@aglyn/plugins-ai 1.0.0-beta.143

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

@@ -0,0 +1,197 @@
+/**
+ * @license
+ * Copyright 2026 Aglyn LLC
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+import type { PluginSubprocessorDeclaration } from '@aglyn/aglyn/plugin-manager/plugin-subprocessors';
+import type { AiModelDescriptor, AiUsage } from './contract';
+/**
+ * The model catalog (AGL-2939): every model id the plugin will route to,
+ * with its provider, its display name, its capabilities and its list rates.
+ * The ONE place a model id is written outside a provider adapter — the
+ * routing table maps a step kind to an id through here, the meter prices
+ * usage by id through here, and a settings form offers ids from here.
+ *
+ * ## Two rates, because one figure was answering two questions
+ *
+ * Every row carries a PROVIDER rate and a BILLED rate (AGL-3015):
+ *
+ * - `providerRates` is what the vendor charges us. It is money leaving the
+ *   company, so cost accounting, the margin surfaces and every staff spend
+ *   meter price usage through this one and no other.
+ * - `billedRates` is what a customer's credits are charged at. The credit
+ *   meter, the band, the overage ladder and the model selector price usage
+ *   through this one, because it is the figure a customer's balance moves by.
+ *
+ * They are equal on nearly every row, and a helper writes the pair from one
+ * figure so an unmarked row cannot drift apart by a typo. A row may be
+ * billed ABOVE its provider rate deliberately — the markup is carried in the
+ * rate rather than in the credit conversion, which keeps a credit a fixed
+ * quantity of billed spend on every model and leaves the band, the ladder
+ * and the invoice on the one arithmetic they were sized with. A row billed
+ * BELOW its provider rate would sell tokens at a loss, and
+ * `billed-rate-is-not-provider-cost.spec.ts` refuses one.
+ *
+ * Both are keyed by model rather than fixed at one model's rates because the
+ * default is an env override: a one-line incident swap to a dearer model
+ * would otherwise keep reporting the cheaper money, and per-org cost would
+ * read as roughly right — the failure mode this whole meter exists to
+ * prevent. An unknown id falls back to the most EXPENSIVE known tier on
+ * purpose, on both rates: an estimate that errs low is worse than one that
+ * errs high, whether the figure is our bill or the customer's.
+ */
+export interface AiTokenRates {
+    inputPerToken: number;
+    outputPerToken: number;
+    cacheReadPerToken: number;
+    cacheWritePerToken: number;
+}
+/** List price per MTok → per-token rates, cache read at 0.1x, write 1.25x. */
+export declare function aiRatesPerMTok(inputPerMTok: number, outputPerMTok: number): AiTokenRates;
+/**
+ * The two rate tables a catalog row carries. Separate fields rather than one
+ * table and a multiplier: a reader that wants our cost has to name it, and
+ * a reader that wants the customer's has to name that, so neither can be
+ * reached by accident.
+ */
+export interface AiCatalogRates {
+    /** What the provider bills us for this model. */
+    providerRates: AiTokenRates;
+    /** What a customer's credits are charged at for this model. */
+    billedRates: AiTokenRates;
+}
+/**
+ * A row billed at exactly the provider's list rate — the usual case, written
+ * once so the pair cannot drift.
+ */
+export declare function aiRatesAtList(inputPerMTok: number, outputPerMTok: number): AiCatalogRates;
+export interface AiCatalogEntry extends AiModelDescriptor, AiCatalogRates {
+    /**
+     * The tier a routing decision reads: `fast` for short, constrained
+     * answers; `balanced` for the assistant and briefs; `deep` for the
+     * dearest, slowest class.
+     */
+    tier: 'fast' | 'balanced' | 'deep';
+    /**
+     * The shortest prefix this model's provider will cache (AGL-2937). A
+     * request whose cached span falls under it is served with the markers
+     * honored and NOTHING cached: no error, no warning, and a usage report
+     * that reads as a permanent cache miss.
+     *
+     * It is a property of the model rather than of the tier or the vendor,
+     * and it does not fall as a generation advances — so the number has to be
+     * written beside the model, and a door that means to cache has to be held
+     * to it (`aiCachedPrefixCaches` in the runtime, and the door table in
+     * `runtime/ai-prompt-cache.spec.ts`). Without it, "this block carries a
+     * breakpoint" is a claim nothing checks, and a prompt can be designed
+     * around a cache it never had.
+     */
+    cacheMinTokens: number;
+}
+/**
+ * The sentinels a docs-only answer is metered under (AGL-2486): not
+ * models, but ids the meter sees, priced at zero so a deflected turn lands
+ * in the same rollup as a served one and a keyless deployment does not
+ * read as the most efficient one on the platform.
+ */
+export declare const AI_METER_SENTINELS: {
+    readonly docsRetrieval: "docs-retrieval";
+    readonly docsLinks: "docs-links";
+    readonly answerCache: "assist-cache";
+};
+/**
+ * A provider's row on the published subprocessor list: every field but the
+ * host, which the registered adapter reports as its `endpointHost`.
+ * `{apiKeyEnv}` anywhere in the wording stands for the adapter's
+ * `apiKeyEnv`, so a credential's name is written in its adapter alone and
+ * never in this module, which the browser bundle carries.
+ */
+export type AiProviderSubprocessorWording = Omit<PluginSubprocessorDeclaration, 'host'>;
+/** A provider the catalog carries models for. */
+export interface AiCatalogProvider {
+    id: string;
+    /** The name a settings form shows. */
+    label: string;
+    /**
+     * The vendor's row on the published subprocessor list, written here and
+     * nowhere else: `aiSubprocessors` derives the plugin's declarations from
+     * it. Absent for a provider whose endpoint an operator names, which has
+     * no fixed recipient to publish.
+     */
+    subprocessor?: AiProviderSubprocessorWording;
+}
+/**
+ * The providers the catalog carries models for, with the name a settings
+ * form shows. A form offers providers from here for the reason it offers
+ * models from here: nothing outside this folder names a vendor. For the
+ * same reason a vendor's subprocessor row is written here too.
+ */
+export declare const AI_CATALOG_PROVIDERS: readonly AiCatalogProvider[];
+export declare const AI_MODEL_CATALOG: readonly AiCatalogEntry[];
+/** The dearest known tier, used when a model id is not in the catalog. */
+export declare const AI_FALLBACK_RATES: AiTokenRates;
+/**
+ * The minimum an unknown model id is assumed to have: the largest one the
+ * catalog knows. It errs the way the rate fallback errs — toward the answer
+ * that costs more — so an unrecognized id reads as "this prompt does not
+ * cache" rather than promising a saving nobody measured.
+ */
+export declare const AI_FALLBACK_CACHE_MIN_TOKENS = 4096;
+/** The shortest prefix a model's provider will cache; the dearest assumption for an unknown id. */
+export declare function aiModelCacheMinTokens(modelId: string): number;
+/** The catalog row for a model id, or `undefined` for an unknown id. */
+export declare function aiCatalogEntry(modelId: string): AiCatalogEntry | undefined;
+/**
+ * What the PROVIDER charges for a model id — our bill. A meter sentinel is
+ * free; an unknown id is priced at the dearest tier.
+ *
+ * The figure every cost and margin surface prices usage at. Answering the
+ * customer's question from here would report a margin the platform does not
+ * earn, and answering ours from `aiBilledRatesForModel` reports one it does
+ * not pay; the two are separate functions so neither is reachable by
+ * forgetting which one was meant.
+ */
+export declare function aiProviderRatesForModel(modelId: string): AiTokenRates;
+/**
+ * What a CUSTOMER'S CREDITS are charged at for a model id. A meter sentinel
+ * is free; an unknown id is priced at the dearest tier.
