching, by construction ───────────────────────────────────────\n *\n * `system` is an ordered list of blocks, and caching is a PREFIX match: a\n * breakpoint covers every block before it, and one byte that differs per\n * request anywhere inside that prefix gives every distinct value its own\n * copy of it. The AGL-2352 rule is that no per-org byte may sit inside a\n * cached prefix — a white-label brand name interpolated into a cached block\n * multiplies the prefix count by the number of brands, each written off its\n * own traffic and expired before it is read. So a block marked `volatile`\n * may not sit at or before the last `cacheBreakpoint`, and the runtime\n * refuses the request before a byte leaves the process. A provider with no\n * cache honors the order and ignores the markers.\n *\n * ── Structured output ─────────────────────────────────────────────────────\n *\n * A `tools` entry is sent strict, with additional properties forbidden on\n * its schema, so a tool call's `input` validates against the schema\n * exactly. Tool choice stays automatic — forced tool use is not portable\n * across models — and the caller's system text names the tool it wants. An\n * adapter delivers `input` as an object; a stream assembles it from\n * fragments and parses the whole once, at the block's end. Nothing\n * string-matches the serialized form: models differ in how they escape it.\n * A provider bounds how much strict schema one request may carry, and an\n * adapter declares those bounds as `toolSchemaLimits`: three bounds a vendor\n * publishes and counts, and a fourth, the size of the grammar the schemas\n * compile to, that none publishes and only a live request settles.\n *\n * ── Thinking ──────────────────────────────────────────────────────────────\n *\n * `thinking: 'off'` and `'adaptive'` are sent when the provider has the\n * control; an omitted field sends NOTHING, leaving the model's own default\n * in force. The omission is deliberate rather than a gap: some models\n * reject an explicit setting, and the copy assistant serves its element\n * mode on such a model. `effort` rides the same rule.\n *\n * ── Pictures ──────────────────────────────────────────────────────────────\n *\n * A user turn may carry pictures beside its text (AGL-2916): `content` is\n * then an ordered list of parts, and a turn with nothing but text stays a\n * string, which every adapter passes through as it always has. A picture is\n * base64 bytes of one of `AI_IMAGE_MEDIA_TYPES`, never a URL, so no provider\n * fetches anything on the platform's behalf. The runtime refuses a picture\n * on an assistant turn, a picture for a model whose descriptor does not say\n * `vision`, and a picture past the size or count bounds, all before any\n * network is touched. Each adapter maps the parts onto its vendor's shape.\n */\n\n/**\n * One system block. `cacheBreakpoint` marks the end of a cacheable prefix;\n * `volatile` declares that the text carries a per-request or per-org byte\n * and therefore may never sit inside one.\n *\n * `site` declares a block that describes ONE site (AGL-3661): its business\n * profile, status and remembered preferences. As a breakpoint it is cached\n * once per site, so it must close the cached span — every platform-wide\n * block sits ahead of it, where every workspace still shares one entry —\n * and the runtime refuses a request that puts another breakpoint after it.\n */\nexport interface AiSystemBlock {\n  text: string\n  cacheBreakpoint?: true\n  volatile?: true\n  site?: true\n}\n\n/** A structured-output tool. `strict` is required, not optional. */\nexport interface AiTool {\n  name: string\n  description: string\n  /** A JSON schema of `type: 'object'`; `additionalProperties` is forced off. */\n  inputSchema: Record<string, unknown>\n  strict: true\n}\n\n/**\n * How much of a request's strict tools a provider can compile (AGL-3096). A\n * provider that constrains decoding to a tool's schema compiles every strict\n * schema of a request together, before the model runs, and refuses a request\n * whose schemas are past its bounds. That refusal is permanent: the schema is\n * the same on every retry. Each bound is a TOTAL over every strict tool one\n * request sends, as `aiToolSchemaCounts` counts them. A bound left out is one\n * the provider does not state.\n *\n * A provider may also refuse a request whose schemas compile to too large a\n * grammar, which none of the published counts can see: a request counting 1\n * strict tool, 0 optional parameters and 1 union was refused for exactly that\n * (AGL-3096). `compiledSchemaBytes` stands in for it, on live evidence.\n */\nexport interface AiToolSchemaLimits {\n  /** The most strict tools one request may send. */\n  strictTools?: number\n  /** The most properties, across every strict schema, left out of their object's `required`. */\n  optionalParameters?: number\n  /** The most schemas, across every strict schema, that are a union: an `anyOf` or `oneOf`, or a list of types. */\n  unionParameters?: number\n  /**\n   * The most bytes of schema a request's grammar may be compiled from, as\n   * `aiToolSchemaCompiledBytes` measures them.\n   *\n   * Unlike the bounds above, this one is EMPIRICAL. A provider refuses a\n   * request whose compiled grammar is too large without publishing the size,\n   * so the number here is not read off a vendor's page: it is the largest a\n   * live request has been watched to carry, and every number above it is\n   * unknown rather than safe. Raising it on an offline measurement alone\n   * puts the guard back where it could not see the break.\n   */\n  compiledSchemaBytes?: number\n}\n\n/** A request's strict tools, counted on the measures `AiToolSchemaLimits` bounds. */\nexport type AiToolSchemaCounts = Required<AiToolSchemaLimits>\n\n/**\n * One schema as a grammar compiler reads it: every local `$ref` written out\n * at each place it is used, and every description, title and `$defs` dropped,\n * since none of them constrains a single token. A schema that refers to\n * itself is written out once around and then left as the reference it is.\n */\nfunction compiledForm(node: unknown, root: Record<string, unknown>, resolving: ReadonlySet<string>): unknown {\n  if (Array.isArray(node)) return node.map((entry) => compiledForm(entry, root, resolving))\n  if (!isSchemaObject(node)) return node\n  const ref = node['$ref']\n  if (typeof ref === 'string') {\n    if (resolving.has(ref)) return { $ref: ref }\n    return compiledForm(localSchemaAt(root, ref), root, new Set([...resolving, ref]))\n  }\n  const written: Record<string, unknown> = {}\n  for (const [key, value] of Object.entries(node)) {\n    if (key === 'description' || key === 'title' || key === '$defs') continue\n    written[key] = compiledForm(value, root, resolving)\n  }\n  return written\n}\n\n/**\n * The bytes of schema a request's grammar is compiled from, over all its\n * strict tools. It is a stand-in for the size of the grammar itself, which\n * nothing outside the provider can measure, and it errs toward more: a `$ref`\n * counts at each place it is used, because a compiler that writes it out\n * there pays for it there.\n */\nexport function aiToolSchemaCompiledBytes(tools: readonly AiTool[]): number {\n  return tools.reduce(\n    (bytes, tool) => bytes + JSON.stringify(compiledForm(tool.inputSchema, tool.inputSchema, new Set())).length,\n    0,\n  )\n}\n\nfunction isSchemaObject(value: unknown): value is Record<string, unknown> {\n  return typeof value === 'object' && value !== null && !Array.isArray(value)\n}\n\n/** The schema a local `$ref` (`#/…`) points at, or `undefined`. */\nfunction localSchemaAt(root: Record<string, unknown>, ref: string): unknown {\n  if (!ref.startsWith('#/')) return undefined\n  return ref\n    .slice(2)\n    .split('/')\n    .map((key) => key.replace(/~1/g, '/').replace(/~0/g, '~'))\n    .reduce<unknown>((node, key) => (isSchemaObject(node) ? node[key] : undefined), root)\n}\n\nfunction countSchema(\n  node: unknown,\n  root: Record<string, unknown>,\n  counts: AiToolSchemaCounts,\n  resolving: ReadonlySet<string>,\n): void {\n  if (!isSchemaObject(node)) return\n  if (Array.isArray(node['anyOf']) || Array.isArray(node['oneOf']) || Array.isArray(node['type'])) {\n    counts.unionParameters += 1\n  }\n  const ref = node['$ref']\n  if (typeof ref === 'string' && !resolving.has(ref)) {\n    countSchema(localSchemaAt(root, ref), root, counts, new Set([...resolving, ref]))\n  }\n  const properties = node['properties']\n  if (isSchemaObject(properties)) {\n    const required = Array.isArray(node['required']) ? node['required'] : []\n    for (const [key, property] of Object.entries(properties)) {\n      if (!required.includes(key)) counts.optionalParameters += 1\n      countSchema(property, root, counts, resolving)\n    }\n  }\n  const items = node['items']\n  for (const item of Array.isArray(items) ? items : [items]) countSchema(item, root, counts, resolving)\n  for (const key of ['anyOf', 'allOf', 'oneOf']) {\n    const branches = node[key]\n    if (Array.isArray(branches)) for (const branch of branches) countSchema(branch, root, counts, resolving)\n  }\n}\n\n/**\n * The strict tools of one request, counted on each measure a provider bounds.\n * Every tool the contract sends is strict. Each schema is walked whole: every\n * property, every `items`, every branch of an `anyOf`, `allOf` or `oneOf`, and\n * a local `$ref` at each place it is used. A union nested in an array's items\n * or in another union's branch counts on its own, as it does for a provider\n * that compiles it; a `$ref` counted where it is used errs toward more.\n */\nexport function aiToolSchemaCounts(tools: readonly AiTool[]): AiToolSchemaCounts {\n  const counts: AiToolSchemaCounts = {\n    strictTools: tools.length,\n    optionalParameters: 0,\n    unionParameters: 0,\n    compiledSchemaBytes: aiToolSchemaCompiledBytes(tools),\n  }\n  for (const tool of tools) countSchema(tool.inputSchema, tool.inputSchema, counts, new Set())\n  return counts\n}\n\n/**\n * Each bound, as one, as many, and where its number comes from — a published\n * bound reads as the provider's own, a measured one as what has been proved.\n */\nconst TOOL_SCHEMA_MEASURES: ReadonlyArray<[keyof AiToolSchemaLimits, string, string, string]> = [\n  ['strictTools', 'strict tool', 'strict tools', 'the provider compiles'],\n  ['optionalParameters', 'optional parameter', 'optional parameters', 'the provider compiles'],\n  ['unionParameters', 'union-typed parameter', 'union-typed parameters', 'the provider compiles'],\n  ['compiledSchemaBytes', 'byte of compiled schema', 'bytes of compiled schema', 'a live request has proved'],\n]\n\n/** Each bound a request's tools are past, as a sentence; empty when every one holds. */\nexport function aiToolSchemaBreaches(\n  tools: readonly AiTool[],\n  limits: AiToolSchemaLimits | undefined,\n): string[] {\n  if (!limits) return []\n  const counts = aiToolSchemaCounts(tools)\n  return TOOL_SCHEMA_MEASURES.flatMap(([measure, one, many, source]) => {\n    const limit = limits[measure]\n    const count = counts[measure]\n    return limit !== undefined && count > limit\n      ? [`${count} ${count === 1 ? one : many} in one request, over the ${limit} ${source}`]\n      : []\n  })\n}\n\nexport type AiEffort = 'low' | 'medium' | 'high'\nexport type AiThinking = 'off' | 'adaptive'\n\n/**\n * The picture formats a request may carry (AGL-2916): the four every model\n * the catalog marks as reading images accepts. A door converts anything else\n * — an AVIF, an SVG — before it attaches the picture, or attaches none.\n */\nexport const AI_IMAGE_MEDIA_TYPES = ['image/jpeg', 'image/png', 'image/webp', 'image/gif'] as const\nexport type AiImageMediaType = (typeof AI_IMAGE_MEDIA_TYPES)[number]\n\n/**\n * The largest picture one part may carry, in decoded bytes: the ceiling the\n * providers behind the contract accept for one image. A door that attaches a\n * picture resizes it far below this; the runtime refuses anything above it\n * before a byte leaves the process, rather than letting the provider do it\n * after the request was paid for in time.