sentry-junior · Oct 26, 2026 · Oct 26, 2026 · Oct 26, 2026 · Oct 26, 2026
diff --git a/packages/docs/src/content/docs/reference/config-and-env.md b/packages/docs/src/content/docs/reference/config-and-env.md
@@ -30,6 +30,7 @@ related:
 | `AI_REASONING_LEVEL`                        | No          | Main-agent reasoning override: `none`, `low`, `medium`, `high`, or `xhigh`. Unset by default. An explicit profile reasoning level takes precedence when routing is enabled. |
 | `AI_FAST_MODEL`                             | No          | Faster model for lightweight tasks and routing/classification passes before the main turn begins. Defaults to `openai/gpt-6-luna`.                                          |
 | `AI_GUARDIAN_MODEL`                         | No          | Model for Guardian action review. Defaults to `openai/gpt-6-luna`.                                                                                                          |
+| `AI_DECISION_MODEL`                         | No          | Decision model for passive reply routing in subscribed threads. Uses the AI SDK decision API. Defaults to `typesafe-ai/jev`.                                                |
 | `AI_HANDOFF_MODEL`                          | No          | Deprecated profile setting. Creates `handoff`. Defaults to `anthropic/claude-opus-5.5` with high reasoning.                                                                 |
 | `AI_MODEL_PROFILES`                         | No          | Deprecated JSON map of profile names to model IDs for env-only setup. Names must match `^[a-z][a-z0-9_-]*$`.                                                                |
 | `AI_EMBEDDING_MODEL`                        | No          | Embedding model for plugin-owned vector retrieval. Defaults to `openai/text-embedding-3-small`; memory v1 stores fixed 1536-dimensional vectors.                            |
@@ -270,6 +271,7 @@ const app = await createApp({
       reasoningLevel: "high",
     },
   },
+  decisionModelId: "typesafe-ai/jev",
   fastModelId: "openai/gpt-6-luna",
   guardianModelId: "openai/gpt-6-luna",
   embeddingModelId: "openai/text-embedding-3-small",

diff --git a/packages/junior-evals/package.json b/packages/junior-evals/package.json
@@ -23,7 +23,7 @@
     "@sentry/junior-plugin-api": "workspace:*",
     "@sentry/junior-sentry": "workspace:*",
     "@sentry/junior-testing": "workspace:*",
-    "ai": "^6.0.190",
+    "ai": "^7.0.153",
     "chat": "4.29.0",
     "hono": "^4.12.27",
     "msw": "^2.14.6",

diff --git a/packages/junior-evals/src/fixture/web.ts b/packages/junior-evals/src/fixture/web.ts
@@ -36,7 +36,7 @@ export interface WebSearchResult {
   url: string;
 }
 
-const SEARCH_REQUEST_URL = "https://ai-gateway.vercel.sh/v3/ai/language-model";
+const SEARCH_REQUEST_URL = "https://ai-gateway.vercel.sh/v4/ai/language-model";
 const SEARCH_TOOL_ID = "gateway.parallel_search";
 
 let searchResults: WebSearchResult[] = [];

diff --git a/packages/junior-evals/src/recording-rules.ts b/packages/junior-evals/src/recording-rules.ts
@@ -46,7 +46,7 @@ const RECORDING_RULES: RecordingRule[] = [
     // image generation and web search before the proxy.
     name: "model",
     match: { method: "POST", url: "https://ai-gateway.vercel.sh/" },
-    // AI SDK gateway requests, such as `/v3/ai/language-model`, name the
+    // AI SDK gateway requests, such as `/v4/ai/language-model`, name the
     // model in a header, not in the body.
     keyHeaders: [
       "ai-language-model-id",

diff --git a/packages/junior/package.json b/packages/junior/package.json
@@ -69,7 +69,7 @@
   },
   "dependencies": {
     "@agentclientprotocol/sdk": "1.3.0",
-    "@ai-sdk/gateway": "^3.0.119",
+    "@ai-sdk/gateway": "^4.0.126",
     "@chat-adapter/slack": "4.29.0",
     "@chat-adapter/state-memory": "4.29.0",
     "@chat-adapter/state-redis": "4.29.0",
@@ -91,7 +91,7 @@
     "@vercel/oidc": "^3.4.1",
     "@vercel/queue": "^0.2.0",
     "@vercel/sandbox": "2.8.0",
-    "ai": "^6.0.190",
+    "ai": "^7.0.153",
     "better-auth": "^1.6.22",
     "chat": "4.29.0",
     "commander": "^14.0.3",