+ *
+ * The figure the credit meter, the band and the model selector price usage
+ * at. It is at or above `aiProviderRatesForModel` on every row.
+ */
+export declare function aiBilledRatesForModel(modelId: string): AiTokenRates;
+/**
+ * What one exchange COST US, at the serving model's provider rates, rounded
+ * to 6dp. Real money, and the only figure a margin may be taken against.
+ */
+export declare function estimateAiProviderCostUsd(usage: AiUsage, modelId: string): number;
+/**
+ * What one exchange DRAWS FROM A CUSTOMER, at the serving model's billed
+ * rates, rounded to 6dp — the figure `assistCreditsFromUsd` turns into
+ * credits. At or above `estimateAiProviderCostUsd` for the same exchange.
+ */
+export declare function estimateAiBilledUsd(usage: AiUsage, modelId: string): number;
+/** Every catalog id a provider serves. */
+export declare function aiModelIdsForProvider(providerId: string): string[];
+/**
+ * What a step asks for, and what the routing table answers with. Every
+ * door names its kind here rather than a model, so a model swap — an
+ * incident override, an org's own setting, a cheaper model for a cheap
+ * step — is one row rather than a search through the doors (AGL-2937).
+ */
+export type AiStepKind = 'assist.chat' | 'copy.element' | 'copy.section' | 'copy.blog' | 'generate.section' | 'job.component' | 'job.crm' | 'job.experiment' | 'job.form' | 'job.insight' | 'job.layout' | 'job.template' | 'job.page' | 'job.products' | 'job.seo' | 'job.text' | 'job.theme' | 'job.plan' | 'job.email' | 'job.campaign' | 'job.workflow';
+/** The tier each step kind is served from when no setting overrides it. */
+export declare const AI_STEP_TIERS: Record<AiStepKind, AiCatalogEntry['tier']>;
+/** The first catalog model of a tier on a provider, or the provider's first model. */
+export declare function aiDefaultModelFor(providerId: string, tier: AiCatalogEntry['tier']): string | undefined;

package/src/lib/providers/catalog.js

@@ -0,0 +1,312 @@
+/**
+ * @license
+ * Copyright 2026 Aglyn LLC
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */ import { _ as _extends } from "@swc/helpers/_/_extends";
+/** List price per MTok → per-token rates, cache read at 0.1x, write 1.25x. */ export function aiRatesPerMTok(inputPerMTok, outputPerMTok) {
+    return {
+        inputPerToken: inputPerMTok / 1000000,
+        outputPerToken: outputPerMTok / 1000000,
+        cacheReadPerToken: inputPerMTok * 0.1 / 1000000,
+        cacheWritePerToken: inputPerMTok * 1.25 / 1000000
+    };
+}
+/**
+ * A row billed at exactly the provider's list rate — the usual case, written
+ * once so the pair cannot drift.
+ */ export function aiRatesAtList(inputPerMTok, outputPerMTok) {
+    return {
+        providerRates: aiRatesPerMTok(inputPerMTok, outputPerMTok),
+        billedRates: aiRatesPerMTok(inputPerMTok, outputPerMTok)
+    };
+}
+/**
+ * The sentinels a docs-only answer is metered under (AGL-2486): not
+ * models, but ids the meter sees, priced at zero so a deflected turn lands
+ * in the same rollup as a served one and a keyless deployment does not
+ * read as the most efficient one on the platform.
+ */ export const AI_METER_SENTINELS = {
+    docsRetrieval: 'docs-retrieval',
+    docsLinks: 'docs-links',
+    answerCache: 'assist-cache'
+};
+const ANTHROPIC = 'anthropic';
+const OPENAI_COMPATIBLE = 'openai-compatible';
+/**
+ * The providers the catalog carries models for, with the name a settings
+ * form shows. A form offers providers from here for the reason it offers
+ * models from here: nothing outside this folder names a vendor. For the
+ * same reason a vendor's subprocessor row is written here too.
+ */ export const AI_CATALOG_PROVIDERS = [
+    {
+        id: ANTHROPIC,
+        label: 'Anthropic',
+        subprocessor: {
+            entity: 'Anthropic, PBC',
+            region: 'United States',
+            purpose: "AI-assisted features in the console and the site editor: the Aglyn Assist helper, including changes it proposes to a page, component, or layout open in the editor; editor assistance (rewriting element copy, drafting blog bodies, generating a section layout); and AI generation, which creates drafts and proposals for a customer's site from a brief or from what the site already holds, such as copy, layouts, templates, automations, search titles and descriptions, theme changes, product descriptions and tags, draft products, and store categories and discounts; explanations of a site's automations, including why an automation's run failed; AI insights, which answer a customer's questions about its own figures, such as site traffic, sales, bookings, forms, campaigns, A/B tests and datasets, and write a weekly summary of them for members who ask for one; and AI assistance in the CRM, which summarizes a contact, company, deal or lead for a user and suggests a next step, drafts a one-to-one email for the user to review and send, and suggests how the columns of a spreadsheet the user imports match CRM fields",
+            publishedOn: '2026-09-17',
+            reason: "Reached through the AI plugin's Anthropic adapter (`libs/plugins/ai/src/lib/providers/anthropic.ts`) by the doors that call the AI runtime. `libs/plugins/ai/src/lib/server/assist-chat.ts` is gated by `release_assist` AND the key, and a generation job's text step by `release_ai_generative`; `libs/plugins/ai/src/lib/server/ai-assist.ts` carries NO release flag, so setting `{apiKeyEnv}` in production is by itself what starts this flow. `assist-anthropic-subprocessor-gate.spec.ts` holds the per-door detail and is the deeper guard for this one vendor.",
+            dataReceived: "What the user submits — a question, instruction or brief, with the earlier messages of the same Assist conversation — and the content of the element, post, section or page being worked on, with the generated response. On Pro and above, the organization's name and the console route and host travel with an Assist question. For an edit the assistant proposes in the besigner, an outline of the open page, component or layout: element and component ids, layer names, shortened setting values and the selected element's styles. For a generation job, the site inventory: the names and addresses of its screens and collections, the names of its components, layouts, templates, forms and datasets with their prop and field names, and the theme's summary, colors and fonts. For a theme change, the site's current theme settings and brand colors as hex values, from the organization's brand settings, the site logo in the media library or a public page the brief links to. For features that review or write search information, the text and structure of the pages concerned. For product copy, the store's name, the product's name, type, description, tags, options and search listing, the store's category names, and the product's first media-library photo as a copy at most 768 px on its longer edge with its metadata stripped; never another media file, a price, stock, an order or a customer. For products, categories and discounts proposed from a brief, the store's name and its existing category names. For an automation drafted from a brief, which of CRM, webhooks and bookings the plan includes; for an explanation, the automation's outline (trigger, conditions, each step's text with the names of the lists, campaigns, workflows, webhooks and datasets it uses and whether each exists, and a workflow's function names and expressions) and, for a failed run, its time, steps and recorded errors, with email addresses removed and never the triggering event's data. For an insight, the figure reports available and aggregate tables: traffic with top page paths, referrers and campaign tags; form views and submissions; revenue, orders and best-selling product names; bookings by service; campaign subjects with delivery, open and click rates; A/B test and variant conversions; and, for a dataset the member can see, field names and types, record and fill counts, number ranges and totals grouped by a value at least three records share. Email addresses and phone numbers are removed, and no individual record is sent. For CRM assistance, the opened contact, company, deal or lead as the CRM shows it: its name and, by kind, job title, company, lifecycle stage, tags, capture history and counts, domain, industry, headcount, pipeline stages, status, amount, dates, lost reason, parties and lead status, with its notes, newest timeline entries and open tasks and deals. Email addresses and phone numbers in that text are replaced, and no email, phone or postal field, marketing consent, custom field value, team member or record id is sent. An email draft adds the request and the record's merge field names; an import sends the field names and types and each column's header and value kind, never a row. No account identifiers, email addresses or authentication tokens."
