fix(ai): support streams without finish reasons

packages/ai/src/protocols/openai-chat.ts

@@ -241,6 +241,8 @@ export interface ParserState {
   readonly reasoningEmitted: boolean
   readonly latestToolIndex?: number
   readonly nextToolIndex: number
+  readonly outputStarted: boolean
+  readonly requireFinishReason: boolean
 }
 
 // =============================================================================
@@ -707,9 +709,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
       Boolean(delta?.content) ||
       reasoning !== undefined ||
       (Array.isArray(delta?.reasoning_details) && delta.reasoning_details.length > 0) ||
-      toolDeltas.some(
-        (tool) => Boolean(tool.id) || Boolean(tool.function?.name) || Boolean(tool.function?.arguments),
-      )
+      toolDeltas.some((tool) => Boolean(tool.id) || Boolean(tool.function?.name) || Boolean(tool.function?.arguments))
     if (state.finishReason !== undefined) {
       if (hasLateContent)
         return yield* ProviderShared.eventError(ADAPTER, "OpenAI Chat received content after the finish reason")
@@ -749,8 +749,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
       const fallback = toolDeltas.length > 1 ? position : (latestToolIndex ?? position)
       const fallbackTool = tools[fallback] ?? pendingTools[fallback]
       const index =
-        tool.index ?? matched ??
-        (tool.id && fallbackTool?.id && fallbackTool.id !== tool.id ? nextToolIndex : fallback)
+        tool.index ?? matched ?? (tool.id && fallbackTool?.id && fallbackTool.id !== tool.id ? nextToolIndex : fallback)
       const current = tools[index]
       const pending = pendingTools[index]
       const id = current?.id ?? pending?.id ?? (tool.id || undefined)
@@ -806,6 +805,8 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
         reasoningEmitted,
         latestToolIndex,
         nextToolIndex,
+        outputStarted: state.outputStarted || hasLateContent,
+        requireFinishReason: state.requireFinishReason,
       },
       events,
     ] as const
@@ -836,6 +837,23 @@ const finishEvents = (state: ParserState): ReadonlyArray<LLMEvent> => {
   return events
 }
 
+const onHalt = (state: ParserState) =>
+  Effect.gen(function* () {
+    if (state.finishReason !== undefined || state.requireFinishReason) return finishEvents(state)
+    if (!state.outputStarted) return []
+    if (Object.keys(state.pendingTools).length > 0)
+      return yield* ProviderShared.eventError(ADAPTER, "OpenAI Chat tool call delta is missing id or name")
+    // Chat has no per-call stop event, so an accepted EOF must finalize every
+    // accumulated tool input before publishing the synthetic terminal reason.
+    const finished = yield* ToolStream.finishAll(ADAPTER, state.tools)
+    return finishEvents({
+      ...state,
+      tools: finished.tools,
+      toolCallEvents: finished.events,
+      finishReason: { normalized: "unknown" },
+    })
+  })
+
 // =============================================================================
 // Protocol And OpenAI Route
 // =============================================================================
@@ -863,9 +881,11 @@ export const protocol = Protocol.make({
       reasoningDetailsObserved: false,
       reasoningEmitted: false,
       nextToolIndex: 0,
+      outputStarted: false,
+      requireFinishReason: request.model.compatibility?.requireFinishReason ?? true,
     }),
     step,
-    onHalt: finishEvents,
+    onHalt,
   },
 })
 

packages/ai/src/route/client.ts

@@ -7,7 +7,7 @@ import { Framing } from "./framing"
 import { HttpTransport } from "./transport"
 import type { HttpMiddleware, Transport, TransportRuntime } from "./transport"
 import { WebSocketExecutor } from "./transport"
-import type { Protocol } from "./protocol"
+import type { Protocol, ProtocolStream } from "./protocol"
 import { applyCachePolicy } from "../cache-policy"
 import * as ProviderShared from "../protocols/shared"
 import type { ProtocolID, ProviderOptions } from "../schema"
@@ -243,28 +243,56 @@ const incompleteStreamError = (route: string) =>
     }),
   })
 
-const requireTerminalEvent = (route: string) => (events: Stream.Stream<LLMEvent, AIError>) =>
+const ensureTerminalEvent = (route: string, required: boolean) => (events: Stream.Stream<LLMEvent, AIError>) =>
   Stream.suspend(() => {
     let terminal = false
+    let output = false
+    const fallback = Stream.suspend(() => {
+      if (terminal) return Stream.empty
+      if (required || !output) return Stream.fail(incompleteStreamError(route))
+      // The compatibility override trusts a clean stream end, but it cannot
+      // recover the provider's omitted reason.
+      const reason = { normalized: "unknown" as const }
+      return Stream.make(LLMEvent.stepFinish({ index: 0, reason }), LLMEvent.finish({ reason }))
+    })
     return events.pipe(
       Stream.mapEffect((event) => {
         if (terminal)
           return Effect.fail(
             ProviderShared.eventError(route, `Provider emitted ${event.type} after the terminal event`),
           )
+        output = true
         if (LLMEvent.is.finish(event) || LLMEvent.is.providerError(event)) terminal = true
         return Effect.succeed(event)
       }),
-      Stream.onEnd(
-        Effect.suspend(() =>
-          terminal
-            ? Effect.void
-            : Effect.fail(incompleteStreamError(route)),
-        ),
-      ),
+      Stream.concat(fallback),
     )
   })
 
