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

2026-05-23 20:06:45 -05:00
parent 0b3a1c2fdf
commit eb84f461b8
22 changed files with 717 additions and 71 deletions
--- a/packages/llm/src/protocols/openai-responses.ts
+++ b/packages/llm/src/protocols/openai-responses.ts
@@ -57,6 +57,11 @@ const OpenAIResponsesReasoningItem = Schema.Struct({
  encrypted_content: optionalNull(Schema.String),
 })

+const OpenAIResponsesItemReference = Schema.Struct({
+  type: Schema.tag("item_reference"),
+  id: Schema.String,
+})
+
 // `function_call_output.output` accepts either a plain string or an ordered
 // array of content items so tools can return images in addition to text.
 // https://platform.openai.com/docs/api-reference/responses/object
@@ -72,6 +77,7 @@ const OpenAIResponsesInputItem = Schema.Union([
  Schema.Struct({ role: Schema.tag("user"), content: Schema.Array(OpenAIResponsesInputContent) }),
  Schema.Struct({ role: Schema.tag("assistant"), content: Schema.Array(OpenAIResponsesOutputText) }),
  OpenAIResponsesReasoningItem,
+  OpenAIResponsesItemReference,
  Schema.Struct({
    type: Schema.tag("function_call"),
    call_id: Schema.String,
@@ -86,6 +92,15 @@ const OpenAIResponsesInputItem = Schema.Union([
 ])
 type OpenAIResponsesInputItem = Schema.Schema.Type<typeof OpenAIResponsesInputItem>

+// Mutable counterpart of the schema reasoning item so `lowerMessages` can fold
+// multiple streamed summary parts into the same item before flushing.
+type OpenAIResponsesReasoningInput = {
+  type: "reasoning"
+  id: string
+  summary: Array<{ type: "summary_text"; text: string }>
+  encrypted_content?: string | null
+}
+
 const OpenAIResponsesTool = Schema.Struct({
  type: Schema.tag("function"),
  name: Schema.String,
@@ -112,7 +127,7 @@ const OpenAIResponsesCoreFields = {
  tool_choice: Schema.optional(OpenAIResponsesToolChoice),
  store: Schema.optional(Schema.Boolean),
  prompt_cache_key: Schema.optional(Schema.String),
-  include: optionalArray(Schema.Literal("reasoning.encrypted_content")),
+  include: optionalArray(OpenAIOptions.OpenAIResponseIncludable),
  reasoning: Schema.optional(
    Schema.Struct({
      effort: Schema.optional(OpenAIOptions.OpenAIReasoningEffort),
@@ -193,6 +208,7 @@ const OpenAIResponsesEvent = Schema.Struct({
  type: Schema.String,
  delta: Schema.optional(Schema.String),
  item_id: Schema.optional(Schema.String),
+  summary_index: Schema.optional(Schema.Number),
  item: Schema.optional(OpenAIResponsesStreamItem),
  response: Schema.optional(
    Schema.StructWithRest(
@@ -216,6 +232,18 @@ interface ParserState {
  readonly tools: ToolStream.State<string>
  readonly hasFunctionCall: boolean
  readonly lifecycle: Lifecycle.State
+  readonly reasoningItems: Readonly<Record<string, ReasoningStreamItem>>
+  readonly store: boolean | undefined
+}
+
+type ReasoningSummaryStatus = "active" | "can-conclude" | "concluded"
+
+interface ReasoningStreamItem {
+  readonly encryptedContent: string | null | undefined
+  // Keyed by OpenAI's numeric `summary_index`. JS object keys coerce to
+  // strings, but typing the map as `Record<number, ...>` documents intent
+  // and matches the wire field.
+  readonly summaryParts: Readonly<Record<number, ReasoningSummaryStatus>>
 }

 const invalid = ProviderShared.invalidRequest
@@ -245,22 +273,21 @@ const lowerToolCall = (part: ToolCallPart): OpenAIResponsesInputItem => ({
  arguments: ProviderShared.encodeJson(part.input),
 })

