fix(session): retry title generation and fall back to the session model

packages/opencode/src/session/prompt.ts

@@ -190,6 +190,39 @@ const layer = Layer.effect(
       return parts
     })
 
+    const generateTitle = Effect.fn("SessionPrompt.generateTitle")(function* (input: {
+      agent: Agent.Info
+      user: SessionV1.User
+      context: SessionV1.WithParts[]
+      subtasks: SessionV1.SubtaskPart[]
+      onlySubtasks: boolean
+      model: Provider.Model
+      small: boolean
+      sessionID: SessionID
+    }) {
+      const msgs = input.onlySubtasks
+        ? [{ role: "user" as const, content: input.subtasks.map((p) => p.prompt).join("\n") }]
+        : yield* MessageV2.toModelMessagesEffect(input.context, input.model)
+      return yield* llm
+        .stream({
+          agent: input.agent,
+          user: input.user,
+          system: [],
+          small: input.small,
+          tools: {},
+          model: input.model,
+          sessionID: input.sessionID,
+          retries: 2,
+          messages: [{ role: "user", content: "Generate a title for this conversation:\n" }, ...msgs],
+        })
+        .pipe(
+          Stream.filter(LLMEvent.is.textDelta),
+          Stream.map((e) => e.text),
+          Stream.mkString,
+          Effect.orDie,
+        )
+    })
+
     const title = Effect.fn("SessionPrompt.ensureTitle")(function* (input: {
       session: Session.Info
       history: SessionV1.WithParts[]
@@ -203,7 +236,6 @@ const layer = Layer.effect(
         m.info.role === "user" && !m.parts.every((p) => "synthetic" in p && p.synthetic)
       const idx = input.history.findIndex(real)
       if (idx === -1) return
-      if (input.history.filter(real).length !== 1) return
 
       const context = input.history.slice(0, idx + 1)
       const firstUser = context[idx]
@@ -219,27 +251,31 @@ const layer = Layer.effect(
         ? yield* provider.getModel(ag.model.providerID, ag.model.modelID)
         : ((yield* provider.getSmallModel(input.providerID)) ??
           (yield* provider.getModel(input.providerID, input.modelID)))
-      const msgs = onlySubtasks
-        ? [{ role: "user" as const, content: subtasks.map((p) => p.prompt).join("\n") }]
-        : yield* MessageV2.toModelMessagesEffect(context, mdl)
-      const text = yield* llm
-        .stream({
-          agent: ag,
-          user: firstInfo,
-          system: [],
-          small: true,
-          tools: {},
-          model: mdl,
-          sessionID: input.session.id,
-          retries: 2,
-          messages: [{ role: "user", content: "Generate a title for this conversation:\n" }, ...msgs],
-        })
-        .pipe(
-          Stream.filter(LLMEvent.is.textDelta),
-          Stream.map((e) => e.text),
-          Stream.mkString,
-          Effect.orDie,
-        )
+      const base = {
+        agent: ag,
+        user: firstInfo,
+        context,
+        subtasks,
+        onlySubtasks,
+        sessionID: input.session.id,
+      }
+      const text = yield* generateTitle({ ...base, model: mdl, small: true }).pipe(
+        Effect.catchCause((cause) =>
+          Effect.gen(function* () {
+            yield* Effect.logWarning(
+              "session title generation with small model failed; falling back to session model",
+              {
+                "session.id": input.session.id,
+                providerID: mdl.providerID,
+                modelID: mdl.id,
+                error: Cause.squash(cause),
+              },
+            )
+            const fallback = yield* provider.getModel(input.providerID, input.modelID)
+            return yield* generateTitle({ ...base, model: fallback, small: false })
+          }),
+        ),
+      )
       const cleaned = text
         .replace(/<think>[\s\S]*?<\/think>\s*/g, "")
         .split("\n")

packages/opencode/test/lib/llm-server.ts

@@ -609,6 +609,8 @@ function isTitleRequest(body: unknown): boolean {
   return JSON.stringify(body).includes("Generate a title for this conversation")
 }
 
+const isTitleHit = (hit: Hit): boolean => isTitleRequest(hit.body)
+
 namespace TestLLMServer {
   export interface Service {
     readonly url: string
@@ -622,6 +624,7 @@ namespace TestLLMServer {
     readonly reason: (value: string, opts?: { text?: string; usage?: Usage }) => Effect.Effect<void>
     readonly fail: (message?: unknown) => Effect.Effect<void>
     readonly error: (status: number, body: unknown) => Effect.Effect<void>
+    readonly titleError: (status: number, body: unknown) => Effect.Effect<void>
     readonly hang: Effect.Effect<void>
     readonly hold: (value: string, wait: PromiseLike<unknown>) => Effect.Effect<void>
     readonly reset: Effect.Effect<void>
@@ -669,13 +672,31 @@ export class TestLLMServer extends Context.Service<TestLLMServer, TestLLMServer.
         return first.item
       }
 