\n */\nexport const AI_IMAGE_MAX_BYTES = 5 * 1024 * 1024\n\n/** The most pictures one request may carry. */\nexport const AI_REQUEST_MAX_IMAGES = 8\n\n/** Text in a message made of parts. */\nexport interface AiTextPart {\n  type: 'text'\n  text: string\n}\n\n/**\n * A picture in a message (AGL-2916): its bytes, base64-encoded, and their\n * format. Only a USER turn carries one, and only to a model whose descriptor\n * says it reads images; the runtime refuses any other request before a byte\n * leaves the process, so a door cannot send a picture by accident to a model\n * that would reject it, or to a turn the model wrote.\n */\nexport interface AiImagePart {\n  type: 'image'\n  mediaType: AiImageMediaType\n  /** The picture's bytes, base64-encoded, with no `data:` prefix. */\n  data: string\n}\n\nexport type AiMessagePart = AiTextPart | AiImagePart\n\nexport interface AiMessage {\n  role: 'user' | 'assistant'\n  /**\n   * The turn's text, or its parts in order — text and pictures — for a door\n   * that shows the model an image. A door with nothing but text sends a\n   * string, which every adapter passes through as it always has.\n   */\n  content: string | readonly AiMessagePart[]\n}\n\n/** A message's text: its parts' text joined, with every picture left out. */\nexport function aiMessageText(message: Pick<AiMessage, 'content'>): string {\n  if (typeof message.content === 'string') return message.content\n  return message.content\n    .filter((part): part is AiTextPart => part.type === 'text')\n    .map((part) => part.text)\n    .join('\\n\\n')\n}\n\n/** Every picture a list of messages carries, in order. */\nexport function aiMessageImages(messages: readonly Pick<AiMessage, 'content'>[]): AiImagePart[] {\n  return messages.flatMap((message) =>\n    typeof message.content === 'string'\n      ? []\n      : message.content.filter((part): part is AiImagePart => part.type === 'image'),\n  )\n}\n\n/** The decoded size of a base64 string, without decoding it. */\nexport function aiBase64Bytes(data: string): number {\n  const padding = data.endsWith('==') ? 2 : data.endsWith('=') ? 1 : 0\n  return Math.max(0, Math.floor((data.length * 3) / 4) - padding)\n}\n\n/** What every door hands the runtime; `provider` and `model` resolve there. */\nexport interface AiRequestBase {\n  model: string\n  system: AiSystemBlock[]\n  messages: AiMessage[]\n  tools?: AiTool[]\n  maxTokens: number\n  effort?: AiEffort\n  thinking?: AiThinking\n  signal?: AbortSignal\n  /**\n   * How long a cache entry the breakpoints write should live, where the\n   * provider offers a choice: the provider's default (five minutes on the\n   * Messages API) when absent. A step a site job runs asks for `1h` through\n   * its routing row (`AiRoutingRow.cacheTtl`, AGL-3830), in production too,\n   * and a development live run asks for it on every request\n   * (`runtime/ai-dev-env.ts`). It changes no answer, only what a request\n   * within the hour pays for the prefix, so a replay key does not read it.\n   */\n  cacheTtl?: '1h'\n  /**\n   * Cache the conversation through its last message as well as the system\n   * blocks (AGL-3809), for a request the door expects to continue: a\n   * generation that looks the site up or is asked again re-sends its whole\n   * first turn, which without this is billed at the full input rate on every\n   * continuation. One more breakpoint, on the last message, counted against\n   * the same limit as the system's; a provider without a cache ignores it.\n   * It changes no answer, so a replay key does not read it.\n   */\n  cacheConversation?: boolean\n}\n\n/** The request an adapter receives: the runtime has already validated it. */\nexport interface AiProviderRequest extends AiRequestBase {\n  /** The provider's own key, resolved by the runtime from the environment or the org's setting. */\n  apiKey: string\n}\n\n/**\n * Token usage in the meter's shape. Cache reads and writes are separate\n * because they are priced separately; a provider with no cache reports 0.\n */\nexport interface AiUsage {\n  inputTokens: number\n  outputTokens: number\n  cacheReadTokens: number\n  cacheWriteTokens: number\n  /**\n   * How many of `cacheWriteTokens` went to an hour-long cache entry\n   * (AGL-3830), where the provider breaks the writes down; absent where none\n   * did or the provider does not say.\n   *\n   * Inside `cacheWriteTokens`, never beside it, like `thinkingTokens` below.\n   * An hour-long write is billed at twice the input price where a rate\n   * table's `cacheWritePerToken` is the five-minute 1.25 times, so a price\n   * reads this to add the difference (`aiCacheWriteHourSurchargeUsd`).\n   */\n  cacheWriteHourTokens?: number\n  /**\n   * How many of `outputTokens` the model spent thinking, where the provider\n   * reports it; absent where it does not (AGL-3143).\n   *\n   * Inside `outputTokens`, never beside it: a rate table prices the four\n   * figures above and would charge these twice if it read this one.\n   *\n   * It is here because it is the figure that says where a generation's\n   * output went. A tool call that comes back far smaller than the tokens it\n   * cost has three shapes, and only this separates the first from the other\n   * two: the model thought and then answered briefly, which is a model\n   * working as asked; a runaway string the partial parse discarded; or\n   * decoding that wrote nothing usable. Kept on every answer, because a\n   * generation burns its budget thinking without ever reaching its ceiling,\n   * and `rawOutput` is only there when it does.\n   */\n  thinkingTokens?: number\n}\n\nexport interface AiToolUse {\n  name: string\n  input: Record<string, unknown>\n}\n\n/** A completed, non-streaming answer. */\nexport interface AiCompletion {\n  kind: 'completion'\n  /** Every text block, joined. */\n  text: string\n  toolUse: AiToolUse[]\n  usage: AiUsage\n  /**\n   * The exchange at the serving model's BILLED rates — what a customer's\n   * credits are drawn from, and never a margin's cost side (AGL-3015). What\n   * the same exchange cost us is `estimateAiProviderCostUsd`, recorded by\n   * the meter; a result travels no further than the credit path, so it\n   * carries the one figure that path needs.\n   */\n  estCostUsd: number\n  stopReason: string | null\n  /**\n   * WHERE THE OUTPUT WENT (AGL-3143). The provider's own output for this\n   * call, as it answered it, kept ONLY when the call stopped at its ceiling\n   * and cut to `AI_RAW_OUTPUT_MAX_CHARS`.\n   *\n   * A tool call truncated mid-answer reaches `toolUse` as whatever the\n   * provider could still parse, and the parsed input alone cannot say where\n   * the tokens went: a runaway string the partial parse discarded and a\n   * decoding stall that wrote nothing usable both leave a small input\n   * against a spent ceiling. This is what tells them apart. These are the\n   * model's own words, so they belong to the trace alone: never to a reader,\n   * and never to a log, which carries the figures that separate the two\n   * shapes — `outputTokens`, `parsedChars`, `rawChars`.\n   */\n  rawOutput?: string\n}\n\n/**\n * Whether a model call stopped because it reached the output ceiling it was\n * asked for (AGL-3042). `max_tokens` is the contract's word, and the\n * OpenAI-compatible adapter maps its endpoint's `length` onto it; `length` is\n * read as well, for an adapter that passes its provider's own word through.\n */\nexport function aiStoppedAtCeiling(stopReason: string | null): boolean {\n  return stopReason === 'max_tokens' || stopReason === 'length'\n}\n\n/** The most characters of raw output a call cut off at its ceiling keeps. */\nexport const AI_RAW_OUTPUT_MAX_CHARS = 20_000\n\n/**\n * A cut-off call's output as the trace keeps it: what the provider answered,\n * serialized where it is not already text, and cut to the bound with an\n * ellipsis so a reader can tell a cut trace from a short answer.\n */\nexport function aiRawOutputOf(output: unknown): string {\n  const text = typeof output === 'string' ? output : (JSON.stringify(output) ?? '')\n  return text.length > AI_RAW_OUTPUT_MAX_CHARS ? `${text.slice(0, AI_RAW_OUTPUT_MAX_CHARS)}…` : text\n}\n\n/**\n * WHERE THE OUTPUT WENT, as a log may carry it (AGL-3143): the figures that\n * separate the shapes a call cut off at its ceiling comes in, and nothing\n * else.\n *\n * Thinking accounts for the spend first, because the thinking budget is\n * drawn from the same ceiling as the answer: a call can spend nearly all of\n * one thinking and still be cut off with a correct, tiny tool input, which\n * is a ceiling to raise rather than a decoder to fix. What thinking does not\n * account for is told apart by how much was written — long, and a runaway\n * string the partial parse discarded; short, and decoding wrote nothing\n * usable.\n *\n * It takes the raw output rather than a count on purpose. The bytes are the\n * model's own words and therefore the site's content, so they belong to the\n * trace alone and never to a log; passing them through one function that can\n * only ever return their LENGTH is what holds that rule at every door,\n * instead of asking each one to remember it.\n */\nexport function aiCutOffFigures(call: {\n  usage: AiUsage\n  /** The tool input that still parsed, or null where none did. */\n  parsed: unknown\n  rawOutput: string | null | undefined\n}): {\n  outputTokens: number\n  thinkingTokens: number | null\n  parsedChars: number\n  rawChars: number | null\n} {\n  const parsed =\n    call.parsed === null || call.parsed === undefined ? '' : (JSON.stringify(call.parsed) ?? '')\n  return {\n    outputTokens: call.usage.outputTokens,\n    thinkingTokens: call.usage.thinkingTokens ?? null,\n    parsedChars: parsed.length,\n    rawChars: call.rawOutput?.length ?? null,\n  }\n}\n\n/**\n * The model declined. Tokens were spent, so the usage is real and the\n * caller meters it; the text is whatever partial answer preceded the stop,\n * usually nothing. Every adapter maps its provider's own decline signal —\n * a refusal stop, a content filter — onto this one kind.\n */\nexport interface AiRefusal {\n  kind: 'refusal'\n  text: string\n  usage: AiUsage\n  estCostUsd: number\n  stopReason: 'refusal'\n}\n\nexport type AiResult = AiCompletion | AiRefusal\n\nexport type AiStreamEvent =\n  | { type: 'delta'; text: string }\n  | { type: 'tool'; name: string; input: Record<string, unknown> }\n  /**\n   * The provider's in-stream error. The stream is still open and ends with\n   * `done`, so the caller decides what the reader is told; the provider's\n   * own words went to the log, not here.\n   */\n  | { type: 'error'; retryable: boolean }\n  | {\n      type: 'done'\n      usage: AiUsage\n      estCostUsd: number\n      stopReason: string | null\n      /**\n       * WHERE THE OUTPUT WENT, for a stream (AGL-3143): the bytes the\n       * stream's tool calls were written in, kept ONLY when the stream\n       * stopped at its ceiling and cut to `AI_RAW_OUTPUT_MAX_CHARS`. The\n       * streaming counterpart of `AiCompletion.rawOutput`.\n       *\n       * A `tool` event carries what `aiToolInputOf` could still parse, which\n       * for a call cut mid-answer is a prefix of the object or nothing at\n       * all; the fragments those events were assembled from are the only\n       * record of what the model actually emitted, and without them a\n       * runaway string and a decoding stall look identical from the\n       * outside. These are the model's own words, so they belong to the\n       * trace alone: never to a reader, and never to a log, which carries\n       * the figures that separate the two shapes.