fix(sse): scope Vertex 404s to a per-model lockout via passthroughModels

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

@@ -30,4 +30,5 @@ export const vertexProvider: RegistryEntry = {
     { id: "claude-opus-4-7", name: "Claude Opus 4.7 (Vertex)" },
     { id: "claude-sonnet-4-6", name: "Claude Sonnet 4.6 (Vertex)" },
   ],
+  passthroughModels: true,
 };

tests/unit/vertex-passthrough-model-lockout.test.ts

@@ -0,0 +1,121 @@
+// Regression guard: after adding passthroughModels: true to Vertex's registry entry, a 404 on
+// one Vertex model (e.g. a stale/synthetic model id) must lock out only that model, not cool
+// down the whole connection — mirrors the existing ollama-cloud/bedrock protection. Before this
+// fix, hasPerModelQuota("vertex", ...) was false, so any 404 on Vertex cooled the whole
+// connection for COOLDOWN_MS.notFound (2 minutes), per errorConfig.ts's generic status_404 rule.
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-vertex-404-"));
+process.env.DATA_DIR = TEST_DATA_DIR;
+
+const core = await import("../../src/lib/db/core.ts");
+const providersDb = await import("../../src/lib/db/providers.ts");
+const auth = await import("../../src/sse/services/auth.ts");
+const accountFallback = await import("../../open-sse/services/accountFallback.ts");
+
+async function resetStorage() {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+  fs.mkdirSync(TEST_DATA_DIR, { recursive: true });
+}
+
+async function seedVertex() {
+  return providersDb.createProviderConnection({
+    provider: "vertex",
+    authType: "apikey",
+    apiKey: "vertex-key",
+    isActive: true,
+    testStatus: "active",
+  });
+}
+
+test.after(() => {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+});
+
+test('hasPerModelQuota("vertex", ...) is true after the passthroughModels registry flag', () => {
+  assert.equal(accountFallback.hasPerModelQuota("vertex", "claude-sonnet-5"), true);
+});
+
+test("404 on one Vertex model locks only that model, connection stays active", async () => {
+  await resetStorage();
+  const conn = await seedVertex();
+
+  const result = await auth.markAccountUnavailable(
+    conn.id,
+    404,
+    "model not found",
+    "vertex",
+    "claude-sonnet-5-high"
+  );
+  assert.equal(result.shouldFallback, true);
+
+  const after = await providersDb.getProviderConnectionById(conn.id);
+  assert.equal(after.testStatus, "active");
+  assert.ok(!after.rateLimitedUntil, "connection must not be rate-limited");
+
+  const lockout = accountFallback.getModelLockoutInfo("vertex", conn.id, "claude-sonnet-5-high");
+  assert.equal(lockout?.reason, "not_found");
+
+  // A sibling model on the same connection must remain immediately eligible.
+  const sibling = accountFallback.getModelLockoutInfo("vertex", conn.id, "gemini-3.1-pro-preview");
+  assert.equal(sibling, null);
+});
+
+test("503 on one Vertex model locks only that model, connection stays active (proves the fix isn't 404-specific)", async () => {
+  // hasPerModelQuota's gate covers status === 404 || status === 429 || status >= 500
+  // (auth.ts:2024) in one shared branch — 502/503/504 keep the model-lockout path (only
+  // the exact 500 is exempted per #5976). This mirrors the 404 test above with a 5xx to
+  // confirm passthroughModels doesn't just fix the specific 404 symptom reported.
+  await resetStorage();
+  const conn = await seedVertex();
+
+  const result = await auth.markAccountUnavailable(
+    conn.id,
+    503,
+    "Service Unavailable",
+    "vertex",
+    "claude-sonnet-5-high"
+  );
+  assert.equal(result.shouldFallback, true);
+
+  const after = await providersDb.getProviderConnectionById(conn.id);
+  assert.equal(after.testStatus, "active");
+  assert.ok(!after.rateLimitedUntil, "connection must not be rate-limited");
+
+  const lockout = accountFallback.getModelLockoutInfo("vertex", conn.id, "claude-sonnet-5-high");
+  assert.equal(lockout?.reason, "server_error");
+
+  const sibling = accountFallback.getModelLockoutInfo("vertex", conn.id, "gemini-3.1-pro-preview");
+  assert.equal(sibling, null);
+});
+
+test("403 PERMISSION_DENIED on Vertex locks only that model too (accepted trade-off, see plan)", async () => {
+  await resetStorage();
+  const conn = await seedVertex();
+
+  // Google Cloud uses the literal "PERMISSION_DENIED" status name for BOTH a
+  // model-specific denial and a connection-wide IAM/API-disabled failure — this
+  // fix cannot distinguish them (no live Vertex credential test in this plan), so
+  // it intentionally treats both as a per-model lockout post-passthroughModels.
+  const result = await auth.markAccountUnavailable(
+    conn.id,
+    403,
+    "PERMISSION_DENIED: the caller does not have permission",
+    "vertex",
+    "claude-sonnet-5-high"
+  );
+  assert.equal(result.shouldFallback, true);
+
+  const after = await providersDb.getProviderConnectionById(conn.id);
+  assert.equal(after.testStatus, "active");
+  assert.ok(!after.rateLimitedUntil, "connection must not be rate-limited");
+
+  const lockout = accountFallback.getModelLockoutInfo("vertex", conn.id, "claude-sonnet-5-high");
+  assert.equal(lockout?.reason, "forbidden");
+});