@aglyn/plugins-ai 1.0.0-beta.250 → 1.0.0-beta.253

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\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  /** The most a batch is waited for before every request in it fails; an hour when absent. */\n  timeoutMs?: number\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","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;AA8FA;;;;;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;AAwDA,2EAA2E,GAC3E,OAAO,SAASG,aAAarF,KAAc;IACzC,MAAMqE,SAASiB,OAAOtF,gBAAAA,QAAS;IAC/B,OAAOsF,OAAOC,QAAQ,CAAClB,WAAWA,SAAS,IAAIV,KAAKE,KAAK,CAACQ,UAAU;AACtE;AAEA;;;;;;;;;CASC,GACD,OAAO,SAASmB,cAAcxF,KAAc;IAC1C,IAAI,OAAOA,UAAU,UAAU;QAC7B,IAAI,CAACA,MAAMyF,IAAI,IAAI,OAAO,CAAC;QAC3B,IAAI;YACF,MAAMpB,SAAkB7D,KAAKkF,KAAK,CAAC1F;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. 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\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;AA8FA;;;;;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/model-choice.d.ts

@@ -78,7 +78,7 @@ export interface AiModelBounds {
     /** The org-wide restriction; `null` for none. */
     orgModels?: readonly string[] | null;
 }
-/** The catalog models a manual pick may name on a provider under these bounds. */
+/** The current catalog models a manual pick may name on a provider under these bounds. */
 export declare function aiModelsSelectable(providerId: string, bounds: AiModelBounds): AiCatalogEntry[];
 export interface AiModelChoice {
     /** The catalog id the request runs, and is metered, on. */

package/src/lib/providers/model-choice.js

@@ -15,7 +15,7 @@
  * limitations under the License.
  */ import { assistCreditsFromUsd } from "../usage/assist-credits.js";
 import { AI_MODEL_CATALOG, AI_STEP_TIERS, aiCatalogEntry, aiBilledRatesForModel, estimateAiBilledUsd } from "./catalog.js";
-import { resolveAiRoute } from "./routing.js";
+import { aiServingModelId, resolveAiRoute } from "./routing.js";
 /**
  * THE MODEL SWITCH (AGL-2942): what a person may pick, what "Auto" means,
  * and what a pick costs.
@@ -105,9 +105,11 @@ import { resolveAiRoute } from "./routing.js";
         cacheReadTokens: 1050,
         cacheWriteTokens: 0
     },
+    // One element's rewrite on the fast tier, counted as Claude Haiku 5.5 counts
+    // it: about 30% more tokens than Haiku 4.5 for the same text (AGL-3781).
     'copy.element': {
-        inputTokens: 350,
-        outputTokens: 120,
+        inputTokens: 455,
+        outputTokens: 156,
         cacheReadTokens: 0,
         cacheWriteTokens: 0
     },
@@ -169,16 +171,17 @@ import { resolveAiRoute } from "./routing.js";
         cacheReadTokens: 5800,
         cacheWriteTokens: 0
     },
-    // The CRM step's exchange (AGL-2917). NOTHING is cached: the step runs on
-    // the fast tier, whose cacheable minimum is far above what its prompt
-    // reaches — the doctrine's field rules, the CRM rules and the tool come to
-    // 650 to 913 tokens, as `runtime/ai-prompt-cache.spec.ts` measures — so
-    // every byte is billed as input. A record's facts ride with them, a
-    // timeline cut to its newest twelve entries, and the typical answer is a
-    // summary and a next step of a few sentences, with no thinking.
+    // The CRM step's exchange (AGL-2917), counted as the fast tier's Claude
+    // Haiku 5.5 counts it, about 30% more tokens than Haiku 4.5 (AGL-3781). The
+    // doctrine's field rules, the CRM rules and the tool come to 650 to 913
+    // tokens, as `runtime/ai-prompt-cache.spec.ts` measures them, which caches on
+    // that model; one question about a record writes it, so it is counted as
+    // input. A record's facts ride with them, a timeline cut to its newest twelve
+    // entries, and the typical answer is a summary and a next step of a few
+    // sentences, with no thinking.
     'job.crm': {
-        inputTokens: 1800,
-        outputTokens: 250,
+        inputTokens: 2340,
+        outputTokens: 325,
         cacheReadTokens: 0,
         cacheWriteTokens: 0
     },
@@ -228,19 +231,20 @@ import { resolveAiRoute } from "./routing.js";
         cacheWriteTokens: 0
     },
     // The SEO step's listing exchange, which a page's or a product's listing is
-    // exactly one of. NOTHING is cached here, whatever the breakpoints say: the
-    // step runs on the fast tier, whose cacheable minimum is four times what
-    // this prompt reaches, so every static byte is billed as input on every
-    // attempt — the doctrine's block in its field scope, the listing rules and
-    // the tool's schema, about 760 tokens together. The page text, capped at
-    // 2,000 characters, rides with the current listing and a dozen other titles
-    // for about another 750. The largest answer the tool accepts is about 600
-    // characters of JSON, and the request asks for no thinking. An audit pass
-    // is heavier: eight pages' findings and text, about 9,000 characters, with
-    // up to about 8,000 back.
+    // exactly one of: the median of 146 production listings of October 8–14
+    // (2,255 input and 213 output tokens on Claude Haiku 4.5), counted as the
+    // fast tier's Claude Haiku 5.5 counts them, about 30% more (AGL-3781). The
+    // doctrine's block in its field scope, the listing rules and the tool's
+    // schema are its prefix, which caches on that model and which a site's next
+    // listing reads; a lone listing writes it, so it is counted as input. The
+    // page text, capped at 2,000 characters, rides with the current listing and a
+    // dozen other titles. The largest answer the tool accepts is about 600
+    // characters of JSON, and the request asks for no thinking. An audit pass is
+    // heavier: eight pages' findings and text, about 9,000 characters, with up to
+    // about 8,000 back.
     'job.seo': {
-        inputTokens: 1500,
-        outputTokens: 250,
+        inputTokens: 2930,
+        outputTokens: 277,
         cacheReadTokens: 0,
         cacheWriteTokens: 0
     },
@@ -331,16 +335,19 @@ import { resolveAiRoute } from "./routing.js";
     }
 };
 /** The fewest measured exchanges a median is taken over. */ export const AI_MEASURED_MEDIAN_MIN_SAMPLES = 3;
+/** Whether a list names a model, a retired id on it naming its successor (AGL-3781). */ function names(list, id) {
+    return list.some((listed)=>aiServingModelId(listed) === id);
+}
 function onLists(entry, bounds) {
     var _bounds_orgModels;
-    if (((_bounds_orgModels = bounds.orgModels) == null ? void 0 : _bounds_orgModels.length) && !bounds.orgModels.includes(entry.id)) return false;
-    if (bounds.allotmentModels && !bounds.allotmentModels.includes(entry.id)) return false;
+    if (((_bounds_orgModels = bounds.orgModels) == null ? void 0 : _bounds_orgModels.length) && !names(bounds.orgModels, entry.id)) return false;
+    if (bounds.allotmentModels && !names(bounds.allotmentModels, entry.id)) return false;
     return true;
 }
-/** The catalog models a manual pick may name on a provider under these bounds. */ export function aiModelsSelectable(providerId, bounds) {
+/** The current catalog models a manual pick may name on a provider under these bounds. */ export function aiModelsSelectable(providerId, bounds) {
     var _AI_PLAN_MODEL_TIERS_bounds_plan;
     const tiers = (_AI_PLAN_MODEL_TIERS_bounds_plan = AI_PLAN_MODEL_TIERS[bounds.plan]) != null ? _AI_PLAN_MODEL_TIERS_bounds_plan : [];
-    return AI_MODEL_CATALOG.filter((entry)=>entry.provider === providerId && tiers.includes(entry.tier) && onLists(entry, bounds));
+    return AI_MODEL_CATALOG.filter((entry)=>entry.provider === providerId && !entry.successor && tiers.includes(entry.tier) && onLists(entry, bounds));
 }
 /**
  * A model's blended BILLED rate, for "the cheapest allowed" and nothing
@@ -362,7 +369,9 @@ function onLists(entry, bounds) {
     var _bounds_orgModels;
     const route = resolveAiRoute(kind, settings);
     if (!route) return null;
-    const wanted = typeof requested === 'string' && requested.trim() && requested.trim() !== AI_MODEL_AUTO ? requested.trim() : null;
+    // A pick of a retired model, from a job made before it retired, is a pick
+    // of the model that serves in its place.