+        }
+    },
+    {
+        id: OPENAI_COMPATIBLE,
+        label: 'OpenAI-compatible endpoint'
+    }
+];
+/**
+ * `vision` is written on every row rather than assumed (AGL-2916): each model
+ * below reads a picture in a user turn, in the four formats
+ * `AI_IMAGE_MEDIA_TYPES` names, as its vendor documents. A row added without
+ * it reads as a model that does not, and is never sent one.
+ */ const anthropicCapabilities = {
+    streaming: true,
+    tools: true,
+    thinking: true,
+    promptCache: true,
+    vision: true
+};
+export const AI_MODEL_CATALOG = [
+    {
+        id: 'claude-sonnet-5',
+        provider: ANTHROPIC,
+        label: 'Claude Sonnet 5',
+        capabilities: anthropicCapabilities,
+        /*
+     * THE ONE ROW BILLED ABOVE ITS PROVIDER RATE (AGL-3015).
+     *
+     * The balanced tier prices most requests, so this markup is most of the
+     * platform's AI margin. It is carried in the rate and not in the credit
+     * conversion for the reason the module head gives: a credit stays a
+     * fixed quantity of billed spend on every model, so the band, the
+     * overage ladder and the invoice keep the arithmetic they were sized
+     * with, and nothing below has to know which model served a request.
+     *
+     * Everything that reasons about what we PAY — the margin surfaces, the
+     * staff spend meters, the discount guardrail's cost of goods — reads
+     * `providerRates`, which is the vendor's published list.
+     */ providerRates: aiRatesPerMTok(2, 10),
+        billedRates: aiRatesPerMTok(3, 15),
+        tier: 'balanced',
+        cacheMinTokens: 1024
+    },
+    _extends({
+        id: 'claude-sonnet-4-6',
+        provider: ANTHROPIC,
+        label: 'Claude Sonnet 4.6',
+        capabilities: anthropicCapabilities
+    }, aiRatesAtList(3, 15), {
+        tier: 'balanced',
+        cacheMinTokens: 1024
+    }),
+    _extends({
+        // Rejects an explicit `thinking` setting; the runtime omits it.
+        id: 'claude-haiku-4-5',
+        provider: ANTHROPIC,
+        label: 'Claude Haiku 4.5',
+        capabilities: _extends({}, anthropicCapabilities, {
+            thinking: false
+        })
+    }, aiRatesAtList(1, 5), {
+        tier: 'fast',
+        // Four times the balanced tier's minimum: the cheapest model per token
+        // is the hardest one to cache for, which is why a short prompt moved
+        // here can cost more per request than it saved per token.
+        cacheMinTokens: 4096
+    }),
+    _extends({
+        id: 'claude-opus-5',
+        provider: ANTHROPIC,
+        label: 'Claude Opus 5',
+        capabilities: anthropicCapabilities
+    }, aiRatesAtList(5, 25), {
+        tier: 'deep',
+        cacheMinTokens: 512
+    }),
+    _extends({
+        id: 'claude-opus-4-8',
+        provider: ANTHROPIC,
+        label: 'Claude Opus 4.8',
+        capabilities: anthropicCapabilities
+    }, aiRatesAtList(5, 25), {
+        tier: 'deep',
+        cacheMinTokens: 1024
+    }),
+    /**
+   * The OpenAI-compatible adapter serves whatever the endpoint behind
+   * `AI_OPENAI_COMPAT_BASE_URL` serves; these are the ids the routing table
+   * may name on it. The figures are the vendor's public list, billed at
+   * list: a deployment that reaches its own gateway at a negotiated price
+   * is paying less than this says, so its margin reads low rather than
+   * high — the direction a cost figure may be wrong in.
+   */ _extends({
+        id: 'gpt-5',
+        provider: OPENAI_COMPATIBLE,
+        label: 'GPT-5',
+        capabilities: {
+            streaming: true,
+            tools: true,
+            thinking: false,
+            promptCache: true,
+            vision: true
+        }
+    }, aiRatesAtList(1.25, 10), {
+        tier: 'balanced',
+        cacheMinTokens: 1024
+    }),
+    _extends({
+        id: 'gpt-5-mini',
+        provider: OPENAI_COMPATIBLE,
+        label: 'GPT-5 mini',
+        capabilities: {
+            streaming: true,
+            tools: true,
+            thinking: false,
+            promptCache: true,
+            vision: true
+        }
+    }, aiRatesAtList(0.25, 2), {
+        tier: 'fast',
+        cacheMinTokens: 1024
+    })
+];
+/** The dearest known tier, used when a model id is not in the catalog. */ export const AI_FALLBACK_RATES = aiRatesPerMTok(10, 50);
+/**
+ * The minimum an unknown model id is assumed to have: the largest one the
+ * catalog knows. It errs the way the rate fallback errs — toward the answer
+ * that costs more — so an unrecognized id reads as "this prompt does not
+ * cache" rather than promising a saving nobody measured.
+ */ export const AI_FALLBACK_CACHE_MIN_TOKENS = 4096;
+/** The shortest prefix a model's provider will cache; the dearest assumption for an unknown id. */ export function aiModelCacheMinTokens(modelId) {
+    var _ref;
+    var _aiCatalogEntry;
+    return (_ref = (_aiCatalogEntry = aiCatalogEntry(modelId)) == null ? void 0 : _aiCatalogEntry.cacheMinTokens) != null ? _ref : AI_FALLBACK_CACHE_MIN_TOKENS;
+}
+const ZERO_RATES = aiRatesPerMTok(0, 0);
+/** The catalog row for a model id, or `undefined` for an unknown id. */ export function aiCatalogEntry(modelId) {
+    return AI_MODEL_CATALOG.find((entry)=>entry.id === modelId);
+}
+const isSentinel = (modelId)=>Object.values(AI_METER_SENTINELS).includes(modelId);
+/**
+ * What the PROVIDER charges for a model id — our bill. A meter sentinel is
+ * free; an unknown id is priced at the dearest tier.
+ *
+ * The figure every cost and margin surface prices usage at. Answering the
+ * customer's question from here would report a margin the platform does not
+ * earn, and answering ours from `aiBilledRatesForModel` reports one it does
+ * not pay; the two are separate functions so neither is reachable by
+ * forgetting which one was meant.
+ */ export function aiProviderRatesForModel(modelId) {
+    var _ref;
+    var _aiCatalogEntry;
+    if (isSentinel(modelId)) return ZERO_RATES;
+    return (_ref = (_aiCatalogEntry = aiCatalogEntry(modelId)) == null ? void 0 : _aiCatalogEntry.providerRates) != null ? _ref : AI_FALLBACK_RATES;
+}
+/**
+ * What a CUSTOMER'S CREDITS are charged at for a model id. A meter sentinel
+ * is free; an unknown id is priced at the dearest tier.
+ *
+ * The figure the credit meter, the band and the model selector price usage
+ * at. It is at or above `aiProviderRatesForModel` on every row.
+ */ export function aiBilledRatesForModel(modelId) {
+    var _ref;
+    var _aiCatalogEntry;
+    if (isSentinel(modelId)) return ZERO_RATES;
+    return (_ref = (_aiCatalogEntry = aiCatalogEntry(modelId)) == null ? void 0 : _aiCatalogEntry.billedRates) != null ? _ref : AI_FALLBACK_RATES;
+}
+function priceUsage(usage, rate) {
+    const raw = usage.inputTokens * rate.inputPerToken + usage.outputTokens * rate.outputPerToken + usage.cacheReadTokens * rate.cacheReadPerToken + usage.cacheWriteTokens * rate.cacheWritePerToken;
+    return Math.round(raw * 1000000) / 1000000;
+}
+/**
+ * What one exchange COST US, at the serving model's provider rates, rounded
+ * to 6dp. Real money, and the only figure a margin may be taken against.
