fix(chat): route fallback models through chat completions (#5121)

Author: rschlaefli

.gitignore

@@ -46,6 +46,9 @@ dumps/**/*.dump.gpg
 # Temporary files
 *.tmp
 *.log
+NOCOMMIT*
+20??-??-??_chatbot*.csv
+output/
 
 # Environment-specific files
 .env.*

AGENTS.md

@@ -215,6 +215,7 @@ Without Traefik, use `http://localhost:<port>` directly. The `*.klicker.com` dom
 - **Assistant UI chat drop targets**: `ComposerPrimitive.AttachmentDropzone` must wrap both normal and edit chat composer roots; it owns the drag/drop capture handlers that prevent native browser file navigation, while Klicker-specific limits stay in local composer code. (`apps/chat/src/components/thread.tsx`)
 - **Deployed chat model registries default attachment support off**: `apps/chat` loads `CHAT_MODEL_REGISTRY_JSON` via `chatModelRegistry.ts`, where omitted `supportsImageAttachments` values default to `false`; if deployment values override the built-in registry, each image-capable model must set the flag explicitly in `deploy/env-uzh-*/values.yaml` or the attach button disappears.
 - **OpenAI Responses storage and tool calls**: GPT-5.5 via the OpenAI Responses API can reference prior response items across tool-call steps; with `store: false`, LiteLLM/Azure can return "item not found" for those references. Keep `CHAT_OPENAI_STORE_RESPONSES=true` in shared/staged deployments that use Responses-compatible OpenAI backends, while local OpenRouter-style setups can leave it false. (`apps/chat/src/app/api/chatbots/[chatbotId]/chat/route.ts`, `deploy/env-uzh-*/values.yaml`)
+- **Chat fallback allow-lists**: Zero-credit course chatbots require a usable fallback model (`gpt-4.1-mini` / `CHAT_FALLBACK_MODEL_ID`) and explicit chatbot `allowedModelIds` must include it. Use `packages/prisma-data/src/scripts/2026-06-15_ensure_chatbot_fallback_model.ts` to audit or fix allow-lists. (`apps/chat/src/lib/server/chatModelRegistry.ts`)
 - **Embedded PWA messaging trust boundary**: For embedded PWA pages, use a parent-initiated `postMessage` handshake to capture `event.origin` and avoid `'*'` target origins; do not add a second per-platform messaging allowlist in page code. Embedding permission remains enforced separately by ingress `frame-ancestors`. (`apps/frontend-pwa/src/pages/course/[courseId]/practiceQuizzes/[id].tsx`, `deploy/charts/klicker-uzh-v3/templates/ingress-frontend-pwa.yaml`)
 - **Local embed harness target**: `util/embed-harness/` is for local verification only and should target the branch-local PWA (`http://127.0.0.1:3101/...`), not `https://pwa.klicker.com/...`, because production CSP / `frame-ancestors` blocks localhost embedding. (`util/embed-harness/`)
 - **Chat PWA login redirects**: `apps/chat/src/app/noLogin/page.tsx` must pass an absolute chat URL to the PWA login `redirect_to`; a relative chatbot path makes the PWA redirect to its own domain and 404. Local chat dev also needs ignored local env values for the backend `APP_SECRET` and `DATABASE_URL` so participant cookies verify and Prisma can load chatbot data.

apps/chat/src/app/api/chatbots/[chatbotId]/chat/route.ts

@@ -3,6 +3,7 @@ import {
   getAllowedReasoningEffortsForModel,
   getAutomaticModelId,
   getChatModelRegistry,
+  type ChatModelConfig,
 } from '@/src/lib/server/chatModelRegistry'
 import { ensureImagePreviewBase64 } from '@/src/lib/server/imagePreview'
 import { getOpenAIResponsesStore } from '@/src/lib/server/openaiResponsesOptions'
@@ -101,7 +102,16 @@ type ModelRouting = {
   baseUrl: string | undefined
 }
 
-function getModel(chatbot: Chatbot, modelId: string) {
+function getOpenAIModel(
+  provider: ReturnType<typeof createOpenAI>,
+  modelConfig: ChatModelConfig
+) {
+  return modelConfig.supportsReasoning
+    ? provider.responses(modelConfig.deploymentId)
+    : provider.chat(modelConfig.deploymentId)
+}
+
+function getModel(chatbot: Chatbot, modelConfig: ChatModelConfig) {
   // Use per-chatbot configuration if available
   const hasCustomKey =
     typeof chatbot.openaiApiKey === 'string' && chatbot.openaiApiKey.length > 0
@@ -136,11 +146,14 @@ function getModel(chatbot: Chatbot, modelId: string) {
     }
 
