@aglyn/plugins-ai 1.0.0-beta.143

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

@@ -0,0 +1,416 @@
+/**
+ * @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 { assistCreditsFromUsd } from "@aglyn/aglyn/app-utils/assist-credits";
+import { AI_MODEL_CATALOG, AI_STEP_TIERS, aiCatalogEntry, aiBilledRatesForModel, estimateAiBilledUsd } from "./catalog.js";
+import { resolveAiRoute } from "./routing.js";
+/**
+ * THE MODEL SWITCH (AGL-2942): what a person may pick, what "Auto" means,
+ * and what a pick costs.
+ *
+ * ## Auto is the routing table
+ *
+ * A request that names no model — or names one it may not have — runs on
+ * `resolveAiRoute(kind)`, the table that maps a step kind to a catalog model
+ * (AGL-2937). Picking a model is an override of that one row for one request,
+ * never a second table.
+ *
+ * ## Three bounds, one intersection
+ *
+ * A manual pick must be served by the active provider, in a tier the org's
+ * plan offers (`AI_PLAN_MODEL_TIERS`), on the org's own restriction when it
+ * set one, and on the allotment allowlists that apply to the request. The
+ * two allowlists bound Auto too: when the routing table's answer is not on
+ * them, Auto runs on the allowed model of the same tier, else the cheapest
+ * allowed one. The plan bounds only a pick — Auto is the platform's choice,
+ * made for the step, and a plan never makes it worse.
+ *
+ * An allowlist that leaves nothing this deployment serves is not a way to
+ * switch AI off by mistake: the choice falls back to the routing table, and
+ * the permission keys are how AI is switched off on purpose.
+ *
+ * ## The price beside each option
+ *
+ * Credits per typical request at each model's list rates, over the step's
+ * measured median exchange when the workspace has enough of them, and over
+ * `AI_STEP_NOMINAL_USAGE` until it does — with the multiple of what Auto
+ * would cost for the same request. Every figure is credits, never dollars.
+ */ /** The value a request sends for "let the platform route it". */ export const AI_MODEL_AUTO = 'auto';
+/**
+ * The catalog tiers a MANUAL pick may name, per plan. Free runs on Auto only:
+ * its taste is a wall, and a pick there would spend the account's credits at
+ * a rate nobody chose on purpose. The dearest tier opens from Scale up, where
+ * the band is large enough that a slower, deeper model is a real choice
+ * rather than a fast way to end the month.
+ */ export const AI_PLAN_MODEL_TIERS = {
+    free: [],
+    starter: [
+        'fast',
+        'balanced'
+    ],
+    pro: [
+        'fast',
+        'balanced'
+    ],
+    business: [
+        'fast',
+        'balanced'
+    ],
+    scale: [
+        'fast',
+        'balanced',
+        'deep'
+    ],
+    advanced: [
+        'fast',
+        'balanced',
+        'deep'
+    ],
+    agency: [
+        'fast',
+        'balanced',
+        'deep'
+    ],
+    enterprise: [
+        'fast',
+        'balanced',
+        'deep'
+    ]
+};
+/** The plans whose admins may restrict the model set org-wide. */ export const AI_MODEL_RESTRICTION_PLANS = [
+    'agency',
+    'enterprise'
+];
+/**
+ * A typical exchange per step kind, before a workspace has measured one: the
+ * doors' own prompt sizes and output ceilings, read conservatively. Replaced
+ * by the workspace's measured median as soon as `aiMedianUsage` has enough
+ * samples, and only ever used for the price beside an option.
+ */ export const AI_STEP_NOMINAL_USAGE = {
+    'assist.chat': {
+        inputTokens: 900,
+        outputTokens: 450,
+        cacheReadTokens: 1050,
+        cacheWriteTokens: 0
+    },
+    'copy.element': {
+        inputTokens: 350,
+        outputTokens: 120,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0
+    },
+    'copy.section': {
+        inputTokens: 700,
+        outputTokens: 1400,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0
+    },
+    'copy.blog': {
+        inputTokens: 600,
+        outputTokens: 1100,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0
+    },
+    'generate.section': {
+        inputTokens: 700,
+        outputTokens: 1400,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0
+    },
+    // The text step's rules are about 500 characters — far under the balanced
+    // tier's cacheable minimum, so they are billed as input every time rather
+    // than read from a cache (`runtime/ai-prompt-cache.spec.ts` measures it).
+    'job.text': {
+        inputTokens: 450,
+        outputTokens: 500,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0
+    },
+    // The theme step: the tool's schema and both system blocks (about 7,500
+    // characters) are the cached prefix, the site's inventory and the brief ride
+    // uncached, and a full-palette answer is about 1,300 characters of JSON with
+    // as much again to think in.
