@opengeni/db 0.36.1 → 3.1.0-canary.0

package/src/session-realtime-context.ts

@@ -237,6 +237,7 @@ export async function flushSessionRealtimeTranscriptTailInTransaction(
       textCodecVersion: schema.sessionRealtimeEntries.textCodecVersion,
       payload: schema.sessionRealtimeEntries.payload,
       payloadCodecVersion: schema.sessionRealtimeEntries.payloadCodecVersion,
+      modelContext: schema.sessionRealtimeEntries.modelContext,
     })
     .from(schema.sessionRealtimeEntries)
     .where(
@@ -258,9 +259,23 @@ export async function flushSessionRealtimeTranscriptTailInTransaction(
   }));
   const rendered = renderSessionRealtimeTail(decodedRows);
   if (!rendered.context) return null;
+  const renderedRows = decodedRows.slice(decodedRows.length - rendered.includedEntryCount);
+  let foundUserTranscript = false;
+  let modelContext: string | null = null;
+  for (let index = renderedRows.length - 1; index >= 0; index -= 1) {
+    const entry = renderedRows[index]!;
+    if (entry.role === "user") {
+      foundUserTranscript = true;
+      modelContext = entry.modelContext;
+      break;
+    }
+  }
+  if (!foundUserTranscript) {
+    modelContext = renderedRows.at(-1)?.modelContext ?? null;
+  }
 
   const [session] = await db
-    .select({ metadata: schema.sessions.metadata })
+    .select({ reasoningEffort: schema.sessions.reasoningEffort })
     .from(schema.sessions)
     .where(
       and(
@@ -270,7 +285,7 @@ export async function flushSessionRealtimeTranscriptTailInTransaction(
     )
     .limit(1);
   if (!session) throw new Error(`Realtime tail session ${input.sessionId} disappeared`);
-  const reasoning = ReasoningEffort.safeParse(session.metadata.reasoningEffort);
+  const reasoningEffort = ReasoningEffort.parse(session.reasoningEffort);
   const admitted = await submitHumanPromptInTransaction(db, {
     accountId: input.accountId,
     workspaceId: input.workspaceId,
@@ -286,13 +301,14 @@ export async function flushSessionRealtimeTranscriptTailInTransaction(
     operationKey: deterministicUuid(`opengeni:session-realtime-tail-flush:${mode.id}`),
     delivery: "steer",
     text: rendered.context,
+    modelContext,
     messagePresentation: {
       kind: "realtime_voice_handoff",
       text: "Voice session ended. Remaining conversation context was sent to the agent.",
       context: rendered.context,
     },
     resources: [],
-    reasoningEffortFallback: reasoning.success ? reasoning.data : "medium",
+    reasoningEffortFallback: reasoningEffort,
     turnMetadata: {
       realtimeTailFlush: {
         source: SESSION_REALTIME_TAIL_SOURCE,

package/src/session-realtime-ledger.ts

@@ -15,7 +15,10 @@ import {
   SessionRealtimeConflictError,
   type AssertSessionRealtimeOwnerInput,
 } from "./session-realtime";
-import { lockSessionEventWriteRows, lockWorkspaceInferenceControl } from "./session-control";
+import {
+  lockSessionEventWriteRows,
+  lockWorkspaceInferenceControlForAdmission,
+} from "./session-control";
 import { submitHumanPromptInTransaction } from "./session-queue-commands";
 import { mirrorSessionRealtimeContextInTransaction } from "./session-realtime-mirror";
 
@@ -155,9 +158,12 @@ export type SessionRealtimeInboundEntryInput = {
   delegationItemId?: string | null | undefined;
   text?: string | null | undefined;
   payload?: Record<string, unknown> | undefined;
+  modelContext?: string | undefined;
 };
 
 export type SyncSessionRealtimeLedgerInput = AssertSessionRealtimeOwnerInput & {
+  /** Request-scoped callers bound the control prefix wait; lifecycle callers omit it. */
+  controlLockTimeoutMs?: number;
   connectionId: string;
   connectionEpoch: number;
   entries?: SessionRealtimeInboundEntryInput[] | undefined;
@@ -1016,12 +1022,34 @@ function canonicalJsonValue(value: unknown): unknown {
   return value;
 }
 
