fix(llm): keep system prompts on premium chat, reject incomplete streams, stop double-charging quota

server/supabase/functions/llm-proxy/index.ts

@@ -66,6 +66,16 @@ async function issueGenerationReceipt(
   return parseGenerationReceiptId(data)
 }
 
+/** 쓰지 않을 공급자 응답 본문을 닫아 연결과 생성을 정리한다. */
+async function discardBody(resp: Response): Promise<void> {
+  if (!resp.body || resp.bodyUsed) return
+  try {
+    await resp.body.cancel()
+  } catch {
+    // 이미 닫힌 스트림 — 무시
+  }
+}
+
 function generationHeaders(generationId: string | null): Record<string, string> {
   return generationId === null ? {} : { [GENERATION_ID_HEADER]: generationId }
 }
@@ -147,27 +157,9 @@ Deno.serve(async (req: Request) => {
       )
     }
 
-    // 3단계: 쿼터 소비 (원자적 base → overage fallback)
-    const policy = getQuotaPolicy(tier, quotaKey)
-    const consume = await consumeQuota(user.id, quotaKey, serviceClient, policy.limit)
-    if (!consume.allowed) {
-      return new Response(
-        JSON.stringify({
-          error: 'quota_exceeded',
-          model: requestedModel,
-          current: consume.current,
-          limit: consume.limit,
-          tier,
-          overage_credits: consume.overageCredits,
-        }),
-        {
-          status: 429,
-          headers: { ...corsHeaders, 'Content-Type': 'application/json' },
-        },
-      )
-    }
-
-    // 실제 Anthropic API 호출 (Prompt Caching 2024-07-31 활성화)
+    // 3단계: 공급자 호출 (Prompt Caching 2024-07-31 활성화)
+    // 쿼터는 공급자가 요청을 받아들인 뒤에만 소비한다. 공급자 5xx/과부하/타임아웃으로
+    // 답을 주지 못한 요청이 쿼터를 깎고, 클라이언트 재시도가 한 번 더 깎는 일을 막는다.
     const systemPayload = buildAnthropicSystemBlocks(body.system)
 
     const anthropicResp = await fetch('https://api.anthropic.com/v1/messages', {
@@ -190,12 +182,54 @@ Deno.serve(async (req: Request) => {
 
     if (!anthropicResp.ok) {
       console.error('Anthropic request failed', { status: anthropicResp.status })
+      await discardBody(anthropicResp)
       return new Response(JSON.stringify({ error: 'provider_request_failed' }), {
         status: 502,
         headers: { ...corsHeaders, 'Content-Type': 'application/json', 'Cache-Control': 'no-store' },
       })
     }
 
+    // 비스트리밍 응답은 형식 검증까지 통과한 뒤에 소비한다.
+    let nonStreamData: unknown = null
+    if (!body.stream || !anthropicResp.body) {
+      nonStreamData = await anthropicResp.json()
+      if (!hasAssistantText(nonStreamData)) {
+        console.error('Anthropic returned an invalid response shape')
+        return new Response(JSON.stringify({ error: 'provider_invalid_response' }), {
+          status: 502,
+          headers: { ...corsHeaders, 'Content-Type': 'application/json', 'Cache-Control': 'no-store' },
+        })
+      }
+    }
+
+    // 4단계: 쿼터 소비 (원자적 base → overage fallback).
+    // 동시 요청이 checkQuota 를 함께 통과했다면 여기서 초과분이 거절된다.
+    const policy = getQuotaPolicy(tier, quotaKey)
+    let consume: Awaited<ReturnType<typeof consumeQuota>>
+    try {
+      consume = await consumeQuota(user.id, quotaKey, serviceClient, policy.limit)
+    } catch (err) {
+      await discardBody(anthropicResp)
+      throw err
+    }
+    if (!consume.allowed) {
+      await discardBody(anthropicResp)
+      return new Response(
+        JSON.stringify({
+          error: 'quota_exceeded',
+          model: requestedModel,
+          current: consume.current,
+          limit: consume.limit,
+          tier,
+          overage_credits: consume.overageCredits,
+        }),
+        {
+          status: 429,
+          headers: { ...corsHeaders, 'Content-Type': 'application/json' },
+        },
+      )
+    }
+
     if (body.stream && anthropicResp.body) {
       const generationId = await issueGenerationReceipt(
         serviceClient,
@@ -215,21 +249,13 @@ Deno.serve(async (req: Request) => {
       })
     }
 
-    const data: unknown = await anthropicResp.json()
-    if (!hasAssistantText(data)) {
-      console.error('Anthropic returned an invalid response shape')
-      return new Response(JSON.stringify({ error: 'provider_invalid_response' }), {
-        status: 502,
-        headers: { ...corsHeaders, 'Content-Type': 'application/json', 'Cache-Control': 'no-store' },
-      })
-    }
     const generationId = await issueGenerationReceipt(
       serviceClient,
       user.id,
       generationPurpose,
       requestedModel,
     )
-    return new Response(JSON.stringify(data), {
+    return new Response(JSON.stringify(nonStreamData), {
       status: 200,
       headers: {
         ...corsHeaders,