+    const wanted = typeof requested === 'string' && requested.trim() && requested.trim() !== AI_MODEL_AUTO ? aiServingModelId(requested.trim()) : null;
     if (wanted && aiModelsSelectable(route.provider.id, bounds).some((entry)=>entry.id === wanted)) {
         return {
             model: wanted,
@@ -373,7 +382,7 @@ function onLists(entry, bounds) {
     const restricted = Boolean((_bounds_orgModels = bounds.orgModels) == null ? void 0 : _bounds_orgModels.length) || Boolean(bounds.allotmentModels);
     let model = route.model;
     if (restricted) {
-        const allowed = AI_MODEL_CATALOG.filter((entry)=>entry.provider === route.provider.id && onLists(entry, bounds));
+        const allowed = AI_MODEL_CATALOG.filter((entry)=>entry.provider === route.provider.id && !entry.successor && onLists(entry, bounds));
         if (allowed.length && !allowed.some((entry)=>entry.id === model)) {
             var _ref;
             var _aiCatalogEntry;

package/src/lib/providers/model-choice.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/model-choice.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 { assistCreditsFromUsd } from '../usage/assist-credits'\nimport type { OrgPlan } from '@aglyn/aglyn/foundation/definitions/org-billing.types'\nimport {\n  AI_MODEL_CATALOG,\n  AI_STEP_TIERS,\n  aiCatalogEntry,\n  aiBilledRatesForModel,\n  estimateAiBilledUsd,\n  type AiCatalogEntry,\n  type AiStepKind,\n} from './catalog'\nimport type { AiUsage } from './contract'\nimport { resolveAiRoute, type AiPluginSettings } from './routing'\n\n/**\n * THE MODEL SWITCH (AGL-2942): what a person may pick, what \"Auto\" means,\n * and what a pick costs.\n *\n * ## Auto is the routing table\n *\n * A request that names no model — or names one it may not have — runs on\n * `resolveAiRoute(kind)`, the table that maps a step kind to a catalog model\n * (AGL-2937). Picking a model is an override of that one row for one request,\n * never a second table.\n *\n * ## Three bounds, one intersection\n *\n * A manual pick must be served by the active provider, in a tier the org's\n * plan offers (`AI_PLAN_MODEL_TIERS`), on the org's own restriction when it\n * set one, and on the allotment allowlists that apply to the request. The\n * two allowlists bound Auto too: when the routing table's answer is not on\n * them, Auto runs on the allowed model of the same tier, else the cheapest\n * allowed one. The plan bounds only a pick — Auto is the platform's choice,\n * made for the step, and a plan never makes it worse.\n *\n * An allowlist that leaves nothing this deployment serves is not a way to\n * switch AI off by mistake: the choice falls back to the routing table, and\n * the permission keys are how AI is switched off on purpose.\n *\n * ## The price beside each option\n *\n * Credits per typical request at each model's list rates, over the step's\n * measured median exchange when the workspace has enough of them, and over\n * `AI_STEP_NOMINAL_USAGE` until it does — with the multiple of what Auto\n * would cost for the same request. Every figure is credits, never dollars.\n */\n\n/** The value a request sends for \"let the platform route it\". */\nexport const AI_MODEL_AUTO = 'auto'\n\n/**\n * The catalog tiers a MANUAL pick may name, per plan. Free runs on Auto only:\n * its taste is a wall, and a pick there would spend the account's credits at\n * a rate nobody chose on purpose. The dearest tier opens from Scale up, where\n * the band is large enough that a slower, deeper model is a real choice\n * rather than a fast way to end the month.\n */\nexport const AI_PLAN_MODEL_TIERS: Record<OrgPlan, readonly AiCatalogEntry['tier'][]> = {\n  free: [],\n  starter: ['fast', 'balanced'],\n  pro: ['fast', 'balanced'],\n  business: ['fast', 'balanced'],\n  scale: ['fast', 'balanced', 'deep'],\n  advanced: ['fast', 'balanced', 'deep'],\n  agency: ['fast', 'balanced', 'deep'],\n  enterprise: ['fast', 'balanced', 'deep'],\n}\n\n/** The plans whose admins may restrict the model set org-wide. */\nexport const AI_MODEL_RESTRICTION_PLANS: readonly OrgPlan[] = ['agency', 'enterprise']\n\n/**\n * A typical exchange per step kind, before a workspace has measured one: the\n * doors' own prompt sizes and output ceilings, read conservatively. Replaced\n * by the workspace's measured median as soon as `aiMedianUsage` has enough\n * samples, and only ever used for the price beside an option.\n */\nexport const AI_STEP_NOMINAL_USAGE: Record<AiStepKind, AiUsage> = {\n  'assist.chat': { inputTokens: 900, outputTokens: 450, cacheReadTokens: 1_050, cacheWriteTokens: 0 },\n  'copy.element': { inputTokens: 350, outputTokens: 120, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  'copy.section': { inputTokens: 700, outputTokens: 1_400, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  'copy.blog': { inputTokens: 600, outputTokens: 1_100, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  'generate.section': { inputTokens: 700, outputTokens: 1_400, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The text step's rules are about 500 characters — far under the balanced\n  // tier's cacheable minimum, so they are billed as input every time rather\n  // than read from a cache (`runtime/ai-prompt-cache.spec.ts` measures it).\n  'job.text': { inputTokens: 450, outputTokens: 500, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The theme step: the tool's schema and both system blocks (about 7,500\n  // characters) are the cached prefix, the site's inventory and the brief ride\n  // uncached, and a full-palette answer is about 1,300 characters of JSON with\n  // as much again to think in.\n  'job.theme': { inputTokens: 500, outputTokens: 850, cacheReadTokens: 2_500, cacheWriteTokens: 0 },\n  // The plan step: the tool's schema, the doctrine and the plan instructions\n  // (about 10,000 characters) are the cached prefix, the site inventory and the\n  // brief ride uncached, and a plan answer is about 600 characters of JSON with\n  // as much again to think in.\n  'job.plan': { inputTokens: 500, outputTokens: 400, cacheReadTokens: 3_600, cacheWriteTokens: 0 },\n  // The component step: the doctrine, the component instructions, the\n  // component surface's catalog and the tool are the cached prefix, the site\n  // inventory, the brief and the confirmed plan ride uncached, and a card-sized\n  // answer — its tree and declared props as JSON — is about 1,700 characters.\n  // The component step's spec measures the prefix and the answer against this.\n  'job.component': { inputTokens: 700, outputTokens: 600, cacheReadTokens: 5_800, cacheWriteTokens: 0 },\n  // The CRM step's exchange (AGL-2917). NOTHING is cached: the step runs on\n  // the fast tier, whose cacheable minimum is far above what its prompt\n  // reaches — the doctrine's field rules, the CRM rules and the tool come to\n  // 650 to 913 tokens, as `runtime/ai-prompt-cache.spec.ts` measures — so\n  // every byte is billed as input. A record's facts ride with them, a\n  // timeline cut to its newest twelve entries, and the typical answer is a\n  // summary and a next step of a few sentences, with no thinking.\n  'job.crm': { inputTokens: 1_800, outputTokens: 250, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // A layout: a tree layout's 1,000 output tokens until AGL-3660, now a\n  // site's header and footer in the layout language — 442 to 583 output\n  // tokens live on 2026-10-07 — with up to 1,500 more for the adaptive\n  // thinking it is asked with, an estimate until a workspace measures it.\n  'job.layout': { inputTokens: 700, outputTokens: 2_000, cacheReadTokens: 5_600, cacheWriteTokens: 0 },\n  'job.template': { inputTokens: 1_200, outputTokens: 700, cacheReadTokens: 6_400, cacheWriteTokens: 0 },\n  // One page section's exchange (AGL-2907), from the modules' own text: the\n  // doctrine (6,281 characters), the page instructions (1,524), the screen\n  // palette catalog (8,353) and the section tool (484) are the cached prefix;\n  // the site inventory (711 on the median golden site), the brief, the plan\n  // and the section line ride uncached; and a golden section answer is 901\n  // characters at the median, asked for with no thinking.\n  // Since AGL-3660 a site's and a build's pages are one answer a page in\n  // the layout language: 683 to 1,745 output tokens live on 2026-10-07, about\n  // 1,150 at the median, with up to 1,500 more for the adaptive thinking it\n  // is asked with — an estimate until a workspace measures it, read high so\n  // the quote beside a model never undersells the exchange.