+function normalizedModelContext(input: SessionRealtimeInboundEntryInput): string | null {
+  const value = input.modelContext?.trim();
+  if (!value) {
+    if (input.modelContext !== undefined) {
+      throw new Error("Realtime modelContext must not be empty");
+    }
+    return null;
+  }
+  if (value.length > 32_768) {
+    throw new Error("Realtime modelContext exceeds the server limit");
+  }
+  if (
+    input.kind !== "delegation_call" &&
+    input.kind !== "user_transcript" &&
+    input.kind !== "assistant_transcript"
+  ) {
+    throw new Error("Realtime modelContext requires a delegation or finalized transcript");
+  }
+  return value;
+}
+
 function inboundReplayMatches(
   row: EntryRow,
   input: SessionRealtimeInboundEntryInput,
   role: "user" | "assistant" | null,
   text: string | null,
   payload: Record<string, unknown>,
+  modelContext: string | null,
 ): boolean {
   const storedText =
     row.text === null ? null : fromPostgresLosslessText(row.text, row.textCodecVersion);
@@ -1034,11 +1062,32 @@ function inboundReplayMatches(
     row.delegationItemId === (input.delegationItemId ?? null) &&
     row.sourceUpdateId === null &&
     storedText === text &&
+    row.modelContext === modelContext &&
     JSON.stringify(canonicalJsonValue(storedPayload)) ===
       JSON.stringify(canonicalJsonValue(payload))
   );
 }
 
