diff --git a/packages/core/src/session/title.ts b/packages/core/src/session/title.ts index 37e2233fef1..2f40d15633c 100644 --- a/packages/core/src/session/title.ts +++ b/packages/core/src/session/title.ts @@ -56,7 +56,8 @@ export const layer = Layer.effect( readonly text: string readonly model: SessionRunnerModel.Resolved }) { - const chunks: string[] = [] + const chunks: Array<{ id: string; text: string }> = [] + const commentary = new Set() let failed = false let usage: SessionUsage.Recorded | undefined const recordUsage = Effect.suspend(() => @@ -76,9 +77,16 @@ export const layer = Layer.effect( messages: [Message.user(input.text)], }) if (prepared.event.result !== undefined) return prepared.event.result + const metadataKey = prepared.request.model.route.providerMetadataKey ?? prepared.request.model.provider yield* llm.stream(prepared.request, prepared.options).pipe( Stream.runForEach((event) => { - if (LLMEvent.is.textDelta(event)) chunks.push(event.text) + // Responses models may emit commentary-phase text, and the phase can arrive only when the block ends. + if ( + (LLMEvent.is.textStart(event) || LLMEvent.is.textEnd(event)) && + event.providerMetadata?.[metadataKey]?.phase === "commentary" + ) + commentary.add(event.id) + if (LLMEvent.is.textDelta(event)) chunks.push({ id: event.id, text: event.text }) if (LLMEvent.is.stepFinish(event)) { const step = SessionUsage.record(event.usage, input.model.cost) usage = usage ? SessionUsage.add(usage, step) : step @@ -95,6 +103,8 @@ export const layer = Layer.effect( yield* recordUsage if (failed) return return chunks + .filter((chunk) => !commentary.has(chunk.id)) + .map((chunk) => chunk.text) .join("") .split("\n") .map((line) => line.trim()) diff --git a/packages/core/test/session-title.test.ts b/packages/core/test/session-title.test.ts index 4e1063bcfa5..c26404e906c 100644 --- a/packages/core/test/session-title.test.ts +++ b/packages/core/test/session-title.test.ts @@ -10,7 +10,7 @@ import { TransportError, type LLMRequest, } from "@opencode/ai/effect" -import { OpenAIChat } from "@opencode/ai/protocols" +import { OpenAIChat, OpenAIResponses } from "@opencode/ai/protocols" import { Agent } from "@opencode/core/agent" import { Database } from "@opencode/core/database/database" import { AppNodeBuilder } from "@opencode/core/effect/app-node-builder" @@ -35,7 +35,7 @@ import { Model } from "@opencode/core/model" import { Provider } from "@opencode/core/provider" import { AbsolutePath } from "@opencode/core/schema" import { Money } from "@opencode/schema/money" -import { Deferred, Effect, Fiber, Layer, Stream } from "effect" +import { Deferred, Effect, Fiber, Layer, Schema, Stream } from "effect" import { testEffect } from "./lib/effect" let requests: LLMRequest[] = [] @@ -84,11 +84,11 @@ const successfulTitle = () => reason: { normalized: "stop" }, }), ) -let titleStream: () => Stream.Stream = successfulTitle +let titleStream: (request: LLMRequest) => Stream.Stream = successfulTitle const client = Layer.mock(LLMClient.Service)({ stream: (request: LLMRequest) => { requests.push(request) - return titleStream() + return titleStream(request) }, generate: () => Effect.die("unused"), }) @@ -237,6 +237,46 @@ it.effect("generates a title from the sole user message and renames the session" }), ) +for (const timing of ["start", "end"] as const) + it.effect(`excludes commentary from titles when the response phase arrives at ${timing}`, () => + Effect.gen(function* () { + yield* enableTitleAgent + const sessionID = Session.ID.make(`ses_title_phase_${timing}`) + yield* insertSession(sessionID) + yield* prompt(sessionID, "Investigate performance across four hosts without making changes") + titleStream = (request) => + Stream.fromIterable([ + { + type: "response.output_item.added", + item: { type: "message", id: "commentary", ...(timing === "start" ? { phase: "commentary" } : {}) }, + }, + { + type: "response.output_text.delta", + item_id: "commentary", + delta: "We need title only. Need capture investigation four machines performance read-only. Title concise.", + }, + { type: "response.output_item.done", item: { type: "message", id: "commentary", phase: "commentary" } }, + { type: "response.output_item.added", item: { type: "message", id: "final", phase: "final_answer" } }, + { type: "response.output_text.delta", item_id: "final", delta: "Read-only performance investigation" }, + { type: "response.output_item.done", item: { type: "message", id: "final", phase: "final_answer" } }, + { type: "response.completed", response: { id: "response_title" } }, + ]).pipe( + Stream.map((event) => Schema.decodeUnknownSync(OpenAIResponses.protocol.stream.event)(JSON.stringify(event))), + Stream.mapAccumEffect( + () => OpenAIResponses.protocol.stream.initial(request), + OpenAIResponses.protocol.stream.step, + ), + ) + + const title = yield* SessionTitle.Service + yield* title.generate(sessionID) + + const store = yield* SessionStore.Service + expect(requests).toHaveLength(1) + expect((yield* store.get(sessionID))?.title).toBe("Read-only performance investigation") + }), + ) + it.effect("runs title hooks instead of context hooks", () => Effect.gen(function* () { yield* enableTitleAgent