feat(tools)!: catalog tools on stock routes; model policy in tenant config (CL-8170)

.changeset/config.json

@@ -34,7 +34,6 @@
     "@corbits/hub-api-client",
     "@corbits/icons",
     "@corbits/inbox",
-    "@corbits/inference-catalog",
     "@corbits/inference-settings",
     "@corbits/insights",
     "@corbits/jimmy-agent",
@@ -43,7 +42,6 @@
     "@corbits/migration-runner",
     "@corbits/mocks",
     "@corbits/notify",
-    "@corbits/ollama-adapter",
     "@corbits/plugins-ui",
     "@corbits/presence",
     "@corbits/process-provisioner",

apps/hub/package.json

@@ -36,7 +36,6 @@
     "@corbits/granola-call-workflow": "workspace:*",
     "@corbits/hub-api-client": "workspace:*",
     "@corbits/inbox": "workspace:*",
-    "@corbits/inference-catalog": "workspace:*",
     "@corbits/insights": "workspace:*",
     "@corbits/last-30-days-research-workflow": "workspace:*",
     "@corbits/mailbox": "github:corbitsdev/corbits-mailbox#118a2cd3b0b2d83abcd57b4bed1698e95ee41b6d",

apps/hub/src/index.ts

@@ -20,7 +20,6 @@ import {
 } from "@intx/db";
 import {
   asset as assetTable,
-  modelPricing,
   tenant as tenantTable,
   user as userTable,
   workflowDefinition,
@@ -105,11 +104,6 @@ import {
   createUsageSink,
   withTurnPartPersistGuard,
 } from "@corbits/insights";
-import {
-  applyInferenceCatalogMigrations,
-  createPostgresBenchModelPolicyStore,
-  createWorkflowCatalogRoutes,
-} from "@corbits/inference-catalog";
 import { generateId } from "@intx/hub-common";
 
 import {
@@ -1706,13 +1700,6 @@ export async function createHub(config: HubConfig) {
         ),
     }),
   );
-  // Bench model policy migrations still run: the store itself is read
-  // below (tenant-session bench-model-policy read), even though the
-  // dedicated read/write mount is gone.
-  await applyInferenceCatalogMigrations(config.databaseUrl);
-  const benchModelPolicy = createPostgresBenchModelPolicyStore(
-    config.databaseUrl,
-  );
   {
     const mailboxApp = new Hono<TenantEnv>();
     mountMailbox(mailboxApp, {
@@ -2210,26 +2197,6 @@ export async function createHub(config: HubConfig) {
       listMcpServers: (tenantId) => listMcpServerConnections(db, tenantId),
     }),
   );
-  // How a running agent asks what this bench can reach for a kind of work
-  // (`@corbits/catalog-tools`' `list_model_concepts` / `pick_models` /
-  // `estimate_run_cost`): the workflow-run-authenticated counterpart to
-  // the tenant-session bench-model-policy mount above. Read-only, and it
-  // takes ports rather than a db handle, so the package never learns the
-  // catalog schema.
-  app.route(
-    "/api/workflow-inference-catalog",
-    createWorkflowCatalogRoutes({
-      authenticator: createWorkflowRunAuthenticator({ db }),
-      listOfferings: (tenantId) => listVisibleOfferings(db, tenantId),
-      listPricing: async (_tenantId, offeringIds) =>
-        offeringIds.length === 0
-          ? []
-          : await db.query.modelPricing.findMany({
-              where: inArray(modelPricing.offeringId, [...offeringIds]),
-            }),
-      getPolicy: (tenantId) => benchModelPolicy.store.getPolicy(tenantId),
-    }),
-  );
   // Notify-to-reconnect for an OAuth-connected credential whose token
   // expired (Hugging Face today — see docs/onboarding-huggingface-connect.md)
   // is now `workflows/credential-expiry` (CL-8181), a schedule-triggered