+ */ export function estimateAiProviderCostUsd(usage, modelId) {
+    return priceUsage(usage, aiProviderRatesForModel(modelId));
+}
+/**
+ * What one exchange DRAWS FROM A CUSTOMER, at the serving model's billed
+ * rates, rounded to 6dp — the figure `assistCreditsFromUsd` turns into
+ * credits. At or above `estimateAiProviderCostUsd` for the same exchange.
+ */ export function estimateAiBilledUsd(usage, modelId) {
+    return priceUsage(usage, aiBilledRatesForModel(modelId));
+}
+/** Every catalog id a provider serves. */ export function aiModelIdsForProvider(providerId) {
+    return AI_MODEL_CATALOG.filter((entry)=>entry.provider === providerId).map((entry)=>entry.id);
+}
+/** The tier each step kind is served from when no setting overrides it. */ export const AI_STEP_TIERS = {
+    'assist.chat': 'balanced',
+    'copy.element': 'fast',
+    'copy.section': 'balanced',
+    'copy.blog': 'balanced',
+    'generate.section': 'balanced',
+    // A reusable component is one structured tree with the typed props it
+    // declares, and choosing what becomes a prop is the judgment the step sells.
+    'job.component': 'balanced',
+    // CRM by AI (AGL-2917): a record's summary and next step, a deal's stage, a
+    // lead's standing, a one-to-one email draft and an import's column matches.
+    // Short answers through a strict tool, held to the facts the CRM reports
+    // about the record the member opened.
+    'job.crm': 'fast',
+    // A/B tests by AI (AGL-2914): variants of one piece of copy, and a
+    // result put into words for a verdict code already reached. Short answers
+    // through a strict tool, but the judgment a variant sells — which ONE idea
+    // to vary so the result says something — is the balanced tier's.
+    'job.experiment': 'balanced',
+    // A form is one small tree held to the building rules and to the contract
+    // its submissions are read by; a re-ask costs more than the tier saves.
+    'job.form': 'balanced',
+    // An insight (AGL-2915) chooses which figures answer a question and says
+    // what they show; the trace holds every number, and the judgment of what is
+    // worth saying is what the fast tier does worse.
+    'job.insight': 'balanced',
+    // A layout or a page template is one structured tree held to every
+    // building rule; the fast tier re-asks more than it saves.
+    'job.layout': 'balanced',
+    'job.template': 'balanced',
+    // A page section (AGL-2907) is held to the whole page's building rules on
+    // every pass, for the same reason.
+    'job.page': 'balanced',
+    // A product's copy, a store's first products, or its categories and
+    // discounts (AGL-2916): judgment about what a photo shows and what a shopper
+    // needs, held to storefront claim rules a re-ask costs more to meet than the
+    // tier saves.
+    'job.products': 'balanced',
+    // SEO fields, alt text and an audit's fixes (AGL-2910): short answers
+    // through a strict tool, held to a length and to the page's own text.
+    'job.seo': 'fast',
+    'job.text': 'balanced',
+    // A theme is one structured answer over a small, fixed control set; the
+    // judgment is in the color choices, which the fast tier makes worse.
+    'job.theme': 'balanced',
+    'job.plan': 'balanced',
+    // An email is one tree held to every building rule, with subject lines
+    // whose judgment is the point of asking; the fast tier re-asks more than it
+    // saves. A campaign is the same email, drafted into a campaign.
+    'job.email': 'balanced',
+    'job.campaign': 'balanced',
+    // An automation is a small structured answer whose judgment is choosing the
+    // trigger and steps a description means, and an explanation reads one
+    // closely; the fast tier re-asks more than it saves on both (AGL-2919).
+    'job.workflow': 'balanced'
+};
+/** The first catalog model of a tier on a provider, or the provider's first model. */ export function aiDefaultModelFor(providerId, tier) {
+    var _onProvider_find;
+    var _this;
+    const onProvider = AI_MODEL_CATALOG.filter((entry)=>entry.provider === providerId);
+    return (_this = (_onProvider_find = onProvider.find((entry)=>entry.tier === tier)) != null ? _onProvider_find : onProvider[0]) == null ? void 0 : _this.id;
+}
+
+//# sourceMappingURL=catalog.js.map

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

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/catalog.ts"],"sourcesContent":["/**\n * @license\n * Copyright 2026 Aglyn LLC\n *\n * Licensed under the Apache License, Version 2.0 (the \"License\");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n *   http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an \"AS IS\" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n */\n\nimport type { PluginSubprocessorDeclaration } from '@aglyn/aglyn/plugin-manager/plugin-subprocessors'\nimport type { AiModelDescriptor, AiUsage } from './contract'\n\n/**\n * The model catalog (AGL-2939): every model id the plugin will route to,\n * with its provider, its display name, its capabilities and its list rates.\n * The ONE place a model id is written outside a provider adapter — the\n * routing table maps a step kind to an id through here, the meter prices\n * usage by id through here, and a settings form offers ids from here.\n *\n * ## Two rates, because one figure was answering two questions\n *\n * Every row carries a PROVIDER rate and a BILLED rate (AGL-3015):\n *\n * - `providerRates` is what the vendor charges us. It is money leaving the\n *   company, so cost accounting, the margin surfaces and every staff spend\n *   meter price usage through this one and no other.\n * - `billedRates` is what a customer's credits are charged at. The credit\n *   meter, the band, the overage ladder and the model selector price usage\n *   through this one, because it is the figure a customer's balance moves by.\n *\n * They are equal on nearly every row, and a helper writes the pair from one\n * figure so an unmarked row cannot drift apart by a typo. A row may be\n * billed ABOVE its provider rate deliberately — the markup is carried in the\n * rate rather than in the credit conversion, which keeps a credit a fixed\n * quantity of billed spend on every model and leaves the band, the ladder\n * and the invoice on the one arithmetic they were sized with. A row billed\n * BELOW its provider rate would sell tokens at a loss, and\n * `billed-rate-is-not-provider-cost.spec.ts` refuses one.