fix(combo): fallback to sibling model on 500 for per-model-quota providers

CHANGELOG.md

@@ -30,6 +30,8 @@
 
 - **cline (empty/false-502 non-streaming responses):** the Cline gateway can wrap OpenAI-compatible chat completions in a `{ success, data: { choices, usage, … } }` envelope. The non-streaming path checked the top-level body for empty content before unwrapping, so a valid Cline response was treated as malformed. The body is now unwrapped via `unwrapClineNonStreamingEnvelope()` right after the provider-envelope unwrap and before the empty-content check, closing the remaining gap from #5956/#5924. Regression guard: `tests/unit/cline-response-envelope.test.ts`. ([#5956](https://github.com/diegosouzapw/OmniRoute/issues/5956) — thanks @KooshaPari)
 
+- **combo (per-model-quota providers exhausted on a single model 500):** for per-model-quota providers (gemini, github, passthrough, compatible) that multiplex many models behind one connection, a model-level `500` (e.g. Gemini "Internal error encountered") wrongly marked the whole connection exhausted, so sibling models on the same connection were skipped and the combo could 502 instead of falling back. `markConnectionLevelExhaustion` now leaves the connection eligible on a `500` for per-model-quota providers (other connection-level statuses — 408/502/503/504/524 — still exhaust correctly), and the retry loop early-returns when a model is already in lockout. Regression guard: `tests/unit/combo/combo-target-exhaustion.test.ts`. ([#5976](https://github.com/diegosouzapw/OmniRoute/pull/5976) — thanks @hartmark)
+
 ### 📝 Maintenance
 
 - **test (deflake `setup-claude`):** `tests/unit/cli/setup-claude.test.ts` failed ~50% of runs with `Unable to deserialize cloned data due to invalid or unsupported version` at file teardown (all subtests passed), randomly reddening `Unit Tests fast-path (2/2)` / `Fast Quality Gates` across the PR→release queue. Root cause: `node --test` streams each file's report to the parent as V8-serialized frames on fd 1 (stdout), and the CLI helper under test (`syncClaudeProfilesFromModels`) prints progress via `console.log` — that stdout output interleaved with the serialized frames and corrupted the stream. The test now silences the stdout-writing `console` methods for the file's duration (no assertion inspects stdout), making it deterministic (15/15 green locally). ([#5959](https://github.com/diegosouzapw/OmniRoute/issues/5959))

open-sse/services/combo.ts

@@ -2102,6 +2102,15 @@ export async function handleComboChat({
             !isTokenLimitBreach &&
             [408, 429, 500, 502, 503, 504].includes(result.status);
           if (retry < maxRetries && isTransient && !providerExhausted) {
+            if (
+              provider &&
+              rawModel &&
+              isModelLocked(provider, targetWithConnection.connectionId || "", rawModel)
+            ) {
+              log.info("COMBO", `Skipping retry for ${modelStr} — model lockout active`);
+              if (i > 0) fallbackCount++;
+              return null;
+            }
             // Record model lockout immediately on the first transient failure —
             // once the model is cooling down, retrying it would waste an upstream
             // call and extend the cooldown via exponential backoff.

open-sse/services/combo/targetExhaustion.ts

@@ -104,7 +104,7 @@ export function applyComboTargetExhaustion(
     if (result.status === 429 && !isTokenLimitBreach && provider && provider !== "unknown") {
       transientRateLimitedProviders.add(provider);
     }
-    markConnectionLevelExhaustion(target, { result, errorText, sets, log, tag });
+    markConnectionLevelExhaustion(target, { result, errorText, sets, log, tag, rawModel });
   }
 
   return providerExhausted;
@@ -118,9 +118,12 @@ export function applyComboTargetExhaustion(
  */
 function markConnectionLevelExhaustion(
   target: ResolvedComboTarget,
-  opts: Pick<ApplyComboTargetExhaustionOptions, "result" | "errorText" | "sets" | "log" | "tag">
+  opts: Pick<
+    ApplyComboTargetExhaustionOptions,
+    "result" | "errorText" | "sets" | "log" | "tag" | "rawModel"
+  >
 ): void {
-  const { result, errorText, sets, log, tag } = opts;
+  const { result, errorText, sets, log, tag, rawModel } = opts;
   const provider = target.provider;
   if (
     !provider ||
@@ -130,7 +133,13 @@ function markConnectionLevelExhaustion(
     // #5085: empty-content 502 is a healthy connection returning no body — model-level, not
     // connection-level. Don't exhaust the provider; let the remaining legs (incl. same-provider)
     // be tried in-request.
-    isEmptyContentFailure(result.status, errorText)
+    isEmptyContentFailure(result.status, errorText) ||
+    // Per-model-quota providers (gemini, github, passthrough, compatible) multiplex models
+    // behind one connection. A model-level 500 (e.g. Gemini "Internal error encountered")
+    // must NOT exhaust the connection — other models on the same connection may still succeed.
+    // Other connection-level statuses (408/502/503/504/524) indicate the connection itself is
+    // bad, so they correctly exhaust even for per-model-quota providers.
+    (result.status === 500 && hasPerModelQuota(provider, rawModel))
   ) {
     return;
   }