+    'job.theme': {
+        inputTokens: 500,
+        outputTokens: 850,
+        cacheReadTokens: 2500,
+        cacheWriteTokens: 0
+    },
+    // The plan step: the tool's schema, the doctrine and the plan instructions
+    // (about 10,000 characters) are the cached prefix, the site inventory and the
+    // brief ride uncached, and a plan answer is about 600 characters of JSON with
+    // as much again to think in.
+    'job.plan': {
+        inputTokens: 500,
+        outputTokens: 400,
+        cacheReadTokens: 3600,
+        cacheWriteTokens: 0
+    },
+    // The component step: the doctrine, the component instructions, the
+    // component surface's catalog and the tool are the cached prefix, the site
+    // inventory, the brief and the confirmed plan ride uncached, and a card-sized
+    // answer — its tree and declared props as JSON — is about 1,700 characters.
+    // The component step's spec measures the prefix and the answer against this.
+    'job.component': {
+        inputTokens: 700,
+        outputTokens: 600,
+        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.
+    'job.crm': {
+        inputTokens: 1800,
+        outputTokens: 250,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0
+    },
+    'job.layout': {
+        inputTokens: 700,
+        outputTokens: 1000,
+        cacheReadTokens: 5600,
+        cacheWriteTokens: 0
+    },
+    'job.template': {
+        inputTokens: 1200,
+        outputTokens: 700,
+        cacheReadTokens: 6400,
+        cacheWriteTokens: 0
+    },
+    // One page section's exchange (AGL-2907), from the modules' own text: the
+    // doctrine (6,281 characters), the page instructions (1,524), the screen
+    // palette catalog (8,353) and the section tool (484) are the cached prefix;
+    // the site inventory (711 on the median golden site), the brief, the plan
+    // and the section line ride uncached; and a golden section answer is 901
+    // characters at the median, asked for with no thinking.
+    'job.page': {
+        inputTokens: 600,
+        outputTokens: 300,
+        cacheReadTokens: 4400,
+        cacheWriteTokens: 0
+    },
+    // One product's copy (AGL-2916), the exchange a bulk job makes once a
+    // product: the doctrine, the copy rules and the tool (about 9,500 characters)
+    // are the cached prefix; the product's facts, the site's categories and a
+    // photo at most 768 pixels on its long edge (about 790 tokens) ride
+    // uncached; and the answer is a description with its search listing, tags,
+    // categories and option names, about 1,300 characters of JSON.
+    'job.products': {
+        inputTokens: 1500,
+        outputTokens: 450,
+        cacheReadTokens: 2400,
+        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.
+    'job.seo': {
+        inputTokens: 1500,
+        outputTokens: 250,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0
+    },
+    // The experiment step (AGL-2914). NOTHING is cached: the A/B test rules
+    // and the acceptable-use block come to about 2,300 characters, under the
+    // balanced tier's cacheable minimum, so they are billed as input every
+    // time. The copy under test, or a test's arms with its verdict, rides with
+    // them, and a four-variant answer is about 800 characters of JSON.
+    'job.experiment': {
+        inputTokens: 900,
+        outputTokens: 300,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0
+    },
+    // The form step: the doctrine, the form instructions, the form surface's
+    // catalog (about 150 tokens) and the tool's schema are the cached prefix, the
+    // site inventory and the brief ride uncached, and a form's answer is 200 to
+    // 400 tokens of JSON with no extended thinking.
+    'job.form': {
+        inputTokens: 700,
+        outputTokens: 450,
+        cacheReadTokens: 3900,
+        cacheWriteTokens: 0
+    },
+    // An insight (AGL-2915): two exchanges, and NOTHING cached — the rules and
+    // the acceptable-use block are under the balanced tier's minimum, so they
+    // are billed as input on both. The read call carries the question and the
+    // readers' catalog (about 2,000 characters) and answers a few reads; the
+    // answer call carries four tables of about ten rows each (about 4,000
+    // characters) and answers three to five insights as JSON.
+    'job.insight': {
+        inputTokens: 2800,
+        outputTokens: 650,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0
+    },
+    // The email step, and the campaign step that shares its generation: the
+    // doctrine, the email instructions, the email palette catalog and the
+    // tool's schema (about 10,300 characters) are the cached prefix, the site
+    // inventory and the brief ride uncached, and the answer is one email's
+    // node map with three subject lines and three preheaders.