\n *\n * Both are keyed by model rather than fixed at one model's rates because the\n * default is an env override: a one-line incident swap to a dearer model\n * would otherwise keep reporting the cheaper money, and per-org cost would\n * read as roughly right — the failure mode this whole meter exists to\n * prevent. An unknown id falls back to the most EXPENSIVE known tier on\n * purpose, on both rates: an estimate that errs low is worse than one that\n * errs high, whether the figure is our bill or the customer's.\n */\n\nexport interface AiTokenRates {\n  inputPerToken: number\n  outputPerToken: number\n  cacheReadPerToken: number\n  cacheWritePerToken: number\n}\n\n/** List price per MTok → per-token rates, cache read at 0.1x, write 1.25x. */\nexport function aiRatesPerMTok(inputPerMTok: number, outputPerMTok: number): AiTokenRates {\n  return {\n    inputPerToken: inputPerMTok / 1_000_000,\n    outputPerToken: outputPerMTok / 1_000_000,\n    cacheReadPerToken: (inputPerMTok * 0.1) / 1_000_000,\n    cacheWritePerToken: (inputPerMTok * 1.25) / 1_000_000,\n  }\n}\n\n/**\n * The two rate tables a catalog row carries. Separate fields rather than one\n * table and a multiplier: a reader that wants our cost has to name it, and\n * a reader that wants the customer's has to name that, so neither can be\n * reached by accident.\n */\nexport interface AiCatalogRates {\n  /** What the provider bills us for this model. */\n  providerRates: AiTokenRates\n  /** What a customer's credits are charged at for this model. */\n  billedRates: AiTokenRates\n}\n\n/**\n * A row billed at exactly the provider's list rate — the usual case, written\n * once so the pair cannot drift.\n */\nexport function aiRatesAtList(\n  inputPerMTok: number,\n  outputPerMTok: number,\n): AiCatalogRates {\n  return {\n    providerRates: aiRatesPerMTok(inputPerMTok, outputPerMTok),\n    billedRates: aiRatesPerMTok(inputPerMTok, outputPerMTok),\n  }\n}\n\nexport interface AiCatalogEntry extends AiModelDescriptor, AiCatalogRates {\n  /**\n   * The tier a routing decision reads: `fast` for short, constrained\n   * answers; `balanced` for the assistant and briefs; `deep` for the\n   * dearest, slowest class.\n   */\n  tier: 'fast' | 'balanced' | 'deep'\n  /**\n   * The shortest prefix this model's provider will cache (AGL-2937). A\n   * request whose cached span falls under it is served with the markers\n   * honored and NOTHING cached: no error, no warning, and a usage report\n   * that reads as a permanent cache miss.\n   *\n   * It is a property of the model rather than of the tier or the vendor,\n   * and it does not fall as a generation advances — so the number has to be\n   * written beside the model, and a door that means to cache has to be held\n   * to it (`aiCachedPrefixCaches` in the runtime, and the door table in\n   * `runtime/ai-prompt-cache.spec.ts`). Without it, \"this block carries a\n   * breakpoint\" is a claim nothing checks, and a prompt can be designed\n   * around a cache it never had.\n   */\n  cacheMinTokens: number\n}\n\n/**\n * The sentinels a docs-only answer is metered under (AGL-2486): not\n * models, but ids the meter sees, priced at zero so a deflected turn lands\n * in the same rollup as a served one and a keyless deployment does not\n * read as the most efficient one on the platform.\n */\nexport const AI_METER_SENTINELS = {\n  docsRetrieval: 'docs-retrieval',\n  docsLinks: 'docs-links',\n  answerCache: 'assist-cache',\n} as const\n\nconst ANTHROPIC = 'anthropic'\nconst OPENAI_COMPATIBLE = 'openai-compatible'\n\n/**\n * A provider's row on the published subprocessor list: every field but the\n * host, which the registered adapter reports as its `endpointHost`.\n * `{apiKeyEnv}` anywhere in the wording stands for the adapter's\n * `apiKeyEnv`, so a credential's name is written in its adapter alone and\n * never in this module, which the browser bundle carries.\n */\nexport type AiProviderSubprocessorWording = Omit<PluginSubprocessorDeclaration, 'host'>\n\n/** A provider the catalog carries models for. */\nexport interface AiCatalogProvider {\n  id: string\n  /** The name a settings form shows. */\n  label: string\n  /**\n   * The vendor's row on the published subprocessor list, written here and\n   * nowhere else: `aiSubprocessors` derives the plugin's declarations from\n   * it. Absent for a provider whose endpoint an operator names, which has\n   * no fixed recipient to publish.\n   */\n  subprocessor?: AiProviderSubprocessorWording\n}\n\n/**\n * The providers the catalog carries models for, with the name a settings\n * form shows. A form offers providers from here for the reason it offers\n * models from here: nothing outside this folder names a vendor. For the\n * same reason a vendor's subprocessor row is written here too.\n */\nexport const AI_CATALOG_PROVIDERS: readonly AiCatalogProvider[] = [\n  {\n    id: ANTHROPIC,\n    label: 'Anthropic',\n    subprocessor: {\n      entity: 'Anthropic, PBC',\n      region: 'United States',\n      purpose:\n        \"AI-assisted features in the console and the site editor: the Aglyn Assist helper, including changes it proposes to a page, component, or layout open in the editor; editor assistance (rewriting element copy, drafting blog bodies, generating a section layout); and AI generation, which creates drafts and proposals for a customer's site from a brief or from what the site already holds, such as copy, layouts, templates, automations, search titles and descriptions, theme changes, product descriptions and tags, draft products, and store categories and discounts; explanations of a site's automations, including why an automation's run failed; AI insights, which answer a customer's questions about its own figures, such as site traffic, sales, bookings, forms, campaigns, A/B tests and datasets, and write a weekly summary of them for members who ask for one; and AI assistance in the CRM, which summarizes a contact, company, deal or lead for a user and suggests a next step, drafts a one-to-one email for the user to review and send, and suggests how the columns of a spreadsheet the user imports match CRM fields\",\n      publishedOn: '2026-09-17',\n      reason:\n        \"Reached through the AI plugin's Anthropic adapter (`libs/plugins/ai/src/lib/providers/anthropic.ts`) by the doors that call the AI runtime. `libs/plugins/ai/src/lib/server/assist-chat.ts` is gated by `release_assist` AND the key, and a generation job's text step by `release_ai_generative`; `libs/plugins/ai/src/lib/server/ai-assist.ts` carries NO release flag, so setting `{apiKeyEnv}` in production is by itself what starts this flow. `assist-anthropic-subprocessor-gate.spec.ts` holds the per-door detail and is the deeper guard for this one vendor.