diff --git a/packages/junior/src/chat/app/services.ts b/packages/junior/src/chat/app/services.ts
@@ -1,4 +1,4 @@
-import { completeObject, completeText } from "@/chat/pi/client";
+import { completeDecision, completeText } from "@/chat/pi/client";
 import { executeAgentRun as executeAgentRunImpl } from "@/chat/agent";
 import type { SandboxEgressTracePropagationConfig } from "@/chat/sandbox/egress/tracing";
 import {
@@ -86,8 +86,8 @@ export function createJuniorRuntimeServices(
     executeTurn: async (run, saveResult) =>
       await executeTurn(agentRunner, run, saveResult),
     subscribedReplyPolicy: createSubscribedReplyPolicy({
-      completeObject:
-        overrides.subscribedReplyPolicy?.completeObject ?? completeObject,
+      completeDecision:
+        overrides.subscribedReplyPolicy?.completeDecision ?? completeDecision,
     }),
     visionContext,
   };

diff --git a/packages/junior/src/chat/config.ts b/packages/junior/src/chat/config.ts
@@ -77,6 +77,8 @@ const DEFAULT_ASSISTANT_LOADING_MESSAGES = [
 ] as const;
 
 export interface BotModelConfig {
+  /** Decision model used by the passive reply router. */
+  decisionModelId?: string;
   /** Model used for embeddings. */
   embeddingModelId?: string;
   /** Model used for low-cost structured and utility calls. */
@@ -94,6 +96,7 @@ export interface BotModelConfig {
 export interface BotConfig {
   contextWindowTokens: number;
   crossActorMidRunMode: CrossActorMidRunMode;
+  decisionModelId: string;
   defaultProfile: ModelProfile;
   embeddingModelId: string;
   fastModelId: string;
@@ -261,6 +264,7 @@ const DEFAULT_HANDOFF_MODEL_ID = resolveGatewayModel(
   DEFAULT_MODEL_PROFILES.handoff.modelId,
 ).id;
 const DEFAULT_WEB_SEARCH_MODEL_ID = resolveGatewayModel("openai/gpt-6-luna").id;
+const DEFAULT_DECISION_MODEL_ID = "openai/gpt-6-luna-decisions";
 const DEFAULT_EMBEDDING_MODEL_ID = "openai/text-embedding-3-small";
 const DEFAULT_IMAGE_GENERATION_MODEL_ID = "google/gemini-3-pro-image";
 const DEFAULT_VISION_MODEL_ID = resolveGatewayModel("openai/gpt-5.6-sol").id;
@@ -502,6 +506,8 @@ function readBotConfig(
       reasoningLevel === undefined
         ? undefined
         : parseTurnReasoningLevel(reasoningLevel),
+    decisionModelId:
+      toOptionalTrimmed(env.AI_DECISION_MODEL) ?? DEFAULT_DECISION_MODEL_ID,
     fastModelId,
     guardianModelId,
     imageGenerationModelId:
@@ -825,6 +831,13 @@ export interface SlackReactionConfig {
 
 /** Apply model overrides from createApp(). */
 export function setBotModelConfig(config: BotModelConfig): void {
+  if (config.decisionModelId !== undefined) {
+    botConfig.decisionModelId = requireModelId(
+      "decisionModelId",
+      config.decisionModelId,
+      toOptionalTrimmed,
+    );
+  }
   if (config.embeddingModelId !== undefined) {
     botConfig.embeddingModelId = requireModelId(
       "embeddingModelId",