+    'job.email': {
+        inputTokens: 600,
+        outputTokens: 1500,
+        cacheReadTokens: 3600,
+        cacheWriteTokens: 0
+    },
+    'job.campaign': {
+        inputTokens: 600,
+        outputTokens: 1500,
+        cacheReadTokens: 3600,
+        cacheWriteTokens: 0
+    },
+    // The workflow step's draft (AGL-2919): the doctrine, the automation
+    // vocabulary and the tool's schema are the cached prefix (4,564 tokens, as
+    // `runtime/ai-prompt-cache.spec.ts` measures it); what the workspace can
+    // run, the site's forms and datasets and the description ride uncached; and
+    // an automation of a few steps is about 1,300 characters of JSON with as
+    // much again to think in. An explanation caches half as much.
+    'job.workflow': {
+        inputTokens: 500,
+        outputTokens: 1000,
+        cacheReadTokens: 4500,
+        cacheWriteTokens: 0
+    }
+};
+/** The fewest measured exchanges a median is taken over. */ export const AI_MEASURED_MEDIAN_MIN_SAMPLES = 3;
+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;
+    return true;
+}
+/** The 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));
+}
+/**
+ * A model's blended BILLED rate, for "the cheapest allowed" and nothing
+ * else.
+ *
+ * Billed and not provider (AGL-3015): the substitution below happens when a
+ * workspace asked for a model its bounds do not allow, and the party it
+ * falls back for is the one whose credits are drawn. "Cheapest" has to mean
+ * the same thing here as in the selector this substitution stands in for,
+ * which sorts on `creditsPerRequest`.
+ */ const blendedRate = (id)=>{
+    const rates = aiBilledRatesForModel(id);
+    return rates.inputPerToken + rates.outputPerToken;
+};
+/**
+ * The model one request runs on, or `null` when no provider is registered —
+ * which every door has refused before it asks.
+ */ export function resolveAiModelChoice(kind, requested, bounds, settings) {
+    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;
+    if (wanted && aiModelsSelectable(route.provider.id, bounds).some((entry)=>entry.id === wanted)) {
+        return {
+            model: wanted,
+            auto: false,
+            declined: null
+        };
+    }
+    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));
+        if (allowed.length && !allowed.some((entry)=>entry.id === model)) {
+            var _ref;
+            var _aiCatalogEntry;
+            const tier = (_ref = (_aiCatalogEntry = aiCatalogEntry(model)) == null ? void 0 : _aiCatalogEntry.tier) != null ? _ref : AI_STEP_TIERS[kind];
+            const sameTier = allowed.find((entry)=>entry.tier === tier);
+            const cheapest = [
+                ...allowed
+            ].sort((a, b)=>blendedRate(a.id) - blendedRate(b.id))[0];
+            model = (sameTier != null ? sameTier : cheapest).id;
+        }
+    }
+    return {
+        model,
+        auto: true,
+        declined: wanted
+    };
+}
+/** The median exchange over measured samples, or `null` below the minimum. */ export function aiMedianUsage(samples) {
+    if (samples.length < AI_MEASURED_MEDIAN_MIN_SAMPLES) return null;
+    const median = (field)=>{
+        const values = samples.map((sample)=>{
+            var _sample_field;
+            return Number((_sample_field = sample[field]) != null ? _sample_field : 0);
+        }).filter((value)=>Number.isFinite(value) && value >= 0).sort((a, b)=>a - b);
+        if (!values.length) return 0;
+        const middle = Math.floor(values.length / 2);
+        return values.length % 2 ? values[middle] : Math.round((values[middle - 1] + values[middle]) / 2);
+    };
+    return {
+        inputTokens: median('inputTokens'),
+        outputTokens: median('outputTokens'),
+        cacheReadTokens: median('cacheReadTokens'),
+        cacheWriteTokens: median('cacheWriteTokens')
+    };
+}
+/** What the selector lists for one step kind, or `null` without a provider. */ export function aiModelOptions(kind, bounds, measured, settings) {
+    var _ref;
+    const route = resolveAiRoute(kind, settings);
+    const auto = resolveAiModelChoice(kind, null, bounds, settings);
+    if (!route || !auto) return null;
+    const typical = measured != null ? measured : AI_STEP_NOMINAL_USAGE[kind];
+    // Billed rates throughout (AGL-3015): every figure below is what the
+    // workspace will be charged for picking this model — credits per request
+    // and the multiple of Auto those credits come to. Our own cost for the
+    // same request is not this selector's subject and would misstate both.