\",\n      dataReceived:\n        \"What the user submits — a question, instruction or brief, with the earlier messages of the same Assist conversation — and the content of the element, post, section or page being worked on, with the generated response. On Pro and above, the organization's name and the console route and host travel with an Assist question. For an edit the assistant proposes in the besigner, an outline of the open page, component or layout: element and component ids, layer names, shortened setting values and the selected element's styles. For a generation job, the site inventory: the names and addresses of its screens and collections, the names of its components, layouts, templates, forms and datasets with their prop and field names, and the theme's summary, colors and fonts. For a theme change, the site's current theme settings and brand colors as hex values, from the organization's brand settings, the site logo in the media library or a public page the brief links to. For features that review or write search information, the text and structure of the pages concerned. For product copy, the store's name, the product's name, type, description, tags, options and search listing, the store's category names, and the product's first media-library photo as a copy at most 768 px on its longer edge with its metadata stripped; never another media file, a price, stock, an order or a customer. For products, categories and discounts proposed from a brief, the store's name and its existing category names. For an automation drafted from a brief, which of CRM, webhooks and bookings the plan includes; for an explanation, the automation's outline (trigger, conditions, each step's text with the names of the lists, campaigns, workflows, webhooks and datasets it uses and whether each exists, and a workflow's function names and expressions) and, for a failed run, its time, steps and recorded errors, with email addresses removed and never the triggering event's data. For an insight, the figure reports available and aggregate tables: traffic with top page paths, referrers and campaign tags; form views and submissions; revenue, orders and best-selling product names; bookings by service; campaign subjects with delivery, open and click rates; A/B test and variant conversions; and, for a dataset the member can see, field names and types, record and fill counts, number ranges and totals grouped by a value at least three records share. Email addresses and phone numbers are removed, and no individual record is sent. For CRM assistance, the opened contact, company, deal or lead as the CRM shows it: its name and, by kind, job title, company, lifecycle stage, tags, capture history and counts, domain, industry, headcount, pipeline stages, status, amount, dates, lost reason, parties and lead status, with its notes, newest timeline entries and open tasks and deals. Email addresses and phone numbers in that text are replaced, and no email, phone or postal field, marketing consent, custom field value, team member or record id is sent. An email draft adds the request and the record's merge field names; an import sends the field names and types and each column's header and value kind, never a row. No account identifiers, email addresses or authentication tokens.\",\n    },\n  },\n  { id: OPENAI_COMPATIBLE, label: 'OpenAI-compatible endpoint' },\n]\n\n/**\n * `vision` is written on every row rather than assumed (AGL-2916): each model\n * below reads a picture in a user turn, in the four formats\n * `AI_IMAGE_MEDIA_TYPES` names, as its vendor documents. A row added without\n * it reads as a model that does not, and is never sent one.\n */\nconst anthropicCapabilities = {\n  streaming: true,\n  tools: true,\n  thinking: true,\n  promptCache: true,\n  vision: true,\n}\n\nexport const AI_MODEL_CATALOG: readonly AiCatalogEntry[] = [\n  {\n    id: 'claude-sonnet-5',\n    provider: ANTHROPIC,\n    label: 'Claude Sonnet 5',\n    capabilities: anthropicCapabilities,\n    /*\n     * THE ONE ROW BILLED ABOVE ITS PROVIDER RATE (AGL-3015).\n     *\n     * The balanced tier prices most requests, so this markup is most of the\n     * platform's AI margin. It is carried in the rate and not in the credit\n     * conversion for the reason the module head gives: a credit stays a\n     * fixed quantity of billed spend on every model, so the band, the\n     * overage ladder and the invoice keep the arithmetic they were sized\n     * with, and nothing below has to know which model served a request.\n     *\n     * Everything that reasons about what we PAY — the margin surfaces, the\n     * staff spend meters, the discount guardrail's cost of goods — reads\n     * `providerRates`, which is the vendor's published list.\n     */\n    providerRates: aiRatesPerMTok(2, 10),\n    billedRates: aiRatesPerMTok(3, 15),\n    tier: 'balanced',\n    cacheMinTokens: 1_024,\n  },\n  {\n    id: 'claude-sonnet-4-6',\n    provider: ANTHROPIC,\n    label: 'Claude Sonnet 4.6',\n    capabilities: anthropicCapabilities,\n    ...aiRatesAtList(3, 15),\n    tier: 'balanced',\n    cacheMinTokens: 1_024,\n  },\n  {\n    // Rejects an explicit `thinking` setting; the runtime omits it.\n    id: 'claude-haiku-4-5',\n    provider: ANTHROPIC,\n    label: 'Claude Haiku 4.5',\n    capabilities: { ...anthropicCapabilities, thinking: false },\n    ...aiRatesAtList(1, 5),\n    tier: 'fast',\n    // Four times the balanced tier's minimum: the cheapest model per token\n    // is the hardest one to cache for, which is why a short prompt moved\n    // here can cost more per request than it saved per token.\n    cacheMinTokens: 4_096,\n  },\n  {\n    id: 'claude-opus-5',\n    provider: ANTHROPIC,\n    label: 'Claude Opus 5',\n    capabilities: anthropicCapabilities,\n    ...aiRatesAtList(5, 25),\n    tier: 'deep',\n    cacheMinTokens: 512,\n  },\n  {\n    id: 'claude-opus-4-8',\n    provider: ANTHROPIC,\n    label: 'Claude Opus 4.8',\n    capabilities: anthropicCapabilities,\n    ...aiRatesAtList(5, 25),\n    tier: 'deep',\n    cacheMinTokens: 1_024,\n  },\n  /**\n   * The OpenAI-compatible adapter serves whatever the endpoint behind\n   * `AI_OPENAI_COMPAT_BASE_URL` serves; these are the ids the routing table\n   * may name on it. The figures are the vendor's public list, billed at\n   * list: a deployment that reaches its own gateway at a negotiated price\n   * is paying less than this says, so its margin reads low rather than\n   * high — the direction a cost figure may be wrong in.