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

package/src/lib/providers/routing.js

@@ -1,3 +1,4 @@
+import { _ as _extends } from "@swc/helpers/_/_extends";
 /**
  * @license
  * Copyright 2026 Aglyn LLC
@@ -14,7 +15,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */ import { AI_SETTING_PLATFORM, AI_STEP_MODEL_SETTING } from "../plugin-config.js";
-import { AI_STEP_TIERS, aiCatalogEntry, aiDefaultModelFor } from "./catalog.js";
+import { AI_MODEL_CATALOG, AI_STEP_TIERS, aiCatalogEntry, aiDefaultModelFor } from "./catalog.js";
 import { aiProviderById, defaultAiProvider } from "./registry.js";
 /**
  * The routing table (AGL-2937, AGL-2939): step kind → provider and model,
@@ -31,6 +32,20 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
  * 3. the catalog's tier for the step kind on the resolved provider.
  */ /** The environment override for every step kind. */ export const AI_DEFAULT_MODEL_ENV = 'AI_DEFAULT_MODEL';
 /** The assistant's own override, honored before the general one. */ export const ASSIST_MODEL_ENV = 'ASSIST_MODEL';
+/**
+ * The model that serves a catalog id (AGL-3781): the id itself, or for a
+ * retired row the successor it names, followed to a current row. An id the
+ * catalog does not know is returned as it is, for the caller to refuse.
+ */ export function aiServingModelId(modelId) {
+    let current = modelId;
+    for(let hops = 0; hops < AI_MODEL_CATALOG.length; hops += 1){
+        var _aiCatalogEntry;
+        const successor = (_aiCatalogEntry = aiCatalogEntry(current)) == null ? void 0 : _aiCatalogEntry.successor;
+        if (!successor) return current;
+        current = successor;
+    }
+    return current;
+}
 /** A settings value that names a real choice, or `undefined` for the platform default. */ function chosen(settings, key) {
     const value = settings == null ? void 0 : settings[key];
     return typeof value === 'string' && value && value !== AI_SETTING_PLATFORM ? value : undefined;
@@ -52,9 +67,12 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
     var _process_env_ASSIST_MODEL_ENV, _process_env_AI_DEFAULT_MODEL_ENV;
     const provider = resolveAiProvider(settings);
     if (!provider) return undefined;
+    // A setting or an override that names a retired model is served by its
+    // successor, never by the retired model itself.
     const servedBy = (modelId)=>{
         var _aiCatalogEntry;
-        return modelId && ((_aiCatalogEntry = aiCatalogEntry(modelId)) == null ? void 0 : _aiCatalogEntry.provider) === provider.id ? modelId : undefined;
+        const serving = modelId ? aiServingModelId(modelId) : undefined;
+        return serving && ((_aiCatalogEntry = aiCatalogEntry(serving)) == null ? void 0 : _aiCatalogEntry.provider) === provider.id ? serving : undefined;
     };
     const model = (_ref = (_ref1 = (_servedBy = servedBy(chosen(settings, AI_STEP_MODEL_SETTING[kind]))) != null ? _servedBy : kind === 'assist.chat' ? servedBy((_process_env_ASSIST_MODEL_ENV = process.env[ASSIST_MODEL_ENV]) == null ? void 0 : _process_env_ASSIST_MODEL_ENV.trim()) : undefined) != null ? _ref1 : servedBy((_process_env_AI_DEFAULT_MODEL_ENV = process.env[AI_DEFAULT_MODEL_ENV]) == null ? void 0 : _process_env_AI_DEFAULT_MODEL_ENV.trim())) != null ? _ref : aiDefaultModelFor(provider.id, AI_STEP_TIERS[kind]);
     return model ? {
@@ -96,7 +114,10 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         }
     },
     'copy.element': {
-        thinking: null,
+        // One element's copy on the fast tier, whose model thinks unless told not
+        // to (AGL-3781): the ceiling holds the copy alone, as it always has, and
+        // the effort is the model's own.
+        thinking: 'off',
         effort: null,
         maxTokens: 1024,
         maxTokensBasis: 'one element of copy, about the length of the text it replaces',
@@ -157,6 +178,9 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
     },
     'job.layout': {
         // The plan was reasoned out and confirmed before the step, so it writes without thinking.
+        // A site's header and footer in the layout language think adaptively, at
+        // low since AGL-3781 (`AI_LAYOUT_LANGUAGE_THINKING`): the live round of
+        // 2026-10-14 built 13 of 13 frames at low, none cut off.
         thinking: 'off',
         effort: null,
         maxTokens: 8000,
@@ -172,8 +196,11 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         }
     },
     'job.template': {
+        // Medium since the golden A/B of 2026-10-14 (AGL-3781): both template
+        // briefs passed at medium as at the default, with about a third fewer
+        // output tokens.
         thinking: 'off',
-        effort: null,
+        effort: 'medium',
         maxTokens: 8000,
         maxTokensBasis: "tighter than the doctrine's default for a template, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way",
         eval: {
@@ -207,9 +234,11 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         // CRM by AI (AGL-2917): short answers through a strict tool, written from
         // the facts the CRM reports about the record the member opened, with no
         // thinking. The ceiling fits an inline door, so a member is answered in
-        // the request that asked.
-        thinking: null,
-        effort: null,
+        // the request that asked, and has no room for thinking: the fast tier's
+        // model thinks unless told not to, so the row says off, and its effort
+        // is the low one a short answer read off given facts needs (AGL-3781).
+        thinking: 'off',
+        effort: 'low',
         maxTokens: 700,
         maxTokensBasis: "the largest answer a CRM tool accepts: sixty columns matched to fields by number, or an email draft at its length limits, at three characters a token with room, as ai-job-crm-step.spec.ts measures it",
         eval: {
@@ -283,7 +312,10 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         // confirmed, so it writes without thinking (AGL-2907). The ceiling here
         // is the MOST a pass may ask for; each pass sizes its own down from it to
         // what fits the time the job beat can give the step, which on the
-        // balanced tier is about half of this.
+        // balanced tier is about half of this. A page in the layout language
+        // thinks adaptively, at low since AGL-3781 (`AI_LAYOUT_LANGUAGE_THINKING`):
+        // the live round of 2026-10-14 built 39 of 39 pages at low, 64 of 65
+        // answers held their rules on the first ask, none cut off.
         thinking: 'off',
         effort: null,
         maxTokens: 2000,
@@ -409,8 +441,12 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         }
     },
     'job.seo': {
-        thinking: null,
-        effort: null,
+        // A listing written from the page's own text to fixed lengths, sized for
+        // the answer alone: no thinking, which the fast tier's model does unless
+        // told not to, at the low effort a short bounded answer needs (AGL-3781).
+        // The audit's site proposal and fix batches run on the same row.
+        thinking: 'off',
+        effort: 'low',
         maxTokens: 1024,
         maxTokensBasis: "the largest listing the tool accepts, at three characters a token with a quarter again, as ai-job-seo-step.spec.ts measures it; an audit's site proposal and fix batches keep ceilings of their own, measured the same way",
         eval: {
@@ -424,8 +460,10 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         }
     },
     'job.text': {
+        // Low since the golden A/B of 2026-10-14 (AGL-3781): the tagline brief
+        // passed at low, its copy scored higher than at the default.
         thinking: 'adaptive',
-        effort: null,
+        effort: 'low',
         maxTokens: 1024,
         maxTokensBasis: 'a few paragraphs of copy; a brief asking for more gets a draft to extend',
         eval: {
@@ -439,8 +477,11 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         }
     },
     'job.theme': {
+        // Medium since the golden A/B of 2026-10-14 (AGL-3781): the theme brief
+        // passed at medium with the default's scores and half its output tokens.
+        // A site start's look runs its own setting (`ai-job-site-look.ts`).
         thinking: 'adaptive',
-        effort: null,
+        effort: 'medium',
         maxTokens: 8000,
         maxTokensBasis: 'the largest answer the theme tool accepts, about 2,500 tokens of JSON as ai-job-theme-step.spec.ts measures it, with as much again to think in',
         eval: {
@@ -454,6 +495,12 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         }
     },
     'job.plan': {
+        // A site's plan is an outline written without thinking (AGL-3594), and a
+        // paid site's asks medium (`AI_PAID_SITE_PLAN_EFFORT`, AGL-3781): the live
+        // round of 2026-10-14 planned 11 of 11 paid site briefs at medium in 16
+        // attempts to the default's 22, $0.129 to $0.164, with no refusal and no
+        // rule broken. Every other plan, and a Free site's on the fast tier, asks
+        // this row's effort — the model's default — until a round holds it lower.
         thinking: 'adaptive',
         effort: null,
         maxTokens: 8000,
@@ -473,6 +520,18 @@ import { aiProviderById, defaultAiProvider } from "./registry.js";
         }
     }
 };
+/**
+ * What a step kind's row asks of its model, as the fields a request carries:
+ * only those the row names, so a row that names none sends none and the
+ * model's own default stands.
