@stackbone/cli 0.7.0 → 0.7.1

package/migrations/0114_memory_person_id.sql

@@ -0,0 +1,16 @@
+-- STA3-33: Studio and the box address a person by this opaque per-agent id
+-- instead of the phone or email in user_id, which leaked into URLs and logs.
+-- It lives on the revision row because an erase bumps the revision and keeps
+-- the row, so a person who is deleted and comes back keeps the same id.
+
+ALTER TABLE stackbone_platform.memory_user_revisions
+  ADD COLUMN person_id uuid NOT NULL DEFAULT gen_random_uuid(),
+  ADD CONSTRAINT memory_user_revisions_person_id_key UNIQUE (person_id);
+
+-- A person with facts or calls but no revision row yet has no id to open.
+INSERT INTO stackbone_platform.memory_user_revisions (agent_id, user_id, revision)
+SELECT agent_id, user_id, 0 FROM (
+  SELECT agent_id, user_id FROM stackbone_platform.memory_facts
+  UNION SELECT agent_id, user_id FROM stackbone_platform.memory_calls
+) AS people
+ON CONFLICT (agent_id, user_id) DO NOTHING;

package/migrations/0115_memory_erased_person_keeps_no_identifier.sql

@@ -0,0 +1,31 @@
+-- STA3-35: an erase keeps the revision row as a tombstone, so an add that
+-- snapshotted before the erase still fails its revision check and cannot write
+-- the erased memory back. The tombstone kept the phone or email in user_id; it
+-- now keeps only the opaque person_id, so user_id is nullable and person_id is
+-- the key. UNIQUE (agent_id, user_id) stays the lookup the SDK's
+-- ON CONFLICT (agent_id, user_id) infers, and Postgres treats NULLs as
+-- distinct, so an agent can hold many erased rows.
+
+ALTER TABLE stackbone_platform.memory_user_revisions
+  DROP CONSTRAINT memory_user_revisions_pkey;
+ALTER TABLE stackbone_platform.memory_user_revisions
+  ALTER COLUMN user_id DROP NOT NULL,
+  ADD CONSTRAINT memory_user_revisions_pkey PRIMARY KEY (person_id),
+  DROP CONSTRAINT memory_user_revisions_person_id_key,
+  ADD CONSTRAINT memory_user_revisions_agent_id_user_id_key UNIQUE (agent_id, user_id);
+
+-- People erased before this change. Only an erase deletes memory_calls, and
+-- every committed add writes a call with its revision bump, so a row with no
+-- facts and no calls on its own agent is an erased tombstone. The rule holds
+-- only here: once adds claim the row at snapshot, a failed add can leave such
+-- a row for a live person.
+UPDATE stackbone_platform.memory_user_revisions AS revisions
+SET user_id = NULL
+WHERE NOT EXISTS (
+    SELECT 1 FROM stackbone_platform.memory_facts AS facts
+    WHERE facts.agent_id = revisions.agent_id AND facts.user_id = revisions.user_id
+  )
+  AND NOT EXISTS (
+    SELECT 1 FROM stackbone_platform.memory_calls AS calls
+    WHERE calls.agent_id = revisions.agent_id AND calls.user_id = revisions.user_id
+  );

package/migrations/0116_rag_hybrid_search.sql

@@ -0,0 +1,14 @@
+-- STA3-57: lexical + hybrid search, opt-in per collection.
+ALTER TABLE stackbone_platform.rag_collections
+  ADD COLUMN IF NOT EXISTS hybrid_enabled boolean NOT NULL DEFAULT false,
+  ADD COLUMN IF NOT EXISTS number_format text NOT NULL DEFAULT 'eu'
+    CONSTRAINT rag_collections_number_format_check CHECK (number_format IN ('eu', 'us'));
+
+-- Nullable: ingest writes the terms only while the collection has hybrid on
+-- (STA3-57 D11). The terms are normalised in TypeScript, so no `unaccent` (D6).
+ALTER TABLE stackbone_platform.rag_chunks
+  ADD COLUMN IF NOT EXISTS lexical tsvector;
+
+-- Built under the migrator's advisory lock, like the 1536 index in 0103.
+CREATE INDEX IF NOT EXISTS rag_chunks_lexical_gin
+  ON stackbone_platform.rag_chunks USING gin (lexical);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stackbone/cli",
-  "version": "0.7.0",
+  "version": "0.7.1",
   "license": "UNLICENSED",
   "type": "module",
   "bin": {