diff --git a/packages/junior/src/chat/pi/client.ts b/packages/junior/src/chat/pi/client.ts
@@ -21,8 +21,13 @@ export {
 } from "@/chat/pi/gateway-auth";
 import {
   embedMany,
+  experimental_decide,
+  Experimental_DecisionRefusalError,
   generateObject,
   NoObjectGeneratedError,
+  type Experimental_DecisionQuestion,
+  type Experimental_DecisionResult,
+  type Experimental_DecisionState,
   type GenerateObjectResult,
   type LanguageModelUsage,
 } from "ai";
@@ -179,6 +184,7 @@ export const GEN_AI_SERVER_ADDRESS = "ai-gateway.vercel.sh";
 export const GEN_AI_SERVER_PORT = 443;
 const GEN_AI_OPERATION_CHAT = "chat" as const;
 const GEN_AI_OPERATION_EMBEDDINGS = "embeddings" as const;
+const GEN_AI_OPERATION_DECIDE = "decide" as const;
 // AI Gateway input rates in USD per million tokens. Keep estimates
 // model-specific so unlisted configured models remain explicitly unknown.
 const EMBEDDING_INPUT_COST_USD_PER_MILLION_TOKENS: Readonly<
@@ -194,6 +200,25 @@ const EMBEDDING_INPUT_COST_USD_PER_MILLION_TOKENS: Readonly<
   "voyage/voyage-3.5": 0.06,
   "voyage/voyage-3.5-lite": 0.02,
 };
+// Decision models bill input tokens only.
+const DECISION_INPUT_COST_USD_PER_MILLION_TOKENS: Readonly<
+  Record<string, number>
+> = {
+  "openai/gpt-6-luna-decisions": 0.1,
+  "typesafe-ai/jev": 0.042,
+};
+
+function decisionCostUsd(
+  modelId: string,
+  inputTokens: number | undefined,
+): number | undefined {
+  const costPerMillionTokens =
+    DECISION_INPUT_COST_USD_PER_MILLION_TOKENS[modelId];
+  return costPerMillionTokens === undefined || inputTokens === undefined
+    ? undefined
+    : (inputTokens * costPerMillionTokens) / 1_000_000;
+}
+
 function embeddingCostUsd(
   modelId: string,
   inputTokens: number,
@@ -552,6 +577,80 @@ export async function completeObject<TSchema extends ZodTypeAny>(params: {
   };
 }
 
+/** Answer typed decision questions with an AI Gateway decision model, such as Jev. */
+export async function completeDecision<
+  const TQuestions extends Record<string, Experimental_DecisionQuestion>,
+>(params: {
+  modelId: string;
+  state: Experimental_DecisionState;
+  questions: TQuestions;
+  signal?: AbortSignal;
+  metadata?: Record<string, unknown>;
+  promptName?: string;
+}): Promise<{
+  answers: Experimental_DecisionResult<TQuestions>["answers"];
+  costUsd?: number;
+}> {
+  const credential = await resolveGatewayCredential();
+  const provider = createGatewayProvider(
+    credential ? { apiKey: credential.token } : {},
+  );
+  try {
+    const result = await withSpan(
+      `${GEN_AI_OPERATION_DECIDE} ${params.modelId}`,
+      "gen_ai.decide",
+      logContextFromMetadata(params.modelId, params.metadata),
+      async (setSpanAttributes) => {
+        const result = await experimental_decide({
+          model: provider.decisionModel(params.modelId),
+          state: params.state,
+          questions: params.questions,
+          ...(params.signal !== undefined
+            ? { abortSignal: params.signal }
+            : undefined),
+        });
+        const costUsd = decisionCostUsd(
+          params.modelId,
+          result.usage.inputTokens,
+        );
+        setSpanAttributes(
+          extractGenAiUsageAttributes({
+            ...result.usage,
+            ...(costUsd !== undefined
+              ? { cost: { input: costUsd, total: costUsd } }
+              : undefined),
+          }),
+        );
+        return { answers: result.answers, costUsd };
+      },
+      {
+        "gen_ai.provider.name": GEN_AI_PROVIDER_NAME,
+        "gen_ai.operation.name": GEN_AI_OPERATION_DECIDE,
+        "gen_ai.request.model": params.modelId,
+        ...(params.promptName
+          ? { "gen_ai.prompt.name": params.promptName }
+          : undefined),
+        "server.address": GEN_AI_SERVER_ADDRESS,
+        "server.port": GEN_AI_SERVER_PORT,
+        "gen_ai.provider.auth_mode": credential?.mode ?? "api_key",
+      },
+    );
+    return {
+      answers: result.answers,
+      ...(result.costUsd !== undefined
+        ? { costUsd: result.costUsd }
+        : undefined),
+    };
+  } catch (error) {
+    throw createProviderError(error, {
+      ...(Experimental_DecisionRefusalError.isInstance(error)
+        ? { kind: "invalid_response" }
+        : undefined),
+      modelId: params.modelId,
+    });
+  }
+}
+
 /** Generate text embeddings through the host-owned AI Gateway provider. */
 export async function embedTexts(params: {
   modelId: string;