\n  'job.page': { inputTokens: 600, outputTokens: 2_650, cacheReadTokens: 4_400, cacheWriteTokens: 0 },\n  // One product's copy (AGL-2916), the exchange a bulk job makes once a\n  // product: the doctrine, the copy rules and the tool (about 9,500 characters)\n  // are the cached prefix; the product's facts, the site's categories and a\n  // photo at most 768 pixels on its long edge (about 790 tokens) ride\n  // uncached; and the answer is a description with its search listing, tags,\n  // categories and option names, about 1,300 characters of JSON.\n  'job.products': { inputTokens: 1_500, outputTokens: 450, cacheReadTokens: 2_400, cacheWriteTokens: 0 },\n  // The SEO step's listing exchange, which a page's or a product's listing is\n  // exactly one of. NOTHING is cached here, whatever the breakpoints say: the\n  // step runs on the fast tier, whose cacheable minimum is four times what\n  // this prompt reaches, so every static byte is billed as input on every\n  // attempt — the doctrine's block in its field scope, the listing rules and\n  // the tool's schema, about 760 tokens together. The page text, capped at\n  // 2,000 characters, rides with the current listing and a dozen other titles\n  // for about another 750. The largest answer the tool accepts is about 600\n  // characters of JSON, and the request asks for no thinking. An audit pass\n  // is heavier: eight pages' findings and text, about 9,000 characters, with\n  // up to about 8,000 back.\n  'job.seo': { inputTokens: 1_500, outputTokens: 250, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The experiment step (AGL-2914). NOTHING is cached: the A/B test rules\n  // and the acceptable-use block come to about 2,300 characters, under the\n  // balanced tier's cacheable minimum, so they are billed as input every\n  // time. The copy under test, or a test's arms with its verdict, rides with\n  // them, and a four-variant answer is about 800 characters of JSON.\n  'job.experiment': { inputTokens: 900, outputTokens: 300, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The form step: the doctrine, the form instructions, the form surface's\n  // catalog (about 150 tokens) and the tool's schema are the cached prefix, the\n  // site inventory and the brief ride uncached, and a form's answer is 200 to\n  // 400 tokens of JSON with no extended thinking.\n  'job.form': { inputTokens: 700, outputTokens: 450, cacheReadTokens: 3_900, cacheWriteTokens: 0 },\n  // An insight (AGL-2915): two exchanges, and NOTHING cached — the rules and\n  // the acceptable-use block are under the balanced tier's minimum, so they\n  // are billed as input on both. The read call carries the question and the\n  // readers' catalog (about 2,000 characters) and answers a few reads; the\n  // answer call carries four tables of about ten rows each (about 4,000\n  // characters) and answers three to five insights as JSON.\n  'job.insight': { inputTokens: 2_800, outputTokens: 650, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The email step, and the campaign step that shares its generation: the\n  // doctrine, the email instructions, the email palette catalog and the\n  // tool's schema (about 10,300 characters) are the cached prefix, the site\n  // inventory and the brief ride uncached, and the answer is one email's\n  // node map with three subject lines and three preheaders.\n  'job.email': { inputTokens: 600, outputTokens: 1_500, cacheReadTokens: 3_600, cacheWriteTokens: 0 },\n  'job.campaign': { inputTokens: 600, outputTokens: 1_500, cacheReadTokens: 3_600, cacheWriteTokens: 0 },\n  // The workflow step's draft (AGL-2919): the doctrine, the automation\n  // vocabulary and the tool's schema are the cached prefix (4,564 tokens, as\n  // `runtime/ai-prompt-cache.spec.ts` measures it); what the workspace can\n  // run, the site's forms and datasets and the description ride uncached; and\n  // an automation of a few steps is about 1,300 characters of JSON with as\n  // much again to think in. An explanation caches half as much.\n  'job.workflow': { inputTokens: 500, outputTokens: 1_000, cacheReadTokens: 4_500, cacheWriteTokens: 0 },\n  // The logic step (AGL-3603): the fields doctrine, one rules block and the\n  // tool's schema are the cached prefix (1,556 tokens for a function, as\n  // `runtime/ai-prompt-cache.spec.ts` measures it); the site's variables and\n  // the request ride uncached, and a function of a few operations is about\n  // 1,200 characters of JSON with as much again to think in.\n  'job.logic': { inputTokens: 700, outputTokens: 900, cacheReadTokens: 1_500, cacheWriteTokens: 0 },\n  // The edit step (AGL-3616): the document doctrine, the job's rules and the\n  // edit rung's protocol and catalog for the document's kind are the cached\n  // prefix (5,970 tokens for a page, as `runtime/ai-prompt-cache.spec.ts`\n  // measures it); the outline of up to sixty elements with their settings\n  // and the request ride uncached, and a change of a few operations is about\n  // 1,500 characters of tool call with as much again to think in.\n  'job.edit': { inputTokens: 2_500, outputTokens: 1_000, cacheReadTokens: 6_000, cacheWriteTokens: 0 },\n}\n\n/** The fewest measured exchanges a median is taken over. */\nexport const AI_MEASURED_MEDIAN_MIN_SAMPLES = 3\n\nexport interface AiModelBounds {\n  plan: OrgPlan\n  /** The allotment allowlists that apply, intersected; `null` for none. */\n  allotmentModels?: readonly string[] | null\n  /** The org-wide restriction; `null` for none. */\n  orgModels?: readonly string[] | null\n}\n\nfunction onLists(entry: AiCatalogEntry, bounds: AiModelBounds): boolean {\n  if (bounds.orgModels?.length && !bounds.orgModels.includes(entry.id)) return false\n  if (bounds.allotmentModels && !bounds.allotmentModels.includes(entry.id)) return false\n  return true\n}\n\n/** The catalog models a manual pick may name on a provider under these bounds. */\nexport function aiModelsSelectable(\n  providerId: string,\n  bounds: AiModelBounds,\n): AiCatalogEntry[] {\n  const tiers = AI_PLAN_MODEL_TIERS[bounds.plan] ?? []\n  return AI_MODEL_CATALOG.filter(\n    (entry) =>\n      entry.provider === providerId && tiers.includes(entry.tier) && onLists(entry, bounds),\n  )\n}\n\n/**\n * A model's blended BILLED rate, for \"the cheapest allowed\" and nothing\n * else.\n *\n * Billed and not provider (AGL-3015): the substitution below happens when a\n * workspace asked for a model its bounds do not allow, and the party it\n * falls back for is the one whose credits are drawn. \"Cheapest\" has to mean\n * the same thing here as in the selector this substitution stands in for,\n * which sorts on `creditsPerRequest`.\n */\nconst blendedRate = (id: string): number => {\n  const rates = aiBilledRatesForModel(id)\n  return rates.inputPerToken + rates.outputPerToken\n}\n\nexport interface AiModelChoice {\n  /** The catalog id the request runs, and is metered, on. */\n  model: string\n  /** True when the routing table decided: Auto, or a pick that was declined. */\n  auto: boolean\n  /** A pick that was asked for and is not allowed, for the envelope. */\n  declined: string | null\n}\n\n/**\n * The model one request runs on, or `null` when no provider is registered —\n * which every door has refused before it asks.\n */\nexport function resolveAiModelChoice(\n  kind: AiStepKind,\n  requested: unknown,\n  bounds: AiModelBounds,\n  settings?: AiPluginSettings,\n): AiModelChoice | null {\n  const route = resolveAiRoute(kind, settings)\n  if (!route) return null\n  const wanted =\n    typeof requested === 'string' && requested.trim() && requested.trim() !