+    const autoCost = estimateAiBilledUsd(typical, auto.model);
+    const priced = (entry)=>{
+        const cost = estimateAiBilledUsd(typical, entry.id);
+        return {
+            id: entry.id,
+            label: entry.label,
+            tier: entry.tier,
+            creditsPerRequest: assistCreditsFromUsd(cost),
+            multiplier: autoCost > 0 ? Math.round(cost / autoCost * 10) / 10 : 1
+        };
+    };
+    const autoEntry = aiCatalogEntry(auto.model);
+    return {
+        auto: {
+            id: AI_MODEL_AUTO,
+            label: autoEntry ? `Auto (${autoEntry.label})` : 'Auto',
+            tier: (_ref = autoEntry == null ? void 0 : autoEntry.tier) != null ? _ref : AI_STEP_TIERS[kind],
+            creditsPerRequest: assistCreditsFromUsd(autoCost),
+            multiplier: 1,
+            model: auto.model
+        },
+        options: aiModelsSelectable(route.provider.id, bounds).map(priced).sort((a, b)=>a.creditsPerRequest - b.creditsPerRequest || a.label.localeCompare(b.label)),
+        measured: measured !== null
+    };
+}
+
+//# sourceMappingURL=model-choice.js.map

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

@@ -0,0 +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 '@aglyn/aglyn/app-utils/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  'job.layout': { inputTokens: 700, outputTokens: 1_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  'job.page': { inputTokens: 600, outputTokens: 300, 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}\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,wCAAuC;AAE5E,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,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,YAAY;QAAEH,aAAa;QAAKC,cAAc;QAAKC,iBAAiB;QAAOC,kBAAkB;IAAE;IAC/F,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;AACvG,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"}

package/src/lib/providers/openai-compatible.d.ts

@@ -0,0 +1,58 @@
+/**
+ * @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 AiProvider, type AiProviderRequest, type AiToolSchemaLimits, type AiUsage } from './contract';
+/**
+ * The OpenAI-compatible adapter (AGL-2939): `POST {base}/chat/completions`
+ * against whatever endpoint `AI_OPENAI_COMPAT_BASE_URL` names — a vendor,
+ * a gateway, a self-hosted model server — with the key in
+ * `AI_OPENAI_COMPAT_API_KEY`. The second adapter behind the contract, and
+ * the proof that the runtime, the gate, the jobs and the meter never
+ * depended on the first vendor's wire shape.
+ *
+ * What the mapping gives up, stated rather than hidden: the endpoint has
+ * no explicit cache breakpoints, so the ordered system blocks become one
+ * system message in order (the runtime's prefix rule still holds, so the
+ * order is cache-friendly for an endpoint that caches prefixes on its own),
+ * and `cacheReadTokens` is whatever `prompt_tokens_details.cached_tokens`
+ * reports. A `content_filter` finish is the endpoint's decline and becomes
+ * an `AiRefusal`, the way a refusal stop does on the other adapter.
+ */
+export declare const OPENAI_COMPATIBLE_PROVIDER_ID = "openai-compatible";
+export declare const OPENAI_COMPAT_BASE_URL_ENV = "AI_OPENAI_COMPAT_BASE_URL";
+export declare const OPENAI_COMPAT_API_KEY_ENV = "AI_OPENAI_COMPAT_API_KEY";
+/** The configured base, with any trailing slash and `/chat/completions` stripped. */
+export declare function openAiCompatibleBaseUrl(): string | null;
+/** The host the endpoint is reached at, for the subprocessor inventory. */
+export declare function openAiCompatibleHost(): string;
+/**
+ * What the strict function shape this adapter sends compiles: every property
+ * of every object listed in `required`, so no optional parameter at all — an
+ * optional field is a union with `null` instead. No other bound is stated:
+ * the endpoint is whatever `AI_OPENAI_COMPAT_BASE_URL` names, and no one
+ * count of tools or unions — nor any grammar size — holds for every endpoint
+ * that speaks the shape. A bound nothing here has watched an endpoint hold to
+ * would be a guess, and a guess in a guard is worse than a gap in one.
+ */
+export declare const OPENAI_COMPATIBLE_TOOL_SCHEMA_LIMITS: AiToolSchemaLimits;
+export declare function openAiCompatibleFailureIsRetryable(status: number | null): boolean;
+/** The endpoint's `usage` object, in the meter's shape. Missing fields read 0. */
+export declare function openAiCompatibleUsageFrom(usage: unknown): AiUsage;
+/** The chat-completions body. Exported so a spec can assert the wire shape. */
+export declare function buildOpenAiCompatibleRequestBody(input: Omit<AiProviderRequest, 'apiKey' | 'signal'> & {
+    stream: boolean;
+}): Record<string, unknown>;
+export declare const openAiCompatibleProvider: AiProvider;