\n       */\n      rawOutput?: string\n    }\n\n/** What a provider says about one of its models; the catalog carries both rates. */\nexport interface AiModelDescriptor {\n  id: string\n  /** The provider id the model is served by. */\n  provider: string\n  label: string\n  capabilities: {\n    streaming: boolean\n    tools: boolean\n    /** Whether the provider accepts an explicit `thinking` setting for this model. */\n    thinking: boolean\n    /** Whether the provider caches a system prefix for this model. */\n    promptCache: boolean\n    /**\n     * Whether the model reads a picture in a user turn (AGL-2916). Absent\n     * reads as no, so a provider registered before the field existed is never\n     * sent an image it did not say it can read.\n     */\n    vision?: boolean\n  }\n}\n\n/** The one sentence a customer may read when a provider fails. */\nexport const AI_UPSTREAM_FAILURE_COPY = 'The AI request failed — try again.'\n\n/**\n * A non-2xx answer from the provider. `message` is the fixed customer-safe\n * sentence; the provider's payload is in the log under `requestId`.\n */\nexport class AiUpstreamError extends UpstreamServiceError {\n  override readonly name = 'AiUpstreamError'\n  /**\n   * A failure about the PLATFORM's account with the provider rather than the\n   * request (AGL-3377): the key was refused, or the account is out of\n   * credit. Every AI feature on the install fails until the operator fixes\n   * it, so the runtime tells them. Each adapter classifies its own vendor's\n   * answers; absent is an ordinary failure.\n   */\n  readonly accountProblem: AiAccountProblem | null\n\n  constructor(\n    status: number | null,\n    retryable: boolean,\n    requestId: string | null,\n    accountProblem: AiAccountProblem | null = null,\n  ) {\n    super(AI_UPSTREAM_FAILURE_COPY, { status, retryable, requestId })\n    this.accountProblem = accountProblem\n  }\n}\n\n/** See {@link AiUpstreamError.accountProblem}. */\nexport type AiAccountProblem = 'credentials' | 'credit'\n\n/**\n * A request the runtime refuses to send. Thrown synchronously before any\n * network is touched, and it names the block at fault: this is a\n * programming error in the caller's prompt assembly, not a condition to\n * handle at runtime.\n */\nexport class AiRequestShapeError extends Error {\n  constructor(message: string) {\n    super(message)\n    this.name = 'AiRequestShapeError'\n  }\n}\n\n/**\n * A provider: one vendor's wire protocol behind the contract. `id` is the\n * routing key (`anthropic`, `openai-compatible`); `models` is what the\n * provider can serve; `complete` and `stream` are the two shapes every door\n * uses. An adapter maps its vendor's errors onto `AiUpstreamError` and its\n * declines onto `AiRefusal`, and reports usage in the meter's shape.\n */\nexport interface AiProvider {\n  readonly id: string\n  readonly label: string\n  /** The environment variable that carries this provider's key, for operator messages. */\n  readonly apiKeyEnv: string\n  /**\n   * The key from the server environment, or `undefined` when the provider\n   * is not configured to answer. Each adapter reads its own variable by\n   * name in its own module, so every key read is one the key-exposure\n   * guard can see (AGL-2240) and no door names a vendor's variable.\n   */\n  readApiKey(): string | undefined\n  /** The host the provider is reached at, for the subprocessor inventory. */\n  readonly endpointHost: string\n  /**\n   * How much of one request's strict tools this provider compiles, where it\n   * states a bound. The doors never read it: a spec holds every tool set a\n   * door sends to every registered provider's limits, so a schema a provider\n   * would refuse is red before it ships rather than a 400 on every request.\n   */\n  readonly toolSchemaLimits?: AiToolSchemaLimits\n  models(): readonly AiModelDescriptor[]\n  complete(request: AiProviderRequest): Promise<AiResult>\n  /**\n   * Resolves once the provider has ACCEPTED the request — after the\n   * headers, before the first token — with an iterator of events. A non-2xx\n   * answer rejects with `AiUpstreamError` before the iterator exists, so a\n   * door can still send a plain error status and hand back its reservation.\n   */\n  stream(request: AiProviderRequest): Promise<AsyncIterable<AiStreamEvent>>\n  /**\n   * Many requests answered together, asynchronously, at the provider's batch\n   * discount, where it has one: each entry is that request's result, or the\n   * error it failed with, in the order asked. Only a development run's live\n   * eval calls it (`runtime/ai-live-batch.ts`); no door does.\n   */\n  completeBatch?(requests: readonly AiProviderRequest[], options?: AiBatchOptions): Promise<Array<AiResult | Error>>\n}\n\n/** How a batch is waited on. */\nexport interface AiBatchOptions {\n  /** Between two looks at a batch still processing; 10 s when absent. */\n  pollMs?: number\n  /**\n   * The most a batch is waited for; an hour when absent. Past it the batch\n   * is canceled, which keeps every answer it already has, and each request\n   * it did not answer comes back as an `AiBatchUnfinishedError` (AGL-3781).\n   */\n  timeoutMs?: number\n  /** How long a canceled batch is given to end, so its answers can be read; 10 minutes when absent. */\n  cancelGraceMs?: number\n}\n\n/**\n * A request a batch did not answer (AGL-3781): canceled when its wait ran\n * out, expired, errored, or missing from the results. Not the request's own\n * failure — the caller answers it with a request of its own rather than\n * losing it, and a failure that request meets is the real one.\n */\nexport class AiBatchUnfinishedError extends Error {\n  override readonly name = 'AiBatchUnfinishedError'\n  constructor(\n    readonly batchId: string,\n    readonly reason: string,\n  ) {\n    super(`batch ${batchId} did not answer this request (${reason})`)\n  }\n}\n\n/** A count off the wire: finite, non-negative, whole; anything else is 0. */\nexport function aiTokenCount(value: unknown): number {\n  const parsed = Number(value ?? 0)\n  return Number.isFinite(parsed) && parsed > 0 ? Math.floor(parsed) : 0\n}\n\n/**\n * A tool call's `input` as an object. A provider sends an object or, for\n * the one shape a stream assembles by hand, a JSON string.\n *\n * A string that does not parse to an object is an EMPTY input, not a throw.\n * The case that produces one is a stream cut off mid-call by the output\n * ceiling, and a throw there would end the stream before its `done` event —\n * the event that carries the usage a door must meter for tokens already\n * spent. An empty input is one every door's own validation refuses.\n */\nexport function aiToolInputOf(value: unknown): Record<string, unknown> {\n  if (typeof value === 'string') {\n    if (!value.trim()) return {}\n    try {\n      const parsed: unknown = JSON.parse(value)\n      return parsed && typeof parsed === 'object' && !Array.isArray(parsed)\n        ? (parsed as Record<string, unknown>)\n        : {}\n    } catch {\n      return {}\n    }\n  }\n  return value && typeof value === 'object' ? (value as Record<string, unknown>) : {}\n}\n"],"names":["UpstreamServiceError","compiledForm","node","root","resolving","Array","isArray","map","entry","isSchemaObject","ref","has","$ref","localSchemaAt","Set","written","key","value","Object","entries","aiToolSchemaCompiledBytes","tools","reduce","bytes","tool","JSON","stringify","inputSchema","length","startsWith","undefined","slice","split","replace","countSchema","counts","unionParameters","properties","required","property","includes","optionalParameters","items","item","branches","branch","aiToolSchemaCounts","strictTools","compiledSchemaBytes","TOOL_SCHEMA_MEASURES","aiToolSchemaBreaches","limits","flatMap","measure","one","many","source","limit","count","AI_IMAGE_MEDIA_TYPES","AI_IMAGE_MAX_BYTES","AI_REQUEST_MAX_IMAGES","aiMessageText","message","content","filter","part","type","text","join","aiMessageImages","messages","aiBase64Bytes","data","padding","endsWith","Math","max","floor","aiStoppedAtCeiling","stopReason","AI_RAW_OUTPUT_MAX_CHARS","aiRawOutputOf","output","aiCutOffFigures","call","parsed","outputTokens","usage","thinkingTokens","parsedChars","rawChars","rawOutput","AI_UPSTREAM_FAILURE_COPY","AiUpstreamError","status","retryable","requestId","accountProblem","name","AiRequestShapeError","Error","AiBatchUnfinishedError","batchId","reason","aiTokenCount","Number","isFinite","aiToolInputOf","trim","parse"],"mappings":"AAAA;;;;;;;;;;;;;;;CAeC,GAED,SAASA,oBAAoB,QAAQ,4BAA2B;AA0HhE;;;;;CAKC,GACD,SAASC,aAAaC,IAAa,EAAEC,IAA6B,EAAEC,SAA8B;IAChG,IAAIC,MAAMC,OAAO,CAACJ,OAAO,OAAOA,KAAKK,GAAG,CAAC,CAACC,QAAUP,aAAaO,OAAOL,MAAMC;IAC9E,IAAI,CAACK,eAAeP,OAAO,OAAOA;IAClC,MAAMQ,MAAMR,IAAI,CAAC,OAAO;IACxB,IAAI,OAAOQ,QAAQ,UAAU;QAC3B,IAAIN,UAAUO,GAAG,CAACD,MAAM,OAAO;YAAEE,MAAMF;QAAI;QAC3C,OAAOT,aAAaY,cAAcV,MAAMO,MAAMP,MAAM,IAAIW,IAAI;eAAIV;YAAWM;SAAI;IACjF;IACA,MAAMK,UAAmC,CAAC;IAC1C,KAAK,MAAM,CAACC,KAAKC,MAAM,IAAIC,OAAOC,OAAO,CAACjB,MAAO;QAC/C,IAAIc,QAAQ,iBAAiBA,QAAQ,WAAWA,QAAQ,SAAS;QACjED,OAAO,CAACC,IAAI,GAAGf,aAAagB,OAAOd,MAAMC;IAC3C;IACA,OAAOW;AACT;AAEA;;;;;;CAMC,GACD,OAAO,SAASK,0BAA0BC,KAAwB;IAChE,OAAOA,MAAMC,MAAM,CACjB,CAACC,OAAOC,OAASD,QAAQE,KAAKC,SAAS,CAACzB,aAAauB,KAAKG,WAAW,EAAEH,KAAKG,WAAW,EAAE,IAAIb,QAAQc,MAAM,EAC3G;AAEJ;AAEA,SAASnB,eAAeQ,KAAc;IACpC,OAAO,OAAOA,UAAU,YAAYA,UAAU,QAAQ,CAACZ,MAAMC,OAAO,CAACW;AACvE;AAEA,iEAAiE,GACjE,SAASJ,cAAcV,IAA6B,EAAEO,GAAW;IAC/D,IAAI,CAACA,IAAImB,UAAU,CAAC,OAAO,OAAOC;IAClC,OAAOpB,IACJqB,KAAK,CAAC,GACNC,KAAK,CAAC,KACNzB,GAAG,CAAC,CAACS,MAAQA,IAAIiB,OAAO,CAAC,OAAO,KAAKA,OAAO,CAAC,OAAO,MACpDX,MAAM,CAAU,CAACpB,MAAMc,MAASP,eAAeP,QAAQA,IAAI,CAACc,IAAI,GAAGc,WAAY3B;AACpF;AAEA,SAAS+B,YACPhC,IAAa,EACbC,IAA6B,EAC7BgC,MAA0B,EAC1B/B,SAA8B;IAE9B,IAAI,CAACK,eAAeP,OAAO;IAC3B,IAAIG,MAAMC,OAAO,CAACJ,IAAI,CAAC,QAAQ,KAAKG,MAAMC,OAAO,CAACJ,IAAI,CAAC,QAAQ,KAAKG,MAAMC,OAAO,CAACJ,IAAI,CAAC,OAAO,GAAG;QAC/FiC,OAAOC,eAAe,IAAI;IAC5B;IACA,MAAM1B,MAAMR,IAAI,CAAC,OAAO;IACxB,IAAI,OAAOQ,QAAQ,YAAY,CAACN,UAAUO,GAAG,CAACD,MAAM;QAClDwB,YAAYrB,cAAcV,MAAMO,MAAMP,MAAMgC,QAAQ,IAAIrB,IAAI;eAAIV;YAAWM;SAAI;IACjF;IACA,MAAM2B,aAAanC,IAAI,CAAC,aAAa;IACrC,IAAIO,eAAe4B,aAAa;QAC9B,MAAMC,WAAWjC,MAAMC,OAAO,CAACJ,IAAI,CAAC,WAAW,IAAIA,IAAI,CAAC,WAAW,GAAG,EAAE;QACxE,KAAK,MAAM,CAACc,KAAKuB,SAAS,IAAIrB,OAAOC,OAAO,CAACkB,YAAa;YACxD,IAAI,CAACC,SAASE,QAAQ,CAACxB,MAAMmB,OAAOM,kBAAkB,IAAI;YAC1DP,YAAYK,UAAUpC,MAAMgC,QAAQ/B;QACtC;IACF;IACA,MAAMsC,QAAQxC,IAAI,CAAC,QAAQ;IAC3B,KAAK,MAAMyC,QAAQtC,MAAMC,OAAO,CAACoC,SAASA,QAAQ;QAACA;KAAM,CAAER,YAAYS,MAAMxC,MAAMgC,QAAQ/B;IAC3F,KAAK,MAAMY,OAAO;QAAC;QAAS;QAAS;KAAQ,CAAE;QAC7C,MAAM4B,WAAW1C,IAAI,CAACc,IAAI;QAC1B,IAAIX,MAAMC,OAAO,CAACsC,WAAW,KAAK,MAAMC,UAAUD,SAAUV,YAAYW,QAAQ1C,MAAMgC,QAAQ/B;IAChG;AACF;AAEA;;;;;;;CAOC,GACD,OAAO,SAAS0C,mBAAmBzB,KAAwB;IACzD,MAAMc,SAA6B;QACjCY,aAAa1B,MAAMO,MAAM;QACzBa,oBAAoB;QACpBL,iBAAiB;QACjBY,qBAAqB5B,0BAA0BC;IACjD;IACA,KAAK,MAAMG,QAAQH,MAAOa,YAAYV,KAAKG,WAAW,EAAEH,KAAKG,WAAW,EAAEQ,QAAQ,IAAIrB;IACtF,OAAOqB;AACT;AAEA;;;CAGC,GACD,MAAMc,uBAA0F;IAC9F;QAAC;QAAe;QAAe;QAAgB;KAAwB;IACvE;QAAC;QAAsB;QAAsB;QAAuB;KAAwB;IAC5F;QAAC;QAAmB;QAAyB;QAA0B;KAAwB;IAC/F;QAAC;QAAuB;QAA2B;QAA4B;KAA4B;CAC5G;AAED,sFAAsF,GACtF,OAAO,SAASC,qBACd7B,KAAwB,EACxB8B,MAAsC;IAEtC,IAAI,CAACA,QAAQ,OAAO,EAAE;IACtB,MAAMhB,SAASW,mBAAmBzB;IAClC,OAAO4B,qBAAqBG,OAAO,CAAC,CAAC,CAACC,SAASC,KAAKC,MAAMC,OAAO;QAC/D,MAAMC,QAAQN,MAAM,CAACE,QAAQ;QAC7B,MAAMK,QAAQvB,MAAM,CAACkB,QAAQ;QAC7B,OAAOI,UAAU3B,aAAa4B,QAAQD,QAClC;YAAC,GAAGC,MAAM,CAAC,EAAEA,UAAU,IAAIJ,MAAMC,KAAK,0BAA0B,EAAEE,MAAM,CAAC,EAAED,QAAQ;SAAC,GACpF,EAAE;IACR;AACF;AAKA;;;;CAIC,GACD,OAAO,MAAMG,uBAAuB;IAAC;IAAc;IAAa;IAAc;CAAY,CAAS;AAGnG;;;;;;CAMC,GACD,OAAO,MAAMC,qBAAqB,IAAI,OAAO,KAAI;AAEjD,6CAA6C,GAC7C,OAAO,MAAMC,wBAAwB,EAAC;AAkCtC,2EAA2E,GAC3E,OAAO,SAASC,cAAcC,OAAmC;IAC/D,IAAI,OAAOA,QAAQC,OAAO,KAAK,UAAU,OAAOD,QAAQC,OAAO;IAC/D,OAAOD,QAAQC,OAAO,CACnBC,MAAM,CAAC,CAACC,OAA6BA,KAAKC,IAAI,KAAK,QACnD5D,GAAG,CAAC,CAAC2D,OAASA,KAAKE,IAAI,EACvBC,IAAI,CAAC;AACV;AAEA,wDAAwD,GACxD,OAAO,SAASC,gBAAgBC,QAA+C;IAC7E,OAAOA,SAASnB,OAAO,CAAC,CAACW,UACvB,OAAOA,QAAQC,OAAO,KAAK,WACvB,EAAE,GACFD,QAAQC,OAAO,CAACC,MAAM,CAAC,CAACC,OAA8BA,KAAKC,IAAI,KAAK;AAE5E;AAEA,8DAA8D,GAC9D,OAAO,SAASK,cAAcC,IAAY;IACxC,MAAMC,UAAUD,KAAKE,QAAQ,CAAC,QAAQ,IAAIF,KAAKE,QAAQ,CAAC,OAAO,IAAI;IACnE,OAAOC,KAAKC,GAAG,CAAC,GAAGD,KAAKE,KAAK,CAAC,AAACL,KAAK7C,MAAM,GAAG,IAAK,KAAK8C;AACzD;AAqHA;;;;;CAKC,GACD,OAAO,SAASK,mBAAmBC,UAAyB;IAC1D,OAAOA,eAAe,gBAAgBA,eAAe;AACvD;AAEA,2EAA2E,GAC3E,OAAO,MAAMC,0BAA0B,MAAM;AAE7C;;;;CAIC,GACD,OAAO,SAASC,cAAcC,MAAe;QACS1D;IAApD,MAAM2C,OAAO,OAAOe,WAAW,WAAWA,UAAU1D,kBAAAA,KAAKC,SAAS,CAACyD,mBAAf1D,kBAA0B;IAC9E,OAAO2C,KAAKxC,MAAM,GAAGqD,0BAA0B,GAAGb,KAAKrC,KAAK,CAAC,GAAGkD,yBAAyB,CAAC,CAAC,GAAGb;AAChG;AAEA;;;;;;;;;;;;;;;;;;CAkBC,GACD,OAAO,SAASgB,gBAAgBC,IAK/B;QAO6D5D,iBAG1C4D;QAENA;IANZ,MAAMC,SACJD,KAAKC,MAAM,KAAK,QAAQD,KAAKC,MAAM,KAAKxD,YAAY,MAAML,kBAAAA,KAAKC,SAAS,CAAC2D,KAAKC,MAAM,aAA1B7D,kBAA+B;IAC3F,OAAO;QACL8D,cAAcF,KAAKG,KAAK,CAACD,YAAY;QACrCE,cAAc,GAAEJ,6BAAAA,KAAKG,KAAK,CAACC,cAAc,YAAzBJ,6BAA6B;QAC7CK,aAAaJ,OAAO1D,MAAM;QAC1B+D,QAAQ,WAAEN,kBAAAA,KAAKO,SAAS,qBAAdP,gBAAgBzD,MAAM,mBAAI;IACtC;AACF;AAwEA,gEAAgE,GAChE,OAAO,MAAMiE,2BAA2B,qCAAoC;AAE5E;;;CAGC,GACD,OAAO,MAAMC,wBAAwB9F;IAWnC,YACE+F,MAAqB,EACrBC,SAAkB,EAClBC,SAAwB,EACxBC,iBAA0C,IAAI,CAC9C;QACA,KAAK,CAACL,0BAA0B;YAAEE;YAAQC;YAAWC;QAAU,SAhB/CE,OAAO;QAiBvB,IAAI,CAACD,cAAc,GAAGA;IACxB;AACF;AAKA;;;;;CAKC,GACD,OAAO,MAAME,4BAA4BC;IACvC,YAAYtC,OAAe,CAAE;QAC3B,KAAK,CAACA;QACN,IAAI,CAACoC,IAAI,GAAG;IACd;AACF;AA8DA;;;;;CAKC,GACD,OAAO,MAAMG,+BAA+BD;IAE1C,YACE,AAASE,OAAe,EACxB,AAASC,MAAc,CACvB;QACA,KAAK,CAAC,CAAC,MAAM,EAAED,QAAQ,8BAA8B,EAAEC,OAAO,CAAC,CAAC,QAHvDD,UAAAA,cACAC,SAAAA,aAHOL,OAAO;IAMzB;AACF;AAEA,2EAA2E,GAC3E,OAAO,SAASM,aAAaxF,KAAc;IACzC,MAAMqE,SAASoB,OAAOzF,gBAAAA,QAAS;IAC/B,OAAOyF,OAAOC,QAAQ,CAACrB,WAAWA,SAAS,IAAIV,KAAKE,KAAK,CAACQ,UAAU;AACtE;AAEA;;;;;;;;;CASC,GACD,OAAO,SAASsB,cAAc3F,KAAc;IAC1C,IAAI,OAAOA,UAAU,UAAU;QAC7B,IAAI,CAACA,MAAM4F,IAAI,IAAI,OAAO,CAAC;QAC3B,IAAI;YACF,MAAMvB,SAAkB7D,KAAKqF,KAAK,CAAC7F;YACnC,OAAOqE,UAAU,OAAOA,WAAW,YAAY,CAACjF,MAAMC,OAAO,CAACgF,UACzDA,SACD,CAAC;QACP,EAAE,eAAM;YACN,OAAO,CAAC;QACV;IACF;IACA,OAAOrE,SAAS,OAAOA,UAAU,WAAYA,QAAoC,CAAC;AACpF"}