+async function delegationModelContextMatches(
+  db: Database,
+  row: EntryRow,
+  modelContext: string | null,
+): Promise<boolean> {
+  if (row.kind !== "delegation_call" || row.turnId === null) return true;
+  const [turn] = await db
+    .select({ modelContext: schema.sessionTurns.modelContext })
+    .from(schema.sessionTurns)
+    .where(
+      and(
+        eq(schema.sessionTurns.workspaceId, row.workspaceId),
+        eq(schema.sessionTurns.sessionId, row.sessionId),
+        eq(schema.sessionTurns.id, row.turnId),
+      ),
+    )
+    .limit(1);
+  return turn?.modelContext === modelContext;
+}
+
 async function appendInvalidDelegationFailure(
   db: Database,
   input: Pick<
@@ -1083,6 +1132,7 @@ async function admitRealtimeDelegationInTransaction(
   accountId: string,
   incoming: SessionRealtimeInboundEntryInput,
   entryId: string,
+  modelContext: string | null,
 ): Promise<{ turnId: string; eventIds: string[]; wakeRevision: number }> {
   const [session] = await db
     .select()
@@ -1097,7 +1147,8 @@ async function admitRealtimeDelegationInTransaction(
   if (!session || session.accountId !== accountId || session.status === "cancelled") {
     throw new SessionRealtimeConflictError("REALTIME_NOT_FOUND", "Session not found");
   }
-  const reasoning = ReasoningEffort.safeParse(session.metadata.reasoningEffort);
+  const sessionReasoning = ReasoningEffort.parse(session.reasoningEffort);
+  const sessionLatency = LatencyMode.parse(session.latencyMode);
   const [latestStarted] = await db
     .select({
       model: schema.sessionTurns.model,
@@ -1142,6 +1193,7 @@ async function admitRealtimeDelegationInTransaction(
     operationKey: incoming.operationId,
     delivery: "steer",
     text: incoming.text!,
+    modelContext,
     messagePresentation: {
       kind: "realtime_voice",
       text: inputTranscript,
@@ -1149,13 +1201,9 @@ async function admitRealtimeDelegationInTransaction(
     },
     resources: [],
     model: latestStarted?.model ?? session.model,
-    reasoningEffort: latestReasoning.success
-      ? latestReasoning.data
-      : reasoning.success
-        ? reasoning.data
-        : "medium",
-    latencyMode: latestLatency.success ? latestLatency.data : "standard",
-    reasoningEffortFallback: reasoning.success ? reasoning.data : "medium",
+    reasoningEffort: latestReasoning.success ? latestReasoning.data : sessionReasoning,
+    latencyMode: latestLatency.success ? latestLatency.data : sessionLatency,
+    reasoningEffortFallback: sessionReasoning,
     turnMetadata: {
       realtimeDelegation: { ...provenance, inputTranscript },
     },
@@ -1277,8 +1325,17 @@ export async function syncSessionRealtimeLedgerInTransaction(
     throw new Error("Realtime ledger batch exceeds the server limit");
   }
   // Realtime sync may admit canonical Steer work. Preserve the global lock
-  // order before owner proof locks the session row.
-  await lockWorkspaceInferenceControl(db, input.workspaceId, "update");
+  // order before owner proof locks the session row. The prefix is shared while
+  // the session branch is active; the delegation Steer below re-enters the same
+  // admission helper and observes this already-held prefix.
+  await lockWorkspaceInferenceControlForAdmission(db, {
+    workspaceId: input.workspaceId,
+    sessionId: input.sessionId,
+    resumePausedBranch: true,
+    ...(input.controlLockTimeoutMs !== undefined
+      ? { lockTimeoutMs: input.controlLockTimeoutMs }
+      : {}),
+  });
   const mode = await assertSessionRealtimeOwnerInTransaction(db, input);
   if (mode.connectionEpoch !== input.connectionEpoch) {
     throw new SessionRealtimeConflictError(
@@ -1359,6 +1416,7 @@ export async function syncSessionRealtimeLedgerInTransaction(
     const payload = boundedPayload(incoming.payload);
     const role = expectedRole(incoming);
     const text = incoming.text ?? null;
+    const modelContext = normalizedModelContext(incoming);
     assertBoundedString(text, SESSION_REALTIME_LEDGER_MAX_TEXT_BYTES, "Realtime text");
     if (
       (incoming.kind === "user_transcript" || incoming.kind === "assistant_transcript") &&
@@ -1392,7 +1450,10 @@ export async function syncSessionRealtimeLedgerInTransaction(
       )
       .limit(1);
     if (existing) {
-      if (!inboundReplayMatches(existing, incoming, role, text, payload)) {
+      if (
+        !inboundReplayMatches(existing, incoming, role, text, payload, modelContext) ||
+        !(await delegationModelContextMatches(db, existing, modelContext))
+      ) {
         throw new SessionRealtimeConflictError(
           incoming.kind === "delegation_call"
             ? "REALTIME_DELEGATION_CHANGED"
@@ -1423,6 +1484,7 @@ export async function syncSessionRealtimeLedgerInTransaction(
         textCodecVersion: LOSSLESS_CONTENT_CODEC_VERSION,
         payload,
         payloadCodecVersion: LOSSLESS_CONTENT_CODEC_VERSION,
+        modelContext,
         createdAt: now,
         updatedAt: now,
       })
@@ -1447,6 +1509,7 @@ export async function syncSessionRealtimeLedgerInTransaction(
           modeRow.accountId,
           incoming,
           entry.id,
+          modelContext,
         );
         const [linked] = await db
           .update(schema.sessionRealtimeEntries)

package/src/session-realtime-mirror.ts

@@ -215,7 +215,7 @@ export function renderRealtimeHumanInputRequestContext(input: {
         lines.push("      </options>");
       }
       lines.push(
-        `      <allow_other>${String(question.allowOther)}</allow_other>`,
+        `      <allow_other>${String(question.kind !== "text")}</allow_other>`,
         "    </question>",
       );
     }