+type ProtocolEvent<Event> = { readonly type: "event"; readonly event: Event } | { readonly type: "halt" }
+
+const parseProtocolEvents = <Event, State>(
+  events: Stream.Stream<Event, AIError>,
+  request: LLMRequest,
+  protocol: { readonly stream: ProtocolStream<unknown, Event, State> },
+) =>
+  events.pipe(
+    Stream.map((event): ProtocolEvent<Event> => ({ type: "event", event })),
+    // A normal halt becomes an in-band parser input so finalization may fail.
+    Stream.concat(Stream.succeed({ type: "halt" } as const)),
+    Stream.mapAccumEffect(
+      () => protocol.stream.initial(request),
+      (state, event) => {
+        if (event.type === "event") return protocol.stream.step(state, event.event)
+        if (!protocol.stream.onHalt) return Effect.succeed([state, []] as const)
+        const events = protocol.stream.onHalt(state)
+        return Effect.isEffect(events)
+          ? events.pipe(Effect.map((events) => [state, events] as const))
+          : Effect.succeed([state, events] as const)
+      },
+    ),
+  )
+
 function makeFromTransport<Body, Prepared, Frame, Event, State>(
   input: MakeTransportInput<Body, Prepared, Frame, Event, State>,
 ): Route<Body, Prepared> {
@@ -329,14 +357,9 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
             Stream.mapEffect(decodeEvent(route)),
             protocol.stream.terminal ? Stream.takeUntil(protocol.stream.terminal) : (stream) => stream,
           )
-        return events.pipe(
-          Stream.mapAccumEffect(
-            () => protocol.stream.initial(request),
-            protocol.stream.step,
-            protocol.stream.onHalt ? { onHalt: protocol.stream.onHalt } : undefined,
-          ),
+        return parseProtocolEvents(events, request, protocol).pipe(
           Stream.catchCause((cause) => Stream.fail(streamError(route, `Failed to read ${route} stream`, cause))),
-          requireTerminalEvent(route),
+          ensureTerminalEvent(route, request.model.compatibility?.requireFinishReason ?? true),
         )
       },
     } satisfies Route<Body, Prepared>

packages/ai/src/route/protocol.ts

@@ -59,8 +59,8 @@ export interface ProtocolStream<Frame, Event, State> {
   readonly step: (state: State, event: Event) => Effect.Effect<readonly [State, ReadonlyArray<LLMEvent>], AIError>
   /** Optional request-completion signal for transports that do not end naturally. */
   readonly terminal?: (event: Event) => boolean
-  /** Optional flush emitted when the framed stream ends. */
-  readonly onHalt?: (state: State) => ReadonlyArray<LLMEvent>
+  /** Optional flush emitted when the framed stream ends successfully. */
+  readonly onHalt?: (state: State) => ReadonlyArray<LLMEvent> | Effect.Effect<ReadonlyArray<LLMEvent>, AIError>
 }
 
 /**

packages/ai/test/adapter.test.ts

@@ -138,6 +138,29 @@ describe("llm route", () => {
     }),
   )
 
+  unterminated.effect("synthesizes an unknown finish when a terminal event is not required", () =>
+    Effect.gen(function* () {
+      const response = yield* (yield* LLMClient.Service).generate(
+        LLMRequest.update(request, {
+          model: updateModel(request.model, { compatibility: { requireFinishReason: false } }),
+        }),
+      )
+
+      expect(response.text).toBe("partial")
+      expect(response.finishReason).toEqual({ normalized: "unknown" })
+      expect(response.events.slice(-2)).toEqual([
+        {
+          type: "step-finish",
+          index: 0,
+          reason: { normalized: "unknown" },
+          usage: undefined,
+          providerMetadata: undefined,
+        },
+        { type: "finish", reason: { normalized: "unknown" }, usage: undefined },
+      ])
+    }),
+  )
+
   it.effect("selects routes by model route value", () =>
     Effect.gen(function* () {
       const prepared = yield* compileRequest(

packages/ai/test/provider/openai-chat.test.ts

@@ -40,6 +40,10 @@ const request = LLM.request({
   generation: { maxTokens: 20, temperature: 0 },
 })
 
+const optionalFinishRequest = LLMRequest.update(request, {
+  model: LanguageModel.update(model, { compatibility: { requireFinishReason: false } }),
+})
+
 describe("OpenAI Chat route", () => {
   it.effect("prepares OpenAI Chat payload", () =>
     Effect.gen(function* () {
@@ -596,6 +600,20 @@ describe("OpenAI Chat route", () => {
     }),
   )
 