package/src/lib/providers/routing.d.ts

@@ -74,6 +74,19 @@ export interface AiRoutingRow {
     /** What the ceiling is sized from. */
     maxTokensBasis: string;
     eval: AiRoutingEval;
+    /**
+     * How long the step's prompt cache lives (AGL-3830): `1h` on every step a
+     * site job runs, in production as in a live run, and the provider's five
+     * minutes on a row that leaves it out. A guided start's routes each write
+     * their prefix once and read it only inside the job, a beat or more apart,
+     * so a five-minute entry is gone before the next job could read it; at an
+     * hour the next site job reads what the last one wrote. The write costs 2x
+     * the input price instead of 1.25x, so a step no job runs leaves it out.
+     * Read by `runValidatedGeneration` into the request's `cacheTtl`; a door
+     * that sends its own request on the same step (the copy assistant's blog
+     * draft) keeps five minutes.
+     */
+    cacheTtl?: '1h';
 }
 export interface AiRoutingEval {
     /** The golden-brief kinds the step answers. */

package/src/lib/providers/routing.js

@@ -80,6 +80,11 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         model
     } : undefined;
 }
+/**
+ * The cache lifetime every step a site job runs asks for (AGL-3830): the
+ * plan, the look, the header and footer, every page and its listing, the
+ * creations, the first posts and products, and the welcome email.
+ */ const AI_SITE_JOB_CACHE_TTL = '1h';
 /**
  * THE ROUTING TABLE (AGL-2937): what each step kind asks of the model, in one
  * place, with the evidence beside it.
@@ -159,7 +164,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 0.9688
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'generate.section': {
         thinking: null,
@@ -193,7 +199,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 0.9833
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'job.template': {
         // Medium since the golden A/B of 2026-10-14 (AGL-3781): both template
@@ -211,7 +218,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 0.9833
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'job.component': {
         // A component's tree and the properties it declares, in one answer, from a
@@ -228,7 +236,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 0.9833
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'job.crm': {
         // CRM by AI (AGL-2917): short answers through a strict tool, written from
@@ -286,7 +295,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 0.9833
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'job.insight': {
         // Readers chosen, then their tables phrased (AGL-2915): short sentences
@@ -328,7 +338,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 0.9833
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'job.products': {
         // A product's copy from its name, text and photo, or a store's first
@@ -348,7 +359,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 0.9922
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'job.email': {
         // One email design from the email palette, written against the brief with
@@ -366,7 +378,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 0.9688
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'job.campaign': {
         // The campaign job generates its email through the email step, so it is
@@ -457,7 +470,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 1
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'job.text': {
         // Low since the golden A/B of 2026-10-14 (AGL-3781): the tagline brief
@@ -492,7 +506,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: 0.9688
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     },
     'job.plan': {
         // A site's plan is an outline written without thinking (AGL-3594), and a
@@ -517,7 +532,8 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
             source: 'authored',
             passRate: 1,
             meanScore: null
-        }
+        },
+        cacheTtl: AI_SITE_JOB_CACHE_TTL
     }
 };
 /**

package/src/lib/providers/routing.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/routing.ts"],"sourcesContent":["/**\n * @license\n * Copyright 2026 Aglyn LLC\n *\n * Licensed under the Apache License, Version 2.0 (the \"License\");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n *   http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an \"AS IS\" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n */\n\nimport { AI_SETTING_PLATFORM, AI_STEP_MODEL_SETTING } from '../plugin-config'\nimport {\n  AI_MODEL_CATALOG,\n  AI_STEP_TIERS,\n  aiCatalogEntry,\n  aiDefaultModelFor,\n  type AiStepKind,\n} from './catalog'\nimport type { AiEvalKind } from '../runtime/ai-eval'\nimport type { AiEffort, AiProvider, AiThinking } from './contract'\nimport { aiProviderById, defaultAiProvider } from './registry'\n\n/**\n * The routing table (AGL-2937, AGL-2939): step kind → provider and model,\n * through the catalog and never a vendor literal.\n *\n * Precedence, dearest override first:\n *\n * 1. the org's `pluginSettings/ai` — a provider, and a model per step kind\n *    that must be one the provider serves;\n * 2. the deployment's environment — `AI_PROVIDER` for the provider,\n *    `AI_DEFAULT_MODEL` for every step, and `ASSIST_MODEL` for the\n *    assistant alone (the incident-response override the chat door has\n *    always honored);\n * 3. the catalog's tier for the step kind on the resolved provider.\n */\n\n/** The environment override for every step kind. */\nexport const AI_DEFAULT_MODEL_ENV = 'AI_DEFAULT_MODEL'\n/** The assistant's own override, honored before the general one. */\nexport const ASSIST_MODEL_ENV = 'ASSIST_MODEL'\n\n/** The org's resolved `pluginSettings/ai` values, as `getPluginConfig` merges them. */\nexport type AiPluginSettings = Record<string, unknown>\n\n/**\n * The model that serves a catalog id (AGL-3781): the id itself, or for a\n * retired row the successor it names, followed to a current row. An id the\n * catalog does not know is returned as it is, for the caller to refuse.\n */\nexport function aiServingModelId(modelId: string): string {\n  let current = modelId\n  for (let hops = 0; hops < AI_MODEL_CATALOG.length; hops += 1) {\n    const successor = aiCatalogEntry(current)?.successor\n    if (!successor) return current\n    current = successor\n  }\n  return current\n}\n\nexport interface AiRoute {\n  provider: AiProvider\n  model: string\n}\n\n/** A settings value that names a real choice, or `undefined` for the platform default. */\nfunction chosen(settings: AiPluginSettings | undefined, key: string): string | undefined {\n  const value = settings?.[key]\n  return typeof value === 'string' && value && value !== AI_SETTING_PLATFORM\n    ? value\n    : undefined\n}\n\n/** The provider a workspace runs on: its own choice when registered, else the platform's. */\nexport function resolveAiProvider(settings?: AiPluginSettings): AiProvider | undefined {\n  const wanted = chosen(settings, 'provider')\n  if (wanted) {\n    const named = aiProviderById(wanted)\n    if (named) return named\n  }\n  return defaultAiProvider()\n}\n\n/**\n * The model a step kind runs on for a workspace, on the resolved provider.\n * A chosen model on another provider is not served — the provider decides\n * what it can run — and falls back to the provider's default for the tier.\n */\nexport function resolveAiRoute(\n  kind: AiStepKind,\n  settings?: AiPluginSettings,\n): AiRoute | undefined {\n  const provider = resolveAiProvider(settings)\n  if (!provider) return undefined\n  // A setting or an override that names a retired model is served by its\n  // successor, never by the retired model itself.\n  const servedBy = (modelId: string | undefined): string | undefined => {\n    const serving = modelId ? aiServingModelId(modelId) : undefined\n    return serving && aiCatalogEntry(serving)?.provider === provider.id ? serving : undefined\n  }\n  const model =\n    servedBy(chosen(settings, AI_STEP_MODEL_SETTING[kind])) ??\n    (kind === 'assist.chat' ? servedBy(process.env[ASSIST_MODEL_ENV]?.trim()) : undefined) ??\n    servedBy(process.env[AI_DEFAULT_MODEL_ENV]?.trim()) ??\n    aiDefaultModelFor(provider.id, AI_STEP_TIERS[kind])\n  return model ? { provider, model } : undefined\n}\n\n/**\n * What one step kind asks of its model, beside the tier it is served from\n * (`AI_STEP_TIERS` in the catalog): whether it thinks and how hard, its\n * answer ceiling, and the eval score that tier holds on the step's golden\n * briefs (AGL-2937).\n */\nexport interface AiRoutingRow {\n  /**\n   * `null` sends no setting: the model's own default, which on every current\n   * model is adaptive thinking. A row whose ceiling holds the answer alone\n   * says `off` (AGL-3781).\n   */\n  thinking: AiThinking | null\n  effort: AiEffort | null\n  /** The answer ceiling, thinking included. */\n  maxTokens: number\n  /** What the ceiling is sized from. */\n  maxTokensBasis: string\n  eval: AiRoutingEval\n}\n\nexport interface AiRoutingEval {\n  /** The golden-brief kinds the step answers. */\n  kinds: readonly AiEvalKind[]\n  /** `answers` scores the step's answers; `plans` scores the plans those kinds' answers carry. */\n  scores: 'answers' | 'plans'\n  /** `authored` while the score is the hand-written references'; `recorded` once a live run records answers. */\n  source: 'authored' | 'recorded'\n  passRate: number\n  /** The mean answer score; `null` for a plan score, which passes or fails. */\n  meanScore: number | null\n}\n\n/**\n * THE ROUTING TABLE (AGL-2937): what each step kind asks of the model, in one\n * place, with the evidence beside it.\n *\n * A step runs on the smallest tier that holds its eval floor. A row moves to a\n * cheaper tier, or a lower effort, only when a recorded eval\n * (`npm run eval:ai-live`) holds the floor there; while a row's score is the\n * authored references', it says so, and the row stays where production has\n * run it. `runtime/ai-eval.spec.ts` recomputes every row's score from the\n * harness, so a row cannot claim a score its briefs do not hold.