== AI_MODEL_AUTO\n      ? requested.trim()\n      : null\n  if (wanted && aiModelsSelectable(route.provider.id, bounds).some((entry) => entry.id === wanted)) {\n    return { model: wanted, auto: false, declined: null }\n  }\n  const restricted = Boolean(bounds.orgModels?.length) || Boolean(bounds.allotmentModels)\n  let model = route.model\n  if (restricted) {\n    const allowed = AI_MODEL_CATALOG.filter(\n      (entry) => entry.provider === route.provider.id && onLists(entry, bounds),\n    )\n    if (allowed.length && !allowed.some((entry) => entry.id === model)) {\n      const tier = aiCatalogEntry(model)?.tier ?? AI_STEP_TIERS[kind]\n      const sameTier = allowed.find((entry) => entry.tier === tier)\n      const cheapest = [...allowed].sort((a, b) => blendedRate(a.id) - blendedRate(b.id))[0]\n      model = (sameTier ?? cheapest).id\n    }\n  }\n  return { model, auto: true, declined: wanted }\n}\n\n/** The median exchange over measured samples, or `null` below the minimum. */\nexport function aiMedianUsage(samples: readonly AiUsage[]): AiUsage | null {\n  if (samples.length < AI_MEASURED_MEDIAN_MIN_SAMPLES) return null\n  const median = (field: keyof AiUsage): number => {\n    const values = samples\n      .map((sample) => Number(sample[field] ?? 0))\n      .filter((value) => Number.isFinite(value) && value >= 0)\n      .sort((a, b) => a - b)\n    if (!values.length) return 0\n    const middle = Math.floor(values.length / 2)\n    return values.length % 2\n      ? values[middle]\n      : Math.round((values[middle - 1] + values[middle]) / 2)\n  }\n  return {\n    inputTokens: median('inputTokens'),\n    outputTokens: median('outputTokens'),\n    cacheReadTokens: median('cacheReadTokens'),\n    cacheWriteTokens: median('cacheWriteTokens'),\n  }\n}\n\nexport interface AiModelOption {\n  /** A catalog id, or `AI_MODEL_AUTO`. */\n  id: string\n  label: string\n  tier: AiCatalogEntry['tier']\n  /** Credits a typical request of this kind costs on the option. */\n  creditsPerRequest: number\n  /** That cost as a multiple of Auto's, to one decimal. */\n  multiplier: number\n}\n\nexport interface AiModelOptions {\n  /** Auto, labeled with the model it routes this kind to under these bounds. */\n  auto: AiModelOption & { model: string }\n  /** The manual picks, cheapest first; empty where the plan offers none. */\n  options: AiModelOption[]\n  /** Whether the price came from the workspace's own measured exchanges. */\n  measured: boolean\n}\n\n/** What the selector lists for one step kind, or `null` without a provider. */\nexport function aiModelOptions(\n  kind: AiStepKind,\n  bounds: AiModelBounds,\n  measured: AiUsage | null,\n  settings?: AiPluginSettings,\n): AiModelOptions | null {\n  const route = resolveAiRoute(kind, settings)\n  const auto = resolveAiModelChoice(kind, null, bounds, settings)\n  if (!route || !auto) return null\n  const typical = measured ?? AI_STEP_NOMINAL_USAGE[kind]\n  // Billed rates throughout (AGL-3015): every figure below is what the\n  // workspace will be charged for picking this model — credits per request\n  // and the multiple of Auto those credits come to. Our own cost for the\n  // same request is not this selector's subject and would misstate both.\n  const autoCost = estimateAiBilledUsd(typical, auto.model)\n  const priced = (entry: AiCatalogEntry): AiModelOption => {\n    const cost = estimateAiBilledUsd(typical, entry.id)\n    return {\n      id: entry.id,\n      label: entry.label,\n      tier: entry.tier,\n      creditsPerRequest: assistCreditsFromUsd(cost),\n      multiplier: autoCost > 0 ? Math.round((cost / autoCost) * 10) / 10 : 1,\n    }\n  }\n  const autoEntry = aiCatalogEntry(auto.model)\n  return {\n    auto: {\n      id: AI_MODEL_AUTO,\n      label: autoEntry ? `Auto (${autoEntry.label})` : 'Auto',\n      tier: autoEntry?.tier ?? AI_STEP_TIERS[kind],\n      creditsPerRequest: assistCreditsFromUsd(autoCost),\n      multiplier: 1,\n      model: auto.model,\n    },\n    options: aiModelsSelectable(route.provider.id, bounds)\n      .map(priced)\n      .sort((a, b) => a.creditsPerRequest - b.creditsPerRequest || a.label.localeCompare(b.label)),\n    measured: measured !== null,\n  }\n}\n"],"names":["assistCreditsFromUsd","AI_MODEL_CATALOG","AI_STEP_TIERS","aiCatalogEntry","aiBilledRatesForModel","estimateAiBilledUsd","resolveAiRoute","AI_MODEL_AUTO","AI_PLAN_MODEL_TIERS","free","starter","pro","business","scale","advanced","agency","enterprise","AI_MODEL_RESTRICTION_PLANS","AI_STEP_NOMINAL_USAGE","inputTokens","outputTokens","cacheReadTokens","cacheWriteTokens","AI_MEASURED_MEDIAN_MIN_SAMPLES","onLists","entry","bounds","orgModels","length","includes","id","allotmentModels","aiModelsSelectable","providerId","tiers","plan","filter","provider","tier","blendedRate","rates","inputPerToken","outputPerToken","resolveAiModelChoice","kind","requested","settings","route","wanted","trim","some","model","auto","declined","restricted","Boolean","allowed","sameTier","find","cheapest","sort","a","b","aiMedianUsage","samples","median","field","values","map","sample","Number","value","isFinite","middle","Math","floor","round","aiModelOptions","measured","typical","autoCost","priced","cost","label","creditsPerRequest","multiplier","autoEntry","options","localeCompare"],"mappings":"AAAA;;;;;;;;;;;;;;;CAeC,GAED,SAASA,oBAAoB,QAAQ,6BAAyB;AAE9D,SACEC,gBAAgB,EAChBC,aAAa,EACbC,cAAc,EACdC,qBAAqB,EACrBC,mBAAmB,QAGd,eAAW;AAElB,SAASC,cAAc,QAA+B,eAAW;AAEjE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA+BC,GAED,+DAA+D,GAC/D,OAAO,MAAMC,gBAAgB,OAAM;AAEnC;;;;;;CAMC,GACD,OAAO,MAAMC,sBAA0E;IACrFC,MAAM,EAAE;IACRC,SAAS;QAAC;QAAQ;KAAW;IAC7BC,KAAK;QAAC;QAAQ;KAAW;IACzBC,UAAU;QAAC;QAAQ;KAAW;IAC9BC,OAAO;QAAC;QAAQ;QAAY;KAAO;IACnCC,UAAU;QAAC;QAAQ;QAAY;KAAO;IACtCC,QAAQ;QAAC;QAAQ;QAAY;KAAO;IACpCC,YAAY;QAAC;QAAQ;QAAY;KAAO;AAC1C,EAAC;AAED,gEAAgE,GAChE,OAAO,MAAMC,6BAAiD;IAAC;IAAU;CAAa,CAAA;AAEtF;;;;;CAKC,GACD,OAAO,MAAMC,wBAAqD;IAChE,eAAe;QAAEC,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAClG,gBAAgB;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC/F,gBAAgB;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAGC,kBAAkB;IAAE;IACjG,aAAa;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC9F,oBAAoB;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAGC,kBAAkB;IAAE;IACrG,0EAA0E;IAC1E,0EAA0E;IAC1E,0EAA0E;IAC1E,YAAY;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC3F,wEAAwE;IACxE,6EAA6E;IAC7E,6EAA6E;IAC7E,6BAA6B;IAC7B,aAAa;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAChG,2EAA2E;IAC3E,8EAA8E;IAC9E,8EAA8E;IAC9E,6BAA6B;IAC7B,YAAY;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAC/F,oEAAoE;IACpE,2EAA2E;IAC3E,8EAA8E;IAC9E,4EAA4E;IAC5E,6EAA6E;IAC7E,iBAAiB;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IACpG,0EAA0E;IAC1E,sEAAsE;IACtE,2EAA2E;IAC3E,wEAAwE;IACxE,oEAAoE;IACpE,yEAAyE;IACzE,gEAAgE;IAChE,WAAW;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC5F,sEAAsE;IACtE,sEAAsE;IACtE,qEAAqE;IACrE,wEAAwE;IACxE,cAAc;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IACnG,gBAAgB;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IACrG,0EAA0E;IAC1E,yEAAyE;IACzE,4EAA4E;IAC5E,0EAA0E;IAC1E,yEAAyE;IACzE,wDAAwD;IACxD,uEAAuE;IACvE,4EAA4E;IAC5E,0EAA0E;IAC1E,0EAA0E;IAC1E,0DAA0D;IAC1D,YAAY;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IACjG,sEAAsE;IACtE,8EAA8E;IAC9E,0EAA0E;IAC1E,oEAAoE;IACpE,2EAA2E;IAC3E,+DAA+D;IAC/D,gBAAgB;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IACrG,4EAA4E;IAC5E,4EAA4E;IAC5E,yEAAyE;IACzE,wEAAwE;IACxE,2EAA2E;IAC3E,yEAAyE;IACzE,4EAA4E;IAC5E,0EAA0E;IAC1E,0EAA0E;IAC1E,2EAA2E;IAC3E,0BAA0B;IAC1B,WAAW;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC5F,wEAAwE;IACxE,yEAAyE;IACzE,uEAAuE;IACvE,2EAA2E;IAC3E,mEAAmE;IACnE,kBAAkB;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IACjG,yEAAyE;IACzE,8EAA8E;IAC9E,4EAA4E;IAC5E,gDAAgD;IAChD,YAAY;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAC/F,2EAA2E;IAC3E,0EAA0E;IAC1E,0EAA0E;IAC1E,yEAAyE;IACzE,sEAAsE;IACtE,0DAA0D;IAC1D,eAAe;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAChG,wEAAwE;IACxE,sEAAsE;IACtE,0EAA0E;IAC1E,uEAAuE;IACvE,0DAA0D;IAC1D,aAAa;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IAClG,gBAAgB;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IACrG,qEAAqE;IACrE,2EAA2E;IAC3E,yEAAyE;IACzE,4EAA4E;IAC5E,yEAAyE;IACzE,8DAA8D;IAC9D,gBAAgB;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IACrG,0EAA0E;IAC1E,uEAAuE;IACvE,2EAA2E;IAC3E,yEAAyE;IACzE,2DAA2D;IAC3D,aAAa;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAChG,2EAA2E;IAC3E,0EAA0E;IAC1E,wEAAwE;IACxE,wEAAwE;IACxE,2EAA2E;IAC3E,gEAAgE;IAChE,YAAY;QAAEH,aAAa;QAAOC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;AACrG,EAAC;AAED,0DAA0D,GAC1D,OAAO,MAAMC,iCAAiC,EAAC;AAU/C,SAASC,QAAQC,KAAqB,EAAEC,MAAqB;QACvDA;IAAJ,IAAIA,EAAAA,oBAAAA,OAAOC,SAAS,qBAAhBD,kBAAkBE,MAAM,KAAI,CAACF,OAAOC,SAAS,CAACE,QAAQ,CAACJ,MAAMK,EAAE,GAAG,OAAO;IAC7E,IAAIJ,OAAOK,eAAe,IAAI,CAACL,OAAOK,eAAe,CAACF,QAAQ,CAACJ,MAAMK,EAAE,GAAG,OAAO;IACjF,OAAO;AACT;AAEA,gFAAgF,GAChF,OAAO,SAASE,mBACdC,UAAkB,EAClBP,MAAqB;QAEPlB;IAAd,MAAM0B,SAAQ1B,mCAAAA,mBAAmB,CAACkB,OAAOS,IAAI,CAAC,YAAhC3B,mCAAoC,EAAE;IACpD,OAAOP,iBAAiBmC,MAAM,CAC5B,CAACX,QACCA,MAAMY,QAAQ,KAAKJ,cAAcC,MAAML,QAAQ,CAACJ,MAAMa,IAAI,KAAKd,QAAQC,OAAOC;AAEpF;AAEA;;;;;;;;;CASC,GACD,MAAMa,cAAc,CAACT;IACnB,MAAMU,QAAQpC,sBAAsB0B;IACpC,OAAOU,MAAMC,aAAa,GAAGD,MAAME,cAAc;AACnD;AAWA;;;CAGC,GACD,OAAO,SAASC,qBACdC,IAAgB,EAChBC,SAAkB,EAClBnB,MAAqB,EACrBoB,QAA2B;QAWApB;IAT3B,MAAMqB,QAAQzC,eAAesC,MAAME;IACnC,IAAI,CAACC,OAAO,OAAO;IACnB,MAAMC,SACJ,OAAOH,cAAc,YAAYA,UAAUI,IAAI,MAAMJ,UAAUI,IAAI,OAAO1C,gBACtEsC,UAAUI,IAAI,KACd;IACN,IAAID,UAAUhB,mBAAmBe,MAAMV,QAAQ,CAACP,EAAE,EAAEJ,QAAQwB,IAAI,CAAC,CAACzB,QAAUA,MAAMK,EAAE,KAAKkB,SAAS;QAChG,OAAO;YAAEG,OAAOH;YAAQI,MAAM;YAAOC,UAAU;QAAK;IACtD;IACA,MAAMC,aAAaC,SAAQ7B,oBAAAA,OAAOC,SAAS,qBAAhBD,kBAAkBE,MAAM,KAAK2B,QAAQ7B,OAAOK,eAAe;IACtF,IAAIoB,QAAQJ,MAAMI,KAAK;IACvB,IAAIG,YAAY;QACd,MAAME,UAAUvD,iBAAiBmC,MAAM,CACrC,CAACX,QAAUA,MAAMY,QAAQ,KAAKU,MAAMV,QAAQ,CAACP,EAAE,IAAIN,QAAQC,OAAOC;QAEpE,IAAI8B,QAAQ5B,MAAM,IAAI,CAAC4B,QAAQN,IAAI,CAAC,CAACzB,QAAUA,MAAMK,EAAE,KAAKqB,QAAQ;;gBACrDhD;YAAb,MAAMmC,gBAAOnC,kBAAAA,eAAegD,2BAAfhD,gBAAuBmC,IAAI,mBAAIpC,aAAa,CAAC0C,KAAK;YAC/D,MAAMa,WAAWD,QAAQE,IAAI,CAAC,CAACjC,QAAUA,MAAMa,IAAI,KAAKA;YACxD,MAAMqB,WAAW;mBAAIH;aAAQ,CAACI,IAAI,CAAC,CAACC,GAAGC,IAAMvB,YAAYsB,EAAE/B,EAAE,IAAIS,YAAYuB,EAAEhC,EAAE,EAAE,CAAC,EAAE;YACtFqB,QAAQ,CAACM,mBAAAA,WAAYE,UAAU7B,EAAE;QACnC;IACF;IACA,OAAO;QAAEqB;QAAOC,MAAM;QAAMC,UAAUL;IAAO;AAC/C;AAEA,4EAA4E,GAC5E,OAAO,SAASe,cAAcC,OAA2B;IACvD,IAAIA,QAAQpC,MAAM,GAAGL,gCAAgC,OAAO;IAC5D,MAAM0C,SAAS,CAACC;QACd,MAAMC,SAASH,QACZI,GAAG,CAAC,CAACC;gBAAkBA;mBAAPC,QAAOD,gBAAAA,MAAM,CAACH,MAAM,YAAbG,gBAAiB;WACxCjC,MAAM,CAAC,CAACmC,QAAUD,OAAOE,QAAQ,CAACD,UAAUA,SAAS,GACrDX,IAAI,CAAC,CAACC,GAAGC,IAAMD,IAAIC;QACtB,IAAI,CAACK,OAAOvC,MAAM,EAAE,OAAO;QAC3B,MAAM6C,SAASC,KAAKC,KAAK,CAACR,OAAOvC,MAAM,GAAG;QAC1C,OAAOuC,OAAOvC,MAAM,GAAG,IACnBuC,MAAM,CAACM,OAAO,GACdC,KAAKE,KAAK,CAAC,AAACT,CAAAA,MAAM,CAACM,SAAS,EAAE,GAAGN,MAAM,CAACM,OAAO,AAAD,IAAK;IACzD;IACA,OAAO;QACLtD,aAAa8C,OAAO;QACpB7C,cAAc6C,OAAO;QACrB5C,iBAAiB4C,OAAO;QACxB3C,kBAAkB2C,OAAO;IAC3B;AACF;AAsBA,6EAA6E,GAC7E,OAAO,SAASY,eACdjC,IAAgB,EAChBlB,MAAqB,EACrBoD,QAAwB,EACxBhC,QAA2B;;IAE3B,MAAMC,QAAQzC,eAAesC,MAAME;IACnC,MAAMM,OAAOT,qBAAqBC,MAAM,MAAMlB,QAAQoB;IACtD,IAAI,CAACC,SAAS,CAACK,MAAM,OAAO;IAC5B,MAAM2B,UAAUD,mBAAAA,WAAY5D,qBAAqB,CAAC0B,KAAK;IACvD,qEAAqE;IACrE,yEAAyE;IACzE,uEAAuE;IACvE,uEAAuE;IACvE,MAAMoC,WAAW3E,oBAAoB0E,SAAS3B,KAAKD,KAAK;IACxD,MAAM8B,SAAS,CAACxD;QACd,MAAMyD,OAAO7E,oBAAoB0E,SAAStD,MAAMK,EAAE;QAClD,OAAO;YACLA,IAAIL,MAAMK,EAAE;YACZqD,OAAO1D,MAAM0D,KAAK;YAClB7C,MAAMb,MAAMa,IAAI;YAChB8C,mBAAmBpF,qBAAqBkF;YACxCG,YAAYL,WAAW,IAAIN,KAAKE,KAAK,CAAC,AAACM,OAAOF,WAAY,MAAM,KAAK;QACvE;IACF;IACA,MAAMM,YAAYnF,eAAeiD,KAAKD,KAAK;IAC3C,OAAO;QACLC,MAAM;YACJtB,IAAIvB;YACJ4E,OAAOG,YAAY,CAAC,MAAM,EAAEA,UAAUH,KAAK,CAAC,CAAC,CAAC,GAAG;YACjD7C,IAAI,UAAEgD,6BAAAA,UAAWhD,IAAI,mBAAIpC,aAAa,CAAC0C,KAAK;YAC5CwC,mBAAmBpF,qBAAqBgF;YACxCK,YAAY;YACZlC,OAAOC,KAAKD,KAAK;QACnB;QACAoC,SAASvD,mBAAmBe,MAAMV,QAAQ,CAACP,EAAE,EAAEJ,QAC5C0C,GAAG,CAACa,QACJrB,IAAI,CAAC,CAACC,GAAGC,IAAMD,EAAEuB,iBAAiB,GAAGtB,EAAEsB,iBAAiB,IAAIvB,EAAEsB,KAAK,CAACK,aAAa,CAAC1B,EAAEqB,KAAK;QAC5FL,UAAUA,aAAa;IACzB;AACF"}
+{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/model-choice.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 { assistCreditsFromUsd } from '../usage/assist-credits'\nimport type { OrgPlan } from '@aglyn/aglyn/foundation/definitions/org-billing.types'\nimport {\n  AI_MODEL_CATALOG,\n  AI_STEP_TIERS,\n  aiCatalogEntry,\n  aiBilledRatesForModel,\n  estimateAiBilledUsd,\n  type AiCatalogEntry,\n  type AiStepKind,\n} from './catalog'\nimport type { AiUsage } from './contract'\nimport { aiServingModelId, resolveAiRoute, type AiPluginSettings } from './routing'\n\n/**\n * THE MODEL SWITCH (AGL-2942): what a person may pick, what \"Auto\" means,\n * and what a pick costs.\n *\n * ## Auto is the routing table\n *\n * A request that names no model — or names one it may not have — runs on\n * `resolveAiRoute(kind)`, the table that maps a step kind to a catalog model\n * (AGL-2937). Picking a model is an override of that one row for one request,\n * never a second table.\n *\n * ## Three bounds, one intersection\n *\n * A manual pick must be served by the active provider, in a tier the org's\n * plan offers (`AI_PLAN_MODEL_TIERS`), on the org's own restriction when it\n * set one, and on the allotment allowlists that apply to the request. The\n * two allowlists bound Auto too: when the routing table's answer is not on\n * them, Auto runs on the allowed model of the same tier, else the cheapest\n * allowed one. The plan bounds only a pick — Auto is the platform's choice,\n * made for the step, and a plan never makes it worse.\n *\n * An allowlist that leaves nothing this deployment serves is not a way to\n * switch AI off by mistake: the choice falls back to the routing table, and\n * the permission keys are how AI is switched off on purpose.\n *\n * ## The price beside each option\n *\n * Credits per typical request at each model's list rates, over the step's\n * measured median exchange when the workspace has enough of them, and over\n * `AI_STEP_NOMINAL_USAGE` until it does — with the multiple of what Auto\n * would cost for the same request. Every figure is credits, never dollars.\n */\n\n/** The value a request sends for \"let the platform route it\". */\nexport const AI_MODEL_AUTO = 'auto'\n\n/**\n * The catalog tiers a MANUAL pick may name, per plan. Free runs on Auto only:\n * its taste is a wall, and a pick there would spend the account's credits at\n * a rate nobody chose on purpose. The dearest tier opens from Scale up, where\n * the band is large enough that a slower, deeper model is a real choice\n * rather than a fast way to end the month.\n */\nexport const AI_PLAN_MODEL_TIERS: Record<OrgPlan, readonly AiCatalogEntry['tier'][]> = {\n  free: [],\n  starter: ['fast', 'balanced'],\n  pro: ['fast', 'balanced'],\n  business: ['fast', 'balanced'],\n  scale: ['fast', 'balanced', 'deep'],\n  advanced: ['fast', 'balanced', 'deep'],\n  agency: ['fast', 'balanced', 'deep'],\n  enterprise: ['fast', 'balanced', 'deep'],\n}\n\n/** The plans whose admins may restrict the model set org-wide. */\nexport const AI_MODEL_RESTRICTION_PLANS: readonly OrgPlan[] = ['agency', 'enterprise']\n\n/**\n * A typical exchange per step kind, before a workspace has measured one: the\n * doors' own prompt sizes and output ceilings, read conservatively. Replaced\n * by the workspace's measured median as soon as `aiMedianUsage` has enough\n * samples, and only ever used for the price beside an option.\n */\nexport const AI_STEP_NOMINAL_USAGE: Record<AiStepKind, AiUsage> = {\n  'assist.chat': { inputTokens: 900, outputTokens: 450, cacheReadTokens: 1_050, cacheWriteTokens: 0 },\n  // One element's rewrite on the fast tier, counted as Claude Haiku 5.5 counts\n  // it: about 30% more tokens than Haiku 4.5 for the same text (AGL-3781).\n  'copy.element': { inputTokens: 455, outputTokens: 156, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  'copy.section': { inputTokens: 700, outputTokens: 1_400, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  'copy.blog': { inputTokens: 600, outputTokens: 1_100, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  'generate.section': { inputTokens: 700, outputTokens: 1_400, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The text step's rules are about 500 characters — far under the balanced\n  // tier's cacheable minimum, so they are billed as input every time rather\n  // than read from a cache (`runtime/ai-prompt-cache.spec.ts` measures it).\n  'job.text': { inputTokens: 450, outputTokens: 500, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The theme step: the tool's schema and both system blocks (about 7,500\n  // characters) are the cached prefix, the site's inventory and the brief ride\n  // uncached, and a full-palette answer is about 1,300 characters of JSON with\n  // as much again to think in.\n  'job.theme': { inputTokens: 500, outputTokens: 850, cacheReadTokens: 2_500, cacheWriteTokens: 0 },\n  // The plan step: the tool's schema, the doctrine and the plan instructions\n  // (about 10,000 characters) are the cached prefix, the site inventory and the\n  // brief ride uncached, and a plan answer is about 600 characters of JSON with\n  // as much again to think in.