fix(email): support legacy AI gateway models

backend/src/modules/email/email.entity-suggestions.ts

@@ -75,17 +75,14 @@ export function buildSuggestionMail(message: any) {
     }
 }
 
-export async function suggestEmailEntities(message: any, allCandidates: EntityCandidate[]) {
-    const central = Boolean(secrets.FEDEO_CENTRAL_SERVICES_ENABLED && centralServicesClient.configured())
-    if (!central && !secrets.OPENAI_API_KEY) {
-        throw Object.assign(new Error("Die KI-Erkennung ist noch nicht konfiguriert."), { statusCode: 503 })
-    }
-    const mail = buildSuggestionMail(message)
-    const candidates = selectCandidates(allCandidates, JSON.stringify(mail))
-    if (!candidates.length) return []
-    const request: any = {
-        model: "gpt-5.6-luna",
-        reasoning_effort: "none",
+export function buildEntitySuggestionRequest(
+    mail: ReturnType<typeof buildSuggestionMail>,
+    candidates: EntityCandidate[],
+    model = "gpt-5.6-luna",
+) {
+    return {
+        model,
+        ...(model.startsWith("gpt-5") ? { reasoning_effort: "none" } : {}),
         store: false,
         max_completion_tokens: 1500,
         response_format: zodResponseFormat(suggestionFormat as any, "email_entity_suggestions"),
@@ -94,9 +91,35 @@ export async function suggestEmailEntities(message: any, allCandidates: EntityCa
             { role: "user", content: JSON.stringify({ mail, candidates }) },
         ],
     }
-    const completion = central
-        ? await centralServicesClient.aiChatCompletions(request)
-        : await new OpenAI({ apiKey: secrets.OPENAI_API_KEY, timeout: 45000, maxRetries: 0 }).chat.completions.create(request)
+}
+
+export function shouldRetryWithLegacyAiModel(error: any) {
+    return [400, 404, 422, 500, 502].includes(Number(error?.status))
+}
+
+export async function suggestEmailEntities(message: any, allCandidates: EntityCandidate[]) {
+    const central = Boolean(secrets.FEDEO_CENTRAL_SERVICES_ENABLED && centralServicesClient.configured())
+    if (!central && !secrets.OPENAI_API_KEY) {
+        throw Object.assign(new Error("Die KI-Erkennung ist noch nicht konfiguriert."), { statusCode: 503 })
+    }
+    const mail = buildSuggestionMail(message)
+    const candidates = selectCandidates(allCandidates, JSON.stringify(mail))
+    if (!candidates.length) return []
+    const request: any = buildEntitySuggestionRequest(mail, candidates)
+    let completion: any
+    if (central) {
+        try {
+            completion = await centralServicesClient.aiChatCompletions(request)
+        } catch (error: any) {
+            if (!shouldRetryWithLegacyAiModel(error)) throw error
+            completion = await centralServicesClient.aiChatCompletions(
+                buildEntitySuggestionRequest(mail, candidates, "gpt-4o-mini"),
+            )
+        }
+    } else {
+        completion = await new OpenAI({ apiKey: secrets.OPENAI_API_KEY, timeout: 45000, maxRetries: 0 })
+            .chat.completions.create(request)
+    }
     const choice = completion.choices?.[0]
     if (choice?.finish_reason !== "stop" || choice?.message?.refusal || !choice?.message?.content) {
         throw new Error("Die KI konnte keine vollständige Analyse liefern. Bitte erneut versuchen.")

backend/tests/emailEntitySuggestions.test.ts

@@ -2,8 +2,10 @@ import assert from "node:assert/strict"
 import test from "node:test"
 import {
     buildSuggestionMail,
+    buildEntitySuggestionRequest,
     dismissEntitySuggestion,
     selectCandidates,
+    shouldRetryWithLegacyAiModel,
     validateSuggestions,
     type EntityCandidate,
 } from "../src/modules/email/email.entity-suggestions"
@@ -69,3 +71,22 @@ test("dismisses only the selected entity suggestion", () => {
     )
     assert.deepEqual(dismissEntitySuggestion(null, "customers", 1), [])
 })
+
+test("uses Luna by default and omits reasoning effort for the legacy fallback", () => {
+    const request = buildEntitySuggestionRequest(buildSuggestionMail({ subject: "Test" }), [customer])
+    assert.equal(request.model, "gpt-5.6-luna")
+    assert.equal(request.reasoning_effort, "none")
+
+    const fallback = buildEntitySuggestionRequest(buildSuggestionMail({ subject: "Test" }), [customer], "gpt-4o-mini")
+    assert.equal(fallback.model, "gpt-4o-mini")
+    assert.equal("reasoning_effort" in fallback, false)
+})
+
+test("retries model compatibility errors without retrying auth, rate-limit, or configuration errors", () => {
+    for (const status of [400, 404, 422, 500, 502]) {
+        assert.equal(shouldRetryWithLegacyAiModel({ status }), true)
+    }
+    for (const status of [401, 403, 429, 503]) {
+        assert.equal(shouldRetryWithLegacyAiModel({ status }), false)
+    }
+})