+ */ export function aiRouteSettings(kind) {
+    const row = AI_ROUTING_TABLE[kind];
+    return _extends({}, row.thinking ? {
+        thinking: row.thinking
+    } : {}, row.effort ? {
+        effort: row.effort
+    } : {});
+}
 /** The model id alone — what a door records on the meter and the job. */ export function aiModelForStep(kind, settings) {
     const route = resolveAiRoute(kind, settings);
     if (!route) {

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

@@ -1 +1 @@
-{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/routing.ts"],"sourcesContent":["/**\n * @license\n * Copyright 2026 Aglyn LLC\n *\n * Licensed under the Apache License, Version 2.0 (the \"License\");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n *   http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an \"AS IS\" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n */\n\nimport { AI_SETTING_PLATFORM, AI_STEP_MODEL_SETTING } from '../plugin-config'\nimport {\n  AI_STEP_TIERS,\n  aiCatalogEntry,\n  aiDefaultModelFor,\n  type AiStepKind,\n} from './catalog'\nimport type { AiEvalKind } from '../runtime/ai-eval'\nimport type { AiEffort, AiProvider, AiThinking } from './contract'\nimport { aiProviderById, defaultAiProvider } from './registry'\n\n/**\n * The routing table (AGL-2937, AGL-2939): step kind → provider and model,\n * through the catalog and never a vendor literal.\n *\n * Precedence, dearest override first:\n *\n * 1. the org's `pluginSettings/ai` — a provider, and a model per step kind\n *    that must be one the provider serves;\n * 2. the deployment's environment — `AI_PROVIDER` for the provider,\n *    `AI_DEFAULT_MODEL` for every step, and `ASSIST_MODEL` for the\n *    assistant alone (the incident-response override the chat door has\n *    always honored);\n * 3. the catalog's tier for the step kind on the resolved provider.\n */\n\n/** The environment override for every step kind. */\nexport const AI_DEFAULT_MODEL_ENV = 'AI_DEFAULT_MODEL'\n/** The assistant's own override, honored before the general one. */\nexport const ASSIST_MODEL_ENV = 'ASSIST_MODEL'\n\n/** The org's resolved `pluginSettings/ai` values, as `getPluginConfig` merges them. */\nexport type AiPluginSettings = Record<string, unknown>\n\nexport interface AiRoute {\n  provider: AiProvider\n  model: string\n}\n\n/** A settings value that names a real choice, or `undefined` for the platform default. */\nfunction chosen(settings: AiPluginSettings | undefined, key: string): string | undefined {\n  const value = settings?.[key]\n  return typeof value === 'string' && value && value !== AI_SETTING_PLATFORM\n    ? value\n    : undefined\n}\n\n/** The provider a workspace runs on: its own choice when registered, else the platform's. */\nexport function resolveAiProvider(settings?: AiPluginSettings): AiProvider | undefined {\n  const wanted = chosen(settings, 'provider')\n  if (wanted) {\n    const named = aiProviderById(wanted)\n    if (named) return named\n  }\n  return defaultAiProvider()\n}\n\n/**\n * The model a step kind runs on for a workspace, on the resolved provider.\n * A chosen model on another provider is not served — the provider decides\n * what it can run — and falls back to the provider's default for the tier.\n */\nexport function resolveAiRoute(\n  kind: AiStepKind,\n  settings?: AiPluginSettings,\n): AiRoute | undefined {\n  const provider = resolveAiProvider(settings)\n  if (!provider) return undefined\n  const servedBy = (modelId: string | undefined): string | undefined =>\n    modelId && aiCatalogEntry(modelId)?.provider === provider.id ? modelId : undefined\n  const model =\n    servedBy(chosen(settings, AI_STEP_MODEL_SETTING[kind])) ??\n    (kind === 'assist.chat' ? servedBy(process.env[ASSIST_MODEL_ENV]?.trim()) : undefined) ??\n    servedBy(process.env[AI_DEFAULT_MODEL_ENV]?.trim()) ??\n    aiDefaultModelFor(provider.id, AI_STEP_TIERS[kind])\n  return model ? { provider, model } : undefined\n}\n\n/**\n * What one step kind asks of its model, beside the tier it is served from\n * (`AI_STEP_TIERS` in the catalog): whether it thinks and how hard, its\n * answer ceiling, and the eval score that tier holds on the step's golden\n * briefs (AGL-2937).\n */\nexport interface AiRoutingRow {\n  /** `null` sends no setting: the model's own default, and all a fast-tier model accepts. */\n  thinking: AiThinking | null\n  effort: AiEffort | null\n  /** The answer ceiling, thinking included. */\n  maxTokens: number\n  /** What the ceiling is sized from. */\n  maxTokensBasis: string\n  eval: AiRoutingEval\n}\n\nexport interface AiRoutingEval {\n  /** The golden-brief kinds the step answers. */\n  kinds: readonly AiEvalKind[]\n  /** `answers` scores the step's answers; `plans` scores the plans those kinds' answers carry. */\n  scores: 'answers' | 'plans'\n  /** `authored` while the score is the hand-written references'; `recorded` once a live run records answers. */\n  source: 'authored' | 'recorded'\n  passRate: number\n  /** The mean answer score; `null` for a plan score, which passes or fails. */\n  meanScore: number | null\n}\n\n/**\n * THE ROUTING TABLE (AGL-2937): what each step kind asks of the model, in one\n * place, with the evidence beside it.\n *\n * A step runs on the smallest tier that holds its eval floor. A row moves to a\n * cheaper tier, or a lower effort, only when a recorded eval\n * (`npm run eval:ai-live`) holds the floor there; while a row's score is the\n * authored references', it says so, and the row stays where production has\n * run it. `runtime/ai-eval.spec.ts` recomputes every row's score from the\n * harness, so a row cannot claim a score its briefs do not hold.\n *\n * A ceiling is sized from the p95 output the staff signals page reports for\n * the kind once production has served it; until then `maxTokensBasis` names\n * what it is sized from. The same spec holds every reference answer under its\n * row's ceiling at three characters a token, a reading that errs high for a\n * tokenizer that splits finer than four.\n */\nexport const AI_ROUTING_TABLE: Readonly<Record<AiStepKind, AiRoutingRow>> = {\n  'assist.chat': {\n    // A scoped, latency-sensitive chat turn: no thinking, the low effort rung.\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis:\n      'a docs-grounded answer of a few short paragraphs; the edit rung raises it for its tool call',\n    eval: { kinds: ['chat'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.element': {\n    thinking: null,\n    effort: null,\n    maxTokens: 1024,\n    maxTokensBasis: 'one element of copy, about the length of the text it replaces',\n    eval: { kinds: ['element'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.section': {\n    thinking: null,\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis: 'a section of four to twelve nodes written as JSON',\n    eval: { kinds: ['section'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.blog': {\n    thinking: null,\n    effort: null,\n    maxTokens: 2048,\n    maxTokensBasis: 'one blog post body in markdown-lite',\n    eval: { kinds: ['blog'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'generate.section': {\n    thinking: null,\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis: 'a section of four to twelve nodes written as JSON',\n    eval: { kinds: ['section'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.layout': {\n    // The plan was reasoned out and confirmed before the step, so it writes without thinking.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"tighter than the doctrine's default for a layout, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['layout'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.template': {\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"tighter than the doctrine's default for a template, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['template'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.component': {\n    // A component's tree and the properties it declares, in one answer, from a\n    // plan reasoned out and confirmed before the step (AGL-2908).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"the ceiling the layout and template steps keep, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['component'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.crm': {\n    // CRM by AI (AGL-2917): short answers through a strict tool, written from\n    // the facts the CRM reports about the record the member opened, with no\n    // thinking. The ceiling fits an inline door, so a member is answered in\n    // the request that asked.\n    thinking: null,\n    effort: null,\n    maxTokens: 700,\n    maxTokensBasis:\n      \"the largest answer a CRM tool accepts: sixty columns matched to fields by number, or an email draft at its length limits, at three characters a token with room, as ai-job-crm-step.spec.ts measures it\",\n    eval: { kinds: ['crm'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9938 },\n  },\n  'job.experiment': {\n    // Variants of one piece of copy, or a result explained against a verdict\n    // code reached before the request was made (AGL-2914). The judgment is\n    // already made either way, so it writes without thinking.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1500,\n    maxTokensBasis:\n      \"four variants at the tool's own lengths — a name, a headline, a body and a rationale each — written as JSON, about 3,000 characters at three characters a token with room; an explanation is a headline, four sentences and a next step, well under it, as ai-job-experiment-step.spec.ts measures both\",\n    eval: { kinds: ['experiment'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.form': {\n    // A form's two halves in one answer: the field declaration and the design\n    // that renders it, as JSON, with no extended thinking (AGL-2913). The step\n    // asks this ceiling, which is the doctrine's own for a form, and registers\n    // the time it takes (AGL-3035).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"the largest form the doctrine's output budget admits, a form's fields, its consent and routing declaration and the design that renders them written as JSON at the wordiest golden's characters a stored byte, under it with room; ai-job-form-step.spec.ts measures it\",\n    eval: { kinds: ['form'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.insight': {\n    // Readers chosen, then their tables phrased (AGL-2915): short sentences\n    // over figures code computed, each held to the rows it cites, so no\n    // extended thinking. The read call before the answer asks for at most\n    // `AI_INSIGHT_READ_MAX_TOKENS` of the same allowance.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1500,\n    maxTokensBasis:\n      'five insights at their 280-character bound, each citing three tables of ten rows, and the gap sentence, written as JSON: about 2,700 characters, under 1,000 tokens at three characters a token, with room',\n    eval: { kinds: ['insight'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.page': {\n    // One section of a page, written against a plan a member already\n    // confirmed, so it writes without thinking (AGL-2907). The ceiling here\n    // is the MOST a pass may ask for; each pass sizes its own down from it to\n    // what fits the time the job beat can give the step, which on the\n    // balanced tier is about half of this.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 2000,\n    maxTokensBasis:\n      \"a section of a page written as JSON, about forty elements at the step's measured 45 tokens an element; a pass lowers it to the worst case that fits AI_JOB_PAGE_STEP_MINIMUM_MS on the model it runs\",\n    eval: { kinds: ['page'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.products': {\n    // A product's copy from its name, text and photo, or a store's first\n    // products, categories and discounts from a brief (AGL-2916): answers held\n    // to the storefront claim rules, written with no extended thinking.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"the largest catalog the tool accepts, twelve proposed products at every bound written as JSON, which the balanced tier asks as much of as fits a beat; one product's copy asks 1,500 and categories with discounts 2,000, as ai-job-products-step.spec.ts measures them\",\n    eval: {\n      kinds: ['product', 'catalog', 'categories'],\n      scores: 'answers',\n      source: 'authored',\n      passRate: 1,\n      meanScore: 0.9922,\n    },\n  },\n  'job.email': {\n    // One email design from the email palette, written against the brief with\n    // no extended thinking (AGL-2912). A campaign job runs the same\n    // generation, so it carries the same ceiling.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"an email's node map written as JSON, with three subject lines and three preheaders beside it\",\n    eval: { kinds: ['email'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.campaign': {\n    // The campaign job generates its email through the email step, so it is\n    // routed identically; the draft campaign it also writes asks nothing of\n    // the model (AGL-2912).