@query-farm/vgi 0.43.0 → 0.44.0

package/src/protocol/handlers/shared.ts

@@ -6,6 +6,7 @@ import type { OverloadContext } from "../../functions/registry.js";
 import { batchToScalarDict, deserializeBatch, serializeBatch, batchFromColumns } from "../../util/arrow/index.js";
 import { toUint8Array } from "../../util/bytes.js";
 import { normalizeSchemaPath } from "../../schema-path.js";
+import { conformRecord } from "../../codec/conform.js";
 
 // The Python vgi-rpc framework wraps ALL non-void unary results in a single
 // "result" column. For ArrowSerializableDataclass types, the result is serialized
@@ -39,7 +40,8 @@ export function normalizeRequestSchemaPaths(params: Record<string, any>): void {
 
 /**
  * Wrap a dict of values into a single "result" Binary column.
- * Builds a 1-row batch from the values using the given schema,
+ * Builds a 1-row batch from the values using the given schema (conformed to
+ * the protocol's container rules, see `conformRecord`),
  * serializes it to IPC bytes, and returns { result: bytes }.
  */
 export function wrapResult(
@@ -47,8 +49,12 @@ export function wrapResult(
   innerSchema: VgiSchema | import("../../arrow/index.js").VgiSchema,
 ): { result: Uint8Array } {
   const a = innerSchema as VgiSchema;
+  // Python's container rules (a non-nullable list/map is never null, list
+  // items never are): `conformRecord` writes `[]` where a handler left a list
+  // out, and an unset list item as its zero value.
+  const row = conformRecord(a, values);
   const batch = batchFromColumns(
-    Object.fromEntries(a.fields.map(f => [f.name, [values[f.name] ?? null]])),
+    Object.fromEntries(a.fields.map(f => [f.name, [row[f.name] ?? null]])),
     a,
   );
   return { result: serializeBatch(batch) };

package/src/protocol/serializers/bind.ts

@@ -72,7 +72,9 @@ export function serializeBindRequest(req: BindRequest): VgiBatch {
   return buildSingleRowBatch(BIND_REQUEST_SCHEMA, {
     function_name: req.function_name,
     arguments: serializeArguments(req.arguments),
-    function_type: req.function_type,
+    // The wire carries the Python enum member NAME ("TABLE"), not the TS
+    // enum's value ("table"); each FunctionType value upper-cases to its name.
+    function_type: req.function_type == null ? req.function_type : String(req.function_type).toUpperCase(),
     input_schema: req.input_schema ? serializeSchema(req.input_schema) : null,
     settings: req.settings ? serializeBatch(req.settings) : null,
     secrets: req.secrets ? serializeBatch(req.secrets) : null,

package/src/protocol/serializers/shared.ts

@@ -8,6 +8,7 @@ import {
   batchFromColumns,
 } from "../../arrow/index.js";
 import { toUint8Array as toUint8ArrayBase } from "../../util/bytes.js";
+import { conformRecord } from "../../codec/conform.js";
 
 /**
  * Coerce a wire value (string / Uint8Array / Buffer / etc.) to a Uint8Array.
@@ -25,7 +26,8 @@ export function toUint8Array(val: any): Uint8Array {
 
 /**
  * Build a single-row RecordBatch from a {fieldName: value} dict matching the
- * given schema. Auto-coerces JS numbers to BigInt for Int64 fields.
+ * given schema. Auto-coerces JS numbers to BigInt for Int64 fields, and
+ * applies the protocol's container rules (`conformRecord`).
  */
 export function buildSingleRowBatch(
   schema: VgiSchema,
@@ -35,8 +37,9 @@ export function buildSingleRowBatch(
   // declares one — vectorFromArray/columnFromArray rejects plain Number for
   // 64-bit ints under both backends.
   const cols: Record<string, any[]> = {};
+  const row = conformRecord(schema, values);
   for (const f of schema.fields) {
-    let val = values[f.name];
+    let val = row[f.name];
     if (isInt(f.type) && (f.type as any).bitWidth === 64 && typeof val === "number") {
       val = BigInt(val);
     }