\n *\n * A ceiling is sized from the p95 output the staff signals page reports for\n * the kind once production has served it; until then `maxTokensBasis` names\n * what it is sized from. The same spec holds every reference answer under its\n * row's ceiling at three characters a token, a reading that errs high for a\n * tokenizer that splits finer than four.\n */\nexport const AI_ROUTING_TABLE: Readonly<Record<AiStepKind, AiRoutingRow>> = {\n  'assist.chat': {\n    // A scoped, latency-sensitive chat turn: no thinking, the low effort rung.\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis:\n      'a docs-grounded answer of a few short paragraphs; the edit rung raises it for its tool call',\n    eval: { kinds: ['chat'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.element': {\n    // One element's copy on the fast tier, whose model thinks unless told not\n    // to (AGL-3781): the ceiling holds the copy alone, as it always has, and\n    // the effort is the model's own.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1024,\n    maxTokensBasis: 'one element of copy, about the length of the text it replaces',\n    eval: { kinds: ['element'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.section': {\n    thinking: null,\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis: 'a section of four to twelve nodes written as JSON',\n    eval: { kinds: ['section'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.blog': {\n    thinking: null,\n    effort: null,\n    maxTokens: 2048,\n    maxTokensBasis: 'one blog post body in markdown-lite',\n    eval: { kinds: ['blog'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'generate.section': {\n    thinking: null,\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis: 'a section of four to twelve nodes written as JSON',\n    eval: { kinds: ['section'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.layout': {\n    // The plan was reasoned out and confirmed before the step, so it writes without thinking.\n    // A site's header and footer in the layout language think adaptively, at\n    // low since AGL-3781 (`AI_LAYOUT_LANGUAGE_THINKING`): the live round of\n    // 2026-10-14 built 13 of 13 frames at low, none cut off.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"tighter than the doctrine's default for a layout, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['layout'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.template': {\n    // Medium since the golden A/B of 2026-10-14 (AGL-3781): both template\n    // briefs passed at medium as at the default, with about a third fewer\n    // output tokens.\n    thinking: 'off',\n    effort: 'medium',\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"tighter than the doctrine's default for a template, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['template'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.component': {\n    // A component's tree and the properties it declares, in one answer, from a\n    // plan reasoned out and confirmed before the step (AGL-2908).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"the ceiling the layout and template steps keep, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['component'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.crm': {\n    // CRM by AI (AGL-2917): short answers through a strict tool, written from\n    // the facts the CRM reports about the record the member opened, with no\n    // thinking. The ceiling fits an inline door, so a member is answered in\n    // the request that asked, and has no room for thinking: the fast tier's\n    // model thinks unless told not to, so the row says off, and its effort\n    // is the low one a short answer read off given facts needs (AGL-3781).\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 700,\n    maxTokensBasis:\n      \"the largest answer a CRM tool accepts: sixty columns matched to fields by number, or an email draft at its length limits, at three characters a token with room, as ai-job-crm-step.spec.ts measures it\",\n    eval: { kinds: ['crm'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9938 },\n  },\n  'job.experiment': {\n    // Variants of one piece of copy, or a result explained against a verdict\n    // code reached before the request was made (AGL-2914). The judgment is\n    // already made either way, so it writes without thinking.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1500,\n    maxTokensBasis:\n      \"four variants at the tool's own lengths — a name, a headline, a body and a rationale each — written as JSON, about 3,000 characters at three characters a token with room; an explanation is a headline, four sentences and a next step, well under it, as ai-job-experiment-step.spec.ts measures both\",\n    eval: { kinds: ['experiment'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.form': {\n    // A form's two halves in one answer: the field declaration and the design\n    // that renders it, as JSON, with no extended thinking (AGL-2913). The step\n    // asks this ceiling, which is the doctrine's own for a form, and registers\n    // the time it takes (AGL-3035).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"the largest form the doctrine's output budget admits, a form's fields, its consent and routing declaration and the design that renders them written as JSON at the wordiest golden's characters a stored byte, under it with room; ai-job-form-step.spec.ts measures it\",\n    eval: { kinds: ['form'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.insight': {\n    // Readers chosen, then their tables phrased (AGL-2915): short sentences\n    // over figures code computed, each held to the rows it cites, so no\n    // extended thinking. The read call before the answer asks for at most\n    // `AI_INSIGHT_READ_MAX_TOKENS` of the same allowance.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1500,\n    maxTokensBasis:\n      'five insights at their 280-character bound, each citing three tables of ten rows, and the gap sentence, written as JSON: about 2,700 characters, under 1,000 tokens at three characters a token, with room',\n    eval: { kinds: ['insight'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.page': {\n    // One section of a page, written against a plan a member already\n    // confirmed, so it writes without thinking (AGL-2907). The ceiling here\n    // is the MOST a pass may ask for; each pass sizes its own down from it to\n    // what fits the time the job beat can give the step, which on the\n    // balanced tier is about half of this. A page in the layout language\n    // thinks adaptively, at low since AGL-3781 (`AI_LAYOUT_LANGUAGE_THINKING`):\n    // the live round of 2026-10-14 built 39 of 39 pages at low, 64 of 65\n    // answers held their rules on the first ask, none cut off.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 2000,\n    maxTokensBasis:\n      \"a section of a page written as JSON, about forty elements at the step's measured 45 tokens an element; a pass lowers it to the worst case that fits AI_JOB_PAGE_STEP_MINIMUM_MS on the model it runs\",\n    eval: { kinds: ['page'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.products': {\n    // A product's copy from its name, text and photo, or a store's first\n    // products, categories and discounts from a brief (AGL-2916): answers held\n    // to the storefront claim rules, written with no extended thinking.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"the largest catalog the tool accepts, twelve proposed products at every bound written as JSON, which the balanced tier asks as much of as fits a beat; one product's copy asks 1,500 and categories with discounts 2,000, as ai-job-products-step.spec.ts measures them\",\n    eval: {\n      kinds: ['product', 'catalog', 'categories'],\n      scores: 'answers',\n      source: 'authored',\n      passRate: 1,\n      meanScore: 0.9922,\n    },\n  },\n  'job.email': {\n    // One email design from the email palette, written against the brief with\n    // no extended thinking (AGL-2912). A campaign job runs the same\n    // generation, so it carries the same ceiling.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"an email's node map written as JSON, with three subject lines and three preheaders beside it\",\n    eval: { kinds: ['email'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.campaign': {\n    // The campaign job generates its email through the email step, so it is\n    // routed identically; the draft campaign it also writes asks nothing of\n    // the model (AGL-2912).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"an email's node map written as JSON, with three subject lines and three preheaders beside it\",\n    eval: { kinds: ['email'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.workflow': {\n    // An automation drafted from a description, or one explained (AGL-2919):\n    // which trigger and which steps a description means is the judgment the\n    // step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 4000,\n    maxTokensBasis:\n      'the largest automation the tool accepts, 6,000 characters written out as JSON, at three characters a token with as much again to think in, as ai-job-workflow-step.spec.ts measures it',\n    eval: { kinds: ['workflow'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.logic': {\n    // A site function or variable proposed, or a function explained\n    // (AGL-3603): which conditions and arithmetic a description means is the\n    // judgment the step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis:\n      'the largest function the tool accepts — twelve parameters, twenty-four operations — is far past any description; a function of eight operations is about 2,400 characters of JSON, at three characters a token with as much again to think in, as ai-job-logic-step.spec.ts measures it',\n    eval: { kinds: ['logic'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.edit': {\n    // A change to a page or a layout the site has (AGL-3616): which elements\n    // a request means, on an outline with no selection, is the judgment the\n    // step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 6144,\n    maxTokensBasis:\n      \"the Assist edit rung's own ceiling for the proposal's tool call (ASSIST_EDIT_MAX_OUTPUT_TOKENS, 4,096: forty new elements written as name/value pairs), with half as much again to think in, as ai-job-edit-step.spec.ts holds it; the whole of it twice, as a re-ask may, still fits one beat on the balanced tier\",\n    eval: { kinds: ['edit'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.seo': {\n    // A listing written from the page's own text to fixed lengths, sized for\n    // the answer alone: no thinking, which the fast tier's model does unless\n    // told not to, at the low effort a short bounded answer needs (AGL-3781).\n    // The audit's site proposal and fix batches run on the same row.\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis:\n      \"the largest listing the tool accepts, at three characters a token with a quarter again, as ai-job-seo-step.spec.ts measures it; an audit's site proposal and fix batches keep ceilings of their own, measured the same way\",\n    eval: { kinds: ['seo'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.text': {\n    // Low since the golden A/B of 2026-10-14 (AGL-3781): the tagline brief\n    // passed at low, its copy scored higher than at the default.\n    thinking: 'adaptive',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis: 'a few paragraphs of copy; a brief asking for more gets a draft to extend',\n    eval: { kinds: ['text'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.theme': {\n    // Medium since the golden A/B of 2026-10-14 (AGL-3781): the theme brief\n    // passed at medium with the default's scores and half its output tokens.\n    // A site start's look runs its own setting (`ai-job-site-look.ts`).