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"an email's node map written as JSON, with three subject lines and three preheaders beside it\",\n    eval: { kinds: ['email'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.workflow': {\n    // An automation drafted from a description, or one explained (AGL-2919):\n    // which trigger and which steps a description means is the judgment the\n    // step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 4000,\n    maxTokensBasis:\n      'the largest automation the tool accepts, 6,000 characters written out as JSON, at three characters a token with as much again to think in, as ai-job-workflow-step.spec.ts measures it',\n    eval: { kinds: ['workflow'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.logic': {\n    // A site function or variable proposed, or a function explained\n    // (AGL-3603): which conditions and arithmetic a description means is the\n    // judgment the step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis:\n      'the largest function the tool accepts — twelve parameters, twenty-four operations — is far past any description; a function of eight operations is about 2,400 characters of JSON, at three characters a token with as much again to think in, as ai-job-logic-step.spec.ts measures it',\n    eval: { kinds: ['logic'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.edit': {\n    // A change to a page or a layout the site has (AGL-3616): which elements\n    // a request means, on an outline with no selection, is the judgment the\n    // step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 6144,\n    maxTokensBasis:\n      \"the Assist edit rung's own ceiling for the proposal's tool call (ASSIST_EDIT_MAX_OUTPUT_TOKENS, 4,096: forty new elements written as name/value pairs), with half as much again to think in, as ai-job-edit-step.spec.ts holds it; the whole of it twice, as a re-ask may, still fits one beat on the balanced tier\",\n    eval: { kinds: ['edit'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.seo': {\n    thinking: null,\n    effort: null,\n    maxTokens: 1024,\n    maxTokensBasis:\n      \"the largest listing the tool accepts, at three characters a token with a quarter again, as ai-job-seo-step.spec.ts measures it; an audit's site proposal and fix batches keep ceilings of their own, measured the same way\",\n    eval: { kinds: ['seo'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.text': {\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 1024,\n    maxTokensBasis: 'a few paragraphs of copy; a brief asking for more gets a draft to extend',\n    eval: { kinds: ['text'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.theme': {\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      'the largest answer the theme tool accepts, about 2,500 tokens of JSON as ai-job-theme-step.spec.ts measures it, with as much again to think in',\n    eval: { kinds: ['theme'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.plan': {\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis: 'a plan of up to twelve screens with sixteen sections each, with room to think',\n    eval: {\n      kinds: ['page', 'template', 'component', 'layout', 'form'],\n      scores: 'plans',\n      source: 'authored',\n      passRate: 1,\n      meanScore: null,\n    },\n  },\n}\n\n/** The model id alone — what a door records on the meter and the job. */\nexport function aiModelForStep(kind: AiStepKind, settings?: AiPluginSettings): string {\n  const route = resolveAiRoute(kind, settings)\n  if (!route) {\n    // Every door gates on the provider being configured before it asks\n    // for a model, so this is a wiring fault rather than a customer path.\n    throw new Error('no AI provider is registered')\n  }\n  return route.model\n}\n"],"names":["AI_SETTING_PLATFORM","AI_STEP_MODEL_SETTING","AI_STEP_TIERS","aiCatalogEntry","aiDefaultModelFor","aiProviderById","defaultAiProvider","AI_DEFAULT_MODEL_ENV","ASSIST_MODEL_ENV","chosen","settings","key","value","undefined","resolveAiProvider","wanted","named","resolveAiRoute","kind","servedBy","process","provider","modelId","id","model","env","trim","AI_ROUTING_TABLE","thinking","effort","maxTokens","maxTokensBasis","eval","kinds","scores","source","passRate","meanScore","aiModelForStep","route","Error"],"mappings":"AAAA;;;;;;;;;;;;;;;CAeC,GAED,SAASA,mBAAmB,EAAEC,qBAAqB,QAAQ,sBAAkB;AAC7E,SACEC,aAAa,EACbC,cAAc,EACdC,iBAAiB,QAEZ,eAAW;AAGlB,SAASC,cAAc,EAAEC,iBAAiB,QAAQ,gBAAY;AAE9D;;;;;;;;;;;;;CAaC,GAED,kDAAkD,GAClD,OAAO,MAAMC,uBAAuB,mBAAkB;AACtD,kEAAkE,GAClE,OAAO,MAAMC,mBAAmB,eAAc;AAU9C,wFAAwF,GACxF,SAASC,OAAOC,QAAsC,EAAEC,GAAW;IACjE,MAAMC,QAAQF,4BAAAA,QAAU,CAACC,IAAI;IAC7B,OAAO,OAAOC,UAAU,YAAYA,SAASA,UAAUZ,sBACnDY,QACAC;AACN;AAEA,2FAA2F,GAC3F,OAAO,SAASC,kBAAkBJ,QAA2B;IAC3D,MAAMK,SAASN,OAAOC,UAAU;IAChC,IAAIK,QAAQ;QACV,MAAMC,QAAQX,eAAeU;QAC7B,IAAIC,OAAO,OAAOA;IACpB;IACA,OAAOV;AACT;AAEA;;;;CAIC,GACD,OAAO,SAASW,eACdC,IAAgB,EAChBR,QAA2B;QAOzBS,MAAAA,OAAAA;QACmCC,+BAC1BA;IAPX,MAAMC,WAAWP,kBAAkBJ;IACnC,IAAI,CAACW,UAAU,OAAOR;IACtB,MAAMM,WAAW,CAACG;YACLnB;eAAXmB,WAAWnB,EAAAA,kBAAAA,eAAemB,6BAAfnB,gBAAyBkB,QAAQ,MAAKA,SAASE,EAAE,GAAGD,UAAUT;;IAC3E,MAAMW,SACJL,QAAAA,SAAAA,YAAAA,SAASV,OAAOC,UAAUT,qBAAqB,CAACiB,KAAK,cAArDC,YACCD,SAAS,gBAAgBC,UAASC,gCAAAA,QAAQK,GAAG,CAACjB,iBAAiB,qBAA7BY,8BAA+BM,IAAI,MAAMb,qBAD5EM,QAEAA,UAASC,oCAAAA,QAAQK,GAAG,CAAClB,qBAAqB,qBAAjCa,kCAAmCM,IAAI,eAFhDP,OAGAf,kBAAkBiB,SAASE,EAAE,EAAErB,aAAa,CAACgB,KAAK;IACpD,OAAOM,QAAQ;QAAEH;QAAUG;IAAM,IAAIX;AACvC;AA+BA;;;;;;;;;;;;;;;;CAgBC,GACD,OAAO,MAAMc,mBAA+D;IAC1E,eAAe;QACb,2EAA2E;QAC3EC,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,gBAAgB;QACdT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,gBAAgB;QACdT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,aAAa;QACXT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,oBAAoB;QAClBT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,cAAc;QACZ,0FAA0F;QAC1FT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAS;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACnG;IACA,gBAAgB;QACdT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAW;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACrG;IACA,iBAAiB;QACf,2EAA2E;QAC3E,8DAA8D;QAC9DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAY;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACtG;IACA,WAAW;QACT,0EAA0E;QAC1E,wEAAwE;QACxE,wEAAwE;QACxE,0BAA0B;QAC1BT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAM;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAChG;IACA,kBAAkB;QAChB,yEAAyE;QACzE,uEAAuE;QACvE,0DAA0D;QAC1DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAa;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAClG;IACA,YAAY;QACV,0EAA0E;QAC1E,2EAA2E;QAC3E,2EAA2E;QAC3E,gCAAgC;QAChCT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,eAAe;QACb,wEAAwE;QACxE,oEAAoE;QACpE,sEAAsE;QACtE,sDAAsD;QACtDT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,YAAY;QACV,iEAAiE;QACjE,wEAAwE;QACxE,0EAA0E;QAC1E,kEAAkE;QAClE,uCAAuC;QACvCT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,gBAAgB;QACd,qEAAqE;QACrE,2EAA2E;QAC3E,oEAAoE;QACpET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YACJC,OAAO;gBAAC;gBAAW;gBAAW;aAAa;YAC3CC,QAAQ;YACRC,QAAQ;YACRC,UAAU;YACVC,WAAW;QACb;IACF;IACA,aAAa;QACX,0EAA0E;QAC1E,gEAAgE;QAChE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,gBAAgB;QACd,wEAAwE;QACxE,wEAAwE;QACxE,wBAAwB;QACxBT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,gBAAgB;QACd,yEAAyE;QACzE,wEAAwE;QACxE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAW;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACrG;IACA,aAAa;QACX,gEAAgE;QAChE,yEAAyE;QACzE,2DAA2D;QAC3DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC7F;IACA,YAAY;QACV,yEAAyE;QACzE,wEAAwE;QACxE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC5F;IACA,WAAW;QACTT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAM;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC3F;IACA,YAAY;QACVT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC5F;IACA,aAAa;QACXT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,YAAY;QACVT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YACJC,OAAO;gBAAC;gBAAQ;gBAAY;gBAAa;gBAAU;aAAO;YAC1DC,QAAQ;YACRC,QAAQ;YACRC,UAAU;YACVC,WAAW;QACb;IACF;AACF,EAAC;AAED,uEAAuE,GACvE,OAAO,SAASC,eAAepB,IAAgB,EAAER,QAA2B;IAC1E,MAAM6B,QAAQtB,eAAeC,MAAMR;IACnC,IAAI,CAAC6B,OAAO;QACV,mEAAmE;QACnE,sEAAsE;QACtE,MAAM,IAAIC,MAAM;IAClB;IACA,OAAOD,MAAMf,KAAK;AACpB"}
+{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/providers/routing.ts"],"sourcesContent":["/**\n * @license\n * Copyright 2026 Aglyn LLC\n *\n * Licensed under the Apache License, Version 2.0 (the \"License\");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n *   http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an \"AS IS\" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n */\n\nimport { AI_SETTING_PLATFORM, AI_STEP_MODEL_SETTING } from '../plugin-config'\nimport {\n  AI_MODEL_CATALOG,\n  AI_STEP_TIERS,\n  aiCatalogEntry,\n  aiDefaultModelFor,\n  type AiStepKind,\n} from './catalog'\nimport type { AiEvalKind } from '../runtime/ai-eval'\nimport type { AiEffort, AiProvider, AiThinking } from './contract'\nimport { aiProviderById, defaultAiProvider } from './registry'\n\n/**\n * The routing table (AGL-2937, AGL-2939): step kind → provider and model,\n * through the catalog and never a vendor literal.\n *\n * Precedence, dearest override first:\n *\n * 1. the org's `pluginSettings/ai` — a provider, and a model per step kind\n *    that must be one the provider serves;\n * 2. the deployment's environment — `AI_PROVIDER` for the provider,\n *    `AI_DEFAULT_MODEL` for every step, and `ASSIST_MODEL` for the\n *    assistant alone (the incident-response override the chat door has\n *    always honored);\n * 3. the catalog's tier for the step kind on the resolved provider.\n */\n\n/** The environment override for every step kind. */\nexport const AI_DEFAULT_MODEL_ENV = 'AI_DEFAULT_MODEL'\n/** The assistant's own override, honored before the general one. */\nexport const ASSIST_MODEL_ENV = 'ASSIST_MODEL'\n\n/** The org's resolved `pluginSettings/ai` values, as `getPluginConfig` merges them. */\nexport type AiPluginSettings = Record<string, unknown>\n\n/**\n * The model that serves a catalog id (AGL-3781): the id itself, or for a\n * retired row the successor it names, followed to a current row. An id the\n * catalog does not know is returned as it is, for the caller to refuse.\n */\nexport function aiServingModelId(modelId: string): string {\n  let current = modelId\n  for (let hops = 0; hops < AI_MODEL_CATALOG.length; hops += 1) {\n    const successor = aiCatalogEntry(current)?.successor\n    if (!successor) return current\n    current = successor\n  }\n  return current\n}\n\nexport interface AiRoute {\n  provider: AiProvider\n  model: string\n}\n\n/** A settings value that names a real choice, or `undefined` for the platform default. */\nfunction chosen(settings: AiPluginSettings | undefined, key: string): string | undefined {\n  const value = settings?.[key]\n  return typeof value === 'string' && value && value !== AI_SETTING_PLATFORM\n    ? value\n    : undefined\n}\n\n/** The provider a workspace runs on: its own choice when registered, else the platform's. */\nexport function resolveAiProvider(settings?: AiPluginSettings): AiProvider | undefined {\n  const wanted = chosen(settings, 'provider')\n  if (wanted) {\n    const named = aiProviderById(wanted)\n    if (named) return named\n  }\n  return defaultAiProvider()\n}\n\n/**\n * The model a step kind runs on for a workspace, on the resolved provider.\n * A chosen model on another provider is not served — the provider decides\n * what it can run — and falls back to the provider's default for the tier.\n */\nexport function resolveAiRoute(\n  kind: AiStepKind,\n  settings?: AiPluginSettings,\n): AiRoute | undefined {\n  const provider = resolveAiProvider(settings)\n  if (!provider) return undefined\n  // A setting or an override that names a retired model is served by its\n  // successor, never by the retired model itself.\n  const servedBy = (modelId: string | undefined): string | undefined => {\n    const serving = modelId ? aiServingModelId(modelId) : undefined\n    return serving && aiCatalogEntry(serving)?.provider === provider.id ? serving : undefined\n  }\n  const model =\n    servedBy(chosen(settings, AI_STEP_MODEL_SETTING[kind])) ??\n    (kind === 'assist.chat' ? servedBy(process.env[ASSIST_MODEL_ENV]?.trim()) : undefined) ??\n    servedBy(process.env[AI_DEFAULT_MODEL_ENV]?.trim()) ??\n    aiDefaultModelFor(provider.id, AI_STEP_TIERS[kind])\n  return model ? { provider, model } : undefined\n}\n\n/**\n * What one step kind asks of its model, beside the tier it is served from\n * (`AI_STEP_TIERS` in the catalog): whether it thinks and how hard, its\n * answer ceiling, and the eval score that tier holds on the step's golden\n * briefs (AGL-2937).\n */\nexport interface AiRoutingRow {\n  /**\n   * `null` sends no setting: the model's own default, which on every current\n   * model is adaptive thinking. A row whose ceiling holds the answer alone\n   * says `off` (AGL-3781).\n   */\n  thinking: AiThinking | null\n  effort: AiEffort | null\n  /** The answer ceiling, thinking included. */\n  maxTokens: number\n  /** What the ceiling is sized from. */\n  maxTokensBasis: string\n  eval: AiRoutingEval\n}\n\nexport interface AiRoutingEval {\n  /** The golden-brief kinds the step answers. */\n  kinds: readonly AiEvalKind[]\n  /** `answers` scores the step's answers; `plans` scores the plans those kinds' answers carry. */\n  scores: 'answers' | 'plans'\n  /** `authored` while the score is the hand-written references'; `recorded` once a live run records answers. */\n  source: 'authored' | 'recorded'\n  passRate: number\n  /** The mean answer score; `null` for a plan score, which passes or fails. */\n  meanScore: number | null\n}\n\n/**\n * THE ROUTING TABLE (AGL-2937): what each step kind asks of the model, in one\n * place, with the evidence beside it.