package/src/rpc-server.ts

@@ -13,7 +13,7 @@
 //   2. The worker's `hostedProtocols()`, in the order returned -- on EVERY
 //      transport.
 //   3. `vgi_rpc.Reflection.v1` -- on every transport.
-//   4. `vgi_rpc.Identity.v1` -- HTTP only, and only when the worker supplies
+//   4. `vgi_rpc.Identity.v2` -- HTTP only, and only when the worker supplies
 //      `resolveToken` and/or `mintGrant` (or grant keys are configured).
 //   5. `vgi.attach_tickets.v1` -- HTTP only, and only when the signing key was
 //      configured explicitly (never a generated one) *and* the worker can
@@ -44,7 +44,7 @@ export type ServerTransport = "pipe" | "unix" | "tcp" | "iroh" | "stream" | "htt
 const TRANSPORTS: ReadonlySet<string> = new Set<ServerTransport>(["pipe", "unix", "tcp", "iroh", "stream", "http"]);
 
 /** Transports that authenticate their callers, and so may host
- *  `vgi_rpc.Identity.v1`. Identity answers questions about *other* callers'
+ *  `vgi_rpc.Identity.v2`. Identity answers questions about *other* callers'
  *  credentials; its allowlist is a list of principals, which a transport with
  *  no caller identity cannot check. */
 const IDENTITY_TRANSPORTS: ReadonlySet<string> = new Set<ServerTransport>(["http"]);
@@ -78,14 +78,14 @@ export interface HostingOptions {
    *
    * Each needs a distinct wire name, which may not be `vgi.v2` and may not
    * use the reserved `vgi_rpc.` prefix: reflection is hosted automatically and
-   * `vgi_rpc.Identity.v1` is enabled through {@link resolveToken} /
+   * `vgi_rpc.Identity.v2` is enabled through {@link resolveToken} /
    * {@link mintGrant}. Violations are a startup error naming this hook.
    */
   hostedProtocols?: () => Iterable<Protocol>;
 
   /**
    * Resolve an opaque bearer credential to the identity it authenticates as.
-   * Supplying it hosts `vgi_rpc.Identity.v1`'s `introspect_token` -- on HTTP,
+   * Supplying it hosts `vgi_rpc.Identity.v2`'s `introspect_token` -- on HTTP,
    * the transport that authenticates callers. Absent, the method is absent
    * (not hosted-and-refusing), which keeps a dependency upgrade from growing a
    * credential-to-identity oracle on every existing worker.
@@ -232,7 +232,7 @@ function validatedHostedProtocols(hosting: HostingOptions, primary: Protocol): P
       throw new Error(
         `${owner} entry ${index} is named '${name}', which claims the reserved '${RESERVED_PREFIX}' prefix. ` +
           "Framework protocols are not supplied through this hook: reflection is hosted automatically, and " +
-          "vgi_rpc.Identity.v1 is enabled by supplying resolveToken and/or mintGrant.",
+          "vgi_rpc.Identity.v2 is enabled by supplying resolveToken and/or mintGrant.",
       );
     }
     if (!/^[A-Za-z_][A-Za-z0-9_.]*$/.test(name)) {
@@ -259,7 +259,7 @@ function validatedHostedProtocols(hosting: HostingOptions, primary: Protocol): P
   return out;
 }
 
-/** Build the `vgi_rpc.Identity.v1` implementation, or `undefined`.
+/** Build the `vgi_rpc.Identity.v2` implementation, or `undefined`.
  *
  *  `undefined` unless the worker supplies at least one hook -- and that is the
  *  point: the server then does not host the protocol at all, rather than
@@ -329,7 +329,7 @@ function resolveIntrospectPrincipals(
   }
   if (principals.length === 0) {
     throw new Error(
-      "This worker supplies resolveToken, which hosts the vgi_rpc.Identity.v1 protocol, but no " +
+      "This worker supplies resolveToken, which hosts the vgi_rpc.Identity.v2 protocol, but no " +
         `introspector allowlist was configured. Set ${INTROSPECT_PRINCIPALS_ENV} (comma-separated) or pass ` +
         "introspectPrincipals.\n\n" +
         "There is no permissive default on purpose: introspection is a separate capability from " +