\n    thinking: 'adaptive',\n    effort: 'medium',\n    maxTokens: 8000,\n    maxTokensBasis:\n      'the largest answer the theme tool accepts, about 2,500 tokens of JSON as ai-job-theme-step.spec.ts measures it, with as much again to think in',\n    eval: { kinds: ['theme'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.plan': {\n    // A site's plan is an outline written without thinking (AGL-3594), and a\n    // paid site's asks medium (`AI_PAID_SITE_PLAN_EFFORT`, AGL-3781): the live\n    // round of 2026-10-14 planned 11 of 11 paid site briefs at medium in 16\n    // attempts to the default's 22, $0.129 to $0.164, with no refusal and no\n    // rule broken. Every other plan, and a Free site's on the fast tier, asks\n    // this row's effort — the model's default — until a round holds it lower.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis: 'a plan of up to twelve screens with sixteen sections each, with room to think',\n    eval: {\n      kinds: ['page', 'template', 'component', 'layout', 'form'],\n      scores: 'plans',\n      source: 'authored',\n      passRate: 1,\n      meanScore: null,\n    },\n  },\n}\n\n/**\n * What a step kind's row asks of its model, as the fields a request carries:\n * only those the row names, so a row that names none sends none and the\n * model's own default stands.\n */\nexport function aiRouteSettings(kind: AiStepKind): { thinking?: AiThinking; effort?: AiEffort } {\n  const row = AI_ROUTING_TABLE[kind]\n  return {\n    ...(row.thinking ? { thinking: row.thinking } : {}),\n    ...(row.effort ? { effort: row.effort } : {}),\n  }\n}\n\n/** The model id alone — what a door records on the meter and the job. */\nexport function aiModelForStep(kind: AiStepKind, settings?: AiPluginSettings): string {\n  const route = resolveAiRoute(kind, settings)\n  if (!route) {\n    // Every door gates on the provider being configured before it asks\n    // for a model, so this is a wiring fault rather than a customer path.\n    throw new Error('no AI provider is registered')\n  }\n  return route.model\n}\n"],"names":["AI_SETTING_PLATFORM","AI_STEP_MODEL_SETTING","AI_MODEL_CATALOG","AI_STEP_TIERS","aiCatalogEntry","aiDefaultModelFor","aiProviderById","defaultAiProvider","AI_DEFAULT_MODEL_ENV","ASSIST_MODEL_ENV","aiServingModelId","modelId","current","hops","length","successor","chosen","settings","key","value","undefined","resolveAiProvider","wanted","named","resolveAiRoute","kind","servedBy","process","provider","serving","id","model","env","trim","AI_ROUTING_TABLE","thinking","effort","maxTokens","maxTokensBasis","eval","kinds","scores","source","passRate","meanScore","aiRouteSettings","row","aiModelForStep","route","Error"],"mappings":";AAAA;;;;;;;;;;;;;;;CAeC,GAED,SAASA,mBAAmB,EAAEC,qBAAqB,QAAQ,sBAAkB;AAC7E,SACEC,gBAAgB,EAChBC,aAAa,EACbC,cAAc,EACdC,iBAAiB,QAEZ,eAAW;AAGlB,SAASC,cAAc,EAAEC,iBAAiB,QAAQ,gBAAY;AAE9D;;;;;;;;;;;;;CAaC,GAED,kDAAkD,GAClD,OAAO,MAAMC,uBAAuB,mBAAkB;AACtD,kEAAkE,GAClE,OAAO,MAAMC,mBAAmB,eAAc;AAK9C;;;;CAIC,GACD,OAAO,SAASC,iBAAiBC,OAAe;IAC9C,IAAIC,UAAUD;IACd,IAAK,IAAIE,OAAO,GAAGA,OAAOX,iBAAiBY,MAAM,EAAED,QAAQ,EAAG;YAC1CT;QAAlB,MAAMW,aAAYX,kBAAAA,eAAeQ,6BAAfR,gBAAyBW,SAAS;QACpD,IAAI,CAACA,WAAW,OAAOH;QACvBA,UAAUG;IACZ;IACA,OAAOH;AACT;AAOA,wFAAwF,GACxF,SAASI,OAAOC,QAAsC,EAAEC,GAAW;IACjE,MAAMC,QAAQF,4BAAAA,QAAU,CAACC,IAAI;IAC7B,OAAO,OAAOC,UAAU,YAAYA,SAASA,UAAUnB,sBACnDmB,QACAC;AACN;AAEA,2FAA2F,GAC3F,OAAO,SAASC,kBAAkBJ,QAA2B;IAC3D,MAAMK,SAASN,OAAOC,UAAU;IAChC,IAAIK,QAAQ;QACV,MAAMC,QAAQjB,eAAegB;QAC7B,IAAIC,OAAO,OAAOA;IACpB;IACA,OAAOhB;AACT;AAEA;;;;CAIC,GACD,OAAO,SAASiB,eACdC,IAAgB,EAChBR,QAA2B;QAWzBS,MAAAA,OAAAA;QACmCC,+BAC1BA;IAXX,MAAMC,WAAWP,kBAAkBJ;IACnC,IAAI,CAACW,UAAU,OAAOR;IACtB,uEAAuE;IACvE,gDAAgD;IAChD,MAAMM,WAAW,CAACf;YAEEP;QADlB,MAAMyB,UAAUlB,UAAUD,iBAAiBC,WAAWS;QACtD,OAAOS,WAAWzB,EAAAA,kBAAAA,eAAeyB,6BAAfzB,gBAAyBwB,QAAQ,MAAKA,SAASE,EAAE,GAAGD,UAAUT;IAClF;IACA,MAAMW,SACJL,QAAAA,SAAAA,YAAAA,SAASV,OAAOC,UAAUhB,qBAAqB,CAACwB,KAAK,cAArDC,YACCD,SAAS,gBAAgBC,UAASC,gCAAAA,QAAQK,GAAG,CAACvB,iBAAiB,qBAA7BkB,8BAA+BM,IAAI,MAAMb,qBAD5EM,QAEAA,UAASC,oCAAAA,QAAQK,GAAG,CAACxB,qBAAqB,qBAAjCmB,kCAAmCM,IAAI,eAFhDP,OAGArB,kBAAkBuB,SAASE,EAAE,EAAE3B,aAAa,CAACsB,KAAK;IACpD,OAAOM,QAAQ;QAAEH;QAAUG;IAAM,IAAIX;AACvC;AAmCA;;;;;;;;;;;;;;;;CAgBC,GACD,OAAO,MAAMc,mBAA+D;IAC1E,eAAe;QACb,2EAA2E;QAC3EC,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,gBAAgB;QACd,0EAA0E;QAC1E,yEAAyE;QACzE,iCAAiC;QACjCT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,gBAAgB;QACdT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,aAAa;QACXT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,oBAAoB;QAClBT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,cAAc;QACZ,0FAA0F;QAC1F,yEAAyE;QACzE,wEAAwE;QACxE,yDAAyD;QACzDT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAS;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACnG;IACA,gBAAgB;QACd,sEAAsE;QACtE,sEAAsE;QACtE,iBAAiB;QACjBT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAW;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACrG;IACA,iBAAiB;QACf,2EAA2E;QAC3E,8DAA8D;QAC9DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAY;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACtG;IACA,WAAW;QACT,0EAA0E;QAC1E,wEAAwE;QACxE,wEAAwE;QACxE,wEAAwE;QACxE,uEAAuE;QACvE,uEAAuE;QACvET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAM;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAChG;IACA,kBAAkB;QAChB,yEAAyE;QACzE,uEAAuE;QACvE,0DAA0D;QAC1DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAa;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAClG;IACA,YAAY;QACV,0EAA0E;QAC1E,2EAA2E;QAC3E,2EAA2E;QAC3E,gCAAgC;QAChCT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,eAAe;QACb,wEAAwE;QACxE,oEAAoE;QACpE,sEAAsE;QACtE,sDAAsD;QACtDT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,YAAY;QACV,iEAAiE;QACjE,wEAAwE;QACxE,0EAA0E;QAC1E,kEAAkE;QAClE,qEAAqE;QACrE,4EAA4E;QAC5E,qEAAqE;QACrE,2DAA2D;QAC3DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,gBAAgB;QACd,qEAAqE;QACrE,2EAA2E;QAC3E,oEAAoE;QACpET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YACJC,OAAO;gBAAC;gBAAW;gBAAW;aAAa;YAC3CC,QAAQ;YACRC,QAAQ;YACRC,UAAU;YACVC,WAAW;QACb;IACF;IACA,aAAa;QACX,0EAA0E;QAC1E,gEAAgE;QAChE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,gBAAgB;QACd,wEAAwE;QACxE,wEAAwE;QACxE,wBAAwB;QACxBT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,gBAAgB;QACd,yEAAyE;QACzE,wEAAwE;QACxE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAW;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACrG;IACA,aAAa;QACX,gEAAgE;QAChE,yEAAyE;QACzE,2DAA2D;QAC3DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC7F;IACA,YAAY;QACV,yEAAyE;QACzE,wEAAwE;QACxE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC5F;IACA,WAAW;QACT,yEAAyE;QACzE,yEAAyE;QACzE,0EAA0E;QAC1E,iEAAiE;QACjET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAM;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC3F;IACA,YAAY;QACV,uEAAuE;QACvE,6DAA6D;QAC7DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC5F;IACA,aAAa;QACX,wEAAwE;QACxE,yEAAyE;QACzE,oEAAoE;QACpET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,YAAY;QACV,yEAAyE;QACzE,2EAA2E;QAC3E,wEAAwE;QACxE,yEAAyE;QACzE,0EAA0E;QAC1E,0EAA0E;QAC1ET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YACJC,OAAO;gBAAC;gBAAQ;gBAAY;gBAAa;gBAAU;aAAO;YAC1DC,QAAQ;YACRC,QAAQ;YACRC,UAAU;YACVC,WAAW;QACb;IACF;AACF,EAAC;AAED;;;;CAIC,GACD,OAAO,SAASC,gBAAgBpB,IAAgB;IAC9C,MAAMqB,MAAMZ,gBAAgB,CAACT,KAAK;IAClC,OAAO,aACDqB,IAAIX,QAAQ,GAAG;QAAEA,UAAUW,IAAIX,QAAQ;IAAC,IAAI,CAAC,GAC7CW,IAAIV,MAAM,GAAG;QAAEA,QAAQU,IAAIV,MAAM;IAAC,IAAI,CAAC;AAE/C;AAEA,uEAAuE,GACvE,OAAO,SAASW,eAAetB,IAAgB,EAAER,QAA2B;IAC1E,MAAM+B,QAAQxB,eAAeC,MAAMR;IACnC,IAAI,CAAC+B,OAAO;QACV,mEAAmE;QACnE,sEAAsE;QACtE,MAAM,IAAIC,MAAM;IAClB;IACA,OAAOD,MAAMjB,KAAK;AACpB"}
+{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/routing.ts"],"sourcesContent":["/**\n * @license\n * Copyright 2026 Aglyn LLC\n *\n * Licensed under the Apache License, Version 2.0 (the \"License\");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n *   http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an \"AS IS\" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n */\n\nimport { AI_SETTING_PLATFORM, AI_STEP_MODEL_SETTING } from '../plugin-config'\nimport {\n  AI_MODEL_CATALOG,\n  AI_STEP_TIERS,\n  aiCatalogEntry,\n  aiDefaultModelFor,\n  type AiStepKind,\n} from './catalog'\nimport type { AiEvalKind } from '../runtime/ai-eval'\nimport type { AiEffort, AiProvider, AiThinking } from './contract'\nimport { aiProviderById, defaultAiProvider } from './registry'\n\n/**\n * The routing table (AGL-2937, AGL-2939): step kind → provider and model,\n * through the catalog and never a vendor literal.\n *\n * Precedence, dearest override first:\n *\n * 1. the org's `pluginSettings/ai` — a provider, and a model per step kind\n *    that must be one the provider serves;\n * 2. the deployment's environment — `AI_PROVIDER` for the provider,\n *    `AI_DEFAULT_MODEL` for every step, and `ASSIST_MODEL` for the\n *    assistant alone (the incident-response override the chat door has\n *    always honored);\n * 3. the catalog's tier for the step kind on the resolved provider.\n */\n\n/** The environment override for every step kind. */\nexport const AI_DEFAULT_MODEL_ENV = 'AI_DEFAULT_MODEL'\n/** The assistant's own override, honored before the general one. */\nexport const ASSIST_MODEL_ENV = 'ASSIST_MODEL'\n\n/** The org's resolved `pluginSettings/ai` values, as `getPluginConfig` merges them. */\nexport type AiPluginSettings = Record<string, unknown>\n\n/**\n * The model that serves a catalog id (AGL-3781): the id itself, or for a\n * retired row the successor it names, followed to a current row. An id the\n * catalog does not know is returned as it is, for the caller to refuse.\n */\nexport function aiServingModelId(modelId: string): string {\n  let current = modelId\n  for (let hops = 0; hops < AI_MODEL_CATALOG.length; hops += 1) {\n    const successor = aiCatalogEntry(current)?.successor\n    if (!successor) return current\n    current = successor\n  }\n  return current\n}\n\nexport interface AiRoute {\n  provider: AiProvider\n  model: string\n}\n\n/** A settings value that names a real choice, or `undefined` for the platform default. */\nfunction chosen(settings: AiPluginSettings | undefined, key: string): string | undefined {\n  const value = settings?.[key]\n  return typeof value === 'string' && value && value !== AI_SETTING_PLATFORM\n    ? value\n    : undefined\n}\n\n/** The provider a workspace runs on: its own choice when registered, else the platform's. */\nexport function resolveAiProvider(settings?: AiPluginSettings): AiProvider | undefined {\n  const wanted = chosen(settings, 'provider')\n  if (wanted) {\n    const named = aiProviderById(wanted)\n    if (named) return named\n  }\n  return defaultAiProvider()\n}\n\n/**\n * The model a step kind runs on for a workspace, on the resolved provider.\n * A chosen model on another provider is not served — the provider decides\n * what it can run — and falls back to the provider's default for the tier.\n */\nexport function resolveAiRoute(\n  kind: AiStepKind,\n  settings?: AiPluginSettings,\n): AiRoute | undefined {\n  const provider = resolveAiProvider(settings)\n  if (!provider) return undefined\n  // A setting or an override that names a retired model is served by its\n  // successor, never by the retired model itself.\n  const servedBy = (modelId: string | undefined): string | undefined => {\n    const serving = modelId ? aiServingModelId(modelId) : undefined\n    return serving && aiCatalogEntry(serving)?.provider === provider.id ? serving : undefined\n  }\n  const model =\n    servedBy(chosen(settings, AI_STEP_MODEL_SETTING[kind])) ??\n    (kind === 'assist.chat' ? servedBy(process.env[ASSIST_MODEL_ENV]?.trim()) : undefined) ??\n    servedBy(process.env[AI_DEFAULT_MODEL_ENV]?.trim()) ??\n    aiDefaultModelFor(provider.id, AI_STEP_TIERS[kind])\n  return model ? { provider, model } : undefined\n}\n\n/**\n * What one step kind asks of its model, beside the tier it is served from\n * (`AI_STEP_TIERS` in the catalog): whether it thinks and how hard, its\n * answer ceiling, and the eval score that tier holds on the step's golden\n * briefs (AGL-2937).\n */\nexport interface AiRoutingRow {\n  /**\n   * `null` sends no setting: the model's own default, which on every current\n   * model is adaptive thinking. A row whose ceiling holds the answer alone\n   * says `off` (AGL-3781).