\n *\n * A step runs on the smallest tier that holds its eval floor. A row moves to a\n * cheaper tier, or a lower effort, only when a recorded eval\n * (`npm run eval:ai-live`) holds the floor there; while a row's score is the\n * authored references', it says so, and the row stays where production has\n * run it. `runtime/ai-eval.spec.ts` recomputes every row's score from the\n * harness, so a row cannot claim a score its briefs do not hold.\n *\n * A ceiling is sized from the p95 output the staff signals page reports for\n * the kind once production has served it; until then `maxTokensBasis` names\n * what it is sized from. The same spec holds every reference answer under its\n * row's ceiling at three characters a token, a reading that errs high for a\n * tokenizer that splits finer than four.\n */\nexport const AI_ROUTING_TABLE: Readonly<Record<AiStepKind, AiRoutingRow>> = {\n  'assist.chat': {\n    // A scoped, latency-sensitive chat turn: no thinking, the low effort rung.\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis:\n      'a docs-grounded answer of a few short paragraphs; the edit rung raises it for its tool call',\n    eval: { kinds: ['chat'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.element': {\n    // One element's copy on the fast tier, whose model thinks unless told not\n    // to (AGL-3781): the ceiling holds the copy alone, as it always has, and\n    // the effort is the model's own.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1024,\n    maxTokensBasis: 'one element of copy, about the length of the text it replaces',\n    eval: { kinds: ['element'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.section': {\n    thinking: null,\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis: 'a section of four to twelve nodes written as JSON',\n    eval: { kinds: ['section'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'copy.blog': {\n    thinking: null,\n    effort: null,\n    maxTokens: 2048,\n    maxTokensBasis: 'one blog post body in markdown-lite',\n    eval: { kinds: ['blog'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'generate.section': {\n    thinking: null,\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis: 'a section of four to twelve nodes written as JSON',\n    eval: { kinds: ['section'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.layout': {\n    // The plan was reasoned out and confirmed before the step, so it writes without thinking.\n    // A site's header and footer in the layout language think adaptively, at\n    // low since AGL-3781 (`AI_LAYOUT_LANGUAGE_THINKING`): the live round of\n    // 2026-10-14 built 13 of 13 frames at low, none cut off.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"tighter than the doctrine's default for a layout, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['layout'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.template': {\n    // Medium since the golden A/B of 2026-10-14 (AGL-3781): both template\n    // briefs passed at medium as at the default, with about a third fewer\n    // output tokens.\n    thinking: 'off',\n    effort: 'medium',\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"tighter than the doctrine's default for a template, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['template'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.component': {\n    // A component's tree and the properties it declares, in one answer, from a\n    // plan reasoned out and confirmed before the step (AGL-2908).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"the ceiling the layout and template steps keep, so one step and its re-ask fit the job beat's budget; the tree is held to its rule 17 budget either way\",\n    eval: { kinds: ['component'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.crm': {\n    // CRM by AI (AGL-2917): short answers through a strict tool, written from\n    // the facts the CRM reports about the record the member opened, with no\n    // thinking. The ceiling fits an inline door, so a member is answered in\n    // the request that asked, and has no room for thinking: the fast tier's\n    // model thinks unless told not to, so the row says off, and its effort\n    // is the low one a short answer read off given facts needs (AGL-3781).\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 700,\n    maxTokensBasis:\n      \"the largest answer a CRM tool accepts: sixty columns matched to fields by number, or an email draft at its length limits, at three characters a token with room, as ai-job-crm-step.spec.ts measures it\",\n    eval: { kinds: ['crm'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9938 },\n  },\n  'job.experiment': {\n    // Variants of one piece of copy, or a result explained against a verdict\n    // code reached before the request was made (AGL-2914). The judgment is\n    // already made either way, so it writes without thinking.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1500,\n    maxTokensBasis:\n      \"four variants at the tool's own lengths — a name, a headline, a body and a rationale each — written as JSON, about 3,000 characters at three characters a token with room; an explanation is a headline, four sentences and a next step, well under it, as ai-job-experiment-step.spec.ts measures both\",\n    eval: { kinds: ['experiment'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.form': {\n    // A form's two halves in one answer: the field declaration and the design\n    // that renders it, as JSON, with no extended thinking (AGL-2913). The step\n    // asks this ceiling, which is the doctrine's own for a form, and registers\n    // the time it takes (AGL-3035).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"the largest form the doctrine's output budget admits, a form's fields, its consent and routing declaration and the design that renders them written as JSON at the wordiest golden's characters a stored byte, under it with room; ai-job-form-step.spec.ts measures it\",\n    eval: { kinds: ['form'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.insight': {\n    // Readers chosen, then their tables phrased (AGL-2915): short sentences\n    // over figures code computed, each held to the rows it cites, so no\n    // extended thinking. The read call before the answer asks for at most\n    // `AI_INSIGHT_READ_MAX_TOKENS` of the same allowance.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 1500,\n    maxTokensBasis:\n      'five insights at their 280-character bound, each citing three tables of ten rows, and the gap sentence, written as JSON: about 2,700 characters, under 1,000 tokens at three characters a token, with room',\n    eval: { kinds: ['insight'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.page': {\n    // One section of a page, written against a plan a member already\n    // confirmed, so it writes without thinking (AGL-2907). The ceiling here\n    // is the MOST a pass may ask for; each pass sizes its own down from it to\n    // what fits the time the job beat can give the step, which on the\n    // balanced tier is about half of this. A page in the layout language\n    // thinks adaptively, at low since AGL-3781 (`AI_LAYOUT_LANGUAGE_THINKING`):\n    // the live round of 2026-10-14 built 39 of 39 pages at low, 64 of 65\n    // answers held their rules on the first ask, none cut off.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 2000,\n    maxTokensBasis:\n      \"a section of a page written as JSON, about forty elements at the step's measured 45 tokens an element; a pass lowers it to the worst case that fits AI_JOB_PAGE_STEP_MINIMUM_MS on the model it runs\",\n    eval: { kinds: ['page'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9833 },\n  },\n  'job.products': {\n    // A product's copy from its name, text and photo, or a store's first\n    // products, categories and discounts from a brief (AGL-2916): answers held\n    // to the storefront claim rules, written with no extended thinking.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis:\n      \"the largest catalog the tool accepts, twelve proposed products at every bound written as JSON, which the balanced tier asks as much of as fits a beat; one product's copy asks 1,500 and categories with discounts 2,000, as ai-job-products-step.spec.ts measures them\",\n    eval: {\n      kinds: ['product', 'catalog', 'categories'],\n      scores: 'answers',\n      source: 'authored',\n      passRate: 1,\n      meanScore: 0.9922,\n    },\n  },\n  'job.email': {\n    // One email design from the email palette, written against the brief with\n    // no extended thinking (AGL-2912). A campaign job runs the same\n    // generation, so it carries the same ceiling.\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"an email's node map written as JSON, with three subject lines and three preheaders beside it\",\n    eval: { kinds: ['email'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.campaign': {\n    // The campaign job generates its email through the email step, so it is\n    // routed identically; the draft campaign it also writes asks nothing of\n    // the model (AGL-2912).\n    thinking: 'off',\n    effort: null,\n    maxTokens: 6000,\n    maxTokensBasis:\n      \"an email's node map written as JSON, with three subject lines and three preheaders beside it\",\n    eval: { kinds: ['email'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.workflow': {\n    // An automation drafted from a description, or one explained (AGL-2919):\n    // which trigger and which steps a description means is the judgment the\n    // step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 4000,\n    maxTokensBasis:\n      'the largest automation the tool accepts, 6,000 characters written out as JSON, at three characters a token with as much again to think in, as ai-job-workflow-step.spec.ts measures it',\n    eval: { kinds: ['workflow'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.logic': {\n    // A site function or variable proposed, or a function explained\n    // (AGL-3603): which conditions and arithmetic a description means is the\n    // judgment the step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 3000,\n    maxTokensBasis:\n      'the largest function the tool accepts — twelve parameters, twenty-four operations — is far past any description; a function of eight operations is about 2,400 characters of JSON, at three characters a token with as much again to think in, as ai-job-logic-step.spec.ts measures it',\n    eval: { kinds: ['logic'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.edit': {\n    // A change to a page or a layout the site has (AGL-3616): which elements\n    // a request means, on an outline with no selection, is the judgment the\n    // step sells, so it thinks before it answers.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 6144,\n    maxTokensBasis:\n      \"the Assist edit rung's own ceiling for the proposal's tool call (ASSIST_EDIT_MAX_OUTPUT_TOKENS, 4,096: forty new elements written as name/value pairs), with half as much again to think in, as ai-job-edit-step.spec.ts holds it; the whole of it twice, as a re-ask may, still fits one beat on the balanced tier\",\n    eval: { kinds: ['edit'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.seo': {\n    // A listing written from the page's own text to fixed lengths, sized for\n    // the answer alone: no thinking, which the fast tier's model does unless\n    // told not to, at the low effort a short bounded answer needs (AGL-3781).\n    // The audit's site proposal and fix batches run on the same row.\n    thinking: 'off',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis:\n      \"the largest listing the tool accepts, at three characters a token with a quarter again, as ai-job-seo-step.spec.ts measures it; an audit's site proposal and fix batches keep ceilings of their own, measured the same way\",\n    eval: { kinds: ['seo'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.text': {\n    // Low since the golden A/B of 2026-10-14 (AGL-3781): the tagline brief\n    // passed at low, its copy scored higher than at the default.\n    thinking: 'adaptive',\n    effort: 'low',\n    maxTokens: 1024,\n    maxTokensBasis: 'a few paragraphs of copy; a brief asking for more gets a draft to extend',\n    eval: { kinds: ['text'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 1 },\n  },\n  'job.theme': {\n    // Medium since the golden A/B of 2026-10-14 (AGL-3781): the theme brief\n    // passed at medium with the default's scores and half its output tokens.\n    // A site start's look runs its own setting (`ai-job-site-look.ts`).