ag-ai 0.4.2 → 0.4.3

package/dist/k8s/opencode.js

@@ -1,8 +1,15 @@
 "use strict";
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.opencodeStack = void 0;
+const node_fs_1 = require("node:fs");
+const node_path_1 = __importDefault(require("node:path"));
 const ag_rpi_mega_1 = require("ag-rpi-mega");
 const common_1 = require("./common");
+/** The runtime's tier table is the single source of truth for which model and effort each tier runs. */
+const MODEL_TIERS = JSON.parse((0, node_fs_1.readFileSync)(node_path_1.default.join(__dirname, '..', '..', 'runtime', 'model-tiers.json'), 'utf8'));
 /** The published OpenCode CLI is installed into a writable emptyDir at pod start. */
 const OPENCODE_PACKAGE = 'opencode-ai@1.18.31';
 const OPENCODE_PREFIX = '/opt/opencode';
@@ -49,19 +56,17 @@ const opencodeStack = (context) => {
             (0, common_1.env)('XDG_CACHE_HOME', '/data/opencode/cache'),
             // The unrestricted /provider response processes thousands of models from
             // unconfigured providers on every request, exceeding the runner's budget.
-            // Each advertised model carries the reasoning effort its consumer asks
-            // for: sol at high for draft copy, 5.5 at high for draft images, and luna
-            // at max for the client's cluster fallback. OpenCode merges these entries
-            // with its model catalogue; it does not replace it.
+            // Each advertised model carries the reasoning effort of the tier that
+            // selects it. OpenCode merges these entries with its model catalogue; it
+            // does not replace it.
             (0, common_1.env)('OPENCODE_CONFIG_CONTENT', JSON.stringify({
                 enabled_providers: ['openai', 'google'],
                 provider: {
                     openai: {
-                        models: {
-                            'gpt-5.6-sol': { options: { reasoningEffort: 'high' } },
-                            'gpt-5.6-luna': { options: { reasoningEffort: 'max' } },
-                            'gpt-5.5': { options: { reasoningEffort: 'high' } },
-                        },
+                        models: Object.fromEntries(Object.values(MODEL_TIERS).map(({ model, reasoningEffort }) => [
+                            model,
+                            { options: { reasoningEffort } },
+                        ])),
                     },
                 },
             })),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ag-ai",
-  "version": "0.4.2",
+  "version": "0.4.3",
   "description": "Provider-neutral AI client and cluster runtime for the rpi-mega cluster.",
   "license": "ISC",
   "author": "admin@gec.dev",

package/runtime/model-tiers.json

@@ -0,0 +1,4 @@
+{
+  "fast": { "model": "gpt-6-luna", "reasoningEffort": "medium" },
+  "quality": { "model": "gpt-6.1-sol", "reasoningEffort": "medium" }
+}

package/runtime/protocol.mjs

@@ -3,6 +3,15 @@
 // differ.
 
 import { createHash } from 'node:crypto';
+import { readFileSync } from 'node:fs';
+
+/**
+ * The only place that names a model: each `prefer` tier maps to one model and
+ * effort. Callers send a tier; the model is resolved when the call is made.
+ */
+const MODEL_TIERS = JSON.parse(
+  readFileSync(new URL('./model-tiers.json', import.meta.url), 'utf8'),
+);
 
 const GENERATION_VERSION = 1;
 const MAX_TEXT_BYTES = 2 * 1024 * 1024;
@@ -257,21 +266,9 @@ const normalizeGenerationRequest = (value, models) => {
   if (prefer !== undefined && prefer !== 'fast' && prefer !== 'quality') {
     throw new GenerationProtocolError(400, 'invalid_request', 'prefer is invalid');
   }
-  let requestedModel;
-  if (request.model !== undefined) {
-    requestedModel = boundedString(request.model, 160, 'model');
-  } else if (prefer === 'fast') {
-    requestedModel = catalog.find((candidate) => candidate.prefer.includes('fast'))?.id;
-    requestedModel ??= catalog.find((candidate) => /(?:mini|luna)$/i.test(candidate.id))?.id;
-    if (!requestedModel) throw unsupported('fast model preference');
-  } else {
-    requestedModel =
-      catalog.find((candidate) => candidate.isDefault && candidate.prefer.includes('quality'))
-        ?.id ??
-      catalog.find((candidate) => candidate.isDefault)?.id ??
-      catalog.find((candidate) => candidate.prefer.includes('quality'))?.id ??
-      catalog[0].id;
-  }
+  const tier = MODEL_TIERS[prefer ?? 'quality'];
+  const requestedModel =
+    request.model === undefined ? tier.model : boundedString(request.model, 160, 'model');
   const model =
     catalog.find((candidate) => candidate.id === requestedModel) ??
     (requestedModel.includes('/') ? undefined : uniqueSuffixModel(catalog, requestedModel));
@@ -290,7 +287,9 @@ const normalizeGenerationRequest = (value, models) => {
   const output = normalizeOutput(request.output, capabilities);
   const reasoningEffort =
     request.reasoningEffort === undefined
-      ? (capabilities.defaultReasoningEffort ?? capabilities.supportedReasoningEfforts[0])
+      ? request.model === undefined
+        ? tier.reasoningEffort
+        : (capabilities.defaultReasoningEffort ?? capabilities.supportedReasoningEfforts[0])
       : boundedString(request.reasoningEffort, 32, 'reasoningEffort');
   if (!reasoningEffort) {
     throw new GenerationProtocolError(
@@ -358,6 +357,7 @@ const publicModelCatalog = (models, { ready = true } = {}) => ({
 
 export {
   GENERATION_VERSION,
+  MODEL_TIERS,
   MAX_INPUT_ATTACHMENT_BYTES,
   MAX_INPUT_TOTAL_BYTES,
   GenerationProtocolError,