     return {
-      model: createOpenAI({
-        baseURL: baseUrl,
-        apiKey: apiKey || 'no-key',
-        fetch: responsesApiFetch,
-      })(modelId),
+      model: getOpenAIModel(
+        createOpenAI({
+          baseURL: baseUrl,
+          apiKey: apiKey || 'no-key',
+          fetch: responsesApiFetch,
+        }),
+        modelConfig
+      ),
       routing,
     }
   }
@@ -153,11 +166,14 @@ function getModel(chatbot: Chatbot, modelId: string) {
   }
 
   return {
-    model: createOpenAI({
-      baseURL: process.env.OPENAI_BASE_URL,
-      apiKey: process.env.OPENAI_API_KEY || 'no-key',
-      fetch: responsesApiFetch,
-    })(modelId),
+    model: getOpenAIModel(
+      createOpenAI({
+        baseURL: process.env.OPENAI_BASE_URL,
+        apiKey: process.env.OPENAI_API_KEY || 'no-key',
+        fetch: responsesApiFetch,
+      }),
+      modelConfig
+    ),
     routing,
   }
 }
@@ -923,7 +939,7 @@ export async function POST(
       ? appliedReasoningEffort
       : undefined
 
-  const { model, routing } = getModel(chatbot, selectedModelConfig.deploymentId)
+  const { model, routing } = getModel(chatbot, selectedModelConfig)
 
   // create image descriptions if images attached
   let imageDescriptionCost: number = 0
@@ -1233,7 +1249,9 @@ export async function POST(
     experimental_telemetry: { isEnabled: true },
     providerOptions: {
       openai: {
-        store: getOpenAIResponsesStore(),
+        ...(selectedModelConfig.supportsReasoning && {
+          store: getOpenAIResponsesStore(),
+        }),
         ...(providerReasoningEffort && {
           reasoningEffort: providerReasoningEffort,
           reasoningSummary: 'auto',

packages/prisma-data/src/scripts/2026-06-15_ensure_chatbot_fallback_model.ts

@@ -0,0 +1,78 @@
+import { prisma } from '@klicker-uzh/prisma'
+
+const DEFAULT_FALLBACK_MODEL_ID = 'gpt-4.1-mini'
+const APPLY_FLAG = '--apply'
+
+const fallbackModelId =
+  process.env.CHAT_FALLBACK_MODEL_ID?.trim() || DEFAULT_FALLBACK_MODEL_ID
+const apply = process.argv.includes(APPLY_FLAG)
+
+async function run() {
+  const [chatbots, zeroCreditCounts] = await Promise.all([
+    prisma.chatbot.findMany({
+      select: {
+        id: true,
+        name: true,
+        modelSelection: true,
+        allowedModelIds: true,
+      },
+      orderBy: { updatedAt: 'desc' },
+    }),
+    prisma.chatUsageCredits.groupBy({
+      by: ['chatbotId'],
+      where: { current: { lte: 0 } },
+      _count: { _all: true },
+    }),
+  ])
+
+  const zeroCreditCountByChatbotId = new Map(
+    zeroCreditCounts.map((entry) => [entry.chatbotId, entry._count._all])
+  )
+  const missingFallback = chatbots.filter(
+    (chatbot) =>
+      chatbot.allowedModelIds.length > 0 &&
+      !chatbot.allowedModelIds.includes(fallbackModelId)
+  )
+
+  console.log(`Fallback model: ${fallbackModelId}`)
+  console.log(`Chatbots scanned: ${chatbots.length}`)
+  console.log(
+    `Explicit allow-lists missing fallback: ${missingFallback.length}`
+  )
+
+  for (const chatbot of missingFallback) {
+    const zeroCreditParticipants =
+      zeroCreditCountByChatbotId.get(chatbot.id) ?? 0
+    console.log(
+      [
+        `- ${chatbot.name}`,
+        `id=${chatbot.id}`,
+        `modelSelection=${chatbot.modelSelection}`,
+        `zeroCreditParticipants=${zeroCreditParticipants}`,
+        `allowedModelIds=${chatbot.allowedModelIds.join(',')}`,
+      ].join(' ')
+    )
+  }
+
+  if (!apply) {
+    console.log(`Dry run only. Re-run with ${APPLY_FLAG} to update rows.`)
+    return
+  }
+
+  for (const chatbot of missingFallback) {
+    await prisma.chatbot.update({
+      where: { id: chatbot.id },
+      data: {
+        allowedModelIds: [...chatbot.allowedModelIds, fallbackModelId],
+      },
+    })
+  }
+
+  console.log(`Updated ${missingFallback.length} chatbot allow-list(s).`)
+}
+
+try {
+  await run()
+} finally {
+  await prisma.$disconnect()
+}