\n    thinking: 'adaptive',\n    effort: 'medium',\n    maxTokens: 8000,\n    maxTokensBasis:\n      'the largest answer the theme tool accepts, about 2,500 tokens of JSON as ai-job-theme-step.spec.ts measures it, with as much again to think in',\n    eval: { kinds: ['theme'], scores: 'answers', source: 'authored', passRate: 1, meanScore: 0.9688 },\n  },\n  'job.plan': {\n    // A site's plan is an outline written without thinking (AGL-3594), and a\n    // paid site's asks medium (`AI_PAID_SITE_PLAN_EFFORT`, AGL-3781): the live\n    // round of 2026-10-14 planned 11 of 11 paid site briefs at medium in 16\n    // attempts to the default's 22, $0.129 to $0.164, with no refusal and no\n    // rule broken. Every other plan, and a Free site's on the fast tier, asks\n    // this row's effort — the model's default — until a round holds it lower.\n    thinking: 'adaptive',\n    effort: null,\n    maxTokens: 8000,\n    maxTokensBasis: 'a plan of up to twelve screens with sixteen sections each, with room to think',\n    eval: {\n      kinds: ['page', 'template', 'component', 'layout', 'form'],\n      scores: 'plans',\n      source: 'authored',\n      passRate: 1,\n      meanScore: null,\n    },\n  },\n}\n\n/**\n * What a step kind's row asks of its model, as the fields a request carries:\n * only those the row names, so a row that names none sends none and the\n * model's own default stands.\n */\nexport function aiRouteSettings(kind: AiStepKind): { thinking?: AiThinking; effort?: AiEffort } {\n  const row = AI_ROUTING_TABLE[kind]\n  return {\n    ...(row.thinking ? { thinking: row.thinking } : {}),\n    ...(row.effort ? { effort: row.effort } : {}),\n  }\n}\n\n/** The model id alone — what a door records on the meter and the job. */\nexport function aiModelForStep(kind: AiStepKind, settings?: AiPluginSettings): string {\n  const route = resolveAiRoute(kind, settings)\n  if (!route) {\n    // Every door gates on the provider being configured before it asks\n    // for a model, so this is a wiring fault rather than a customer path.\n    throw new Error('no AI provider is registered')\n  }\n  return route.model\n}\n"],"names":["AI_SETTING_PLATFORM","AI_STEP_MODEL_SETTING","AI_MODEL_CATALOG","AI_STEP_TIERS","aiCatalogEntry","aiDefaultModelFor","aiProviderById","defaultAiProvider","AI_DEFAULT_MODEL_ENV","ASSIST_MODEL_ENV","aiServingModelId","modelId","current","hops","length","successor","chosen","settings","key","value","undefined","resolveAiProvider","wanted","named","resolveAiRoute","kind","servedBy","process","provider","serving","id","model","env","trim","AI_ROUTING_TABLE","thinking","effort","maxTokens","maxTokensBasis","eval","kinds","scores","source","passRate","meanScore","aiRouteSettings","row","aiModelForStep","route","Error"],"mappings":";AAAA;;;;;;;;;;;;;;;CAeC,GAED,SAASA,mBAAmB,EAAEC,qBAAqB,QAAQ,sBAAkB;AAC7E,SACEC,gBAAgB,EAChBC,aAAa,EACbC,cAAc,EACdC,iBAAiB,QAEZ,eAAW;AAGlB,SAASC,cAAc,EAAEC,iBAAiB,QAAQ,gBAAY;AAE9D;;;;;;;;;;;;;CAaC,GAED,kDAAkD,GAClD,OAAO,MAAMC,uBAAuB,mBAAkB;AACtD,kEAAkE,GAClE,OAAO,MAAMC,mBAAmB,eAAc;AAK9C;;;;CAIC,GACD,OAAO,SAASC,iBAAiBC,OAAe;IAC9C,IAAIC,UAAUD;IACd,IAAK,IAAIE,OAAO,GAAGA,OAAOX,iBAAiBY,MAAM,EAAED,QAAQ,EAAG;YAC1CT;QAAlB,MAAMW,aAAYX,kBAAAA,eAAeQ,6BAAfR,gBAAyBW,SAAS;QACpD,IAAI,CAACA,WAAW,OAAOH;QACvBA,UAAUG;IACZ;IACA,OAAOH;AACT;AAOA,wFAAwF,GACxF,SAASI,OAAOC,QAAsC,EAAEC,GAAW;IACjE,MAAMC,QAAQF,4BAAAA,QAAU,CAACC,IAAI;IAC7B,OAAO,OAAOC,UAAU,YAAYA,SAASA,UAAUnB,sBACnDmB,QACAC;AACN;AAEA,2FAA2F,GAC3F,OAAO,SAASC,kBAAkBJ,QAA2B;IAC3D,MAAMK,SAASN,OAAOC,UAAU;IAChC,IAAIK,QAAQ;QACV,MAAMC,QAAQjB,eAAegB;QAC7B,IAAIC,OAAO,OAAOA;IACpB;IACA,OAAOhB;AACT;AAEA;;;;CAIC,GACD,OAAO,SAASiB,eACdC,IAAgB,EAChBR,QAA2B;QAWzBS,MAAAA,OAAAA;QACmCC,+BAC1BA;IAXX,MAAMC,WAAWP,kBAAkBJ;IACnC,IAAI,CAACW,UAAU,OAAOR;IACtB,uEAAuE;IACvE,gDAAgD;IAChD,MAAMM,WAAW,CAACf;YAEEP;QADlB,MAAMyB,UAAUlB,UAAUD,iBAAiBC,WAAWS;QACtD,OAAOS,WAAWzB,EAAAA,kBAAAA,eAAeyB,6BAAfzB,gBAAyBwB,QAAQ,MAAKA,SAASE,EAAE,GAAGD,UAAUT;IAClF;IACA,MAAMW,SACJL,QAAAA,SAAAA,YAAAA,SAASV,OAAOC,UAAUhB,qBAAqB,CAACwB,KAAK,cAArDC,YACCD,SAAS,gBAAgBC,UAASC,gCAAAA,QAAQK,GAAG,CAACvB,iBAAiB,qBAA7BkB,8BAA+BM,IAAI,MAAMb,qBAD5EM,QAEAA,UAASC,oCAAAA,QAAQK,GAAG,CAACxB,qBAAqB,qBAAjCmB,kCAAmCM,IAAI,eAFhDP,OAGArB,kBAAkBuB,SAASE,EAAE,EAAE3B,aAAa,CAACsB,KAAK;IACpD,OAAOM,QAAQ;QAAEH;QAAUG;IAAM,IAAIX;AACvC;AAmCA;;;;;;;;;;;;;;;;CAgBC,GACD,OAAO,MAAMc,mBAA+D;IAC1E,eAAe;QACb,2EAA2E;QAC3EC,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,gBAAgB;QACd,0EAA0E;QAC1E,yEAAyE;QACzE,iCAAiC;QACjCT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,gBAAgB;QACdT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,aAAa;QACXT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,oBAAoB;QAClBT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,cAAc;QACZ,0FAA0F;QAC1F,yEAAyE;QACzE,wEAAwE;QACxE,yDAAyD;QACzDT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAS;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACnG;IACA,gBAAgB;QACd,sEAAsE;QACtE,sEAAsE;QACtE,iBAAiB;QACjBT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAW;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACrG;IACA,iBAAiB;QACf,2EAA2E;QAC3E,8DAA8D;QAC9DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAY;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACtG;IACA,WAAW;QACT,0EAA0E;QAC1E,wEAAwE;QACxE,wEAAwE;QACxE,wEAAwE;QACxE,uEAAuE;QACvE,uEAAuE;QACvET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAM;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAChG;IACA,kBAAkB;QAChB,yEAAyE;QACzE,uEAAuE;QACvE,0DAA0D;QAC1DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAa;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAClG;IACA,YAAY;QACV,0EAA0E;QAC1E,2EAA2E;QAC3E,2EAA2E;QAC3E,gCAAgC;QAChCT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,eAAe;QACb,wEAAwE;QACxE,oEAAoE;QACpE,sEAAsE;QACtE,sDAAsD;QACtDT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAU;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACpG;IACA,YAAY;QACV,iEAAiE;QACjE,wEAAwE;QACxE,0EAA0E;QAC1E,kEAAkE;QAClE,qEAAqE;QACrE,4EAA4E;QAC5E,qEAAqE;QACrE,2DAA2D;QAC3DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACjG;IACA,gBAAgB;QACd,qEAAqE;QACrE,2EAA2E;QAC3E,oEAAoE;QACpET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YACJC,OAAO;gBAAC;gBAAW;gBAAW;aAAa;YAC3CC,QAAQ;YACRC,QAAQ;YACRC,UAAU;YACVC,WAAW;QACb;IACF;IACA,aAAa;QACX,0EAA0E;QAC1E,gEAAgE;QAChE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,gBAAgB;QACd,wEAAwE;QACxE,wEAAwE;QACxE,wBAAwB;QACxBT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,gBAAgB;QACd,yEAAyE;QACzE,wEAAwE;QACxE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAW;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IACrG;IACA,aAAa;QACX,gEAAgE;QAChE,yEAAyE;QACzE,2DAA2D;QAC3DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC7F;IACA,YAAY;QACV,yEAAyE;QACzE,wEAAwE;QACxE,8CAA8C;QAC9CT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC5F;IACA,WAAW;QACT,yEAAyE;QACzE,yEAAyE;QACzE,0EAA0E;QAC1E,iEAAiE;QACjET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAM;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC3F;IACA,YAAY;QACV,uEAAuE;QACvE,6DAA6D;QAC7DT,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YAAEC,OAAO;gBAAC;aAAO;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAE;IAC5F;IACA,aAAa;QACX,wEAAwE;QACxE,yEAAyE;QACzE,oEAAoE;QACpET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBACE;QACFC,MAAM;YAAEC,OAAO;gBAAC;aAAQ;YAAEC,QAAQ;YAAWC,QAAQ;YAAYC,UAAU;YAAGC,WAAW;QAAO;IAClG;IACA,YAAY;QACV,yEAAyE;QACzE,2EAA2E;QAC3E,wEAAwE;QACxE,yEAAyE;QACzE,0EAA0E;QAC1E,0EAA0E;QAC1ET,UAAU;QACVC,QAAQ;QACRC,WAAW;QACXC,gBAAgB;QAChBC,MAAM;YACJC,OAAO;gBAAC;gBAAQ;gBAAY;gBAAa;gBAAU;aAAO;YAC1DC,QAAQ;YACRC,QAAQ;YACRC,UAAU;YACVC,WAAW;QACb;IACF;AACF,EAAC;AAED;;;;CAIC,GACD,OAAO,SAASC,gBAAgBpB,IAAgB;IAC9C,MAAMqB,MAAMZ,gBAAgB,CAACT,KAAK;IAClC,OAAO,aACDqB,IAAIX,QAAQ,GAAG;QAAEA,UAAUW,IAAIX,QAAQ;IAAC,IAAI,CAAC,GAC7CW,IAAIV,MAAM,GAAG;QAAEA,QAAQU,IAAIV,MAAM;IAAC,IAAI,CAAC;AAE/C;AAEA,uEAAuE,GACvE,OAAO,SAASW,eAAetB,IAAgB,EAAER,QAA2B;IAC1E,MAAM+B,QAAQxB,eAAeC,MAAMR;IACnC,IAAI,CAAC+B,OAAO;QACV,mEAAmE;QACnE,sEAAsE;QACtE,MAAM,IAAIC,MAAM;IAClB;IACA,OAAOD,MAAMjB,KAAK;AACpB"}

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

@@ -22,6 +22,7 @@
  * system module into a server that will never use one. Every answer is `off`
  * in a deployed server.
  */
+import type { AiEffort } from '../providers/contract';
 export type AiDevReplayMode = 'off' | 'replay' | 'refresh';
 /** The variable that turns the replay cache on, off, or to refresh. */
 export declare const AI_REPLAY_ENV = "AGLYN_AI_REPLAY";
@@ -34,6 +35,17 @@ export declare const AI_LIVE_LAUNCHER_ENV = "AI_EVAL_LIVE_LAUNCHER";
 export type AiDevEnv = Readonly<Record<string, string | undefined>>;
 /** Whether this process is a deployed server, where no development layer may run. */
 export declare function aiIsDeployedRuntime(env?: AiDevEnv): boolean;
+/** The variable every live spec is started with. */
+export declare const AI_LIVE_RUN_ENV = "AGLYN_LIVE_AI";
+/**
+ * Whether this process is a development run that asked for the live
+ * provider: through the launcher, or a live spec started with
+ * `AGLYN_LIVE_AI=1` (AGL-3781). Every development layer is on by default in
+ * such a run — a live spec started without the launcher's mark used to send
+ * every request again, at the five-minute cache and full price — and each can
+ * still be turned off by its own variable.
+ */
+export declare function aiUnderLiveRun(env?: AiDevEnv): boolean;
 /** The replay mode this process runs under; `off` in any deployed server. */
 export declare function aiDevReplayMode(env?: AiDevEnv): AiDevReplayMode;
 /**
@@ -49,5 +61,50 @@ export declare function aiDevCacheTtl(env?: AiDevEnv): '1h' | undefined;
 export declare function aiWarnIfReplayAskedInDeployment(env?: AiDevEnv): void;
 /** The variable that turns batch mode on (`runtime/ai-live-batch.ts`). */
 export declare const AI_LIVE_BATCH_ENV = "AGLYN_LIVE_AI_BATCH";
-/** Whether this process batches its live requests. */
+/**
+ * Whether this process batches its live requests: what `AGLYN_LIVE_AI_BATCH`
+ * says, and in a live run that says nothing, yes (AGL-3781). Nobody waits on
+ * an eval's answers, and a batch is half price on every token, cache reads
+ * and writes included; `AGLYN_LIVE_AI_BATCH=0` is for the quick single run an
+ * iteration wants back in seconds.
+ */
 export declare function aiLiveBatchWanted(env?: AiDevEnv): boolean;
+/** The variable that bounds how long a live run waits on one batch (AGL-3781). */
+export declare const AI_LIVE_BATCH_WAIT_ENV = "AGLYN_LIVE_AI_BATCH_WAIT_MS";
+/** How long a batch is waited on when the run names no wait. */
+export declare const AI_LIVE_BATCH_DEFAULT_WAIT_MS: number;
+/**
+ * How long a live run waits on one batch (AGL-3781): what
+ * `AGLYN_LIVE_AI_BATCH_WAIT_MS` says, or half an hour. Most batches end in
+ * minutes. One that has not by then is canceled, which keeps every answer it
+ * already has, and the requests it left unanswered are asked on their own at
+ * full price — so a slow batch costs a round part of its discount, never the
+ * round.
+ */
+export declare function aiLiveBatchWaitMs(env?: AiDevEnv): number;
+/**
+ * The most a batched live run may take: a run's requests go as a few batches
+ * one after another — the answers, their re-asks, the grades — and each is
+ * waited on for `aiLiveBatchWaitMs`, then given time to end its cancel and to
+ * have what it left asked on its own. A spec that timed out would keep none
+ * of the answers it already paid for, so the bound is the worst case.
+ */
+export declare const AI_LIVE_BATCH_TIMEOUT_MS: number;
+/** A live spec's timeout: `baseMs`, or the batched run's when this run batches. */
+export declare function aiLiveTimeoutMs(baseMs: number, env?: AiDevEnv): number;
+/**
+ * The variable an effort A/B names its arm with (AGL-3781): a step kind's
+ * effort for this run, as `job.form:medium,job.text:low`, or `*:medium` for
+ * every kind. A development run only, so the defaults stay where production
+ * runs them until a recorded round shows a lower effort holds.
+ */
+export declare const AI_EFFORT_ARM_ENV = "AGLYN_AI_EFFORT";
+/**
+ * Every step kind this development run's A/B arm names, with its effort, in
+ * kind order — `*` for every kind it does not name. A pair whose effort is no
+ * level is ignored; a kind named twice keeps its first effort, and `*` its
+ * last. Empty when the run names no arm, and in any deployed server.
+ */
+export declare function aiDevEffortArm(env?: AiDevEnv): Record<string, AiEffort>;
+/** The effort this development run's A/B arm asks `step` at, or `undefined` for the step's own. */
+export declare function aiDevEffortFor(step: string | undefined, env?: AiDevEnv): AiEffort | undefined;

package/src/lib/runtime/ai-dev-env.js

@@ -27,8 +27,18 @@
 /** Whether this process is a deployed server, where no development layer may run. */ export function aiIsDeployedRuntime(env = process.env) {
     return env['NODE_ENV'] === 'production' || env['VERCEL_ENV'] === 'production' || env['VERCEL_ENV'] === 'preview' || Boolean(env['K_SERVICE']);
 }
-/** Whether this process is a development run that asked for the live provider through the launcher. */ function underLiveLauncher(env) {
-    return Boolean(env[AI_LIVE_LAUNCHER_ENV]);
+/** The variable every live spec is started with. */ export const AI_LIVE_RUN_ENV = 'AGLYN_LIVE_AI';
+/**
+ * Whether this process is a development run that asked for the live
+ * provider: through the launcher, or a live spec started with
+ * `AGLYN_LIVE_AI=1` (AGL-3781). Every development layer is on by default in
+ * such a run — a live spec started without the launcher's mark used to send
+ * every request again, at the five-minute cache and full price — and each can
+ * still be turned off by its own variable.