\n  'job.plan': { inputTokens: 500, outputTokens: 400, cacheReadTokens: 3_600, cacheWriteTokens: 0 },\n  // The component step: the doctrine, the component instructions, the\n  // component surface's catalog and the tool are the cached prefix, the site\n  // inventory, the brief and the confirmed plan ride uncached, and a card-sized\n  // answer — its tree and declared props as JSON — is about 1,700 characters.\n  // The component step's spec measures the prefix and the answer against this.\n  'job.component': { inputTokens: 700, outputTokens: 600, cacheReadTokens: 5_800, cacheWriteTokens: 0 },\n  // The CRM step's exchange (AGL-2917), counted as the fast tier's Claude\n  // Haiku 5.5 counts it, about 30% more tokens than Haiku 4.5 (AGL-3781). The\n  // doctrine's field rules, the CRM rules and the tool come to 650 to 913\n  // tokens, as `runtime/ai-prompt-cache.spec.ts` measures them, which caches on\n  // that model; one question about a record writes it, so it is counted as\n  // input. A record's facts ride with them, a timeline cut to its newest twelve\n  // entries, and the typical answer is a summary and a next step of a few\n  // sentences, with no thinking.\n  'job.crm': { inputTokens: 2_340, outputTokens: 325, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // A layout: a tree layout's 1,000 output tokens until AGL-3660, now a\n  // site's header and footer in the layout language — 442 to 583 output\n  // tokens live on 2026-10-07 — with up to 1,500 more for the adaptive\n  // thinking it is asked with, an estimate until a workspace measures it.\n  'job.layout': { inputTokens: 700, outputTokens: 2_000, cacheReadTokens: 5_600, cacheWriteTokens: 0 },\n  'job.template': { inputTokens: 1_200, outputTokens: 700, cacheReadTokens: 6_400, cacheWriteTokens: 0 },\n  // One page section's exchange (AGL-2907), from the modules' own text: the\n  // doctrine (6,281 characters), the page instructions (1,524), the screen\n  // palette catalog (8,353) and the section tool (484) are the cached prefix;\n  // the site inventory (711 on the median golden site), the brief, the plan\n  // and the section line ride uncached; and a golden section answer is 901\n  // characters at the median, asked for with no thinking.\n  // Since AGL-3660 a site's and a build's pages are one answer a page in\n  // the layout language: 683 to 1,745 output tokens live on 2026-10-07, about\n  // 1,150 at the median, with up to 1,500 more for the adaptive thinking it\n  // is asked with — an estimate until a workspace measures it, read high so\n  // the quote beside a model never undersells the exchange.\n  'job.page': { inputTokens: 600, outputTokens: 2_650, cacheReadTokens: 4_400, cacheWriteTokens: 0 },\n  // One product's copy (AGL-2916), the exchange a bulk job makes once a\n  // product: the doctrine, the copy rules and the tool (about 9,500 characters)\n  // are the cached prefix; the product's facts, the site's categories and a\n  // photo at most 768 pixels on its long edge (about 790 tokens) ride\n  // uncached; and the answer is a description with its search listing, tags,\n  // categories and option names, about 1,300 characters of JSON.\n  'job.products': { inputTokens: 1_500, outputTokens: 450, cacheReadTokens: 2_400, cacheWriteTokens: 0 },\n  // The SEO step's listing exchange, which a page's or a product's listing is\n  // exactly one of: the median of 146 production listings of October 8–14\n  // (2,255 input and 213 output tokens on Claude Haiku 4.5), counted as the\n  // fast tier's Claude Haiku 5.5 counts them, about 30% more (AGL-3781). The\n  // doctrine's block in its field scope, the listing rules and the tool's\n  // schema are its prefix, which caches on that model and which a site's next\n  // listing reads; a lone listing writes it, so it is counted as input. The\n  // page text, capped at 2,000 characters, rides with the current listing and a\n  // dozen other titles. The largest answer the tool accepts is about 600\n  // characters of JSON, and the request asks for no thinking. An audit pass is\n  // heavier: eight pages' findings and text, about 9,000 characters, with up to\n  // about 8,000 back.\n  'job.seo': { inputTokens: 2_930, outputTokens: 277, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The experiment step (AGL-2914). NOTHING is cached: the A/B test rules\n  // and the acceptable-use block come to about 2,300 characters, under the\n  // balanced tier's cacheable minimum, so they are billed as input every\n  // time. The copy under test, or a test's arms with its verdict, rides with\n  // them, and a four-variant answer is about 800 characters of JSON.\n  'job.experiment': { inputTokens: 900, outputTokens: 300, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The form step: the doctrine, the form instructions, the form surface's\n  // catalog (about 150 tokens) and the tool's schema are the cached prefix, the\n  // site inventory and the brief ride uncached, and a form's answer is 200 to\n  // 400 tokens of JSON with no extended thinking.\n  'job.form': { inputTokens: 700, outputTokens: 450, cacheReadTokens: 3_900, cacheWriteTokens: 0 },\n  // An insight (AGL-2915): two exchanges, and NOTHING cached — the rules and\n  // the acceptable-use block are under the balanced tier's minimum, so they\n  // are billed as input on both. The read call carries the question and the\n  // readers' catalog (about 2,000 characters) and answers a few reads; the\n  // answer call carries four tables of about ten rows each (about 4,000\n  // characters) and answers three to five insights as JSON.\n  'job.insight': { inputTokens: 2_800, outputTokens: 650, cacheReadTokens: 0, cacheWriteTokens: 0 },\n  // The email step, and the campaign step that shares its generation: the\n  // doctrine, the email instructions, the email palette catalog and the\n  // tool's schema (about 10,300 characters) are the cached prefix, the site\n  // inventory and the brief ride uncached, and the answer is one email's\n  // node map with three subject lines and three preheaders.\n  'job.email': { inputTokens: 600, outputTokens: 1_500, cacheReadTokens: 3_600, cacheWriteTokens: 0 },\n  'job.campaign': { inputTokens: 600, outputTokens: 1_500, cacheReadTokens: 3_600, cacheWriteTokens: 0 },\n  // The workflow step's draft (AGL-2919): the doctrine, the automation\n  // vocabulary and the tool's schema are the cached prefix (4,564 tokens, as\n  // `runtime/ai-prompt-cache.spec.ts` measures it); what the workspace can\n  // run, the site's forms and datasets and the description ride uncached; and\n  // an automation of a few steps is about 1,300 characters of JSON with as\n  // much again to think in. An explanation caches half as much.\n  'job.workflow': { inputTokens: 500, outputTokens: 1_000, cacheReadTokens: 4_500, cacheWriteTokens: 0 },\n  // The logic step (AGL-3603): the fields doctrine, one rules block and the\n  // tool's schema are the cached prefix (1,556 tokens for a function, as\n  // `runtime/ai-prompt-cache.spec.ts` measures it); the site's variables and\n  // the request ride uncached, and a function of a few operations is about\n  // 1,200 characters of JSON with as much again to think in.\n  'job.logic': { inputTokens: 700, outputTokens: 900, cacheReadTokens: 1_500, cacheWriteTokens: 0 },\n  // The edit step (AGL-3616): the document doctrine, the job's rules and the\n  // edit rung's protocol and catalog for the document's kind are the cached\n  // prefix (5,970 tokens for a page, as `runtime/ai-prompt-cache.spec.ts`\n  // measures it); the outline of up to sixty elements with their settings\n  // and the request ride uncached, and a change of a few operations is about\n  // 1,500 characters of tool call with as much again to think in.