-const lowerReasoning = (part: ReasoningPart, store: boolean | undefined): OpenAIResponsesInputItem | undefined => {
+const lowerReasoning = (part: ReasoningPart): OpenAIResponsesReasoningInput | undefined => {
  const openai = part.providerMetadata?.openai
-  if (!ProviderShared.isRecord(openai) || typeof openai.itemId !== "string") return undefined
-  // With store:false, OpenAI only accepts previous reasoning items when the
-  // encrypted state is present. Bare rs_* ids point to non-persisted items.
-  if (store === false && typeof openai.reasoningEncryptedContent !== "string") return undefined
+  if (!ProviderShared.isRecord(openai) || typeof openai.itemId !== "string" || openai.itemId.length === 0)
+    return undefined
+  const encryptedContent =
+    typeof openai.reasoningEncryptedContent === "string"
+      ? openai.reasoningEncryptedContent
+      : openai.reasoningEncryptedContent === null
+        ? null
+        : undefined
  return {
    type: "reasoning",
    id: openai.itemId,
    summary: part.text.length > 0 ? [{ type: "summary_text", text: part.text }] : [],
-    encrypted_content:
-      typeof openai.reasoningEncryptedContent === "string"
-        ? openai.reasoningEncryptedContent
-        : openai.reasoningEncryptedContent === null
-          ? null
-          : undefined,
+    encrypted_content: encryptedContent,
  }
 }

@@ -310,6 +337,8 @@ const lowerMessages = Effect.fn("OpenAIResponses.lowerMessages")(function* (requ

    if (message.role === "assistant") {
      const content: TextPart[] = []
+      const reasoningItems: Record<string, OpenAIResponsesReasoningInput> = {}
+      const reasoningReferences = new Set<string>()
      const flushText = () => {
        if (content.length === 0) return
        input.push({ role: "assistant", content: content.map((part) => ({ type: "output_text", text: part.text })) })
@@ -322,8 +351,21 @@ const lowerMessages = Effect.fn("OpenAIResponses.lowerMessages")(function* (requ
        }
        if (part.type === "reasoning") {
          flushText()
-          const reasoning = lowerReasoning(part, store)
-          if (reasoning) input.push(reasoning)
+          const reasoning = lowerReasoning(part)
+          if (!reasoning) continue
+          if (store !== false && reasoning.id) {
+            if (!reasoningReferences.has(reasoning.id)) input.push({ type: "item_reference", id: reasoning.id })
+            reasoningReferences.add(reasoning.id)
+            continue
+          }
+          const existing = reasoningItems[reasoning.id]
+          if (existing) {
+            existing.summary.push(...reasoning.summary)
+            if (typeof reasoning.encrypted_content === "string") existing.encrypted_content = reasoning.encrypted_content
+            continue
+          }
+          reasoningItems[reasoning.id] = reasoning
+          input.push(reasoning)
          continue
        }
        if (part.type === "tool-call") {
@@ -352,7 +394,14 @@ const lowerMessages = Effect.fn("OpenAIResponses.lowerMessages")(function* (requ
    }
  }

-  return input
+  // With store:false, OpenAI only accepts previous reasoning items when the
+  // complete item has encrypted state. Summary blocks for one item may carry
+  // that state only on the last block, so filter after they have been joined.
+  return store === false
+    ? input.filter(
+        (item) => !("type" in item) || item.type !== "reasoning" || typeof item.encrypted_content === "string",
+      )
+    : input
 })

 const lowerOptions = Effect.fn("OpenAIResponses.lowerOptions")(function* (request: LLMRequest) {
@@ -362,14 +411,14 @@ const lowerOptions = Effect.fn("OpenAIResponses.lowerOptions")(function* (reques
  if (effort && !OpenAIOptions.isReasoningEffort(effort))
    return yield* invalid(`OpenAI Responses does not support reasoning effort ${effort}`)
  const summary = OpenAIOptions.reasoningSummary(request)
-  const encryptedState = OpenAIOptions.encryptedReasoning(request)
+  const include = OpenAIOptions.include(request)
  const verbosity = OpenAIOptions.textVerbosity(request)
  const instructions = OpenAIOptions.instructions(request)
  return {
    ...(instructions ? { instructions } : {}),
    ...(store !== undefined ? { store } : {}),
    ...(promptCacheKey ? { prompt_cache_key: promptCacheKey } : {}),
-    ...(encryptedState ? { include: ["reasoning.encrypted_content"] as const } : {}),
+    ...(include ? { include } : {}),
    ...(effort || summary ? { reasoning: { effort, summary } } : {}),
    ...(verbosity ? { text: { verbosity } } : {}),
  }
@@ -517,24 +566,53 @@ const onOutputTextDelta = (state: ParserState, event: OpenAIResponsesEvent): Ste
 const onReasoningDelta = (state: ParserState, event: OpenAIResponsesEvent): StepResult => {
  if (!event.delta) return [state, NO_EVENTS]
  const events: LLMEvent[] = []
+  const itemID = event.item_id ?? "reasoning-0"
+  const id =
+    event.summary_index !== undefined || state.reasoningItems[itemID]
+      ? `${itemID}:${event.summary_index ?? 0}`
+      : itemID
  return [
    {
      ...state,
-      lifecycle: Lifecycle.reasoningDelta(state.lifecycle, events, event.item_id ?? "reasoning-0", event.delta),
+      lifecycle: Lifecycle.reasoningDelta(state.lifecycle, events, id, event.delta),
    },
    events,
  ]
 }