+ */ export function aiUnderLiveRun(env = process.env) {
+    var _env_AI_LIVE_RUN_ENV;
+    if (aiIsDeployedRuntime(env)) return false;
+    return Boolean(env[AI_LIVE_LAUNCHER_ENV]) || String((_env_AI_LIVE_RUN_ENV = env[AI_LIVE_RUN_ENV]) != null ? _env_AI_LIVE_RUN_ENV : '').trim() === '1';
 }
 /** The replay mode this process runs under; `off` in any deployed server. */ export function aiDevReplayMode(env = process.env) {
     var _env_AI_REPLAY_ENV;
@@ -37,7 +47,7 @@
     if (raw === 'refresh') return 'refresh';
     if (raw === '1' || raw === 'on' || raw === 'true' || raw === 'replay') return 'replay';
     if (raw === '0' || raw === 'off' || raw === 'false') return 'off';
-    return underLiveLauncher(env) ? 'replay' : 'off';
+    return aiUnderLiveRun(env) ? 'replay' : 'off';
 }
 /**
  * The prompt-cache lifetime a development run asks for. An agent re-runs a
@@ -53,7 +63,7 @@
     const raw = String((_env_AI_DEV_CACHE_TTL_ENV = env[AI_DEV_CACHE_TTL_ENV]) != null ? _env_AI_DEV_CACHE_TTL_ENV : '').trim().toLowerCase();
     if (raw === '1h') return '1h';
     if (raw === '5m') return undefined;
-    return underLiveLauncher(env) ? '1h' : undefined;
+    return aiUnderLiveRun(env) ? '1h' : undefined;
 }
 /** A deployed server that was handed the variable says so once, and runs without it. */ let warnedDeployed = false;
 export function aiWarnIfReplayAskedInDeployment(env = process.env) {
@@ -63,10 +73,80 @@ export function aiWarnIfReplayAskedInDeployment(env = process.env) {
     console.error(`[ai] ${AI_REPLAY_ENV} / ${AI_DEV_CACHE_TTL_ENV} are development-only and are ignored in a deployed server`);
 }
 /** The variable that turns batch mode on (`runtime/ai-live-batch.ts`). */ export const AI_LIVE_BATCH_ENV = 'AGLYN_LIVE_AI_BATCH';
-/** Whether this process batches its live requests. */ export function aiLiveBatchWanted(env = process.env) {
+/**
+ * Whether this process batches its live requests: what `AGLYN_LIVE_AI_BATCH`
+ * says, and in a live run that says nothing, yes (AGL-3781). Nobody waits on
+ * an eval's answers, and a batch is half price on every token, cache reads
+ * and writes included; `AGLYN_LIVE_AI_BATCH=0` is for the quick single run an
+ * iteration wants back in seconds.
+ */ export function aiLiveBatchWanted(env = process.env) {
     var _env_AI_LIVE_BATCH_ENV;
     if (aiIsDeployedRuntime(env)) return false;
-    return /^(?:1|true|yes|on)$/i.test(String((_env_AI_LIVE_BATCH_ENV = env[AI_LIVE_BATCH_ENV]) != null ? _env_AI_LIVE_BATCH_ENV : '').trim());
+    const raw = String((_env_AI_LIVE_BATCH_ENV = env[AI_LIVE_BATCH_ENV]) != null ? _env_AI_LIVE_BATCH_ENV : '').trim();
+    if (/^(?:1|true|yes|on)$/i.test(raw)) return true;
+    if (/^(?:0|false|no|off)$/i.test(raw)) return false;
+    return aiUnderLiveRun(env);
+}
+/** The variable that bounds how long a live run waits on one batch (AGL-3781). */ export const AI_LIVE_BATCH_WAIT_ENV = 'AGLYN_LIVE_AI_BATCH_WAIT_MS';
+/** How long a batch is waited on when the run names no wait. */ export const AI_LIVE_BATCH_DEFAULT_WAIT_MS = 30 * 60000;
+/**
+ * How long a live run waits on one batch (AGL-3781): what
+ * `AGLYN_LIVE_AI_BATCH_WAIT_MS` says, or half an hour. Most batches end in
+ * minutes. One that has not by then is canceled, which keeps every answer it
+ * already has, and the requests it left unanswered are asked on their own at
+ * full price — so a slow batch costs a round part of its discount, never the
+ * round.
+ */ export function aiLiveBatchWaitMs(env = process.env) {
+    var _env_AI_LIVE_BATCH_WAIT_ENV;
+    const raw = String((_env_AI_LIVE_BATCH_WAIT_ENV = env[AI_LIVE_BATCH_WAIT_ENV]) != null ? _env_AI_LIVE_BATCH_WAIT_ENV : '').trim();
+    const parsed = raw ? Number(raw) : Number.NaN;
+    return Number.isFinite(parsed) && parsed >= 0 ? parsed : AI_LIVE_BATCH_DEFAULT_WAIT_MS;
+}
+/**
+ * The most a batched live run may take: a run's requests go as a few batches
+ * one after another — the answers, their re-asks, the grades — and each is
+ * waited on for `aiLiveBatchWaitMs`, then given time to end its cancel and to
+ * have what it left asked on its own. A spec that timed out would keep none
+ * of the answers it already paid for, so the bound is the worst case.
+ */ export const AI_LIVE_BATCH_TIMEOUT_MS = 5 * 60 * 60000;
+/** A live spec's timeout: `baseMs`, or the batched run's when this run batches. */ export function aiLiveTimeoutMs(baseMs, env = process.env) {
+    if (!aiLiveBatchWanted(env)) return baseMs;
+    // Four batches in a row, each waited on, canceled and finished by hand.
+    const longWait = 4 * (aiLiveBatchWaitMs(env) + 20 * 60000);
+    return Math.max(baseMs, AI_LIVE_BATCH_TIMEOUT_MS, longWait);
+}
+/**
+ * The variable an effort A/B names its arm with (AGL-3781): a step kind's
+ * effort for this run, as `job.form:medium,job.text:low`, or `*:medium` for
+ * every kind. A development run only, so the defaults stay where production
+ * runs them until a recorded round shows a lower effort holds.
+ */ export const AI_EFFORT_ARM_ENV = 'AGLYN_AI_EFFORT';
+const AI_EFFORT_LEVELS = [
+    'low',
+    'medium',
+    'high'
+];
+/**
+ * Every step kind this development run's A/B arm names, with its effort, in
+ * kind order — `*` for every kind it does not name. A pair whose effort is no
+ * level is ignored; a kind named twice keeps its first effort, and `*` its
+ * last. Empty when the run names no arm, and in any deployed server.
+ */ export function aiDevEffortArm(env = process.env) {
+    var _env_AI_EFFORT_ARM_ENV;
+    if (aiIsDeployedRuntime(env)) return {};
+    const raw = String((_env_AI_EFFORT_ARM_ENV = env[AI_EFFORT_ARM_ENV]) != null ? _env_AI_EFFORT_ARM_ENV : '').trim();
+    const arm = {};
+    for (const pair of raw ? raw.split(',') : []){
+        const [kind, level] = pair.split(':').map((part)=>part.trim());
+        if (!kind || !AI_EFFORT_LEVELS.includes(level)) continue;
+        if (kind === '*' || !(kind in arm)) arm[kind] = level;
+    }
+    return Object.fromEntries(Object.entries(arm).sort(([a], [b])=>a < b ? -1 : a > b ? 1 : 0));
+}
+/** The effort this development run's A/B arm asks `step` at, or `undefined` for the step's own. */ export function aiDevEffortFor(step, env = process.env) {
+    var _ref;
+    const arm = aiDevEffortArm(env);
+    return (_ref = step !== undefined && step !== '*' ? arm[step] : undefined) != null ? _ref : arm['*'];
 }
 
 //# sourceMappingURL=ai-dev-env.js.map

package/src/lib/runtime/ai-dev-env.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/runtime/ai-dev-env.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\n/**\n * Which development layers this process runs (AGL-3660): the replay cache,\n * the hour-long prompt cache and batch mode, each read from the environment.\n * Kept apart from `ai-dev-replay.ts`, which reads and writes files, so the\n * runtime asks these questions on every request without loading a file\n * system module into a server that will never use one. Every answer is `off`\n * in a deployed server.\n */\n\nexport type AiDevReplayMode = 'off' | 'replay' | 'refresh'\n\n/** The variable that turns the replay cache on, off, or to refresh. */\nexport const AI_REPLAY_ENV = 'AGLYN_AI_REPLAY'\n/** Where the answers are kept, when not under the repository's `.cache/ai-replay`. */\nexport const AI_REPLAY_DIR_ENV = 'AGLYN_AI_REPLAY_DIR'\n/** The prompt-cache lifetime a development run asks for: `1h` or `5m`. */\nexport const AI_DEV_CACHE_TTL_ENV = 'AGLYN_AI_CACHE_TTL'\n/** The mark the live-eval launcher leaves, which `jest.setup.js` also reads. */\nexport const AI_LIVE_LAUNCHER_ENV = 'AI_EVAL_LIVE_LAUNCHER'\n\nexport type AiDevEnv = Readonly<Record<string, string | undefined>>\n\n/** Whether this process is a deployed server, where no development layer may run. */\nexport function aiIsDeployedRuntime(env: AiDevEnv = process.env): boolean {\n  return (\n    env['NODE_ENV'] === 'production' ||\n    env['VERCEL_ENV'] === 'production' ||\n    env['VERCEL_ENV'] === 'preview' ||\n    Boolean(env['K_SERVICE'])\n  )\n}\n\n/** Whether this process is a development run that asked for the live provider through the launcher. */\nfunction underLiveLauncher(env: AiDevEnv): boolean {\n  return Boolean(env[AI_LIVE_LAUNCHER_ENV])\n}\n\n/** The replay mode this process runs under; `off` in any deployed server. */\nexport function aiDevReplayMode(env: AiDevEnv = process.env): AiDevReplayMode {\n  if (aiIsDeployedRuntime(env)) return 'off'\n  const raw = String(env[AI_REPLAY_ENV] ?? '').trim().toLowerCase()\n  if (raw === 'refresh') return 'refresh'\n  if (raw === '1' || raw === 'on' || raw === 'true' || raw === 'replay') return 'replay'\n  if (raw === '0' || raw === 'off' || raw === 'false') return 'off'\n  return underLiveLauncher(env) ? 