+      const pullMatch = (hit: Hit) => {
+        const index = list.findIndex((entry) => entry.match && entry.match(hit))
+        if (index === -1) return
+        const first = list[index]
+        list = [...list.slice(0, index), ...list.slice(index + 1)]
+        return first.item
+      }
+
       const handle = Effect.fn("TestLLMServer.handle")(function* (mode: "chat" | "responses") {
         const req = yield* HttpServerRequest.HttpServerRequest
         const body = yield* req.json.pipe(Effect.orElseSucceed(() => ({})))
         const current = hit(req.originalUrl, body)
         if (isTitleRequest(body)) {
           hits = [...hits, current]
           yield* notify()
+          const override = pullMatch(current)
+          if (override) {
+            if (override.type !== "sse") return fail(override)
+            if (mode === "responses") return send(responses(override, modelFrom(body)))
+            if (override.reset) {
+              yield* reset(override)
+              return HttpServerResponse.empty()
+            }
+            return send(override)
+          }
           const auto: Sse = { type: "sse", head: [role()], tail: [textLine("E2E Title"), finishLine("stop")] }
           if (mode === "responses") return send(responses(auto, modelFrom(body)))
           return send(auto)
@@ -750,6 +771,9 @@ export class TestLLMServer extends Context.Service<TestLLMServer, TestLLMServer.
         error: Effect.fn("TestLLMServer.error")(function* (status: number, body: unknown) {
           queue(httpError(status, body))
         }),
+        titleError: Effect.fn("TestLLMServer.titleError")(function* (status: number, body: unknown) {
+          queueMatch(isTitleHit, httpError(status, body))
+        }),
         hang: Effect.gen(function* () {
           queue(reply().hang().item())
         }).pipe(Effect.withSpan("TestLLMServer.hang")),

packages/opencode/test/session/prompt.test.ts

@@ -2468,3 +2468,79 @@ noLLMServer.instance(
     }),
   30_000,
 )
+
+// Session title generation
+
+const isTitleBody = (body: Record<string, unknown>) =>
+  JSON.stringify(body).includes("Generate a title for this conversation")
+
+it.instance("title generation falls back to a second attempt when the title request errors", () =>
+  Effect.gen(function* () {
+    const { llm } = yield* useServerConfig(providerCfg)
+    const prompt = yield* SessionPrompt.Service
+    const sessions = yield* Session.Service
+    const chat = yield* sessions.create({})
+
+    yield* llm.titleError(400, { error: { message: "title request failed" } })
+    yield* llm.text("world")
+    yield* prompt.prompt({
+      sessionID: chat.id,
+      agent: "build",
+      noReply: true,
+      parts: [{ type: "text", text: "hello" }],
+    })
+    yield* prompt.loop({ sessionID: chat.id })
+
+    yield* pollWithTimeout(
+      Effect.gen(function* () {
+        const session = yield* sessions.get(chat.id)
+        return session.title === "E2E Title" ? true : undefined
+      }),
+      "title fallback never set the session title",
+    )
+    expect(yield* llm.pending).toBe(0)
+
+    const titleBodies = (yield* llm.inputs).filter(isTitleBody)
+    expect(titleBodies).toHaveLength(2)
+  }),
+)
+
+it.instance("title generation retries on a later message while the title is still default", () =>
+  Effect.gen(function* () {
+    const { llm } = yield* useServerConfig(providerCfg)
+    const prompt = yield* SessionPrompt.Service
+    const sessions = yield* Session.Service
+    const chat = yield* sessions.create({})
+
+    yield* llm.titleError(400, { error: { message: "title request failed" } })
+    yield* llm.titleError(400, { error: { message: "title request failed" } })
+    yield* llm.text("world one")
+    yield* prompt.prompt({
+      sessionID: chat.id,
+      agent: "build",
+      noReply: true,
+      parts: [{ type: "text", text: "first" }],
+    })
+    yield* prompt.loop({ sessionID: chat.id })
+    yield* llm.wait(3)
+    expect((yield* sessions.get(chat.id)).title).toMatch(/^New session - /)
+
+    yield* llm.text("world two")
+    yield* prompt.prompt({
+      sessionID: chat.id,
+      agent: "build",
+      noReply: true,
+      parts: [{ type: "text", text: "second" }],
+    })
+    yield* prompt.loop({ sessionID: chat.id })
+
+    yield* pollWithTimeout(
+      Effect.gen(function* () {
+        const session = yield* sessions.get(chat.id)
+        return session.title === "E2E Title" ? true : undefined
+      }),
+      "title generation did not retry on a later message",
+    )
+    expect(yield* llm.pending).toBe(0)
+  }),
+)