\n   */\n  {\n    id: 'gpt-5',\n    provider: OPENAI_COMPATIBLE,\n    label: 'GPT-5',\n    capabilities: { streaming: true, tools: true, thinking: false, promptCache: true, vision: true },\n    ...aiRatesAtList(1.25, 10),\n    tier: 'balanced',\n    cacheMinTokens: 1_024,\n  },\n  {\n    id: 'gpt-5-mini',\n    provider: OPENAI_COMPATIBLE,\n    label: 'GPT-5 mini',\n    capabilities: { streaming: true, tools: true, thinking: false, promptCache: true, vision: true },\n    ...aiRatesAtList(0.25, 2),\n    tier: 'fast',\n    cacheMinTokens: 1_024,\n  },\n]\n\n/** The dearest known tier, used when a model id is not in the catalog. */\nexport const AI_FALLBACK_RATES: AiTokenRates = aiRatesPerMTok(10, 50)\n\n/**\n * The minimum an unknown model id is assumed to have: the largest one the\n * catalog knows. It errs the way the rate fallback errs — toward the answer\n * that costs more — so an unrecognized id reads as \"this prompt does not\n * cache\" rather than promising a saving nobody measured.\n */\nexport const AI_FALLBACK_CACHE_MIN_TOKENS = 4_096\n\n/** The shortest prefix a model's provider will cache; the dearest assumption for an unknown id. */\nexport function aiModelCacheMinTokens(modelId: string): number {\n  return aiCatalogEntry(modelId)?.cacheMinTokens ?? AI_FALLBACK_CACHE_MIN_TOKENS\n}\n\nconst ZERO_RATES = aiRatesPerMTok(0, 0)\n\n/** The catalog row for a model id, or `undefined` for an unknown id. */\nexport function aiCatalogEntry(modelId: string): AiCatalogEntry | undefined {\n  return AI_MODEL_CATALOG.find((entry) => entry.id === modelId)\n}\n\nconst isSentinel = (modelId: string): boolean =>\n  (Object.values(AI_METER_SENTINELS) as string[]).includes(modelId)\n\n/**\n * What the PROVIDER charges for a model id — our bill. A meter sentinel is\n * free; an unknown id is priced at the dearest tier.\n *\n * The figure every cost and margin surface prices usage at. Answering the\n * customer's question from here would report a margin the platform does not\n * earn, and answering ours from `aiBilledRatesForModel` reports one it does\n * not pay; the two are separate functions so neither is reachable by\n * forgetting which one was meant.\n */\nexport function aiProviderRatesForModel(modelId: string): AiTokenRates {\n  if (isSentinel(modelId)) return ZERO_RATES\n  return aiCatalogEntry(modelId)?.providerRates ?? AI_FALLBACK_RATES\n}\n\n/**\n * What a CUSTOMER'S CREDITS are charged at for a model id. A meter sentinel\n * is free; an unknown id is priced at the dearest tier.\n *\n * The figure the credit meter, the band and the model selector price usage\n * at. It is at or above `aiProviderRatesForModel` on every row.\n */\nexport function aiBilledRatesForModel(modelId: string): AiTokenRates {\n  if (isSentinel(modelId)) return ZERO_RATES\n  return aiCatalogEntry(modelId)?.billedRates ?? AI_FALLBACK_RATES\n}\n\nfunction priceUsage(usage: AiUsage, rate: AiTokenRates): number {\n  const raw =\n    usage.inputTokens * rate.inputPerToken +\n    usage.outputTokens * rate.outputPerToken +\n    usage.cacheReadTokens * rate.cacheReadPerToken +\n    usage.cacheWriteTokens * rate.cacheWritePerToken\n  return Math.round(raw * 1_000_000) / 1_000_000\n}\n\n/**\n * What one exchange COST US, at the serving model's provider rates, rounded\n * to 6dp. Real money, and the only figure a margin may be taken against.\n */\nexport function estimateAiProviderCostUsd(usage: AiUsage, modelId: string): number {\n  return priceUsage(usage, aiProviderRatesForModel(modelId))\n}\n\n/**\n * What one exchange DRAWS FROM A CUSTOMER, at the serving model's billed\n * rates, rounded to 6dp — the figure `assistCreditsFromUsd` turns into\n * credits. At or above `estimateAiProviderCostUsd` for the same exchange.\n */\nexport function estimateAiBilledUsd(usage: AiUsage, modelId: string): number {\n  return priceUsage(usage, aiBilledRatesForModel(modelId))\n}\n\n/** Every catalog id a provider serves. */\nexport function aiModelIdsForProvider(providerId: string): string[] {\n  return AI_MODEL_CATALOG.filter((entry) => entry.provider === providerId).map(\n    (entry) => entry.id,\n  )\n}\n\n/**\n * What a step asks for, and what the routing table answers with. Every\n * door names its kind here rather than a model, so a model swap — an\n * incident override, an org's own setting, a cheaper model for a cheap\n * step — is one row rather than a search through the doors (AGL-2937).\n */\nexport type AiStepKind =\n  | 'assist.chat'\n  | 'copy.element'\n  | 'copy.section'\n  | 'copy.blog'\n  | 'generate.section'\n  | 'job.component'\n  | 'job.crm'\n  | 'job.experiment'\n  | 'job.form'\n  | 'job.insight'\n  | 'job.layout'\n  | 'job.template'\n  | 'job.page'\n  | 'job.products'\n  | 'job.seo'\n  | 'job.text'\n  | 'job.theme'\n  | 'job.plan'\n  | 'job.email'\n  | 'job.campaign'\n  | 'job.workflow'\n\n/** The tier each step kind is served from when no setting overrides it. */\nexport const AI_STEP_TIERS: Record<AiStepKind, AiCatalogEntry['tier']> = {\n  'assist.chat': 'balanced',\n  'copy.element': 'fast',\n  'copy.section': 'balanced',\n  'copy.blog': 'balanced',\n  'generate.section': 'balanced',\n  // A reusable component is one structured tree with the typed props it\n  // declares, and choosing what becomes a prop is the judgment the step sells.\n  'job.component': 'balanced',\n  // CRM by AI (AGL-2917): a record's summary and next step, a deal's stage, a\n  // lead's standing, a one-to-one email draft and an import's column matches.\n  // Short answers through a strict tool, held to the facts the CRM reports\n  // about the record the member opened.\n  'job.crm': 'fast',\n  // A/B tests by AI (AGL-2914): variants of one piece of copy, and a\n  // result put into words for a verdict code already reached. Short answers\n  // through a strict tool, but the judgment a variant sells — which ONE idea\n  // to vary so the result says something — is the balanced tier's.\n  'job.experiment': 'balanced',\n  // A form is one small tree held to the building rules and to the contract\n  // its submissions are read by; a re-ask costs more than the tier saves.\n  'job.form': 'balanced',\n  // An insight (AGL-2915) chooses which figures answer a question and says\n  // what they show; the trace holds every number, and the judgment of what is\n  // worth saying is what the fast tier does worse.\n  'job.insight': 'balanced',\n  // A layout or a page template is one structured tree held to every\n  // building rule; the fast tier re-asks more than it saves.\n  'job.layout': 'balanced',\n  'job.template': 'balanced',\n  // A page section (AGL-2907) is held to the whole page's building rules on\n  // every pass, for the same reason.