'replay' : 'off'\n}\n\n/**\n * The prompt-cache lifetime a development run asks for. An agent re-runs a\n * live eval every ten to forty minutes while it iterates, which is past the\n * default five-minute entry and inside an hour, so under the launcher the\n * breakpoints ask for the hour: the second run reads the prefix the first one\n * wrote instead of writing it again. Production keeps the default — its\n * steps run a minute apart on the beat, inside five minutes — and never reads\n * this: `undefined` there.\n */\nexport function aiDevCacheTtl(env: AiDevEnv = process.env): '1h' | undefined {\n  if (aiIsDeployedRuntime(env)) return undefined\n  const raw = String(env[AI_DEV_CACHE_TTL_ENV] ?? '').trim().toLowerCase()\n  if (raw === '1h') return '1h'\n  if (raw === '5m') return undefined\n  return underLiveLauncher(env) ? '1h' : undefined\n}\n\n/** A deployed server that was handed the variable says so once, and runs without it. */\nlet warnedDeployed = false\nexport function aiWarnIfReplayAskedInDeployment(env: AiDevEnv = process.env): void {\n  if (warnedDeployed || !aiIsDeployedRuntime(env)) return\n  if (!env[AI_REPLAY_ENV] && !env[AI_DEV_CACHE_TTL_ENV]) return\n  warnedDeployed = true\n  console.error(\n    `[ai] ${AI_REPLAY_ENV} / ${AI_DEV_CACHE_TTL_ENV} are development-only and are ignored in a deployed server`,\n  )\n}\n\n/** The variable that turns batch mode on (`runtime/ai-live-batch.ts`). */\nexport const AI_LIVE_BATCH_ENV = 'AGLYN_LIVE_AI_BATCH'\n\n/** Whether this process batches its live requests. */\nexport function aiLiveBatchWanted(env: AiDevEnv = process.env): boolean {\n  if (aiIsDeployedRuntime(env)) return false\n  return /^(?:1|true|yes|on)$/i.test(String(env[AI_LIVE_BATCH_ENV] ?? '').trim())\n}\n"],"names":["AI_REPLAY_ENV","AI_REPLAY_DIR_ENV","AI_DEV_CACHE_TTL_ENV","AI_LIVE_LAUNCHER_ENV","aiIsDeployedRuntime","env","process","Boolean","underLiveLauncher","aiDevReplayMode","raw","String","trim","toLowerCase","aiDevCacheTtl","undefined","warnedDeployed","aiWarnIfReplayAskedInDeployment","console","error","AI_LIVE_BATCH_ENV","aiLiveBatchWanted","test"],"mappings":"AAAA;;;;;;;;;;;;;;;CAeC,GAED;;;;;;;CAOC,GAID,qEAAqE,GACrE,OAAO,MAAMA,gBAAgB,kBAAiB;AAC9C,oFAAoF,GACpF,OAAO,MAAMC,oBAAoB,sBAAqB;AACtD,wEAAwE,GACxE,OAAO,MAAMC,uBAAuB,qBAAoB;AACxD,8EAA8E,GAC9E,OAAO,MAAMC,uBAAuB,wBAAuB;AAI3D,mFAAmF,GACnF,OAAO,SAASC,oBAAoBC,MAAgBC,QAAQD,GAAG;IAC7D,OACEA,GAAG,CAAC,WAAW,KAAK,gBACpBA,GAAG,CAAC,aAAa,KAAK,gBACtBA,GAAG,CAAC,aAAa,KAAK,aACtBE,QAAQF,GAAG,CAAC,YAAY;AAE5B;AAEA,qGAAqG,GACrG,SAASG,kBAAkBH,GAAa;IACtC,OAAOE,QAAQF,GAAG,CAACF,qBAAqB;AAC1C;AAEA,2EAA2E,GAC3E,OAAO,SAASM,gBAAgBJ,MAAgBC,QAAQD,GAAG;QAEtCA;IADnB,IAAID,oBAAoBC,MAAM,OAAO;IACrC,MAAMK,MAAMC,QAAON,qBAAAA,GAAG,CAACL,cAAc,YAAlBK,qBAAsB,IAAIO,IAAI,GAAGC,WAAW;IAC/D,IAAIH,QAAQ,WAAW,OAAO;IAC9B,IAAIA,QAAQ,OAAOA,QAAQ,QAAQA,QAAQ,UAAUA,QAAQ,UAAU,OAAO;IAC9E,IAAIA,QAAQ,OAAOA,QAAQ,SAASA,QAAQ,SAAS,OAAO;IAC5D,OAAOF,kBAAkBH,OAAO,WAAW;AAC7C;AAEA;;;;;;;;CAQC,GACD,OAAO,SAASS,cAAcT,MAAgBC,QAAQD,GAAG;QAEpCA;IADnB,IAAID,oBAAoBC,MAAM,OAAOU;IACrC,MAAML,MAAMC,QAAON,4BAAAA,GAAG,CAACH,qBAAqB,YAAzBG,4BAA6B,IAAIO,IAAI,GAAGC,WAAW;IACtE,IAAIH,QAAQ,MAAM,OAAO;IACzB,IAAIA,QAAQ,MAAM,OAAOK;IACzB,OAAOP,kBAAkBH,OAAO,OAAOU;AACzC;AAEA,sFAAsF,GACtF,IAAIC,iBAAiB;AACrB,OAAO,SAASC,gCAAgCZ,MAAgBC,QAAQD,GAAG;IACzE,IAAIW,kBAAkB,CAACZ,oBAAoBC,MAAM;IACjD,IAAI,CAACA,GAAG,CAACL,cAAc,IAAI,CAACK,GAAG,CAACH,qBAAqB,EAAE;IACvDc,iBAAiB;IACjBE,QAAQC,KAAK,CACX,CAAC,KAAK,EAAEnB,cAAc,GAAG,EAAEE,qBAAqB,0DAA0D,CAAC;AAE/G;AAEA,wEAAwE,GACxE,OAAO,MAAMkB,oBAAoB,sBAAqB;AAEtD,oDAAoD,GACpD,OAAO,SAASC,kBAAkBhB,MAAgBC,QAAQD,GAAG;QAEjBA;IAD1C,IAAID,oBAAoBC,MAAM,OAAO;IACrC,OAAO,uBAAuBiB,IAAI,CAACX,QAAON,yBAAAA,GAAG,CAACe,kBAAkB,YAAtBf,yBAA0B,IAAIO,IAAI;AAC9E"}
+{"version":3,"sources":["../../../../../../../libs/plugins/ai/src/lib/runtime/ai-dev-env.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\n/**\n * Which development layers this process runs (AGL-3660): the replay cache,\n * the hour-long prompt cache and batch mode, each read from the environment.\n * Kept apart from `ai-dev-replay.ts`, which reads and writes files, so the\n * runtime asks these questions on every request without loading a file\n * system module into a server that will never use one. Every answer is `off`\n * in a deployed server.\n */\n\nimport type { AiEffort } from '../providers/contract'\n\nexport type AiDevReplayMode = 'off' | 'replay' | 'refresh'\n\n/** The variable that turns the replay cache on, off, or to refresh. */\nexport const AI_REPLAY_ENV = 'AGLYN_AI_REPLAY'\n/** Where the answers are kept, when not under the repository's `.cache/ai-replay`. */\nexport const AI_REPLAY_DIR_ENV = 'AGLYN_AI_REPLAY_DIR'\n/** The prompt-cache lifetime a development run asks for: `1h` or `5m`. */\nexport const AI_DEV_CACHE_TTL_ENV = 'AGLYN_AI_CACHE_TTL'\n/** The mark the live-eval launcher leaves, which `jest.setup.js` also reads. */\nexport const AI_LIVE_LAUNCHER_ENV = 'AI_EVAL_LIVE_LAUNCHER'\n\nexport type AiDevEnv = Readonly<Record<string, string | undefined>>\n\n/** Whether this process is a deployed server, where no development layer may run. */\nexport function aiIsDeployedRuntime(env: AiDevEnv = process.env): boolean {\n  return (\n    env['NODE_ENV'] === 'production' ||\n    env['VERCEL_ENV'] === 'production' ||\n    env['VERCEL_ENV'] === 'preview' ||\n    Boolean(env['K_SERVICE'])\n  )\n}\n\n/** The variable every live spec is started with. */\nexport const AI_LIVE_RUN_ENV = 'AGLYN_LIVE_AI'\n\n/**\n * Whether this process is a development run that asked for the live\n * provider: through the launcher, or a live spec started with\n * `AGLYN_LIVE_AI=1` (AGL-3781). Every development layer is on by default in\n * such a run — a live spec started without the launcher's mark used to send\n * every request again, at the five-minute cache and full price — and each can\n * still be turned off by its own variable.\n */\nexport function aiUnderLiveRun(env: AiDevEnv = process.env): boolean {\n  if (aiIsDeployedRuntime(env)) return false\n  return Boolean(env[AI_LIVE_LAUNCHER_ENV]) || String(env[AI_LIVE_RUN_ENV] ?? '').trim() === '1'\n}\n\n/** The replay mode this process runs under; `off` in any deployed server. */\nexport function aiDevReplayMode(env: AiDevEnv = process.env): AiDevReplayMode {\n  if (aiIsDeployedRuntime(env)) return 'off'\n  const raw = String(env[AI_REPLAY_ENV] ?? '').trim().toLowerCase()\n  if (raw === 'refresh') return 'refresh'\n  if (raw === '1' || raw === 'on' || raw === 'true' || raw === 'replay') return 'replay'\n  if (raw === '0' || raw === 'off' || raw === 'false') return 'off'\n  return aiUnderLiveRun(env) ? 'replay' : 'off'\n}\n\n/**\n * The prompt-cache lifetime a development run asks for. An agent re-runs a\n * live eval every ten to forty minutes while it iterates, which is past the\n * default five-minute entry and inside an hour, so under the launcher the\n * breakpoints ask for the hour: the second run reads the prefix the first one\n * wrote instead of writing it again. Production keeps the default — its\n * steps run a minute apart on the beat, inside five minutes — and never reads\n * this: `undefined` there.\n */\nexport function aiDevCacheTtl(env: AiDevEnv = process.env): '1h' | undefined {\n  if (aiIsDeployedRuntime(env)) return undefined\n  const raw = String(env[AI_DEV_CACHE_TTL_ENV] ?? '').trim().toLowerCase()\n  if (raw === '1h') return '1h'\n  if (raw === '5m') return undefined\n  return aiUnderLiveRun(env) ? '1h' : undefined\n}\n\n/** A deployed server that was handed the variable says so once, and runs without it. */\nlet warnedDeployed = false\nexport function aiWarnIfReplayAskedInDeployment(env: AiDevEnv = process.env): void {\n  if (warnedDeployed || !aiIsDeployedRuntime(env)) return\n  if (!env[AI_REPLAY_ENV] && !env[AI_DEV_CACHE_TTL_ENV]) return\n  warnedDeployed = true\n  console.error(\n    `[ai] ${AI_REPLAY_ENV} / ${AI_DEV_CACHE_TTL_ENV} are development-only and are ignored in a deployed server`,\n  )\n}\n\n/** The variable that turns batch mode on (`runtime/ai-live-batch.ts`). */\nexport const AI_LIVE_BATCH_ENV = 'AGLYN_LIVE_AI_BATCH'\n\n/**\n * Whether this process batches its live requests: what `AGLYN_LIVE_AI_BATCH`\n * says, and in a live run that says nothing, yes (AGL-3781). Nobody waits on\n * an eval's answers, and a batch is half price on every token, cache reads\n * and writes included; `AGLYN_LIVE_AI_BATCH=0` is for the quick single run an\n * iteration wants back in seconds.\n */\nexport function aiLiveBatchWanted(env: AiDevEnv = process.env): boolean {\n  if (aiIsDeployedRuntime(env)) return false\n  const raw = String(env[AI_LIVE_BATCH_ENV] ?? '').trim()\n  if (/^(?:1|true|yes|on)$/i.test(raw)) return true\n  if (/^(?:0|false|no|off)$/i.test(raw)) return false\n  return aiUnderLiveRun(env)\n}\n\n/** The variable that bounds how long a live run waits on one batch (AGL-3781). */\nexport const AI_LIVE_BATCH_WAIT_ENV = 'AGLYN_LIVE_AI_BATCH_WAIT_MS'\n\n/** How long a batch is waited on when the run names no wait. */\nexport const AI_LIVE_BATCH_DEFAULT_WAIT_MS = 30 * 60_000\n\n/**\n * How long a live run waits on one batch (AGL-3781): what\n * `AGLYN_LIVE_AI_BATCH_WAIT_MS` says, or half an hour. Most batches end in\n * minutes. One that has not by then is canceled, which keeps every answer it\n * already has, and the requests it left unanswered are asked on their own at\n * full price — so a slow batch costs a round part of its discount, never the\n * round.