-// The summary done event does not carry encrypted continuation state. Finish the
-// common reasoning block when the full reasoning item arrives in output_item.done.
 const onReasoningDone = (state: ParserState, _event: OpenAIResponsesEvent): StepResult => [state, NO_EVENTS]

 const reasoningMetadata = (item: OpenAIResponsesStreamItem & { id: string }) =>
  openaiMetadata({ itemId: item.id, reasoningEncryptedContent: item.encrypted_content ?? null })

+// OpenAI Responses streams reasoning items in a stable order:
+//   `output_item.added` (reasoning) →
+//     `reasoning_summary_part.added` (index=0) →
+//     `reasoning_summary_text.delta` →
+//     `reasoning_summary_part.done` (index=0) →
+//     (repeat for index>0) →
+//   `output_item.done` (reasoning).
+// The handlers below rely on this ordering: `onOutputItemAdded` seeds the
+// per-item entry, `onReasoningSummaryPartAdded` for `summary_index === 0`
+// short-circuits when the entry already exists, and higher-index handlers
+// fold against the same entry. Behaviour for out-of-order events is
+// best-effort, not guaranteed.
 const onOutputItemAdded = (state: ParserState, event: OpenAIResponsesEvent): StepResult => {
  const item = event.item
+  if (item && isReasoningItem(item)) {
+    const events: LLMEvent[] = []
+    return [
+      {
+        ...state,
+        lifecycle: Lifecycle.reasoningStart(state.lifecycle, events, `${item.id}:0`, reasoningMetadata(item)),
+        reasoningItems: {
+          ...state.reasoningItems,
+          [item.id]: { encryptedContent: item.encrypted_content, summaryParts: { 0: "active" } },
+        },
+      },
+      events,
+    ]
+  }
  if (item?.type !== "function_call" || !item.id) return [state, NO_EVENTS]
  const providerMetadata = openaiMetadata({ itemId: item.id })
  const events: LLMEvent[] = []
@@ -555,6 +633,103 @@ const onOutputItemAdded = (state: ParserState, event: OpenAIResponsesEvent): Ste
  ]
 }