\n   */\n  thinking: AiThinking | null\n  effort: AiEffort | null\n  /** The answer ceiling, thinking included. */\n  maxTokens: number\n  /** What the ceiling is sized from. */\n  maxTokensBasis: string\n  eval: AiRoutingEval\n  /**\n   * How long the step's prompt cache lives (AGL-3830): `1h` on every step a\n   * site job runs, in production as in a live run, and the provider's five\n   * minutes on a row that leaves it out. A guided start's routes each write\n   * their prefix once and read it only inside the job, a beat or more apart,\n   * so a five-minute entry is gone before the next job could read it; at an\n   * hour the next site job reads what the last one wrote. The write costs 2x\n   * the input price instead of 1.25x, so a step no job runs leaves it out.\n   * Read by `runValidatedGeneration` into the request's `cacheTtl`; a door\n   * that sends its own request on the same step (the copy assistant's blog\n   * draft) keeps five minutes.\n   */\n  cacheTtl?: '1h'\n}\n\n/**\n * The cache lifetime every step a site job runs asks for (AGL-3830): the\n * plan, the look, the header and footer, every page and its listing, the\n * creations, the first posts and products, and the welcome email.\n */\nconst AI_SITE_JOB_CACHE_TTL = '1h' as const\n\nexport interface AiRoutingEval {\n  /** The golden-brief kinds the step answers. */\n  kinds: readonly AiEvalKind[]\n  /** `answers` scores the step's answers; `plans` scores the plans those kinds' answers carry. */\n  scores: 'answers' | 'plans'\n  /** `authored` while the score is the hand-written references'; `recorded` once a live run records answers. */\n  source: 'authored' | 'recorded'\n  passRate: number\n  /** The mean answer score; `null` for a plan score, which passes or fails. */\n  meanScore: number | null\n}\n\n/**\n * THE ROUTING TABLE (AGL-2937): what each step kind asks of the model, in one\n * place, with the evidence beside it.\n *\n * A step runs on the smallest tier that holds its eval floor. A row moves to a\n * cheaper tier, or a lower effort, only when a recorded eval\n * (`npm run eval:ai-live`) holds the floor there; while a row's score is the\n * authored references', it says so, and the row stays where production has\n * run it. `runtime/ai-eval.spec.ts` recomputes every row's score from the\n * harness, so a row cannot claim a score its briefs do not hold.\n *\n * A ceiling is sized from the p95 output the staff signals page reports for\n * the kind once production has served it; until then `maxTokensBasis` names\n * what it is sized from. The same spec holds every reference answer under its\n * row's ceiling at three characters a token, a reading that errs high for a\n * tokenizer that splits finer than four.\n */\nexport const AI_ROUTING_TABLE: Readonly<Record<AiStepKind, AiRoutingRow>> = {\n  'assist.chat': {\n    // A scoped, latency-sensitive chat turn: no thinking, the low effort rung.\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis:\n      'a docs-grounded answer of a few short paragraphs; the edit rung raises it for its tool call',\n    eval: { kinds: ['chat'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.element': {\n    // One element's copy on the fast tier, whose model thinks unless told not\n    // to (AGL-3781): the ceiling holds the copy alone, as it always has, and\n    // the effort is the model's own.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1024,\n    maxTokensBasis: 'one element of copy, about the length of the text it replaces',\n    eval: { kinds: ['element'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.section': {\n    thinking: null,\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis: 'a section of four to twelve nodes written as JSON',\n    eval: { kinds: ['section'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.blog': {\n    thinking: null,\n    effort: null,\n    maxTokens: 2048,\n    maxTokensBasis: 'one blog post body in markdown-lite',\n    eval: { kinds: ['blog'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'generate.section': {\n    thinking: null,\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis: 'a section of four to twelve nodes written as JSON',\n    eval: { kinds: ['section'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.layout': {\n    // The plan was reasoned out and confirmed before the step, so it writes without thinking.\n    // A site's header and footer in the layout language think adaptively, at\n    // low since AGL-3781 (`AI_LAYOUT_LANGUAGE_THINKING`): the live round of\n    // 2026-10-14 built 13 of 13 frames at low, none cut off.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"tighter than the doctrine's default for a layout, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['layout'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'job.template': {\n    // Medium since the golden A/B of 2026-10-14 (AGL-3781): both template\n    // briefs passed at medium as at the default, with about a third fewer\n    // output tokens.\n    thinking: 'off',\n    effort: 'medium',\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"tighter than the doctrine's default for a template, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['template'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'job.component': {\n    // A component's tree and the properties it declares, in one answer, from a\n    // plan reasoned out and confirmed before the step (AGL-2908).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"the ceiling the layout and template steps keep, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['component'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'job.crm': {\n    // CRM by AI (AGL-2917): short answers through a strict tool, written from\n    // the facts the CRM reports about the record the member opened, with no\n    // thinking. The ceiling fits an inline door, so a member is answered in\n    // the request that asked, and has no room for thinking: the fast tier's\n    // model thinks unless told not to, so the row says off, and its effort\n    // is the low one a short answer read off given facts needs (AGL-3781).\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 700,\n    maxTokensBasis:\n      \"the largest answer a CRM tool accepts: sixty columns matched to fields by number, or an email draft at its length limits, at three characters a token with room, as ai-job-crm-step.spec.ts measures it\",\n    eval: { kinds: ['crm'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9938 },\n  },\n  'job.experiment': {\n    // Variants of one piece of copy, or a result explained against a verdict\n    // code reached before the request was made (AGL-2914). The judgment is\n    // already made either way, so it writes without thinking.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1500,\n    maxTokensBasis:\n      \"four variants at the tool's own lengths — a name, a headline, a body and a rationale each — written as JSON, about 3,000 characters at three characters a token with room; an explanation is a headline, four sentences and a next step, well under it, as ai-job-experiment-step.spec.ts measures both\",\n    eval: { kinds: ['experiment'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.form': {\n    // A form's two halves in one answer: the field declaration and the design\n    // that renders it, as JSON, with no extended thinking (AGL-2913). The step\n    // asks this ceiling, which is the doctrine's own for a form, and registers\n    // the time it takes (AGL-3035).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"the largest form the doctrine's output budget admits, a form's fields, its consent and routing declaration and the design that renders them written as JSON at the wordiest golden's characters a stored byte, under it with room; ai-job-form-step.spec.ts measures it\",\n    eval: { kinds: ['form'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'job.insight': {\n    // Readers chosen, then their tables phrased (AGL-2915): short sentences\n    // over figures code computed, each held to the rows it cites, so no\n    // extended thinking. The read call before the answer asks for at most\n    // `AI_INSIGHT_READ_MAX_TOKENS` of the same allowance.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1500,\n    maxTokensBasis:\n      'five insights at their 280-character bound, each citing three tables of ten rows, and the gap sentence, written as JSON: about 2,700 characters, under 1,000 tokens at three characters a token, with room',\n    eval: { kinds: ['insight'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.page': {\n    // One section of a page, written against a plan a member already\n    // confirmed, so it writes without thinking (AGL-2907). The ceiling here\n    // is the MOST a pass may ask for; each pass sizes its own down from it to\n    // what fits the time the job beat can give the step, which on the\n    // balanced tier is about half of this. A page in the layout language\n    // thinks adaptively, at low since AGL-3781 (`AI_LAYOUT_LANGUAGE_THINKING`):\n    // the live round of 2026-10-14 built 39 of 39 pages at low, 64 of 65\n    // answers held their rules on the first ask, none cut off.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 2000,\n    maxTokensBasis:\n      \"a section of a page written as JSON, about forty elements at the step's measured 45 tokens an element; a pass lowers it to the worst case that fits AI_JOB_PAGE_STEP_MINIMUM_MS on the model it runs\",\n    eval: { kinds: ['page'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'job.products': {\n    // A product's copy from its name, text and photo, or a store's first\n    // products, categories and discounts from a brief (AGL-2916): answers held\n    // to the storefront claim rules, written with no extended thinking.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"the largest catalog the tool accepts, twelve proposed products at every bound written as JSON, which the balanced tier asks as much of as fits a beat; one product's copy asks 1,500 and categories with discounts 2,000, as ai-job-products-step.spec.ts measures them\",\n    eval: {\n      kinds: ['product', 'catalog', 'categories'],\n      scores: 'answers',\n      source: 'authored',\n      passRate: 1,\n      meanScore: 0.9922,\n    },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'job.email': {\n    // One email design from the email palette, written against the brief with\n    // no extended thinking (AGL-2912). A campaign job runs the same\n    // generation, so it carries the same ceiling.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"an email's node map written as JSON, with three subject lines and three preheaders beside it\",\n    eval: { kinds: ['email'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'job.campaign': {\n    // The campaign job generates its email through the email step, so it is\n    // routed identically; the draft campaign it also writes asks nothing of\n    // the model (AGL-2912).