\n  'job.edit': { inputTokens: 2_500, outputTokens: 1_000, cacheReadTokens: 6_000, cacheWriteTokens: 0 },\n}\n\n/** The fewest measured exchanges a median is taken over. */\nexport const AI_MEASURED_MEDIAN_MIN_SAMPLES = 3\n\nexport interface AiModelBounds {\n  plan: OrgPlan\n  /** The allotment allowlists that apply, intersected; `null` for none. */\n  allotmentModels?: readonly string[] | null\n  /** The org-wide restriction; `null` for none. */\n  orgModels?: readonly string[] | null\n}\n\n/** Whether a list names a model, a retired id on it naming its successor (AGL-3781). */\nfunction names(list: readonly string[], id: string): boolean {\n  return list.some((listed) => aiServingModelId(listed) === id)\n}\n\nfunction onLists(entry: AiCatalogEntry, bounds: AiModelBounds): boolean {\n  if (bounds.orgModels?.length && !names(bounds.orgModels, entry.id)) return false\n  if (bounds.allotmentModels && !names(bounds.allotmentModels, entry.id)) return false\n  return true\n}\n\n/** The current catalog models a manual pick may name on a provider under these bounds. */\nexport function aiModelsSelectable(\n  providerId: string,\n  bounds: AiModelBounds,\n): AiCatalogEntry[] {\n  const tiers = AI_PLAN_MODEL_TIERS[bounds.plan] ?? []\n  return AI_MODEL_CATALOG.filter(\n    (entry) =>\n      entry.provider === providerId &&\n      !entry.successor &&\n      tiers.includes(entry.tier) &&\n      onLists(entry, bounds),\n  )\n}\n\n/**\n * A model's blended BILLED rate, for \"the cheapest allowed\" and nothing\n * else.\n *\n * Billed and not provider (AGL-3015): the substitution below happens when a\n * workspace asked for a model its bounds do not allow, and the party it\n * falls back for is the one whose credits are drawn. \"Cheapest\" has to mean\n * the same thing here as in the selector this substitution stands in for,\n * which sorts on `creditsPerRequest`.\n */\nconst blendedRate = (id: string): number => {\n  const rates = aiBilledRatesForModel(id)\n  return rates.inputPerToken + rates.outputPerToken\n}\n\nexport interface AiModelChoice {\n  /** The catalog id the request runs, and is metered, on. */\n  model: string\n  /** True when the routing table decided: Auto, or a pick that was declined. */\n  auto: boolean\n  /** A pick that was asked for and is not allowed, for the envelope. */\n  declined: string | null\n}\n\n/**\n * The model one request runs on, or `null` when no provider is registered —\n * which every door has refused before it asks.\n */\nexport function resolveAiModelChoice(\n  kind: AiStepKind,\n  requested: unknown,\n  bounds: AiModelBounds,\n  settings?: AiPluginSettings,\n): AiModelChoice | null {\n  const route = resolveAiRoute(kind, settings)\n  if (!route) return null\n  // A pick of a retired model, from a job made before it retired, is a pick\n  // of the model that serves in its place.\n  const wanted =\n    typeof requested === 'string' && requested.trim() && requested.trim() !== AI_MODEL_AUTO\n      ? aiServingModelId(requested.trim())\n      : null\n  if (wanted && aiModelsSelectable(route.provider.id, bounds).some((entry) => entry.id === wanted)) {\n    return { model: wanted, auto: false, declined: null }\n  }\n  const restricted = Boolean(bounds.orgModels?.length) || Boolean(bounds.allotmentModels)\n  let model = route.model\n  if (restricted) {\n    const allowed = AI_MODEL_CATALOG.filter(\n      (entry) => entry.provider === route.provider.id && !entry.successor && onLists(entry, bounds),\n    )\n    if (allowed.length && !allowed.some((entry) => entry.id === model)) {\n      const tier = aiCatalogEntry(model)?.tier ?? AI_STEP_TIERS[kind]\n      const sameTier = allowed.find((entry) => entry.tier === tier)\n      const cheapest = [...allowed].sort((a, b) => blendedRate(a.id) - blendedRate(b.id))[0]\n      model = (sameTier ?? cheapest).id\n    }\n  }\n  return { model, auto: true, declined: wanted }\n}\n\n/** The median exchange over measured samples, or `null` below the minimum. */\nexport function aiMedianUsage(samples: readonly AiUsage[]): AiUsage | null {\n  if (samples.length < AI_MEASURED_MEDIAN_MIN_SAMPLES) return null\n  const median = (field: keyof AiUsage): number => {\n    const values = samples\n      .map((sample) => Number(sample[field] ?? 0))\n      .filter((value) => Number.isFinite(value) && value >= 0)\n      .sort((a, b) => a - b)\n    if (!values.length) return 0\n    const middle = Math.floor(values.length / 2)\n    return values.length % 2\n      ? values[middle]\n      : Math.round((values[middle - 1] + values[middle]) / 2)\n  }\n  return {\n    inputTokens: median('inputTokens'),\n    outputTokens: median('outputTokens'),\n    cacheReadTokens: median('cacheReadTokens'),\n    cacheWriteTokens: median('cacheWriteTokens'),\n  }\n}\n\nexport interface AiModelOption {\n  /** A catalog id, or `AI_MODEL_AUTO`. */\n  id: string\n  label: string\n  tier: AiCatalogEntry['tier']\n  /** Credits a typical request of this kind costs on the option. */\n  creditsPerRequest: number\n  /** That cost as a multiple of Auto's, to one decimal. */\n  multiplier: number\n}\n\nexport interface AiModelOptions {\n  /** Auto, labeled with the model it routes this kind to under these bounds. */\n  auto: AiModelOption & { model: string }\n  /** The manual picks, cheapest first; empty where the plan offers none. */\n  options: AiModelOption[]\n  /** Whether the price came from the workspace's own measured exchanges. */\n  measured: boolean\n}\n\n/** What the selector lists for one step kind, or `null` without a provider. */\nexport function aiModelOptions(\n  kind: AiStepKind,\n  bounds: AiModelBounds,\n  measured: AiUsage | null,\n  settings?: AiPluginSettings,\n): AiModelOptions | null {\n  const route = resolveAiRoute(kind, settings)\n  const auto = resolveAiModelChoice(kind, null, bounds, settings)\n  if (!route || !auto) return null\n  const typical = measured ?? AI_STEP_NOMINAL_USAGE[kind]\n  // Billed rates throughout (AGL-3015): every figure below is what the\n  // workspace will be charged for picking this model — credits per request\n  // and the multiple of Auto those credits come to. Our own cost for the\n  // same request is not this selector's subject and would misstate both.\n  const autoCost = estimateAiBilledUsd(typical, auto.model)\n  const priced = (entry: AiCatalogEntry): AiModelOption => {\n    const cost = estimateAiBilledUsd(typical, entry.id)\n    return {\n      id: entry.id,\n      label: entry.label,\n      tier: entry.tier,\n      creditsPerRequest: assistCreditsFromUsd(cost),\n      multiplier: autoCost > 0 ? Math.round((cost / autoCost) * 10) / 10 : 1,\n    }\n  }\n  const autoEntry = aiCatalogEntry(auto.model)\n  return {\n    auto: {\n      id: AI_MODEL_AUTO,\n      label: autoEntry ? `Auto (${autoEntry.label})` : 'Auto',\n      tier: autoEntry?.tier ?? AI_STEP_TIERS[kind],\n      creditsPerRequest: assistCreditsFromUsd(autoCost),\n      multiplier: 1,\n      model: auto.model,\n    },\n    options: aiModelsSelectable(route.provider.id, bounds)\n      .map(priced)\n      .sort((a, b) => a.creditsPerRequest - b.creditsPerRequest || a.label.localeCompare(b.label)),\n    measured: measured !== null,\n  }\n}\n"],"names":["assistCreditsFromUsd","AI_MODEL_CATALOG","AI_STEP_TIERS","aiCatalogEntry","aiBilledRatesForModel","estimateAiBilledUsd","aiServingModelId","resolveAiRoute","AI_MODEL_AUTO","AI_PLAN_MODEL_TIERS","free","starter","pro","business","scale","advanced","agency","enterprise","AI_MODEL_RESTRICTION_PLANS","AI_STEP_NOMINAL_USAGE","inputTokens","outputTokens","cacheReadTokens","cacheWriteTokens","AI_MEASURED_MEDIAN_MIN_SAMPLES","names","list","id","some","listed","onLists","entry","bounds","orgModels","length","allotmentModels","aiModelsSelectable","providerId","tiers","plan","filter","provider","successor","includes","tier","blendedRate","rates","inputPerToken","outputPerToken","resolveAiModelChoice","kind","requested","settings","route","wanted","trim","model","auto","declined","restricted","Boolean","allowed","sameTier","find","cheapest","sort","a","b","aiMedianUsage","samples","median","field","values","map","sample","Number","value","isFinite","middle","Math","floor","round","aiModelOptions","measured","typical","autoCost","priced","cost","label","creditsPerRequest","multiplier","autoEntry","options","localeCompare"],"mappings":"AAAA;;;;;;;;;;;;;;;CAeC,GAED,SAASA,oBAAoB,QAAQ,6BAAyB;AAE9D,SACEC,gBAAgB,EAChBC,aAAa,EACbC,cAAc,EACdC,qBAAqB,EACrBC,mBAAmB,QAGd,eAAW;AAElB,SAASC,gBAAgB,EAAEC,cAAc,QAA+B,eAAW;AAEnF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA+BC,GAED,+DAA+D,GAC/D,OAAO,MAAMC,gBAAgB,OAAM;AAEnC;;;;;;CAMC,GACD,OAAO,MAAMC,sBAA0E;IACrFC,MAAM,EAAE;IACRC,SAAS;QAAC;QAAQ;KAAW;IAC7BC,KAAK;QAAC;QAAQ;KAAW;IACzBC,UAAU;QAAC;QAAQ;KAAW;IAC9BC,OAAO;QAAC;QAAQ;QAAY;KAAO;IACnCC,UAAU;QAAC;QAAQ;QAAY;KAAO;IACtCC,QAAQ;QAAC;QAAQ;QAAY;KAAO;IACpCC,YAAY;QAAC;QAAQ;QAAY;KAAO;AAC1C,EAAC;AAED,gEAAgE,GAChE,OAAO,MAAMC,6BAAiD;IAAC;IAAU;CAAa,CAAA;AAEtF;;;;;CAKC,GACD,OAAO,MAAMC,wBAAqD;IAChE,eAAe;QAAEC,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAClG,6EAA6E;IAC7E,yEAAyE;IACzE,gBAAgB;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC/F,gBAAgB;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAGC,kBAAkB;IAAE;IACjG,aAAa;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC9F,oBAAoB;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAGC,kBAAkB;IAAE;IACrG,0EAA0E;IAC1E,0EAA0E;IAC1E,0EAA0E;IAC1E,YAAY;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC3F,wEAAwE;IACxE,6EAA6E;IAC7E,6EAA6E;IAC7E,6BAA6B;IAC7B,aAAa;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAChG,2EAA2E;IAC3E,8EAA8E;IAC9E,8EAA8E;IAC9E,6BAA6B;IAC7B,YAAY;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAC/F,oEAAoE;IACpE,2EAA2E;IAC3E,8EAA8E;IAC9E,4EAA4E;IAC5E,6EAA6E;IAC7E,iBAAiB;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IACpG,wEAAwE;IACxE,4EAA4E;IAC5E,wEAAwE;IACxE,8EAA8E;IAC9E,yEAAyE;IACzE,8EAA8E;IAC9E,wEAAwE;IACxE,+BAA+B;IAC/B,WAAW;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC5F,sEAAsE;IACtE,sEAAsE;IACtE,qEAAqE;IACrE,wEAAwE;IACxE,cAAc;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IACnG,gBAAgB;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IACrG,0EAA0E;IAC1E,yEAAyE;IACzE,4EAA4E;IAC5E,0EAA0E;IAC1E,yEAAyE;IACzE,wDAAwD;IACxD,uEAAuE;IACvE,4EAA4E;IAC5E,0EAA0E;IAC1E,0EAA0E;IAC1E,0DAA0D;IAC1D,YAAY;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IACjG,sEAAsE;IACtE,8EAA8E;IAC9E,0EAA0E;IAC1E,oEAAoE;IACpE,2EAA2E;IAC3E,+DAA+D;IAC/D,gBAAgB;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IACrG,4EAA4E;IAC5E,wEAAwE;IACxE,0EAA0E;IAC1E,2EAA2E;IAC3E,wEAAwE;IACxE,4EAA4E;IAC5E,0EAA0E;IAC1E,8EAA8E;IAC9E,uEAAuE;IACvE,6EAA6E;IAC7E,8EAA8E;IAC9E,oBAAoB;IACpB,WAAW;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAC5F,wEAAwE;IACxE,yEAAyE;IACzE,uEAAuE;IACvE,2EAA2E;IAC3E,mEAAmE;IACnE,kBAAkB;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IACjG,yEAAyE;IACzE,8EAA8E;IAC9E,4EAA4E;IAC5E,gDAAgD;IAChD,YAAY;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAC/F,2EAA2E;IAC3E,0EAA0E;IAC1E,0EAA0E;IAC1E,yEAAyE;IACzE,sEAAsE;IACtE,0DAA0D;IAC1D,eAAe;QAAEH,aAAa;QAAOC,cAAc;QAAKC,iBAAiB;QAAGC,kBAAkB;IAAE;IAChG,wEAAwE;IACxE,sEAAsE;IACtE,0EAA0E;IAC1E,uEAAuE;IACvE,0DAA0D;IAC1D,aAAa;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IAClG,gBAAgB;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IACrG,qEAAqE;IACrE,2EAA2E;IAC3E,yEAAyE;IACzE,4EAA4E;IAC5E,yEAAyE;IACzE,8DAA8D;IAC9D,gBAAgB;QAAEH,aAAa;QAAKC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;IACrG,0EAA0E;IAC1E,uEAAuE;IACvE,2EAA2E;IAC3E,yEAAyE;IACzE,2DAA2D;IAC3D,aAAa;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAChG,2EAA2E;IAC3E,0EAA0E;IAC1E,wEAAwE;IACxE,wEAAwE;IACxE,2EAA2E;IAC3E,gEAAgE;IAChE,YAAY;QAAEH,aAAa;QAAOC,cAAc;QAAOC,iBAAiB;QAAOC,kBAAkB;IAAE;AACrG,EAAC;AAED,0DAA0D,GAC1D,OAAO,MAAMC,iCAAiC,EAAC;AAU/C,sFAAsF,GACtF,SAASC,MAAMC,IAAuB,EAAEC,EAAU;IAChD,OAAOD,KAAKE,IAAI,CAAC,CAACC,SAAWvB,iBAAiBuB,YAAYF;AAC5D;AAEA,SAASG,QAAQC,KAAqB,EAAEC,MAAqB;QACvDA;IAAJ,IAAIA,EAAAA,oBAAAA,OAAOC,SAAS,qBAAhBD,kBAAkBE,MAAM,KAAI,CAACT,MAAMO,OAAOC,SAAS,EAAEF,MAAMJ,EAAE,GAAG,OAAO;IAC3E,IAAIK,OAAOG,eAAe,IAAI,CAACV,MAAMO,OAAOG,eAAe,EAAEJ,MAAMJ,EAAE,GAAG,OAAO;IAC/E,OAAO;AACT;AAEA,wFAAwF,GACxF,OAAO,SAASS,mBACdC,UAAkB,EAClBL,MAAqB;QAEPvB;IAAd,MAAM6B,SAAQ7B,mCAAAA,mBAAmB,CAACuB,OAAOO,IAAI,CAAC,YAAhC9B,mCAAoC,EAAE;IACpD,OAAOR,iBAAiBuC,MAAM,CAC5B,CAACT,QACCA,MAAMU,QAAQ,KAAKJ,cACnB,CAACN,MAAMW,SAAS,IAChBJ,MAAMK,QAAQ,CAACZ,MAAMa,IAAI,KACzBd,QAAQC,OAAOC;AAErB;AAEA;;;;;;;;;CASC,GACD,MAAMa,cAAc,CAAClB;IACnB,MAAMmB,QAAQ1C,sBAAsBuB;IACpC,OAAOmB,MAAMC,aAAa,GAAGD,MAAME,cAAc;AACnD;AAWA;;;CAGC,GACD,OAAO,SAASC,qBACdC,IAAgB,EAChBC,SAAkB,EAClBnB,MAAqB,EACrBoB,QAA2B;QAaApB;IAX3B,MAAMqB,QAAQ9C,eAAe2C,MAAME;IACnC,IAAI,CAACC,OAAO,OAAO;IACnB,0EAA0E;IAC1E,yCAAyC;IACzC,MAAMC,SACJ,OAAOH,cAAc,YAAYA,UAAUI,IAAI,MAAMJ,UAAUI,IAAI,OAAO/C,gBACtEF,iBAAiB6C,UAAUI,IAAI,MAC/B;IACN,IAAID,UAAUlB,mBAAmBiB,MAAMZ,QAAQ,CAACd,EAAE,EAAEK,QAAQJ,IAAI,CAAC,CAACG,QAAUA,MAAMJ,EAAE,KAAK2B,SAAS;QAChG,OAAO;YAAEE,OAAOF;YAAQG,MAAM;YAAOC,UAAU;QAAK;IACtD;IACA,MAAMC,aAAaC,SAAQ5B,oBAAAA,OAAOC,SAAS,qBAAhBD,kBAAkBE,MAAM,KAAK0B,QAAQ5B,OAAOG,eAAe;IACtF,IAAIqB,QAAQH,MAAMG,KAAK;IACvB,IAAIG,YAAY;QACd,MAAME,UAAU5D,iBAAiBuC,MAAM,CACrC,CAACT,QAAUA,MAAMU,QAAQ,KAAKY,MAAMZ,QAAQ,CAACd,EAAE,IAAI,CAACI,MAAMW,SAAS,IAAIZ,QAAQC,OAAOC;QAExF,IAAI6B,QAAQ3B,MAAM,IAAI,CAAC2B,QAAQjC,IAAI,CAAC,CAACG,QAAUA,MAAMJ,EAAE,KAAK6B,QAAQ;;gBACrDrD;YAAb,MAAMyC,gBAAOzC,kBAAAA,eAAeqD,2BAAfrD,gBAAuByC,IAAI,mBAAI1C,aAAa,CAACgD,KAAK;YAC/D,MAAMY,WAAWD,QAAQE,IAAI,CAAC,CAAChC,QAAUA,MAAMa,IAAI,KAAKA;YACxD,MAAMoB,WAAW;mBAAIH;aAAQ,CAACI,IAAI,CAAC,CAACC,GAAGC,IAAMtB,YAAYqB,EAAEvC,EAAE,IAAIkB,YAAYsB,EAAExC,EAAE,EAAE,CAAC,EAAE;YACtF6B,QAAQ,CAACM,mBAAAA,WAAYE,UAAUrC,EAAE;QACnC;IACF;IACA,OAAO;QAAE6B;QAAOC,MAAM;QAAMC,UAAUJ;IAAO;AAC/C;AAEA,4EAA4E,GAC5E,OAAO,SAASc,cAAcC,OAA2B;IACvD,IAAIA,QAAQnC,MAAM,GAAGV,gCAAgC,OAAO;IAC5D,MAAM8C,SAAS,CAACC;QACd,MAAMC,SAASH,QACZI,GAAG,CAAC,CAACC;gBAAkBA;mBAAPC,QAAOD,gBAAAA,MAAM,CAACH,MAAM,YAAbG,gBAAiB;WACxClC,MAAM,CAAC,CAACoC,QAAUD,OAAOE,QAAQ,CAACD,UAAUA,SAAS,GACrDX,IAAI,CAAC,CAACC,GAAGC,IAAMD,IAAIC;QACtB,IAAI,CAACK,OAAOtC,MAAM,EAAE,OAAO;QAC3B,MAAM4C,SAASC,KAAKC,KAAK,CAACR,OAAOtC,MAAM,GAAG;QAC1C,OAAOsC,OAAOtC,MAAM,GAAG,IACnBsC,MAAM,CAACM,OAAO,GACdC,KAAKE,KAAK,CAAC,AAACT,CAAAA,MAAM,CAACM,SAAS,EAAE,GAAGN,MAAM,CAACM,OAAO,AAAD,IAAK;IACzD;IACA,OAAO;QACL1D,aAAakD,OAAO;QACpBjD,cAAciD,OAAO;QACrBhD,iBAAiBgD,OAAO;QACxB/C,kBAAkB+C,OAAO;IAC3B;AACF;AAsBA,6EAA6E,GAC7E,OAAO,SAASY,eACdhC,IAAgB,EAChBlB,MAAqB,EACrBmD,QAAwB,EACxB/B,QAA2B;;IAE3B,MAAMC,QAAQ9C,eAAe2C,MAAME;IACnC,MAAMK,OAAOR,qBAAqBC,MAAM,MAAMlB,QAAQoB;IACtD,IAAI,CAACC,SAAS,CAACI,MAAM,OAAO;IAC5B,MAAM2B,UAAUD,mBAAAA,WAAYhE,qBAAqB,CAAC+B,KAAK;IACvD,qEAAqE;IACrE,yEAAyE;IACzE,uEAAuE;IACvE,uEAAuE;IACvE,MAAMmC,WAAWhF,oBAAoB+E,SAAS3B,KAAKD,KAAK;IACxD,MAAM8B,SAAS,CAACvD;QACd,MAAMwD,OAAOlF,oBAAoB+E,SAASrD,MAAMJ,EAAE;QAClD,OAAO;YACLA,IAAII,MAAMJ,EAAE;YACZ6D,OAAOzD,MAAMyD,KAAK;YAClB5C,MAAMb,MAAMa,IAAI;YAChB6C,mBAAmBzF,qBAAqBuF;YACxCG,YAAYL,WAAW,IAAIN,KAAKE,KAAK,CAAC,AAACM,OAAOF,WAAY,MAAM,KAAK;QACvE;IACF;IACA,MAAMM,YAAYxF,eAAesD,KAAKD,KAAK;IAC3C,OAAO;QACLC,MAAM;YACJ9B,IAAInB;YACJgF,OAAOG,YAAY,CAAC,MAAM,EAAEA,UAAUH,KAAK,CAAC,CAAC,CAAC,GAAG;YACjD5C,IAAI,UAAE+C,6BAAAA,UAAW/C,IAAI,mBAAI1C,aAAa,CAACgD,KAAK;YAC5CuC,mBAAmBzF,qBAAqBqF;YACxCK,YAAY;YACZlC,OAAOC,KAAKD,KAAK;QACnB;QACAoC,SAASxD,mBAAmBiB,MAAMZ,QAAQ,CAACd,EAAE,EAAEK,QAC5CyC,GAAG,CAACa,QACJrB,IAAI,CAAC,CAACC,GAAGC,IAAMD,EAAEuB,iBAAiB,GAAGtB,EAAEsB,iBAAiB,IAAIvB,EAAEsB,KAAK,CAACK,aAAa,CAAC1B,EAAEqB,KAAK;QAC5FL,UAAUA,aAAa;IACzB;AACF"}

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

@@ -37,6 +37,12 @@ export declare const AI_DEFAULT_MODEL_ENV = "AI_DEFAULT_MODEL";
 export declare const ASSIST_MODEL_ENV = "ASSIST_MODEL";
 /** The org's resolved `pluginSettings/ai` values, as `getPluginConfig` merges them. */
 export type AiPluginSettings = Record<string, unknown>;
+/**
+ * The model that serves a catalog id (AGL-3781): the id itself, or for a
+ * retired row the successor it names, followed to a current row. An id the
+ * catalog does not know is returned as it is, for the caller to refuse.
+ */
+export declare function aiServingModelId(modelId: string): string;
 export interface AiRoute {
     provider: AiProvider;
     model: string;
@@ -56,7 +62,11 @@ export declare function resolveAiRoute(kind: AiStepKind, settings?: AiPluginSett
  * briefs (AGL-2937).
  */
 export interface AiRoutingRow {
-    /** `null` sends no setting: the model's own default, and all a fast-tier model accepts. */
+    /**
+     * `null` sends no setting: the model's own default, which on every current
+     * model is adaptive thinking. A row whose ceiling holds the answer alone
+     * says `off` (AGL-3781).
+     */
     thinking: AiThinking | null;
     effort: AiEffort | null;
     /** The answer ceiling, thinking included. */
@@ -94,5 +104,14 @@ export interface AiRoutingEval {
  * tokenizer that splits finer than four.
  */
 export declare const AI_ROUTING_TABLE: Readonly<Record<AiStepKind, AiRoutingRow>>;
+/**
+ * What a step kind's row asks of its model, as the fields a request carries:
+ * only those the row names, so a row that names none sends none and the
+ * model's own default stands.
+ */
+export declare function aiRouteSettings(kind: AiStepKind): {
+    thinking?: AiThinking;
+    effort?: AiEffort;
+};
 /** The model id alone — what a door records on the meter and the job. */
 export declare function aiModelForStep(kind: AiStepKind, settings?: AiPluginSettings): string;