From 0c8e2261508e96ec56f1807000bc57ca7966eff8 Mon Sep 17 00:00:00 2001
From: Reinis <support.motiondevlab@proton.me>
Date: Sat, 10 Oct 2026 14:09:20 +0300
Subject: [PATCH] fix(proxy): single-count 5xx last-attempt fall-through

Guard the 5xx branch recordRequest with if (attempt < maxAttempts - 1) so the final attempt counts once via the tail recorder (same pattern as f85cb2e throttle-403). Display-only: add throttle-403 legend chip + retitle other-4xx. Plan: 2026-10-10-5xx-last-attempt-double-count (amended: breaker 1/0 first-5xx-only feed).
---
 src/dashboard/public/app.js |  3 ++-
 src/proxy/server.ts         | 18 +++++++++++-------
 2 files changed, 13 insertions(+), 8 deletions(-)

diff --git a/src/dashboard/public/app.js b/src/dashboard/public/app.js
index 6ff30c5..ceb4d17 100644
--- a/src/dashboard/public/app.js
+++ b/src/dashboard/public/app.js
@@ -1421,7 +1421,8 @@ async function renderFailoverTuning() {
             <span class="chip chip-red" title="Server-side failure. Always feeds the breaker (streak + window). An exact 500 on Zen chat also triggers one translated retry via /responses on the same key.">5xx feeds breaker</span>
             <span class="chip chip-yellow" title="Burst rate limit (no quota markers). Returned to the client verbatim — no same-request key-burning — but counts toward the breaker when Burst-failover is on, so the NEXT request spills.">burst-429 feeds breaker</span>
             <span class="chip chip-yellow" title="Quota exhaustion (402, or 429 with quota markers). Parks the key in cooldown for the upstream-supplied wait and fails over to the next key within the SAME request.">quota-429 → cooldown + failover</span>
-            <span class="chip chip-muted" title="400/403/404 and similar (dashboard probes included). Counted in REQ/ERR for honest stats, but say nothing about key health — they neither feed nor reset the breaker.">other 4xx count-only</span>
+            <span class="chip chip-yellow" title="Empty-body / throttle-phrased 403 (volume gating, not bad auth). Parks the key in a short cooldown (60s fallback, honors Retry-After) with NO quota signal, and fails over to the next key within the SAME request. Genuine auth 403s stay count-only.">throttle-403 → short cooldown + failover</span>
+            <span class="chip chip-muted" title="400/401/404 and genuine auth 403s (bad key, revoked, quota wording). Counted in REQ/ERR for honest stats, but say nothing about key health — they neither feed nor reset the breaker.">other 4xx count-only</span>
             <span class="chip chip-muted" title="Fetch threw or the client disconnected (statusCode 0: connect timeouts, aborts, hung-upstream kills). Counted as an error, but never trips — a cancelled turn must not exile a healthy key.">transport errors never trip</span>
             <span class="chip chip-muted" title="DNS down / no route / Wi-Fi off: the request never left the machine. Counted in REQ/ERR for honesty, but nothing else burns — no breaker feed, no failover cycling. The client gets HTTP 503 'Local network unreachable' instead of 'All API keys failed'.">local outage never burns</span>
           </div>
diff --git a/src/proxy/server.ts b/src/proxy/server.ts
index bf4a93d..d21908b 100644
--- a/src/proxy/server.ts
+++ b/src/proxy/server.ts
@@ -558,13 +558,17 @@ export class ProxyServer {
           if (!repeatedServerError) {
             this.keyManager.markError(key.id)
           }
-          this.keyManager.recordRequest(key.id, {
-            statusCode: upstreamRes.status,
-            durationMs: duration,
-            model: prepared.model,
-            sessionId: upstreamSessionId ?? sessionKey ?? null,
-            successful: false,
-          })
+          if (attempt < maxAttempts - 1) {
+            // Counted here; on the last attempt the tail recordRequest
+            // below counts the verbatim 5xx once (no double-count).
+            this.keyManager.recordRequest(key.id, {
+              statusCode: upstreamRes.status,
+              durationMs: duration,
+              model: prepared.model,
+              sessionId: upstreamSessionId ?? sessionKey ?? null,
+              successful: false,
+            })
+          }
           attemptedKeyIds.add(key.id)
 
           if (circuitState === CircuitState.OPEN) {