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"an email's node map written as JSON, with three subject lines and three preheaders beside it\",\n    eval: { kinds: ['email'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.workflow': {\n    // An automation drafted from a description, or one explained (AGL-2919):\n    // which trigger and which steps a description means is the judgment the\n    // step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 4000,\n    maxTokensBasis:\n      'the largest automation the tool accepts, 6,000 characters written out as JSON, at three characters a token with as much again to think in, as ai-job-workflow-step.spec.ts measures it',\n    eval: { kinds: ['workflow'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.logic': {\n    // A site function or variable proposed, or a function explained\n    // (AGL-3603): which conditions and arithmetic a description means is the\n    // judgment the step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis:\n      'the largest function the tool accepts — twelve parameters, twenty-four operations — is far past any description; a function of eight operations is about 2,400 characters of JSON, at three characters a token with as much again to think in, as ai-job-logic-step.spec.ts measures it',\n    eval: { kinds: ['logic'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.edit': {\n    // A change to a page or a layout the site has (AGL-3616): which elements\n    // a request means, on an outline with no selection, is the judgment the\n    // step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 6144,\n    maxTokensBasis:\n      \"the Assist edit rung's own ceiling for the proposal's tool call (ASSIST_EDIT_MAX_OUTPUT_TOKENS, 4,096: forty new elements written as name/value pairs), with half as much again to think in, as ai-job-edit-step.spec.ts holds it; the whole of it twice, as a re-ask may, still fits one beat on the balanced tier\",\n    eval: { kinds: ['edit'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.seo': {\n    // A listing written from the page's own text to fixed lengths, sized for\n    // the answer alone: no thinking, which the fast tier's model does unless\n    // told not to, at the low effort a short bounded answer needs (AGL-3781).\n    // The audit's site proposal and fix batches run on the same row.\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis:\n      \"the largest listing the tool accepts, at three characters a token with a quarter again, as ai-job-seo-step.spec.ts measures it; an audit's site proposal and fix batches keep ceilings of their own, measured the same way\",\n    eval: { kinds: ['seo'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'job.text': {\n    // Low since the golden A/B of 2026-10-14 (AGL-3781): the tagline brief\n    // passed at low, its copy scored higher than at the default.\n    thinking: 'adaptive',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis: 'a few paragraphs of copy; a brief asking for more gets a draft to extend',\n    eval: { kinds: ['text'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.theme': {\n    // Medium since the golden A/B of 2026-10-14 (AGL-3781): the theme brief\n    // passed at medium with the default's scores and half its output tokens.\n    // A site start's look runs its own setting (`ai-job-site-look.ts`).\n    thinking: 'adaptive',\n    effort: 'medium',\n    maxTokens: 8000,\n    maxTokensBasis:\n      'the largest answer the theme tool accepts, about 2,500 tokens of JSON as ai-job-theme-step.spec.ts measures it, with as much again to think in',\n    eval: { kinds: ['theme'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n  'job.plan': {\n    // A site's plan is an outline written without thinking (AGL-3594), and a\n    // paid site's asks medium (`AI_PAID_SITE_PLAN_EFFORT`, AGL-3781): the live\n    // round of 2026-10-14 planned 11 of 11 paid site briefs at medium in 16\n    // attempts to the default's 22, $0.129 to $0.164, with no refusal and no\n    // rule broken. Every other plan, and a Free site's on the fast tier, asks\n    // this row's effort — the model's default — until a round holds it lower.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis: 'a plan of up to twelve screens with sixteen sections each, with room to think',\n    eval: {\n      kinds: ['page', 'template', 'component', 'layout', 'form'],\n      scores: 'plans',\n      source: 'authored',\n      passRate: 1,\n      meanScore: null,\n    },\n    cacheTtl: AI_SITE_JOB_CACHE_TTL,\n  },\n}\n\n/**\n * What a step kind's row asks of its model, as the fields a request carries:\n * only those the row names, so a row that names none sends none and the\n * model's own default stands.\n */\nexport function aiRouteSettings(kind: AiStepKind): { thinking?: AiThinking; effort?: AiEffort } {\n  const row = AI_ROUTING_TABLE[kind]\n  return {\n    ...(row.thinking ? { thinking: row.thinking } : {}),\n    ...(row.effort ? { effort: row.effort } : {}),\n  }\n}\n\n/** The model id alone — what a door records on the meter and the job. */\nexport function aiModelForStep(kind: AiStepKind, settings?: AiPluginSettings): string {\n  const route = resolveAiRoute(kind, settings)\n  if (!route) {\n    // Every door gates on the provider being configured before it asks\n    // for a model, so this is a wiring fault rather than a customer path.\n    throw new Error('no AI provider is registered')\n  }\n  return route.model\n}\n"],"names":["AI_SETTING_PLATFORM","AI_STEP_MODEL_SETTING","AI_MODEL_CATALOG","AI_STEP_TIERS","aiCatalogEntry","aiDefaultModelFor","aiProviderById","defaultAiProvider","AI_DEFAULT_MODEL_ENV","ASSIST_MODEL_ENV","aiServingModelId","modelId","current","hops","length","successor","chosen","settings","key","value","undefined","resolveAiProvider","wanted","named","resolveAiRoute","kind","servedBy","process","provider","serving","id","model","env","trim","AI_SITE_JOB_CACHE_TTL","AI_ROUTING_TABLE","thinking","effort","maxTokens","maxTokensBasis","eval","kinds","scores","source","passRate","meanScore","cacheTtl","aiRouteSettings","row","aiModelForStep","route","Error"],"mappings":";AAAA;;;;;;;;;;;;;;;CAeC,GAED,SAASA,mBAAmB,EAAEC,qBAAqB,QAAQ,sBAAkB;AAC7E,SACEC,gBAAgB,EAChBC,aAAa,EACbC,cAAc,EACdC,iBAAiB,QAEZ,eAAW;AAGlB,SAASC,cAAc,EAAEC,iBAAiB,QAAQ,gBAAY;AAE9D;;;;;;;;;;;;;CAaC,GAED,kDAAkD,GAClD,OAAO,MAAMC,uBAAuB,mBAAkB;AACtD,kEAAkE,GAClE,OAAO,MAAMC,mBAAmB,eAAc;AAK9C;;;;CAIC,GACD,OAAO,SAASC,iBAAiBC,OAAe;IAC9C,IAAIC,UAAUD;IACd,IAAK,IAAIE,OAAO,GAAGA,OAAOX,iBAAiBY,MAAM,EAAED,QAAQ,EAAG;YAC1CT;QAAlB,MAAMW,aAAYX,kBAAAA,eAAeQ,6BAAfR,gBAAyBW,SAAS;QACpD,IAAI,CAACA,WAAW,OAAOH;QACvBA,UAAUG;IACZ;IACA,OAAOH;AACT;AAOA,wFAAwF,GACxF,SAASI,OAAOC,QAAsC,EAAEC,GAAW;IACjE,MAAMC,QAAQF,4BAAAA,QAAU,CAACC,IAAI;IAC7B,OAAO,OAAOC,UAAU,YAAYA,SAASA,UAAUnB,sBACnDmB,QACAC;AACN;AAEA,2FAA2F,GAC3F,OAAO,SAASC,kBAAkBJ,QAA2B;IAC3D,MAAMK,SAASN,OAAOC,UAAU;IAChC,IAAIK,QAAQ;QACV,MAAMC,QAAQjB,eAAegB;QAC7B,IAAIC,OAAO,OAAOA;IACpB;IACA,OAAOhB;AACT;AAEA;;;;CAIC,GACD,OAAO,SAASiB,eACdC,IAAgB,EAChBR,QAA2B;QAWzBS,MAAAA,OAAAA;QACmCC,+BAC1BA;IAXX,MAAMC,WAAWP,kBAAkBJ;IACnC,IAAI,CAACW,UAAU,OAAOR;IACtB,uEAAuE;IACvE,gDAAgD;IAChD,MAAMM,WAAW,CAACf;YAEEP;QADlB,MAAMyB,UAAUlB,UAAUD,iBAAiBC,WAAWS;QACtD,OAAOS,WAAWzB,EAAAA,kBAAAA,eAAeyB,6BAAfzB,gBAAyBwB,QAAQ,MAAKA,SAASE,EAAE,GAAGD,UAAUT;IAClF;IACA,MAAMW,SACJL,QAAAA,SAAAA,YAAAA,SAASV,OAAOC,UAAUhB,qBAAqB,CAACwB,KAAK,cAArDC,YACCD,SAAS,gBAAgBC,UAASC,gCAAAA,QAAQK,GAAG,CAACvB,iBAAiB,qBAA7BkB,8BAA+BM,IAAI,MAAMb,qBAD5EM,QAEAA,UAASC,oCAAAA,QAAQK,GAAG,CAACxB,qBAAqB,qBAAjCmB,kCAAmCM,IAAI,eAFhDP,OAGArB,kBAAkBuB,SAASE,EAAE,EAAE3B,aAAa,CAACsB,KAAK;IACpD,OAAOM,QAAQ;QAAEH;QAAUG;IAAM,IAAIX;AACvC;AAoCA;;;;CAIC,GACD,MAAMc,wBAAwB;AAc9B;;;;;;;;;;;;;;;;CAgBC,GACD,OAAO,MAAMC,mBAA+D;IAC1E,eAAe;QACb,2EAA2E;QAC3EC,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,gBAAgB;QACd,0EAA0E;QAC1E,yEAAyE;QACzE,iCAAiC;QACjCT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,gBAAgB;QACdT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,aAAa;QACXT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;QAC/FC,UAAUZ;IACZ;IACA,oBAAoB;QAClBE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,cAAc;QACZ,0FAA0F;QAC1F,yEAAyE;QACzE,wEAAwE;QACxE,yDAAyD;QACzDT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAS;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;QACjGC,UAAUZ;IACZ;IACA,gBAAgB;QACd,sEAAsE;QACtE,sEAAsE;QACtE,iBAAiB;QACjBE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAW;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;QACnGC,UAAUZ;IACZ;IACA,iBAAiB;QACf,2EAA2E;QAC3E,8DAA8D;QAC9DE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAY;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;QACpGC,UAAUZ;IACZ;IACA,WAAW;QACT,0EAA0E;QAC1E,wEAAwE;QACxE,wEAAwE;QACxE,wEAAwE;QACxE,uEAAuE;QACvE,uEAAuE;QACvEE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAM;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAChG;IACA,kBAAkB;QAChB,yEAAyE;QACzE,uEAAuE;QACvE,0DAA0D;QAC1DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAa;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAClG;IACA,YAAY;QACV,0EAA0E;QAC1E,2EAA2E;QAC3E,2EAA2E;QAC3E,gCAAgC;QAChCT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;QAC/FC,UAAUZ;IACZ;IACA,eAAe;QACb,wEAAwE;QACxE,oEAAoE;QACpE,sEAAsE;QACtE,sDAAsD;QACtDE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,YAAY;QACV,iEAAiE;QACjE,wEAAwE;QACxE,0EAA0E;QAC1E,kEAAkE;QAClE,qEAAqE;QACrE,4EAA4E;QAC5E,qEAAqE;QACrE,2DAA2D;QAC3DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;QAC/FC,UAAUZ;IACZ;IACA,gBAAgB;QACd,qEAAqE;QACrE,2EAA2E;QAC3E,oEAAoE;QACpEE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YACJC,OAAO;gBAAC;gBAAW;gBAAW;aAAa;YAC3CC,QAAQ;YACRC,QAAQ;YACRC,UAAU;YACVC,WAAW;QACb;QACAC,UAAUZ;IACZ;IACA,aAAa;QACX,0EAA0E;QAC1E,gEAAgE;QAChE,8CAA8C;QAC9CE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;QAChGC,UAAUZ;IACZ;IACA,gBAAgB;QACd,wEAAwE;QACxE,wEAAwE;QACxE,wBAAwB;QACxBE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,gBAAgB;QACd,yEAAyE;QACzE,wEAAwE;QACxE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAW;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACrG;IACA,aAAa;QACX,gEAAgE;QAChE,yEAAyE;QACzE,2DAA2D;QAC3DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC7F;IACA,YAAY;QACV,yEAAyE;QACzE,wEAAwE;QACxE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC5F;IACA,WAAW;QACT,yEAAyE;QACzE,yEAAyE;QACzE,0EAA0E;QAC1E,iEAAiE;QACjET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAM;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;QACzFC,UAAUZ;IACZ;IACA,YAAY;QACV,uEAAuE;QACvE,6DAA6D;QAC7DE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC5F;IACA,aAAa;QACX,wEAAwE;QACxE,yEAAyE;QACzE,oEAAoE;QACpET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;QAChGC,UAAUZ;IACZ;IACA,YAAY;QACV,yEAAyE;QACzE,2EAA2E;QAC3E,wEAAwE;QACxE,yEAAyE;QACzE,0EAA0E;QAC1E,0EAA0E;QAC1EE,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YACJC,OAAO;gBAAC;gBAAQ;gBAAY;gBAAa;gBAAU;aAAO;YAC1DC,QAAQ;YACRC,QAAQ;YACRC,UAAU;YACVC,WAAW;QACb;QACAC,UAAUZ;IACZ;AACF,EAAC;AAED;;;;CAIC,GACD,OAAO,SAASa,gBAAgBtB,IAAgB;IAC9C,MAAMuB,MAAMb,gBAAgB,CAACV,KAAK;IAClC,OAAO,aACDuB,IAAIZ,QAAQ,GAAG;QAAEA,UAAUY,IAAIZ,QAAQ;IAAC,IAAI,CAAC,GAC7CY,IAAIX,MAAM,GAAG;QAAEA,QAAQW,IAAIX,MAAM;IAAC,IAAI,CAAC;AAE/C;AAEA,uEAAuE,GACvE,OAAO,SAASY,eAAexB,IAAgB,EAAER,QAA2B;IAC1E,MAAMiC,QAAQ1B,eAAeC,MAAMR;IACnC,IAAI,CAACiC,OAAO;QACV,mEAAmE;QACnE,sEAAsE;QACtE,MAAM,IAAIC,MAAM;IAClB;IACA,OAAOD,MAAMnB,KAAK;AACpB"}

package/src/lib/runtime/ai-dev-env.d.ts

@@ -53,11 +53,21 @@ export declare function aiDevReplayMode(env?: AiDevEnv): AiDevReplayMode;
  * live eval every ten to forty minutes while it iterates, which is past the
  * default five-minute entry and inside an hour, so under the launcher the
  * breakpoints ask for the hour: the second run reads the prefix the first one
- * wrote instead of writing it again. Production keeps the default — its
- * steps run a minute apart on the beat, inside five minutes — and never reads
- * this: `undefined` there.
+ * wrote instead of writing it again. A deployed server never reads this:
+ * `undefined` there, where the lifetime is the route's own
+ * (`aiRequestCacheTtl`).
  */
 export declare function aiDevCacheTtl(env?: AiDevEnv): '1h' | undefined;
+/**
+ * The lifetime a request goes out with: the one its door asked for, which a
+ * site job's routes set to the hour through their routing rows (AGL-3830),
+ * else this process's own. A deployed server sends what the door asked for
+ * and nothing else. In development `AGLYN_AI_CACHE_TTL=5m` turns every
+ * request back to five minutes, the route's hour included, for a one-off
+ * sweep that will not run again within the hour (docs/AI_CACHE_AUDIT_2026-10.md,
+ * lever 5).
+ */
+export declare function aiRequestCacheTtl(requested: '1h' | undefined, env?: AiDevEnv): '1h' | undefined;
 export declare function aiWarnIfReplayAskedInDeployment(env?: AiDevEnv): void;
 /** The variable that turns batch mode on (`runtime/ai-live-batch.ts`). */
 export declare const AI_LIVE_BATCH_ENV = "AGLYN_LIVE_AI_BATCH";