\n  'job.page': 'balanced',\n  // A product's copy, a store's first products, or its categories and\n  // discounts (AGL-2916): judgment about what a photo shows and what a shopper\n  // needs, held to storefront claim rules a re-ask costs more to meet than the\n  // tier saves.\n  'job.products': 'balanced',\n  // SEO fields, alt text and an audit's fixes (AGL-2910): short answers\n  // through a strict tool, held to a length and to the page's own text.\n  'job.seo': 'fast',\n  'job.text': 'balanced',\n  // A theme is one structured answer over a small, fixed control set; the\n  // judgment is in the color choices, which the fast tier makes worse.\n  'job.theme': 'balanced',\n  'job.plan': 'balanced',\n  // An email is one tree held to every building rule, with subject lines\n  // whose judgment is the point of asking; the fast tier re-asks more than it\n  // saves. A campaign is the same email, drafted into a campaign.\n  'job.email': 'balanced',\n  'job.campaign': 'balanced',\n  // An automation is a small structured answer whose judgment is choosing the\n  // trigger and steps a description means, and an explanation reads one\n  // closely; the fast tier re-asks more than it saves on both (AGL-2919).\n  'job.workflow': 'balanced',\n}\n\n/** The first catalog model of a tier on a provider, or the provider's first model. */\nexport function aiDefaultModelFor(providerId: string, tier: AiCatalogEntry['tier']): string | undefined {\n  const onProvider = AI_MODEL_CATALOG.filter((entry) => entry.provider === providerId)\n  return (onProvider.find((entry) => entry.tier === tier) ?? onProvider[0])?.id\n}\n"],"names":["aiRatesPerMTok","inputPerMTok","outputPerMTok","inputPerToken","outputPerToken","cacheReadPerToken","cacheWritePerToken","aiRatesAtList","providerRates","billedRates","AI_METER_SENTINELS","docsRetrieval","docsLinks","answerCache","ANTHROPIC","OPENAI_COMPATIBLE","AI_CATALOG_PROVIDERS","id","label","subprocessor","entity","region","purpose","publishedOn","reason","dataReceived","anthropicCapabilities","streaming","tools","thinking","promptCache","vision","AI_MODEL_CATALOG","provider","capabilities","tier","cacheMinTokens","AI_FALLBACK_RATES","AI_FALLBACK_CACHE_MIN_TOKENS","aiModelCacheMinTokens","modelId","aiCatalogEntry","ZERO_RATES","find","entry","isSentinel","Object","values","includes","aiProviderRatesForModel","aiBilledRatesForModel","priceUsage","usage","rate","raw","inputTokens","outputTokens","cacheReadTokens","cacheWriteTokens","Math","round","estimateAiProviderCostUsd","estimateAiBilledUsd","aiModelIdsForProvider","providerId","filter","map","AI_STEP_TIERS","aiDefaultModelFor","onProvider"],"mappings":"AAAA;;;;;;;;;;;;;;;CAeC;AAgDD,4EAA4E,GAC5E,OAAO,SAASA,eAAeC,YAAoB,EAAEC,aAAqB;IACxE,OAAO;QACLC,eAAeF,eAAe;QAC9BG,gBAAgBF,gBAAgB;QAChCG,mBAAmB,AAACJ,eAAe,MAAO;QAC1CK,oBAAoB,AAACL,eAAe,OAAQ;IAC9C;AACF;AAeA;;;CAGC,GACD,OAAO,SAASM,cACdN,YAAoB,EACpBC,aAAqB;IAErB,OAAO;QACLM,eAAeR,eAAeC,cAAcC;QAC5CO,aAAaT,eAAeC,cAAcC;IAC5C;AACF;AA0BA;;;;;CAKC,GACD,OAAO,MAAMQ,qBAAqB;IAChCC,eAAe;IACfC,WAAW;IACXC,aAAa;AACf,EAAU;AAEV,MAAMC,YAAY;AAClB,MAAMC,oBAAoB;AAyB1B;;;;;CAKC,GACD,OAAO,MAAMC,uBAAqD;IAChE;QACEC,IAAIH;QACJI,OAAO;QACPC,cAAc;YACZC,QAAQ;YACRC,QAAQ;YACRC,SACE;YACFC,aAAa;YACbC,QACE;YACFC,cACE;QACJ;IACF;IACA;QAAER,IAAIF;QAAmBG,OAAO;IAA6B;CAC9D,CAAA;AAED;;;;;CAKC,GACD,MAAMQ,wBAAwB;IAC5BC,WAAW;IACXC,OAAO;IACPC,UAAU;IACVC,aAAa;IACbC,QAAQ;AACV;AAEA,OAAO,MAAMC,mBAA8C;IACzD;QACEf,IAAI;QACJgB,UAAUnB;QACVI,OAAO;QACPgB,cAAcR;QACd;;;;;;;;;;;;;KAaC,GACDlB,eAAeR,eAAe,GAAG;QACjCS,aAAaT,eAAe,GAAG;QAC/BmC,MAAM;QACNC,gBAAgB;IAClB;IACA;QACEnB,IAAI;QACJgB,UAAUnB;QACVI,OAAO;QACPgB,cAAcR;OACXnB,cAAc,GAAG;QACpB4B,MAAM;QACNC,gBAAgB;;IAElB;QACE,gEAAgE;QAChEnB,IAAI;QACJgB,UAAUnB;QACVI,OAAO;QACPgB,cAAc,aAAKR;YAAuBG,UAAU;;OACjDtB,cAAc,GAAG;QACpB4B,MAAM;QACN,uEAAuE;QACvE,qEAAqE;QACrE,0DAA0D;QAC1DC,gBAAgB;;IAElB;QACEnB,IAAI;QACJgB,UAAUnB;QACVI,OAAO;QACPgB,cAAcR;OACXnB,cAAc,GAAG;QACpB4B,MAAM;QACNC,gBAAgB;;IAElB;QACEnB,IAAI;QACJgB,UAAUnB;QACVI,OAAO;QACPgB,cAAcR;OACXnB,cAAc,GAAG;QACpB4B,MAAM;QACNC,gBAAgB;;IAElB;;;;;;;GAOC,GACD;QACEnB,IAAI;QACJgB,UAAUlB;QACVG,OAAO;QACPgB,cAAc;YAAEP,WAAW;YAAMC,OAAO;YAAMC,UAAU;YAAOC,aAAa;YAAMC,QAAQ;QAAK;OAC5FxB,cAAc,MAAM;QACvB4B,MAAM;QACNC,gBAAgB;;IAElB;QACEnB,IAAI;QACJgB,UAAUlB;QACVG,OAAO;QACPgB,cAAc;YAAEP,WAAW;YAAMC,OAAO;YAAMC,UAAU;YAAOC,aAAa;YAAMC,QAAQ;QAAK;OAC5FxB,cAAc,MAAM;QACvB4B,MAAM;QACNC,gBAAgB;;CAEnB,CAAA;AAED,wEAAwE,GACxE,OAAO,MAAMC,oBAAkCrC,eAAe,IAAI,IAAG;AAErE;;;;;CAKC,GACD,OAAO,MAAMsC,+BAA+B,KAAK;AAEjD,iGAAiG,GACjG,OAAO,SAASC,sBAAsBC,OAAe;;QAC5CC;IAAP,gBAAOA,kBAAAA,eAAeD,6BAAfC,gBAAyBL,cAAc,mBAAIE;AACpD;AAEA,MAAMI,aAAa1C,eAAe,GAAG;AAErC,sEAAsE,GACtE,OAAO,SAASyC,eAAeD,OAAe;IAC5C,OAAOR,iBAAiBW,IAAI,CAAC,CAACC,QAAUA,MAAM3B,EAAE,KAAKuB;AACvD;AAEA,MAAMK,aAAa,CAACL,UAClB,AAACM,OAAOC,MAAM,CAACrC,oBAAiCsC,QAAQ,CAACR;AAE3D;;;;;;;;;CASC,GACD,OAAO,SAASS,wBAAwBT,OAAe;;QAE9CC;IADP,IAAII,WAAWL,UAAU,OAAOE;IAChC,gBAAOD,kBAAAA,eAAeD,6BAAfC,gBAAyBjC,aAAa,mBAAI6B;AACnD;AAEA;;;;;;CAMC,GACD,OAAO,SAASa,sBAAsBV,OAAe;;QAE5CC;IADP,IAAII,WAAWL,UAAU,OAAOE;IAChC,gBAAOD,kBAAAA,eAAeD,6BAAfC,gBAAyBhC,WAAW,mBAAI4B;AACjD;AAEA,SAASc,WAAWC,KAAc,EAAEC,IAAkB;IACpD,MAAMC,MACJF,MAAMG,WAAW,GAAGF,KAAKlD,aAAa,GACtCiD,MAAMI,YAAY,GAAGH,KAAKjD,cAAc,GACxCgD,MAAMK,eAAe,GAAGJ,KAAKhD,iBAAiB,GAC9C+C,MAAMM,gBAAgB,GAAGL,KAAK/C,kBAAkB;IAClD,OAAOqD,KAAKC,KAAK,CAACN,MAAM,WAAa;AACvC;AAEA;;;CAGC,GACD,OAAO,SAASO,0BAA0BT,KAAc,EAAEZ,OAAe;IACvE,OAAOW,WAAWC,OAAOH,wBAAwBT;AACnD;AAEA;;;;CAIC,GACD,OAAO,SAASsB,oBAAoBV,KAAc,EAAEZ,OAAe;IACjE,OAAOW,WAAWC,OAAOF,sBAAsBV;AACjD;AAEA,wCAAwC,GACxC,OAAO,SAASuB,sBAAsBC,UAAkB;IACtD,OAAOhC,iBAAiBiC,MAAM,CAAC,CAACrB,QAAUA,MAAMX,QAAQ,KAAK+B,YAAYE,GAAG,CAC1E,CAACtB,QAAUA,MAAM3B,EAAE;AAEvB;AA+BA,yEAAyE,GACzE,OAAO,MAAMkD,gBAA4D;IACvE,eAAe;IACf,gBAAgB;IAChB,gBAAgB;IAChB,aAAa;IACb,oBAAoB;IACpB,sEAAsE;IACtE,6EAA6E;IAC7E,iBAAiB;IACjB,4EAA4E;IAC5E,4EAA4E;IAC5E,yEAAyE;IACzE,sCAAsC;IACtC,WAAW;IACX,mEAAmE;IACnE,0EAA0E;IAC1E,2EAA2E;IAC3E,iEAAiE;IACjE,kBAAkB;IAClB,0EAA0E;IAC1E,wEAAwE;IACxE,YAAY;IACZ,yEAAyE;IACzE,4EAA4E;IAC5E,iDAAiD;IACjD,eAAe;IACf,mEAAmE;IACnE,2DAA2D;IAC3D,cAAc;IACd,gBAAgB;IAChB,0EAA0E;IAC1E,mCAAmC;IACnC,YAAY;IACZ,oEAAoE;IACpE,6EAA6E;IAC7E,6EAA6E;IAC7E,cAAc;IACd,gBAAgB;IAChB,sEAAsE;IACtE,sEAAsE;IACtE,WAAW;IACX,YAAY;IACZ,wEAAwE;IACxE,qEAAqE;IACrE,aAAa;IACb,YAAY;IACZ,uEAAuE;IACvE,4EAA4E;IAC5E,gEAAgE;IAChE,aAAa;IACb,gBAAgB;IAChB,4EAA4E;IAC5E,sEAAsE;IACtE,wEAAwE;IACxE,gBAAgB;AAClB,EAAC;AAED,oFAAoF,GACpF,OAAO,SAASC,kBAAkBJ,UAAkB,EAAE7B,IAA4B;QAExEkC;QAAAA;IADR,MAAMA,aAAarC,iBAAiBiC,MAAM,CAAC,CAACrB,QAAUA,MAAMX,QAAQ,KAAK+B;IACzE,QAAQK,SAAAA,mBAAAA,WAAW1B,IAAI,CAAC,CAACC,QAAUA,MAAMT,IAAI,KAAKA,iBAA1CkC,mBAAmDA,UAAU,CAAC,EAAE,qBAAjE,AAACA,MAAmEpD,EAAE;AAC/E"}