fix(llm): split OpenAI reasoning summary blocks (#29000)

Author: rekram1-node

packages/llm/src/protocols/openai-responses.ts

@@ -24,16 +24,24 @@ export const textDelta = (state: State, events: LLMEvent[], id: string, text: st
   return { ...stepped, text: new Set([...stepped.text, id]) }
 }
 
-export const reasoningDelta = (state: State, events: LLMEvent[], id: string, text: string): State => {
+export const reasoningStart = (
+  state: State,
+  events: LLMEvent[],
+  id: string,
+  providerMetadata?: ProviderMetadata,
+): State => {
+  if (state.reasoning.has(id)) return state
   const stepped = stepStart(state, events)
-  if (stepped.reasoning.has(id)) {
-    events.push(LLMEvent.reasoningDelta({ id, text }))
-    return stepped
-  }
-  events.push(LLMEvent.reasoningStart({ id }), LLMEvent.reasoningDelta({ id, text }))
+  events.push(LLMEvent.reasoningStart({ id, providerMetadata }))
   return { ...stepped, reasoning: new Set([...stepped.reasoning, id]) }
 }
 
+export const reasoningDelta = (state: State, events: LLMEvent[], id: string, text: string): State => {
+  const started = reasoningStart(state, events, id)
+  events.push(LLMEvent.reasoningDelta({ id, text }))
+  return started
+}
+
 export const reasoningEnd = (
   state: State,
   events: LLMEvent[],

packages/llm/src/protocols/utils/lifecycle.ts

@@ -7,12 +7,28 @@ export const OpenAIReasoningEfforts = ReasoningEfforts.filter(
 )
 export type OpenAIReasoningEffort = (typeof OpenAIReasoningEfforts)[number]
 
+// Mirrors OpenAI's `ResponseIncludable` union from the official SDK. Keep this
+// in lockstep with `openai-node/src/resources/responses/responses.ts`.
+export const OpenAIResponseIncludables = [
+  "file_search_call.results",
+  "web_search_call.results",
+  "web_search_call.action.sources",
+  "message.input_image.image_url",
+  "computer_call_output.output.image_url",
+  "code_interpreter_call.outputs",
+  "reasoning.encrypted_content",
+  "message.output_text.logprobs",
+] as const
+export type OpenAIResponseIncludable = (typeof OpenAIResponseIncludables)[number]
+
 const REASONING_EFFORTS = new Set<string>(ReasoningEfforts)
 const OPENAI_REASONING_EFFORTS = new Set<string>(OpenAIReasoningEfforts)
 const TEXT_VERBOSITY = new Set<string>(["low", "medium", "high"])
+const INCLUDABLES = new Set<string>(OpenAIResponseIncludables)
 
 export const OpenAIReasoningEffort = Schema.Literals(OpenAIReasoningEfforts)
 export const OpenAITextVerbosity = TextVerbosity
+export const OpenAIResponseIncludable = Schema.Literals(OpenAIResponseIncludables)
 
 const isAnyReasoningEffort = (effort: unknown): effort is ReasoningEffort =>
   typeof effort === "string" && REASONING_EFFORTS.has(effort)
@@ -35,12 +51,20 @@ export const reasoningEffort = (request: LLMRequest): ReasoningEffort | undefine
   return isAnyReasoningEffort(value) ? value : undefined
 }
 
-export const reasoningSummary = (request: LLMRequest): "auto" | undefined => {
-  return options(request)?.reasoningSummary === "auto" ? "auto" : undefined
-}
+export const reasoningSummary = (request: LLMRequest): "auto" | undefined =>
+  options(request)?.reasoningSummary === "auto" ? "auto" : undefined
 
-export const encryptedReasoning = (request: LLMRequest) =>
-  options(request)?.includeEncryptedReasoning === true ? true : undefined
+// Resolve the OpenAI Responses `include` field. Filters out unknown
+// includable values defensively so a typo in upstream config drops the
+// invalid entry instead of poisoning the wire body. An empty array (either
+// passed directly or produced by filtering) is treated as "no include" and
+// returns undefined so the request body omits the field entirely.
+export const include = (request: LLMRequest): ReadonlyArray<OpenAIResponseIncludable> | undefined => {
+  const value = options(request)?.include
+  if (!Array.isArray(value)) return undefined
+  const filtered = value.filter((entry): entry is OpenAIResponseIncludable => INCLUDABLES.has(entry))
+  return filtered.length > 0 ? filtered : undefined
+}
 