package/voice-worker/main.js

@@ -557,6 +557,7 @@ var STUDIO_RAG_COLLECTION_REGEX = /^[A-Za-z0-9][A-Za-z0-9._\-/:]{0,127}$/;
 var ragCollectionNameParamSchema = z7.string().min(1).max(128).regex(STUDIO_RAG_COLLECTION_REGEX, "collection must be 1..128 ASCII identifier-safe chars");
 var ragDistanceSchema = z7.enum(["cosine", "l2", "ip"]);
 var ragIndexKindSchema = z7.enum(["hnsw", "exact"]);
+var ragNumberFormatSchema = z7.enum(["eu", "us"]);
 var ragCollectionSchema = z7.object({
   name: z7.string(),
   dimensions: z7.number().int().min(1),
@@ -570,7 +571,10 @@ var ragCollectionSchema = z7.object({
   // Maximum `created_at` across the collection's documents. Null when the
   // collection has zero documents (an empty collection row created before its
   // first ingest).
-  last_ingested_at: z7.iso.datetime().nullable()
+  last_ingested_at: z7.iso.datetime().nullable(),
+  // STA3-57. Optional so a list from a box older than hybrid search still parses.
+  hybrid_enabled: z7.boolean().optional(),
+  number_format: ragNumberFormatSchema.optional()
 });
 var ragCollectionListResponseSchema = z7.object({
   items: z7.array(ragCollectionSchema)
@@ -617,6 +621,51 @@ var ragDocumentSchema = z7.object({
   pending_version: z7.number().int().min(1).nullable().optional()
 });
 var ragFilterSchema = jsonObject();
+var ragDateBoundSchema = z7.string().length(10).regex(/^[0-9]{4}-[0-9]{2}-[0-9]{2}$/).refine((value) => {
+  const year = Number(value.slice(0, 4));
+  const month = Number(value.slice(5, 7));
+  const day = Number(value.slice(8, 10));
+  const date = /* @__PURE__ */ new Date(0);
+  date.setUTCFullYear(year, month - 1, day);
+  return year >= 1 && year <= 9999 && date.getUTCFullYear() === year && date.getUTCMonth() === month - 1 && date.getUTCDate() === day;
+}, "date must be a real YYYY-MM-DD day in years 0001\u20139999");
+var ragRangeFieldSchema = z7.string().min(1).refine((field) => !field.includes("."), {
+  message: "field must be a literal top-level metadata key without dots"
+});
+var ragRangeSchema = z7.discriminatedUnion("type", [
+  z7.strictObject({
+    field: ragRangeFieldSchema,
+    type: z7.literal("number"),
+    gt: z7.number().finite().optional(),
+    gte: z7.number().finite().optional(),
+    lt: z7.number().finite().optional(),
+    lte: z7.number().finite().optional()
+  }),
+  z7.strictObject({
+    field: ragRangeFieldSchema,
+    type: z7.literal("date"),
+    gt: ragDateBoundSchema.optional(),
+    gte: ragDateBoundSchema.optional(),
+    lt: ragDateBoundSchema.optional(),
+    lte: ragDateBoundSchema.optional()
+  })
+]).superRefine((range, ctx) => {
+  const lower = range.gt ?? range.gte;
+  const upper = range.lt ?? range.lte;
+  if (lower === void 0 && upper === void 0) {
+    ctx.addIssue({ code: "custom", message: "range must have at least one bound" });
+  }
+  if (range.gt !== void 0 && range.gte !== void 0) {
+    ctx.addIssue({ code: "custom", message: "gt and gte cannot be combined", path: ["gte"] });
+  }
+  if (range.lt !== void 0 && range.lte !== void 0) {
+    ctx.addIssue({ code: "custom", message: "lt and lte cannot be combined", path: ["lte"] });
+  }
+  if (lower !== void 0 && upper !== void 0 && lower > upper) {
+    ctx.addIssue({ code: "custom", message: "lower bound must not exceed upper bound" });
+  }
+});
+var ragRangesSchema = z7.array(ragRangeSchema).max(10);
 var listRagDocsQuerySchema = forwardPageQuerySchema.extend({
   limit: limitParamSchema(STUDIO_RAG_DOCS_LIMIT_MAX).default(STUDIO_RAG_DOCS_LIMIT_DEFAULT),
   // Filter is sent JSON-encoded inside a single query param so the UI can
@@ -655,7 +704,9 @@ var ragQueryRequestSchema = z7.object({
   model: z7.string().min(1).optional(),