apps/sidecar/src/workflow-substrate-factory/step-env.ts

@@ -446,8 +446,9 @@ export function createSidecarStepBuildEnv(
       hubRoutinesUrl: deps.hubArtifactsUrl,
       hubConnectionsUrl: deps.hubArtifactsUrl,
       // And once more under the key `@corbits/catalog-tools` declares
-      // (`requires: ["hubCatalogUrl", "sidecarToken", "address"]`) for the
-      // inference catalog's own run-authenticated surface.
+      // (`requires: ["hubCatalogUrl", "tenantId", "sidecarToken",
+      // "address"]`) for the stock tenant model-discovery and tenant reads
+      // it resolves a model chain from.
       hubCatalogUrl: deps.hubArtifactsUrl,
       hubAgentDirectoryUrl: deps.hubArtifactsUrl,
       hubChatUrl: deps.hubArtifactsUrl,
@@ -466,7 +467,7 @@ export function createSidecarStepBuildEnv(
       // The run's own tenant/principal, for the tool bundles that address
       // stock `/api/tenants/:tenantId/*` routes with the run bearer
       // (`@corbits/access-tools`, `@corbits/connections-tools`,
-      // `@corbits/workflow-authoring-tools`).
+      // `@corbits/workflow-authoring-tools`, `@corbits/catalog-tools`).
       tenantId: deps.tenantId,
       principalId: deps.principalId,
     };

packages/catalog-tools/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@corbits/catalog-tools",
   "description": "The list_model_concepts, pick_models and estimate_run_cost tools as an @intx/agent bundle: an agent asks for a model by what the work needs and gets back an ordered, priced chain of what this bench can actually reach — never a model named by hand.",
-  "version": "0.0.6",
+  "version": "0.1.0",
   "license": "LGPL-2.1-or-later",
   "type": "module",
   "exports": {
@@ -25,6 +25,7 @@
     "test": "bun test"
   },
   "dependencies": {
+    "@corbits/inference-catalog": "0.1.0",
     "arktype": "catalog:"
   },
   "peerDependencies": {

packages/catalog-tools/src/chain.ts

@@ -0,0 +1,67 @@
+// Chain resolution, in the workflow child.
+//
+// The two stock reads happen once per tool call and are then used for every
+// question that call asks — `list_model_concepts` ranks every concept off
+// one catalog read rather than one read per concept, which is what the
+// deleted hub route did per request.
+import {
+  catalogOfferingsFrom,
+  resolveModelChain,
+  type BenchModelPolicy,
+  type ChainNeed,
+  type ChainOrder,
+  type DiscoveredModel,
+  type ModelChain,
+} from "@corbits/inference-catalog";
+
+import {
+  fetchCatalog,
+  fetchModelPolicy,
+  type CatalogToolClientConfig,
+} from "./client";
+
+/** What this bench can reach and what it is willing to spend, read once. */
+export type BenchCatalog = {
+  readonly models: readonly DiscoveredModel[];
+  readonly policy: BenchModelPolicy;
+};
+
+export async function readBenchCatalog(
+  config: CatalogToolClientConfig,
+): Promise<BenchCatalog> {
+  const [models, policy] = await Promise.all([
+    fetchCatalog(config),
+    fetchModelPolicy(config),
+  ]);
+  return { models, policy };
+}
+
+export function chainFor(
+  bench: BenchCatalog,
+  need: ChainNeed,
+  order?: ChainOrder,
+  limit?: number,
+): ModelChain {
+  const { offerings, pricing } = catalogOfferingsFrom(bench.models);
+  return resolveModelChain({
+    need,
+    offerings,
+    pricing,
+    policy: bench.policy,
+    ...(order !== undefined ? { order } : {}),
+    ...(limit !== undefined ? { limit } : {}),
+  });
+}
+
+/** The one line that explains an otherwise puzzling chain: empty, entirely
+ * unpriced, or partly over this bench's ceiling. */
+export function chainNote(chain: ModelChain): string | null {
+  if (chain.entries.length === 0) return "nothing on this bench can do that";
+  if (chain.entries.every((entry) => !entry.price.known)) {
+    return "nothing on this bench is priced yet, so cheapest-first is ordered by catalog priority";
+  }
+  if (chain.entries.some((entry) => entry.overCeiling)) {
+    return "some of these cost more than this bench's ceiling for that kind of work";
+  }
+  return null;
+}