+  it.effect("accepts text and usage without a finish reason when configured", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        deltaChunk({ role: "assistant", content: "Hello" }),
+        usageChunk({ prompt_tokens: 5, completion_tokens: 1, total_tokens: 6 }),
+      )
+      const response = yield* LLMClient.generate(optionalFinishRequest).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.text).toBe("Hello")
+      expect(response.finishReason).toEqual({ normalized: "unknown" })
+      expect(response.usage).toMatchObject({ inputTokens: 5, outputTokens: 1, totalTokens: 6 })
+    }),
+  )
+
   it.effect("parses and replays OpenAI-compatible reasoning fields", () =>
     Effect.gen(function* () {
       const fields = ["reasoning_content", "reasoning", "reasoning_text"] as const
@@ -1145,21 +1163,89 @@ describe("OpenAI Chat route", () => {
     }),
   )
 
-  it.effect("fails on malformed stream events", () =>
+  it.effect("finalizes a streamed tool call without a finish reason when configured", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        deltaChunk({
+          role: "assistant",
+          tool_calls: [{ index: 0, id: "call_1", function: { name: "lookup", arguments: '{"query"' } }],
+        }),
+        deltaChunk({ tool_calls: [{ index: 0, function: { arguments: ':"weather"}' } }] }),
+      )
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(optionalFinishRequest, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.toolCalls).toMatchObject([{ id: "call_1", name: "lookup", input: { query: "weather" } }])
+      expect(response.finishReason).toEqual({ normalized: "unknown" })
+    }),
+  )
+
+  it.effect("settles malformed tool input without a finish reason when configured", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        deltaChunk({
+          role: "assistant",
+          tool_calls: [{ index: 0, id: "call_1", function: { name: "lookup", arguments: '{"query"' } }],
+        }),
+      )
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(optionalFinishRequest, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.events.filter(LLMEvent.is.toolInputError)).toMatchObject([
+        { id: "call_1", name: "lookup", raw: '{"query"' },
+      ])
+      expect(response.toolCalls).toEqual([])
+      expect(response.finishReason).toEqual({ normalized: "unknown" })
+    }),
+  )
+
+  it.effect("rejects incomplete tool identity without a finish reason when configured", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(deltaChunk({ tool_calls: [{ index: 0, id: "call_1", function: { arguments: "{}" } }] }))
+      const error = yield* LLMClient.generate(optionalFinishRequest).pipe(
+        Effect.provide(fixedResponse(body)),
+        Effect.flip,
+      )
+
+      expect(error.message).toContain("OpenAI Chat tool call delta is missing id or name")
+    }),
+  )
+
+  it.effect("rejects an empty stream when a finish reason is not required", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(optionalFinishRequest).pipe(
+        Effect.provide(fixedResponse(sseEvents())),
+        Effect.flip,
+      )
+
+      expect(error.reason).toMatchObject({ _tag: "InvalidProviderOutput", classification: "incomplete-stream" })
+    }),
+  )
+
+  it.effect("fails on malformed stream events when a finish reason is not required", () =>
     Effect.gen(function* () {
       const body = sseEvents(deltaChunk({ content: 123 }))
-      const error = yield* LLMClient.generate(request).pipe(Effect.provide(fixedResponse(body)), Effect.flip)
+      const error = yield* LLMClient.generate(optionalFinishRequest).pipe(
+        Effect.provide(fixedResponse(body)),
+        Effect.flip,
+      )
 
       expect(error.message).toContain("Invalid openai/openai-chat stream event")
     }),
   )
 
-  it.effect("surfaces transport errors that occur mid-stream", () =>
+  it.effect("surfaces transport errors when a finish reason is not required", () =>
     Effect.gen(function* () {
       const layer = truncatedStream([
         `data: ${JSON.stringify(deltaChunk({ role: "assistant", content: "Hello" }))}\n\n`,
       ])
-      const error = yield* LLMClient.generate(request).pipe(Effect.provide(layer), Effect.flip)
+      const error = yield* LLMClient.generate(optionalFinishRequest).pipe(Effect.provide(layer), Effect.flip)
 
       expect(error.message).toContain("Failed to read openai/openai-chat stream")
     }),

packages/core/test/session-runner.test.ts

@@ -2440,6 +2440,30 @@ describe("SessionRunnerLLM", () => {
     }),
   )
 
+  it.effect("continues after an unknown finish containing a local tool call", () =>
+    Effect.gen(function* () {
+      const session = yield* setup
+      yield* admit(session, "Echo this")
+      yield* TestLLM.push(
+        TestLLM.complete(
+          { reason: { normalized: "unknown" } },
+          LLMEvent.toolCall({ id: "call-echo", name: "echo", input: { text: "hello" } }),
+        ),
+        TestLLM.text("Done", "text-final"),
+      )
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(executions).toEqual(["hello"])
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Echo this" },
+        { type: "assistant", finish: "unknown", content: [{ type: "tool", state: { status: "completed" } }] },
+        { type: "assistant", finish: "stop", content: [{ type: "text", text: "Done" }] },
+      ])
+    }),
+  )
+
   it.effect("reloads a model switch before a tool-driven continuation step", () =>
     Effect.gen(function* () {
       const session = yield* setup