   topK: z7.coerce.number().int().min(1).max(STUDIO_RAG_TOPK_MAX).default(STUDIO_RAG_TOPK_DEFAULT),
   filter: ragFilterSchema.optional(),
-  preview: z7.object({ docId: z7.string().min(1), version: ragDocumentVersionNumberSchema }).optional()
+  ranges: ragRangesSchema.optional(),
+  preview: z7.object({ docId: z7.string().min(1), version: ragDocumentVersionNumberSchema }).optional(),
+  hybrid: z7.boolean().optional()
 }).refine((value) => value.embedding !== void 0 || value.text !== void 0, {
   message: "either `embedding` or `text` must be provided",
   path: ["embedding"]
@@ -667,6 +718,9 @@ var ragQueryHitSchema = z7.object({
   // query)` over `stackbone_rag_chunks`, so higher is closer. We surface the
   // raw value; the UI formats it.
   score: z7.number(),
+  // The reciprocal rank fusion score a hybrid search ordered by (STA3-57 D10).
+  // Absent when only the search by meaning ran, and from a box older than it.
+  mixed_score: z7.number().optional(),
   content: z7.string(),
   metadata: jsonObject()
 });
@@ -710,7 +764,10 @@ var createRagCollectionRequestSchema = z7.object({
   name: ragCollectionNameParamSchema,
   metadata: jsonObject().optional(),
   embedding_provider: pickedModelProviderIdSchema.optional(),
-  embedding_model: z7.string().trim().min(1).optional()
+  embedding_model: z7.string().trim().min(1).optional(),
+  // Omitted, the box creates the collection with hybrid search off and `eu`.
+  hybrid_enabled: z7.boolean().optional(),
+  number_format: ragNumberFormatSchema.optional()
 }).refine(
   (value) => value.embedding_provider === void 0 === (value.embedding_model === void 0),
   {
@@ -718,6 +775,17 @@ var createRagCollectionRequestSchema = z7.object({
     path: ["embedding_model"]
   }
 );
+var updateRagCollectionRequestSchema = z7.strictObject({
+  hybrid_enabled: z7.boolean().optional(),
+  number_format: ragNumberFormatSchema.optional()
+}).refine((value) => value.hybrid_enabled !== void 0 || value.number_format !== void 0, {
+  message: "Send hybrid_enabled, number_format or both"
+});
+var updateRagCollectionResponseSchema = z7.object({
+  name: z7.string(),
+  hybrid_enabled: z7.boolean(),
+  number_format: ragNumberFormatSchema
+});
 var deleteRagCollectionResponseSchema = z7.object({
   name: z7.string(),
   deleted: z7.literal(true)
@@ -858,7 +926,14 @@ var effectiveVoiceSettingsObjectSchema = z8.strictObject({
    */
   fastModelRouting: z8.boolean(),
   /** The box runs the search before the model's turn, so the worker never reads it. */
-  rag: ragSchema
+  rag: ragSchema,
+  /**
+   * What the agent says when an input guardrail stops the caller's sentence (FDE-66):
+   * a policy message written for a chat reader would leak the rule if a TTS read it.
+   * The box answers the turn itself, so the worker never reads it. Not nullable: a
+   * blocked turn always says something.
+   */
+  refusalPhrase: z8.string().trim().min(1).max(300)
 });
 var checkCrossFieldRules = (value, ctx) => {
   const min = value.turn?.minEndpointingDelayMs;
@@ -889,7 +964,8 @@ var effectiveVoiceSettingsSchema = effectiveVoiceSettingsObjectSchema.superRefin
 var callVoiceSettingsObjectSchema = effectiveVoiceSettingsObjectSchema.omit({
   recording: true,
   fastModelRouting: true,
-  rag: true
+  rag: true,
+  refusalPhrase: true
 });
 var callVoiceSettingsSchema = callVoiceSettingsObjectSchema.superRefine(checkCrossFieldRules);
 var VOICE_SETTINGS_LEAF_PATHS = [
@@ -918,7 +994,8 @@ var VOICE_SETTINGS_LEAF_PATHS = [
   "rag.minScore",
   "rag.minWords",
   "rag.topK",
-  "rag.timeoutMs"
+  "rag.timeoutMs",
+  "refusalPhrase"
 ];
 var voiceSettingOriginSchema = z8.enum(["studio", "code", "default"]);
 var effectiveVoiceSettingsViewSchema = z8.strictObject({