+const onReasoningSummaryPartAdded = (state: ParserState, event: OpenAIResponsesEvent): StepResult => {
+  if (!event.item_id || event.summary_index === undefined) return [state, NO_EVENTS]
+  const item = state.reasoningItems[event.item_id] ?? { encryptedContent: undefined, summaryParts: {} }
+  if (event.summary_index === 0) {
+    if (state.reasoningItems[event.item_id]) return [state, NO_EVENTS]
+    const events: LLMEvent[] = []
+    return [
+      {
+        ...state,
+        lifecycle: Lifecycle.reasoningStart(
+          state.lifecycle,
+          events,
+          `${event.item_id}:0`,
+          openaiMetadata({ itemId: event.item_id, reasoningEncryptedContent: null }),
+        ),
+        reasoningItems: {
+          ...state.reasoningItems,
+          [event.item_id]: { ...item, summaryParts: { 0: "active" } },
+        },
+      },
+      events,
+    ]
+  }
+
+  const events: LLMEvent[] = []
+  const closed = Object.entries(item.summaryParts)
+    .filter((entry) => entry[1] === "can-conclude")
+    .reduce(
+      (lifecycle, entry) =>
+        Lifecycle.reasoningEnd(
+          lifecycle,
+          events,
+          `${event.item_id}:${entry[0]}`,
+          openaiMetadata({ itemId: event.item_id }),
+        ),
+      state.lifecycle,
+    )
+  return [
+    {
+      ...state,
+      lifecycle: Lifecycle.reasoningStart(
+        closed,
+        events,
+        `${event.item_id}:${event.summary_index}`,
+        openaiMetadata({ itemId: event.item_id, reasoningEncryptedContent: item.encryptedContent ?? null }),
+      ),
+      reasoningItems: {
+        ...state.reasoningItems,
+        [event.item_id]: {
+          ...item,
+          summaryParts: {
+            ...Object.fromEntries(
+              Object.entries(item.summaryParts).map((entry) =>
+                entry[1] === "can-conclude" ? [entry[0], "concluded" as const] : entry,
+              ),
+            ),
+            [event.summary_index]: "active",
+          },
+        },
+      },
+    },
+    events,
+  ]
+}
+
+const onReasoningSummaryPartDone = (state: ParserState, event: OpenAIResponsesEvent): StepResult => {
+  if (!event.item_id || event.summary_index === undefined) return [state, NO_EVENTS]
+  const item = state.reasoningItems[event.item_id]
+  if (!item) return [state, NO_EVENTS]
+  const events: LLMEvent[] = []
+  return [
+    {
+      ...state,
+      lifecycle:
+        state.store !== false
+          ? Lifecycle.reasoningEnd(
+              state.lifecycle,
+              events,
+              `${event.item_id}:${event.summary_index}`,
+              openaiMetadata({ itemId: event.item_id }),
+            )
+          : state.lifecycle,
+      reasoningItems: {
+        ...state.reasoningItems,
+        [event.item_id]: {
+          ...item,
+          summaryParts: {
+            ...item.summaryParts,
+            [event.summary_index]: state.store !== false ? "concluded" : "can-conclude",
+          },
+        },
+      },
+    },
+    events,
+  ]
+}
+
 const onFunctionCallArgumentsDelta = Effect.fn("OpenAIResponses.onFunctionCallArgumentsDelta")(function* (
  state: ParserState,
  event: OpenAIResponsesEvent,
@@ -615,6 +790,17 @@ const onOutputItemDone = Effect.fn("OpenAIResponses.onOutputItemDone")(function*
  if (isReasoningItem(item)) {
    const events: LLMEvent[] = []
    const providerMetadata = reasoningMetadata(item)
+    const reasoningItem = state.reasoningItems[item.id]
+    if (reasoningItem) {
+      const lifecycle = Object.entries(reasoningItem.summaryParts)
+        .filter((entry) => entry[1] === "active" || entry[1] === "can-conclude")
+        .reduce(
+          (lifecycle, entry) => Lifecycle.reasoningEnd(lifecycle, events, `${item.id}:${entry[0]}`, providerMetadata),
+          state.lifecycle,
+        )
+      const { [item.id]: _removed, ...reasoningItems } = state.reasoningItems
+      return [{ ...state, lifecycle, reasoningItems }, events] satisfies StepResult
+    }
    if (!state.lifecycle.reasoning.has(item.id)) {
      const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
      events.push(LLMEvent.reasoningStart({ id: item.id, providerMetadata }))
@@ -683,6 +869,10 @@ const step = (state: ParserState, event: OpenAIResponsesEvent) => {
    event.type === "response.reasoning_summary_text.done"
  )
    return Effect.succeed(onReasoningDone(state, event))
+  if (event.type === "response.reasoning_summary_part.added")
+    return Effect.succeed(onReasoningSummaryPartAdded(state, event))
+  if (event.type === "response.reasoning_summary_part.done")
+    return Effect.succeed(onReasoningSummaryPartDone(state, event))
  if (event.type === "response.output_item.added") return Effect.succeed(onOutputItemAdded(state, event))
  if (event.type === "response.function_call_arguments.delta") return onFunctionCallArgumentsDelta(state, event)
  if (event.type === "response.output_item.done") return onOutputItemDone(state, event)
@@ -709,7 +899,13 @@ export const protocol = Protocol.make({
  },
  stream: {
    event: Protocol.jsonEvent(OpenAIResponsesEvent),
-    initial: () => ({ hasFunctionCall: false, tools: ToolStream.empty<string>(), lifecycle: Lifecycle.initial() }),
+    initial: (request) => ({
+      hasFunctionCall: false,
+      tools: ToolStream.empty<string>(),
+      lifecycle: Lifecycle.initial(),
+      reasoningItems: {},
+      store: OpenAIOptions.store(request),
+    }),
    step,
    terminal: (event) => TERMINAL_TYPES.has(event.type),
  },
--- a/packages/llm/src/protocols/utils/lifecycle.ts
+++ b/packages/llm/src/protocols/utils/lifecycle.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[],
--- a/packages/llm/src/protocols/utils/openai-options.ts
+++ b/packages/llm/src/protocols/utils/openai-options.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
--- a/packages/llm/src/providers/openai-options.ts
+++ b/packages/llm/src/providers/openai-options.ts
@@ -1,5 +1,8 @@
 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
@@ -7,7 +10,10 @@ export interface OpenAIOptionsInput {
  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"
--- a/packages/llm/src/providers/openai.ts
+++ b/packages/llm/src/providers/openai.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")

--- a/packages/llm/src/route/client.ts
+++ b/packages/llm/src/route/client.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,
          ),
--- a/packages/llm/src/route/protocol.ts
+++ b/packages/llm/src/route/protocol.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. */