 export const promptCacheKey = (request: LLMRequest) => {
   const value = options(request)?.promptCacheKey

packages/llm/src/protocols/utils/openai-options.ts

@@ -1,13 +1,19 @@
 import type { ProviderOptions, ReasoningEffort, TextVerbosity } from "../schema"
 import { mergeProviderOptions } from "../schema"
+import type { OpenAIResponseIncludable } from "../protocols/utils/openai-options"
+
+export type { OpenAIResponseIncludable } from "../protocols/utils/openai-options"
 
 export interface OpenAIOptionsInput {
   readonly [key: string]: unknown
   readonly store?: boolean
   readonly promptCacheKey?: string
   readonly reasoningEffort?: ReasoningEffort
   readonly reasoningSummary?: "auto"
-  readonly includeEncryptedReasoning?: boolean
+  // OpenAI Responses `include` wire field. Mirrors the official SDK's
+  // `ResponseIncludable[]` union exactly so AI SDK callers and direct
+  // native-SDK callers share one shape and no translation is required.
+  readonly include?: ReadonlyArray<OpenAIResponseIncludable>
   readonly textVerbosity?: TextVerbosity
 }
 
@@ -25,7 +31,7 @@ const openAIProviderOptions = (options: OpenAIOptionsInput | undefined): Provide
       promptCacheKey: options?.promptCacheKey,
       reasoningEffort: options?.reasoningEffort,
       reasoningSummary: options?.reasoningSummary,
-      includeEncryptedReasoning: options?.includeEncryptedReasoning,
+      include: options?.include,
       textVerbosity: options?.textVerbosity,
     }),
   )
@@ -42,6 +48,12 @@ export const gpt5DefaultOptions = (
   return openAIProviderOptions({
     reasoningEffort: "medium",
     reasoningSummary: "auto",
+    // GPT-5 reasoning models are configured stateless (`store: false`) by
+    // `openAIDefaultOptions` below, so the only way a follow-up turn can
+    // carry reasoning state is via the encrypted reasoning include. Without
+    // this, callers using the default model facade get reasoning summaries
+    // they cannot replay statelessly.
+    include: ["reasoning.encrypted_content"],
     textVerbosity:
       options.textVerbosity === true && id.includes("gpt-5.") && !id.includes("codex") && !id.includes("-chat")
         ? "low"

packages/llm/src/providers/openai-options.ts

@@ -5,7 +5,7 @@ import * as OpenAIChat from "../protocols/openai-chat"
 import * as OpenAIResponses from "../protocols/openai-responses"
 import { withOpenAIOptions, type OpenAIProviderOptionsInput } from "./openai-options"
 
-export type { OpenAIOptionsInput } from "./openai-options"
+export type { OpenAIOptionsInput, OpenAIResponseIncludable } from "./openai-options"
 
 export const id = ProviderID.make("openai")
 

packages/llm/src/providers/openai.ts

@@ -283,7 +283,7 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
           )
         return events.pipe(
           Stream.mapAccumEffect(
-            protocol.stream.initial,
+            () => protocol.stream.initial(request),
             protocol.stream.step,
             protocol.stream.onHalt ? { onHalt: protocol.stream.onHalt } : undefined,
           ),

packages/llm/src/route/client.ts

@@ -52,8 +52,8 @@ export interface ProtocolBody<Body> {
 export interface ProtocolStream<Frame, Event, State> {
   /** Schema for one decoded streaming event, decoded from a transport frame. */
   readonly event: Schema.Codec<Event, Frame>
-  /** Initial parser state. Called once per response. */
-  readonly initial: () => State
+  /** Initial parser state. Called once per response with the resolved request. */
+  readonly initial: (request: LLMRequest) => State
   /** Translate one event into emitted `LLMEvent`s plus the next state. */
   readonly step: (state: State, event: Event) => Effect.Effect<readonly [State, ReadonlyArray<LLMEvent>], LLMError>
   /** Optional request-completion signal for transports that do not end naturally. */

packages/llm/src/route/protocol.ts

@@ -97,7 +97,7 @@ export function continuationRequest(input: {
     tools: features.has("tool-call") ? [continuationTool] : [],
     cache: "none",
     providerOptions: features.has("encrypted-reasoning")
-      ? { openai: { store: false, includeEncryptedReasoning: true, reasoningSummary: "auto" } }
+      ? { openai: { store: false, include: ["reasoning.encrypted_content"], reasoningSummary: "auto" } }
       : undefined,
     generation: { maxTokens: 80, temperature: 0 },
   })