fix(providers): scope nvidia NIM 404s to the single failing model (#6773) (#6888)

* fix(providers): scope nvidia NIM 404s to the single failing model (#6773)

The nvidia registry entry multiplexes 17 models from 9 different upstream
vendors (z-ai/, minimaxai/, deepseek-ai/, qwen/, mistralai/, stepfun-ai/,
moonshotai/, openai/, nvidia/) behind one connection, but was missing
passthroughModels: true — unlike 34 other multi-model registries
(modelscope, synthetic, kilo-gateway, etc). Without it, hasPerModelQuota
returns false for nvidia, so a 404 on a single stale/renamed model falls
through checkFallbackError's generic catch-all as a connection-wide
cooldown instead of being scoped to just that model, poisoning all 17
nvidia models for the cooldown window.

Add passthroughModels: true to the nvidia registry entry so 404/429s on
one model lock out only that model.

Regression test: tests/unit/nvidia-passthrough-models-6773.test.ts

* fix(quality): register nvidia passthrough test in stryker tap.testFiles

check-mutation-test-coverage.mjs --strict flagged
tests/unit/nvidia-passthrough-models-6773.test.ts as an unregistered
covering test for accountFallback.ts (Fast Quality Gates).

Co-authored-by: Diego Rodrigues de Sa e Souza <souzamiriamrodrigues790@gmail.com>

---------

Co-authored-by: Diego Rodrigues de Sa e Souza <souzamiriamrodrigues790@gmail.com>

changelog.d/fixes/6773-6773-nim-404.md

@@ -0,0 +1 @@
+- fix(providers): scope nvidia NIM 404s to the single failing model instead of cooling down the whole connection (#6773)

open-sse/config/providers/registry/nvidia/index.ts

@@ -8,6 +8,13 @@ export const nvidiaProvider: RegistryEntry = {
   baseUrl: "https://integrate.api.nvidia.com/v1/chat/completions",
   authType: "apikey",
   authHeader: "bearer",
+  // #6773: nvidia multiplexes 17 models from 9 different upstream vendors
+  // (z-ai/, minimaxai/, deepseek-ai/, qwen/, mistralai/, stepfun-ai/,
+  // moonshotai/, openai/, nvidia/) behind ONE connection — mark it passthrough
+  // so a single stale/renamed model's 404 locks out only that model instead
+  // of cooling down the whole connection (see accountFallback.ts
+  // hasPerModelQuota doc comment; matches modelscope/synthetic/kilo-gateway).
+  passthroughModels: true,
   models: [
     // #6108: z-ai/glm-5.1 EOL'd 2026-07-02 (direct probe returns 410) — dropped.
     { id: "z-ai/glm-5.2", name: "GLM 5.2" },

stryker.conf.json

@@ -196,6 +196,7 @@
       "tests/unit/model-lockout-max-cooldown.test.ts",
       "tests/unit/no-memory-header.test.ts",
       "tests/unit/non-streaming-sse-terminal-typescan-4459.test.ts",
+      "tests/unit/nvidia-passthrough-models-6773.test.ts",
       "tests/unit/oauth-providers-config.test.ts",
       "tests/unit/oauth-redirect-uri-mismatch.test.ts",
       "tests/unit/observability-fase04.test.ts",

tests/unit/nvidia-passthrough-models-6773.test.ts

@@ -0,0 +1,79 @@
+/**
+ * Regression test for #6773 — NVIDIA NIM models listed available:true but 404 at router.
+ *
+ * Root cause: the `nvidia` provider registry entry multiplexes many distinct
+ * third-party vendor models (z-ai/, minimaxai/, deepseek-ai/, qwen/,
+ * mistralai/, stepfun-ai/, moonshotai/, openai/, nvidia/) behind ONE base URL
+ * and ONE API key connection — architecturally identical to `modelscope`,
+ * `synthetic`, and `kilo-gateway`, which all set `passthroughModels: true` so
+ * that a single model's 404/429 stays scoped to that model instead of cooling
+ * down the whole connection (see accountFallback.ts `hasPerModelQuota` doc
+ * comment). Without the flag, a single upstream 404 for one (possibly
+ * stale/renamed) model poisons ALL nvidia models for the connection-cooldown
+ * duration — matching the issue's "All 17 behave the same" symptom.
+ */
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const accountFallback = await import("../../open-sse/services/accountFallback.ts");
+const providerRegistry = await import("../../open-sse/config/providerRegistry.ts");
+
+test("#6773: nvidia registry entry sets passthroughModels", () => {
+  const entry = providerRegistry.getRegistryEntry("nvidia");
+  assert.equal(
+    entry?.passthroughModels,
+    true,
+    "nvidia multiplexes many third-party vendor models behind one connection " +
+      "(z-ai/, minimaxai/, deepseek-ai/, qwen/, mistralai/, stepfun-ai/, " +
+      "moonshotai/, openai/, nvidia/) — it should set passthroughModels: true " +
+      "like modelscope/synthetic/kilo-gateway, so a single stale model 404 " +
+      "does not cool down the whole connection for all other models"
+  );
+});
+
+test("#6773: hasPerModelQuota('nvidia') is true, so a 404 on one nvidia model is model-scoped", () => {
+  assert.equal(
+    accountFallback.hasPerModelQuota("nvidia", "z-ai/glm-5.2"),
+    true,
+    "expected nvidia to use per-model lockout (like gemini/github/codex/compatible " +
+      "providers) so a 404 on one model doesn't cool down the other nvidia models"
+  );
+});
+
+test("#6773: checkFallbackError + lockModelIfPerModelQuota scope a single-model 404 to just that model for nvidia", () => {
+  // A plain upstream 404 (e.g. one stale/renamed nvidia model id) falls through
+  // checkFallbackError's generic catch-all: shouldFallback=true with a non-zero
+  // connection cooldown. With hasPerModelQuota=true, lockModelIfPerModelQuota
+  // now scopes that cooldown to just the one failing model instead of the
+  // whole connection.
+  const result = accountFallback.checkFallbackError(
+    404,
+    "Not Found",
+    0,
+    "z-ai/glm-5.2",
+    "nvidia",
+    null,
+    null,
+    null
+  );
+  assert.equal(result.shouldFallback, true, "404 triggers a connection-level fallback/cooldown");
+  assert.ok(
+    (result.cooldownMs ?? 0) > 0,
+    "the connection-level cooldown is non-zero, so it also blocks the other nvidia models" +
+      " unless it gets scoped to just this model below"
+  );
+
+  const locked = accountFallback.lockModelIfPerModelQuota(
+    "nvidia",
+    "conn-6773",
+    "z-ai/glm-5.2",
+    "unknown",
+    result.cooldownMs ?? 30_000
+  );
+  assert.equal(
+    locked,
+    true,
+    "expected the 404 to be scoped to just this one model (per-model lockout), " +
+      "not the whole connection"
+  );
+});