\n */\nexport function aiLiveBatchWaitMs(env: AiDevEnv = process.env): number {\n  const raw = String(env[AI_LIVE_BATCH_WAIT_ENV] ?? '').trim()\n  const parsed = raw ? Number(raw) : Number.NaN\n  return Number.isFinite(parsed) && parsed >= 0 ? parsed : AI_LIVE_BATCH_DEFAULT_WAIT_MS\n}\n\n/**\n * The most a batched live run may take: a run's requests go as a few batches\n * one after another — the answers, their re-asks, the grades — and each is\n * waited on for `aiLiveBatchWaitMs`, then given time to end its cancel and to\n * have what it left asked on its own. A spec that timed out would keep none\n * of the answers it already paid for, so the bound is the worst case.\n */\nexport const AI_LIVE_BATCH_TIMEOUT_MS = 5 * 60 * 60_000\n\n/** A live spec's timeout: `baseMs`, or the batched run's when this run batches. */\nexport function aiLiveTimeoutMs(baseMs: number, env: AiDevEnv = process.env): number {\n  if (!aiLiveBatchWanted(env)) return baseMs\n  // Four batches in a row, each waited on, canceled and finished by hand.\n  const longWait = 4 * (aiLiveBatchWaitMs(env) + 20 * 60_000)\n  return Math.max(baseMs, AI_LIVE_BATCH_TIMEOUT_MS, longWait)\n}\n\n/**\n * The variable an effort A/B names its arm with (AGL-3781): a step kind's\n * effort for this run, as `job.form:medium,job.text:low`, or `*:medium` for\n * every kind. A development run only, so the defaults stay where production\n * runs them until a recorded round shows a lower effort holds.\n */\nexport const AI_EFFORT_ARM_ENV = 'AGLYN_AI_EFFORT'\n\nconst AI_EFFORT_LEVELS: readonly AiEffort[] = ['low', 'medium', 'high']\n\n/**\n * Every step kind this development run's A/B arm names, with its effort, in\n * kind order — `*` for every kind it does not name. A pair whose effort is no\n * level is ignored; a kind named twice keeps its first effort, and `*` its\n * last. Empty when the run names no arm, and in any deployed server.\n */\nexport function aiDevEffortArm(env: AiDevEnv = process.env): Record<string, AiEffort> {\n  if (aiIsDeployedRuntime(env)) return {}\n  const raw = String(env[AI_EFFORT_ARM_ENV] ?? '').trim()\n  const arm: Record<string, AiEffort> = {}\n  for (const pair of raw ? raw.split(',') : []) {\n    const [kind, level] = pair.split(':').map((part) => part.trim())\n    if (!kind || !AI_EFFORT_LEVELS.includes(level as AiEffort)) continue\n    if (kind === '*' || !(kind in arm)) arm[kind] = level as AiEffort\n  }\n  return Object.fromEntries(Object.entries(arm).sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0)))\n}\n\n/** The effort this development run's A/B arm asks `step` at, or `undefined` for the step's own. */\nexport function aiDevEffortFor(step: string | undefined, env: AiDevEnv = process.env): AiEffort | undefined {\n  const arm = aiDevEffortArm(env)\n  return (step !== undefined && step !== '*' ? arm[step] : undefined) ?? arm['*']\n}\n"],"names":["AI_REPLAY_ENV","AI_REPLAY_DIR_ENV","AI_DEV_CACHE_TTL_ENV","AI_LIVE_LAUNCHER_ENV","aiIsDeployedRuntime","env","process","Boolean","AI_LIVE_RUN_ENV","aiUnderLiveRun","String","trim","aiDevReplayMode","raw","toLowerCase","aiDevCacheTtl","undefined","warnedDeployed","aiWarnIfReplayAskedInDeployment","console","error","AI_LIVE_BATCH_ENV","aiLiveBatchWanted","test","AI_LIVE_BATCH_WAIT_ENV","AI_LIVE_BATCH_DEFAULT_WAIT_MS","aiLiveBatchWaitMs","parsed","Number","NaN","isFinite","AI_LIVE_BATCH_TIMEOUT_MS","aiLiveTimeoutMs","baseMs","longWait","Math","max","AI_EFFORT_ARM_ENV","AI_EFFORT_LEVELS","aiDevEffortArm","arm","pair","split","kind","level","map","part","includes","Object","fromEntries","entries","sort","a","b","aiDevEffortFor","step"],"mappings":"AAAA;;;;;;;;;;;;;;;CAeC,GAED;;;;;;;CAOC,GAMD,qEAAqE,GACrE,OAAO,MAAMA,gBAAgB,kBAAiB;AAC9C,oFAAoF,GACpF,OAAO,MAAMC,oBAAoB,sBAAqB;AACtD,wEAAwE,GACxE,OAAO,MAAMC,uBAAuB,qBAAoB;AACxD,8EAA8E,GAC9E,OAAO,MAAMC,uBAAuB,wBAAuB;AAI3D,mFAAmF,GACnF,OAAO,SAASC,oBAAoBC,MAAgBC,QAAQD,GAAG;IAC7D,OACEA,GAAG,CAAC,WAAW,KAAK,gBACpBA,GAAG,CAAC,aAAa,KAAK,gBACtBA,GAAG,CAAC,aAAa,KAAK,aACtBE,QAAQF,GAAG,CAAC,YAAY;AAE5B;AAEA,kDAAkD,GAClD,OAAO,MAAMG,kBAAkB,gBAAe;AAE9C;;;;;;;CAOC,GACD,OAAO,SAASC,eAAeJ,MAAgBC,QAAQD,GAAG;QAEJA;IADpD,IAAID,oBAAoBC,MAAM,OAAO;IACrC,OAAOE,QAAQF,GAAG,CAACF,qBAAqB,KAAKO,QAAOL,uBAAAA,GAAG,CAACG,gBAAgB,YAApBH,uBAAwB,IAAIM,IAAI,OAAO;AAC7F;AAEA,2EAA2E,GAC3E,OAAO,SAASC,gBAAgBP,MAAgBC,QAAQD,GAAG;QAEtCA;IADnB,IAAID,oBAAoBC,MAAM,OAAO;IACrC,MAAMQ,MAAMH,QAAOL,qBAAAA,GAAG,CAACL,cAAc,YAAlBK,qBAAsB,IAAIM,IAAI,GAAGG,WAAW;IAC/D,IAAID,QAAQ,WAAW,OAAO;IAC9B,IAAIA,QAAQ,OAAOA,QAAQ,QAAQA,QAAQ,UAAUA,QAAQ,UAAU,OAAO;IAC9E,IAAIA,QAAQ,OAAOA,QAAQ,SAASA,QAAQ,SAAS,OAAO;IAC5D,OAAOJ,eAAeJ,OAAO,WAAW;AAC1C;AAEA;;;;;;;;CAQC,GACD,OAAO,SAASU,cAAcV,MAAgBC,QAAQD,GAAG;QAEpCA;IADnB,IAAID,oBAAoBC,MAAM,OAAOW;IACrC,MAAMH,MAAMH,QAAOL,4BAAAA,GAAG,CAACH,qBAAqB,YAAzBG,4BAA6B,IAAIM,IAAI,GAAGG,WAAW;IACtE,IAAID,QAAQ,MAAM,OAAO;IACzB,IAAIA,QAAQ,MAAM,OAAOG;IACzB,OAAOP,eAAeJ,OAAO,OAAOW;AACtC;AAEA,sFAAsF,GACtF,IAAIC,iBAAiB;AACrB,OAAO,SAASC,gCAAgCb,MAAgBC,QAAQD,GAAG;IACzE,IAAIY,kBAAkB,CAACb,oBAAoBC,MAAM;IACjD,IAAI,CAACA,GAAG,CAACL,cAAc,IAAI,CAACK,GAAG,CAACH,qBAAqB,EAAE;IACvDe,iBAAiB;IACjBE,QAAQC,KAAK,CACX,CAAC,KAAK,EAAEpB,cAAc,GAAG,EAAEE,qBAAqB,0DAA0D,CAAC;AAE/G;AAEA,wEAAwE,GACxE,OAAO,MAAMmB,oBAAoB,sBAAqB;AAEtD;;;;;;CAMC,GACD,OAAO,SAASC,kBAAkBjB,MAAgBC,QAAQD,GAAG;QAExCA;IADnB,IAAID,oBAAoBC,MAAM,OAAO;IACrC,MAAMQ,MAAMH,QAAOL,yBAAAA,GAAG,CAACgB,kBAAkB,YAAtBhB,yBAA0B,IAAIM,IAAI;IACrD,IAAI,uBAAuBY,IAAI,CAACV,MAAM,OAAO;IAC7C,IAAI,wBAAwBU,IAAI,CAACV,MAAM,OAAO;IAC9C,OAAOJ,eAAeJ;AACxB;AAEA,gFAAgF,GAChF,OAAO,MAAMmB,yBAAyB,8BAA6B;AAEnE,8DAA8D,GAC9D,OAAO,MAAMC,gCAAgC,KAAK,MAAM;AAExD;;;;;;;CAOC,GACD,OAAO,SAASC,kBAAkBrB,MAAgBC,QAAQD,GAAG;QACxCA;IAAnB,MAAMQ,MAAMH,QAAOL,8BAAAA,GAAG,CAACmB,uBAAuB,YAA3BnB,8BAA+B,IAAIM,IAAI;IAC1D,MAAMgB,SAASd,MAAMe,OAAOf,OAAOe,OAAOC,GAAG;IAC7C,OAAOD,OAAOE,QAAQ,CAACH,WAAWA,UAAU,IAAIA,SAASF;AAC3D;AAEA;;;;;;CAMC,GACD,OAAO,MAAMM,2BAA2B,IAAI,KAAK,MAAM;AAEvD,iFAAiF,GACjF,OAAO,SAASC,gBAAgBC,MAAc,EAAE5B,MAAgBC,QAAQD,GAAG;IACzE,IAAI,CAACiB,kBAAkBjB,MAAM,OAAO4B;IACpC,wEAAwE;IACxE,MAAMC,WAAW,IAAKR,CAAAA,kBAAkBrB,OAAO,KAAK,KAAK;IACzD,OAAO8B,KAAKC,GAAG,CAACH,QAAQF,0BAA0BG;AACpD;AAEA;;;;;CAKC,GACD,OAAO,MAAMG,oBAAoB,kBAAiB;AAElD,MAAMC,mBAAwC;IAAC;IAAO;IAAU;CAAO;AAEvE;;;;;CAKC,GACD,OAAO,SAASC,eAAelC,MAAgBC,QAAQD,GAAG;QAErCA;IADnB,IAAID,oBAAoBC,MAAM,OAAO,CAAC;IACtC,MAAMQ,MAAMH,QAAOL,yBAAAA,GAAG,CAACgC,kBAAkB,YAAtBhC,yBAA0B,IAAIM,IAAI;IACrD,MAAM6B,MAAgC,CAAC;IACvC,KAAK,MAAMC,QAAQ5B,MAAMA,IAAI6B,KAAK,CAAC,OAAO,EAAE,CAAE;QAC5C,MAAM,CAACC,MAAMC,MAAM,GAAGH,KAAKC,KAAK,CAAC,KAAKG,GAAG,CAAC,CAACC,OAASA,KAAKnC,IAAI;QAC7D,IAAI,CAACgC,QAAQ,CAACL,iBAAiBS,QAAQ,CAACH,QAAoB;QAC5D,IAAID,SAAS,OAAO,CAAEA,CAAAA,QAAQH,GAAE,GAAIA,GAAG,CAACG,KAAK,GAAGC;IAClD;IACA,OAAOI,OAAOC,WAAW,CAACD,OAAOE,OAAO,CAACV,KAAKW,IAAI,CAAC,CAAC,CAACC,EAAE,EAAE,CAACC,EAAE,GAAMD,IAAIC,IAAI,CAAC,IAAID,IAAIC,IAAI,IAAI;AAC7F;AAEA,iGAAiG,GACjG,OAAO,SAASC,eAAeC,IAAwB,EAAElD,MAAgBC,QAAQD,GAAG;QAE1EkD;IADR,MAAMf,MAAMD,eAAelC;IAC3B,QAAQkD,OAAAA,SAASvC,aAAauC,SAAS,MAAMf,GAAG,CAACe,KAAK,GAAGvC,qBAAjDuC,OAA+Df,GAAG,CAAC,IAAI;AACjF"}

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

@@ -108,9 +108,13 @@ export declare function resetAiDevWarmPrefixes(): void;
  * the live ones cost at list, and how many batches carried them. A table
  * whose `live` is 0 proved nothing new — every prompt in it had been answered
  * before — and is not a live pass; `AGLYN_AI_REPLAY=refresh` asks again.
+ * `batchedRequests` are the ones a batch answered, at half price;
+ * `batchFallbacks` are the ones a batch left unanswered past its wait and
+ * that were asked on their own, at full price (AGL-3781). Both count in `live`.
  */
 export declare function aiLiveRunLedger(env?: Env): AiDevReplayStats & {
     batches: number;
     batchedRequests: number;
+    batchFallbacks: number;
     allReplayed: boolean;
 };

package/src/lib/runtime/ai-dev-replay.js

@@ -300,12 +300,16 @@ export async function aiDevWarmPrefixFirst(prefixKey, ttlMs, send, now = Date.no
  * the live ones cost at list, and how many batches carried them. A table
  * whose `live` is 0 proved nothing new — every prompt in it had been answered
  * before — and is not a live pass; `AGLYN_AI_REPLAY=refresh` asks again.
+ * `batchedRequests` are the ones a batch answered, at half price;
+ * `batchFallbacks` are the ones a batch left unanswered past its wait and
+ * that were asked on their own, at full price (AGL-3781). Both count in `live`.
  */ export function aiLiveRunLedger(env = process.env) {
     const replay = aiDevReplayStats(env);
     const batch = aiLiveBatchStats();
     return _extends({}, replay, {
         batches: batch.batches,
         batchedRequests: batch.requests,
+        batchFallbacks: batch.fallbacks,
         allReplayed: replay.live === 0 && replay.replayed > 0
     });
 }